ANTONELLI Et Al - New Frontiers in the Economics of Innovaiton and New Technology

New Frontiers in the Economics ofInnovation and New Technology

New Frontiers in theEconomics of Innovationand New TechnologyEssays in Honour of Paul A. David

Edited by

Cristiano AntonelliProfessor of Economics and Director of the Department of Economics atthe University of Torino, Italy

Dominique ForayProfessor of Economics of Innovation and Director of the College ofManagement of Technology at the Ecole Polytechnique Fédérale deLausanne, Switzerland

Bronwyn H. HallProfessor of Economics at the University of California at Berkeley,Research Associate at the National Bureau of Economic Research,Cambridge, Massachusetts, and the Institute of Fiscal Studies, London

W. Edward SteinmuellerProfessor of Information and Communication Technology Policy, SPRU –Science and Technology Policy Research, University of Sussex, Falmer,Brighton, UK

Edward ElgarCheltenham, UK • Northampton, MA, USA

© Cristiano Antonelli, Dominique Foray, Bronwyn H. Hall and W. EdwardSteinmueller, 2006

All rights reserved. No part of this publication may be reproduced, stored ina retrieval system or transmitted in any form or by any means, electronic,mechanical or photocopying, recording, or otherwise without the priorpermission of the publisher.

Published byEdward Elgar Publishing LimitedGlensanda HouseMontpellier ParadeCheltenhamGlos GL50 1UAUK

Edward Elgar Publishing, Inc.136 West StreetSuite 202NorthamptonMassachusetts 01060USA

A catalogue record for this bookis available from the British Library

Library of Congress Cataloguing in Publication Data

New frontiers in the economics of innovation and new technology: essays in honour of Paul A. David/edited by Cristiano Antonelli . . . [et al.]

p. cm.Includes bibliographical references and index.1. Innovative technology–Economic aspects. 2. Diffusion of innovations–

Economic aspects. I. Antonelli, Cristiano. II. David, Paul A.HC79.T4N472 2005338’.064–dc22 2005046147

ISBN-13: 978 1 84376 631 5ISBN-10: 1 84376 631 0

Printed and bound in Great Britain by MPG Books Ltd, Bodmin, Cornwall

Contents

List of contributors vii

PART I GENERAL INTRODUCTION

1 The economics of innovation: between �� and �� 3Cristiano Antonelli, Dominique Foray, Bronwyn H. Hall andW. Edward Steinmueller

PART II PATH DEPENDENCE IN TECHNICAL CHANGE

2 Competing technologies, technological monopolies and the rateof convergence to a stable market structure 23Andrea P. Bassanini and Giovanni Dosi

3 Path dependence, localised technological change and the questfor dynamic efficiency 51Cristiano Antonelli

4 A history-friendly model of innovation, market structure andregulation in the age of random screening of thepharmaceutical industry 70Franco Malerba and Luigi Orsenigo

5 Path dependence and diversification in corporate technologicalhistories 118John Cantwell

6 Is the world flat or round? Mapping changes in the taste for art 158G.M. Peter Swann

7 Waves and cycles: explorations in the pure theory of price forfine art 188Robin Cowan

PART III THE ECONOMICS OF KNOWLEDGE

8 Learning in the knowledge-based economy: the future as viewedfrom the past 207W. Edward Steinmueller

v

9 The economics of open technology: collective organisationand individual claims in the ‘fabrique lyonnaise’ during theold regime 239Dominique Foray and Liliane Hilaire Perez

10 Measurement and explanation of the intensity of co-publicationin scientific research: an analysis at the laboratory level 255Jacques Mairesse and Laure Turner

11 Epistemic communities and communities of practice in theknowledge-based firm 296Patrick Cohendet and Ash Amin

12 Markets for technology: ‘panda’s thumbs’, ‘calypso policies’and other institutional considerations 323Ashish Arora, Andrea Fosfuri and Alfonso Gambardella

13 The key characteristics of sectoral knowledge bases: aninternational comparison 361Stefano Brusoni and Aldo Geuna

PART IV THE DIFFUSION OF NEW TECHNOLOGIES

14 Uncovering general purpose technologies with patent data 389Bronwyn H. Hall and Manuel Trajtenberg

15 Equilibrium, epidemic and catastrophe: diffusion ofinnovations with network effects 427Luís M.B. Cabral

16 Technological diffusion under uncertainty: a real optionsmodel applied to the comparative international diffusion ofrobot technology 438Paul Stoneman and Otto Toivanen

PART V POSTSCRIPT

17 An appreciation of Paul David’s work 471Dominique Foray

Index 475

vi Contents

Contributors

Ash Amin, University of Durham, UK

Cristiano Antonelli, University of Turin, Italy

Ashish Arora, Heinz School of Public Policy and Management, Pittsburgh,USA

Andrea P. Bassanini, OECD, Paris, France

Stefano Brusoni, University of Bocconi, Milan, Italy

Luís M.B. Cabral, New York University, USA

John Cantwell, Rutgers University, Newark, USA

Patrick Cohendet, Université Louis Pasteur, Strasbourg, France

Robin Cowan, Universiteit Maastricht, the Netherlands

Giovanni Dosi, Sant’Anna School of Advanced Studies, Pisa, Italy

Dominique Foray, Ecole Polytechnique Fédérale, Lausanne, Switzerland

Andrea Fosfuri, Universidad Carlos III, Madrid, Spain

Alfonso Gambardella, University of Bocconi, Milan, Italy

Aldo Geuna, SPRU – Science and Technology Policy Research, Universityof Sussex, UK

Bronwyn H. Hall, University of California, Berkeley, USA

Liliane Hilaire Perez, Conservatoire National des Arts et Métiers, France

Jacques Mairesse, Institut National de la Statistique et des EtudesEconomiques, France

Franco Malerba, University of Bocconi, Milan, Italy

Luigi Orsenigo, University of Bocconi, Milan, Italy

W. Edward Steinmueller, SPRU – Science and Technology Policy Research,University of Sussex, UK

Paul Stoneman, Warwick Business School, UK

vii

G.M. Peter Swann, Nottingham University Business School, UK

Otto Toivanen, Helsinki School of Economics, Finland

Manuel Trajtenberg, Tel Aviv University, Israel

Laure Turner, Ecole Nationale de la Statistique et de l’AdministrationEconomique, France

viii Contributors

PART I

General Introduction

1. The economics of innovation:between �� and ��

Christiano Antonelli, Dominique Foray,Bronwyn H. Hall and W. Edward Steinmueller

1 PRELUDE

The birth and infancy of innovation economics, as a specific area of acad-emic study and empirical research, and one that had applications to realproblems of industrial growth, firm strategy and economic policy, occurredat the same time that the boundaries of economic theory were expandinghorizontally in the 1950s and 1960s. During that period, economic theorybegan to offer a detailed and interpretative framework for the study of eco-nomics, assimilated the Keynesian heresy as a structural variation in behav-ioural modelling, and foraged for new areas of research in which to applyits analytical categories and its powerful mechanisms of systematic analy-sis. This newly found confidence in the methodological validity of indi-vidual decision-making and marginal calculations was based uponrigorous assumptions about the rational behaviour of agents. The additionof a new, large macroeconomic area based upon the paradigm of generaleconomic equilibrium, encouraged economists to venture still further intonew areas of application.

The extraordinary heuristic capacity of economic theory supported boldclaims of explanatory competence and scientific primacy for the field. It isnot surprising that those years are sometimes spoken of as economic imperi-alism. The new territories that economic theory colonised included the eco-nomics of health and education, of risk and insurance, of uncertainty andinformation, and of marriage and the family. Economics and related disci-plines claimed to be able rationally to assess choice, in all of its manifesta-tions, by processing the alternatives through the machinery of finance,probability and opportunity cost. At the same time there was a progressivespecialisation of competencies. Areas which up to then had been thoughtof as interchangeable, such as international and regional, developed anddeveloping economies or emerging and mature industries, became distinct.

3

There are many reasons why this phase can be metaphorically referred toas the ��. Just as the Greek heroes started their campaign of con-version of the barbarians to the superiority of Greek culture and ideologyby winning the Golden Fleece, economists were convinced that theirinsights had become universal. In particular, their insights could beapplied to the origin and spread of innovation and, consequently, to theissues of technical progress, science, the university and, even, the creativeprocess.

The result was that economists asked themselves not only what creates anation’s wealth, but also why are certain countries more innovative thanothers or why do certain historic periods appear to be more fertile anddynamic than others? The analysis of what determines innovative activity,identified as a specific form of economic action, was distinguished from theanalysis of the effects of introducing innovation.

In this way, innovation economics invaded the territory that had alreadybeen subdivided by industrial economics, the theory of the firm, regionaleconomics while not ignoring important points of international economicsand, above all, public economics. The claim of scholars of innovation tothe ownership of the Golden Fleece was certainly at hand: the rate ofgrowth of economic systems and their share of the international market,wage differentials and rates of profit were increasingly linked to innovativecapacity. Innovative capacity was (and still is) perceived as one of the fun-damental source of a nation’s wealth and more specifically the ever chang-ing, if not increasing, differences in that wealth.

The wealth of knowledge and the abundance of recipes, diagnoses andtherapies, deriving from the economics of innovation have tempted manyof its followers to advocate the ��, a contest for the capital citieswaged from what is still often regarded as the hinterland. Studying andinvestigating the multifaceted field of innovation economics has, in fact,more than once unearthed some awkward and significant results; resultsthat cannot easily be ignored and that are, at least in some circles, embar-rassing for claimants to the Golden Fleece of a true knowledge of themeans and ends by which economic results are produced. This is true, aboveall, in the dominant theory that the imposing expedition had advanced and,indeed, financed.

Technological knowledge does not appear to be an exogenous flow intothe economic system, as had been assumed for the purpose of ‘simplifying’analysis. Neither does it appear that the choice of factors of production orindividuals’ preferences are necessarily reversible. The workings of thewhole economic system seem to be far from such Newtonian physics.Instead, the emerging model of economic life is one suffused with non-ergodic dynamic elements, where the path taken shapes the destination

4 General introduction

reached, where rate and direction are interdependent and where causes areentangled with effects in uncomfortable new patterns. Correspondingly,examining the origins and spread of innovation has made it possible tounderstand clearly how incomplete and indeed inaccurate it is to assumethat economic agents act in a perfectly rational way with the corollaries ofperfect cognition and predictive capacity. There is a notable absence in thereal world of the race of super-rational agents endowed with rationalexpectations and able, thereby, to anticipate a whole range of future actionsincluding innovation and new patterns of consumers’ preference.

But an even fiercer battle remains to be waged. For, if technologicalknowledge is no longer exogenous, but instead strongly influenced by theunfolding of economic life, the same hypotheses regarding the workings ofthe market must also be questioned. Firms do not limit themselves toadjusting outputs to prices, but also struggle to survive or race tosupremacy through innovation. In such a world, it can be understood thatit is no longer possible to imagine a single general economic equilibrium, itis necessary to speak more of a range or, perhaps, even more precisely, of aseries of possible general economic equilibria. When the evolution of con-sumers’ preferences are also recognised as being endogenous and depend-ent upon experience, it seems equally doubtful that we are living in the bestavailable world, and that we might at the same time know how long or howfar our journey might be to a better one. As innovative capacity is stronglyinfluenced by the processes of accumulation of technical knowledge, thedirection that we take in attempting to add to this accumulation affectswhat destination can be reached. And through this uncertainty, choice,imagination and inspiration are reintroduced into economic life.

The weapons for the �� are by now ready. Will our heroesmanage to complete the journey which returns them to Athens? And, aboveall, will Athens be able to appreciate the new language, the plates and goldand silver that the Argonauts will bring with them? Will Athens be able torecognise that its past certainties were limited and fragile, and were, aboveall, based on static and incoherent assumptions, although they were thesource of great daring and farsightedness?

Economics of innovation enters the twenty-first century unable to decidebetween the modest yet reassuring temptation to consolidate long-held cer-tainties and the far bolder and more tenuous goal of rewriting the model.Regardless, it has begun its return from the interior to the country of itsbirth, eager to show that new truths are to be discovered that are less certainbut more complex and plausible than those made by the forefather con-querors.

This is the context into which the lifelong contribution of Paul Davidcan be better appreciated and valued. David has built essential pathways

The economics of innovation 5

in both directions; in the �� as well as in the ��. Thecontribution of David has been fundamental in the �� to build-ing economics of innovation as a new discipline and area of specialistexpertise and competence. In the same fruitful career, Paul David hascontributed some of the most powerful analytical tools by which the�� may be organised.

In this introduction we review David’s contributions to three importantareas in the economics of innovation: the economics of knowledge, the roleof path dependence in the evolution of economic equilibria and thediffusion of new technologies. We have necessarily been selective in ourchoice of contributions to discuss, lest this introduction become longerthan the volume it introduces! Nevertheless we hope that our brief surveyswill give the reader an idea of his many faceted contributions to this liter-ature, at the same time that it places the chapters in the context in whichthey were written.

2 THE ECONOMICS OF KNOWLEDGE

While ‘knowledge’ in a very broad sense has always been at the heart ofPaul David’s work, at both microscopic and macroscopic levels, he came toexplore this topic more systematically in the early 1990s. Starting with theanalysis of the peculiar properties of knowledge and information as an eco-nomic good, he proceeds to the historical and normative analysis ofresource allocation mechanisms in the field of knowledge production anddistribution and, more generally, socio-economic institutions that can berelied upon to produce, mediate and use knowledge efficiently. The intel-lectual journey of Paul David in this field consists of a systematic explor-ation of the three-dimensional space in which ‘knowledge-products’ aredistributed.1 The first dimension of that space is the continuum betweensecrecy and full disclosure; the second is the spectrum of asset ownershipstatus ranging from legally enforced private property rights to pure publicgoods; and the third is the dimension along which codification appears atone extreme and tacitness at the other.

How do institutions, technologies and economic factors determine thelocation of knowledge in this three-dimensional space, and what are theimplications of this on allocative efficiency in the domain of knowledgeproduction and use? These two questions are raised repeatedly in PaulDavid’s work. The argument that knowledge and information have theproperties of public goods creates the theoretical framework in which thesequestions are addressed.


Knowledge, the Lighthouse and the Economist

David has made it very clear that saying a good (for example, knowledge)is a public good, on the basis of the properties of non-excludability andnon-rivalry, does not imply that this good must necessarily be produced bythe state, that markets for it do not exist or that its private production isimpossible. It simply means that, considering the properties of the good, itis not possible to rely exclusively on a system of competitive markets toefficiently guarantee its production. Considering the example of the light-house as a paradigmatic case of a public good, David shows that the factthat the lighthouse service was once provided by the private sector in theUK, as Coase documented in a famous paper, does not mean that the light-house is a private good.2 Private markets function in this case because anagent is granted local monopoly on the right to collect a tax in exchange forthe service provided. In the same way, the creation of a private monopolyon new knowledge (a patent) enables the market to produce that good. Butin both cases the remedy is imperfect, for the owner of the monopoly willnot supply the ‘light’ (of the lighthouse or of the knowledge) at a price(harbour tax or royalties) equivalent to the negligible cost of making thesegoods available to additional users (the marginal cost of use of existingknowledge is nil, as it is in the case of using the harbour’s lighthouse).3

What are the consequences of such a clear and strict position on the eco-nomic nature of knowledge? In the domain of the production of new ideas(a very broad domain ranging from scientific discovery and technologicaland engineering innovations to intellectual creation), the fundamentalproblem for the allocative efficiency of competitive markets arises from theexternalities that exist because of the public good nature of ideas. It is,therefore, crucial to analyse the historical emergence and the allocativeefficiency of various kinds of institutions devised to correct or to providealternatives to such market failures. The 3 Ps figure (public procurement,patronage and private property) then provides a framework for such aninvestigation.4

From the New Economics of Science . . .

In this direction, the main contribution of David (notably with ParthaDasgupta) has been the rigorous and systematic exploration of the eco-nomics of ‘open science’ – including both a detailed analysis of the normof openness which is ‘incentive-compatible with a collegiate reputationalreward system based upon accepted claims to priority’,5 and an analysis ofthe historical emergence of this institution.6 Both analytical and modellingapproaches show how efficient such a system is, for it ensures the rapid and


complete diffusion of new knowledge while preserving a certain level ofincentive.7 Moreover, complete disclosure functions as a sort of ‘qualityassurance’ in so far as published results can be reproduced and verified byother members of the community. Given the fact that by most measures theproductivity of scientific research as organised under these institutionalprinciples has been outstanding, it is clear that the economics of openscience provides a framework to study and assess other kinds of institutionwith similar positive effects on the growth of knowledge. The economics offree/libre open source software is, for example, a direct extension of thework done on science. Today, David is in the forefront of research on thisparticular class of social systems in which high rates of innovation are cor-related with rich spillovers, implying that private agents do not always relyon exclusivity and excludability mechanisms to capture private benefitsfrom their intellectual creative work.

. . . to the Comedy of the Commons

Owing to the peculiar features of knowledge, the production of know-ledge has the potential to create a ‘combinatorial explosion’. This good isdifficult to control and can be used and reused infinitely to produce otherknowledge which is in turn non-excludable, non-rival and cumulative, andso forth. In many cases knowledge is also deliberately disclosed andorganised in order to facilitate its access and reproduction by others.All these processes give rise to the creation and expansion of a ‘know-ledge commons’. ‘Knowledge commons’ are not subject to the classictragedy of the commons, a parable describing the case where exhaustibleresources (such as a pasture or a shoal of fish) are subject to destructionby unregulated access and exploitation. Knowledge may be used concur-rently by many, without diminishing its availability to any of the users,and will not become ‘depleted’ through intensive use. As David recentlywrote, contradicting the American poet Robert Frost’s elegy to NewEngland civility, good fences do not make good neighbours: ‘informationis not like forage, depleted by use for consumption; data sets are notsubject to being “overgrazed” but instead are likely to be enriched andrendered more accurate and more fully documented the more researchersare allowed to comb through them’.8 The properties of non-excludability,non-rivalry and cumulativeness have features akin to quasi-infiniteincreasing returns.

Thus, the commons is not tragic, but comedic, in the classical sense of astory with a happy ending. Managing and protecting the ‘knowledgecommons’ requires social regulations that are entirely different from thesocial arrangements used to regulate ecological systems of exhaustible


resources. In this respect, Paul David has devoted a great deal of time andintellectual creativity to conceive of remedies to the current tendency tostrengthen intellectual property systems, especially as they apply to scien-tific research and scientific databases.

‘Knowledge on Line’: the Economics of Learning-by-doing

Learning-by-doing has been a key form of technical change analysed byPaul David in his studies in economic history, allowing him to address theissues of localised learning, the importance of history, and the policy impli-cations of supporting infant industries (see the section on path depend-ence).9 Apart from these works on economic history and path dependence,the most interesting contribution of David related to learning-by-doing isperhaps his emphasis on the fact that the economics of learning-by-doingappears to be an area in which the conflict between static and dynamicefficiency is particularly important. There is a tension between the normalperformance expected in the course of ordinary operations and the learningaspect:

In most instances of learning-by-doing, the feedback from experience to inferredunderstanding is severely constrained. The doers have limited facilities for accur-ately observing and recording process outcomes, or for hypothesizing about thestructure of the processes they are trying to control. Advances in knowledge thatare empirically grounded upon inferences from trial-and-error in a myopiccontrol process cannot be a big help when they are restricted in both the numberof trials they can undertake, and the states of the world they can imagine asworth considering.10

In a similar vein, David has been a pioneer in building the concept ofexperimental or explicitly cognitive learning-by-doing.11 Such a processconsists in performing experiments during the production of goods or ser-vices, or in the case considered by David and Sanderson, the production (ornon-production) of children. Through these experiments new options arespawned and variety emerges. This is learning based on an experimentalconcept, where data is collected so that the best strategy for future activitiescan be selected. Technical and organisational changes are then introducedas a consequence of learning-by-doing. In other words, explicitly cognitivelearning-by-doing consists of ‘on line’ experiments. The possibility ofmoving on to explicitly cognitive learning in activities other than ‘crafttrades’ represents an important transition in the historical emergence of theknowledge-based economy. As long as an activity remains fundamentallyreliant on learning processes that are procedures of routine adaptation andleave no room for deliberate planning of experiments during economic


activity, the gap between those who deliberately produce knowledge andthose who use and exploit it remains wide. When an activity moves on tohigher forms of learning where the individual can plan experimentsand draw conclusions, knowledge production becomes far more collectivelydistributed.

The Increasing Use of Codification . . .

Paul David has been a leader in the analysis of the economic significanceof knowledge codification both at the macro and micro levels. Codifiedknowledge serves inter alia as a storage depository, as a reference point andpossibly as an authority.12 As such, codification is a state in which know-ledge is presented (improving memory, learning and communication), a toolfor constructing new knowledge and a means to facilitate co-ordination. Inpointing out the important distinction between knowledge that is codifiable(in the sense of articulable) and that which actually is codified, and in focus-ing analytical attention upon the endogenous boundary between what is andwhat is not codified at a particular point in time, David and colleagues havehelped to persuade economists to ‘colonize’ this new area (previously leftopen to other social sciences) with their tools and concepts. At the macro-economic level, Abramovitz and David have analysed knowledge codifica-tion as both the driving force behind the expansion of the knowledge baseand its favourite form; in short, the most salient characteristic of moderneconomic growth.13

. . . and the Trap of ‘Taciturnity’

Investment in codification is suboptimal due to public good problems andhigh fixed costs. Any lack of attention paid to complementary componentsof a codified knowledge base (continuity of languages, software enablingaccess to older files) runs the risk of irremediably altering the codified know-ledge base and diminishing private and social returns from codificationinvestments.

One policy failure can, therefore, be seen in the lack of provision ofincentives to codify resulting in the building of excessive stocks of know-ledge that are left in a ‘tacit’ form. These accumulations of tacit knowledgeare distributed throughout the private sector in a way that makes themmore costly to locate, to appraise and to transfer. Policies that encouragestrategies of excessive tacitness by some firms thus tend to reduce incen-tives for other firms to invest in searching for existing codified solutions totheir scientific and technological problems. A result may be excessive insu-larity and a waste of resources. In lowering the private rate of return on


monitoring and information searches over a wide field, such strategies andpolicies will reinforce excess concentration of research and development(R&D) in certain areas, and contribute to the under-utilisation of existingstocks of knowledge. ‘Taciturnity’ (a word proposed by Paul in thiscontext) describes a deliberate failure to express feasibly codifiable infor-mation for strategic or cost reasons and may therefore create private andsocial inefficiencies.14

Information and Communication Technologies as a Knowledge Instrumentand a Mischievous Break

A central theme in Paul David’s work on knowledge is the information andcommunication technology (ICT) revolution in so far as it involves tech-nologies geared to the production and dissemination of knowledge andinformation. Apart from the productivity paradox issue, which is mainly a‘diffusion story’ (see section 4 in this chapter), David has developed sys-tematic analyses of the ways in which ICTs interact with organisational andinstitutional changes to profoundly transform the organisation of eco-nomic activities dealing with knowledge creation.

E-science is a particular field which warrants in-depth investigation inorder to evaluate how far the system of knowledge production and use istransformed through the full realisation of the potential of ICTs as know-ledge instruments. Scientists are now reaching the step of building andusing a comprehensive virtual federation of resources. ‘Comprehensive’means that the extent and nature of the resources available through thecyber-infrastructure (people, data, information, computational tools, spe-cialised instruments and facilities) could approach functional completenessfor a specific community of practice. However, institutional and organisa-tional issues have to be addressed so that old institutions and organisationsare not an impediment to the full and efficient deployment of resources andpotentialities in an e-science environment.15

Paul David is certainly a great ‘techno-optimist’ and places strong hopein the advances of new ICTs to improve the ways people are organisingtheir private and professional activities. However, he also knows very wellthat any success in collaboration and interaction with colleagues is contin-gent on ‘emotional trust’, namely, a sense of shared identity and familiar-ity. And this is not going to emerge spontaneously from long-distancecollaboration even if mediated through the best of the present ICT infra-structure. This is why Paul applies to himself a practical recommendationfor enhancing trust in geographically distributed teams, which is to increasetravel early in the history of the project and to travel again each time emo-tional trust requires some further support!


Public–Private Interactions and the Transferability of Knowledge

Direct transfers of knowledge between academic science communities andthe proprietary R&D organisations of the private business sector are espe-cially problematic to institutionalise. This is because the coexistence oftwo reward systems within any single organisation makes the behavioursof the participants difficult to anticipate, and tends to undermine the for-mation of coherent cultural norms which promote co-operation amongteam members. The main issue here is to maintain a proper balancebetween, on the one hand, the requirements of openness and autonomyof investigation (as these are required for the rapid growth of the stock ofknowledge) and, on the other, the need for delays and restrictions uponthe full disclosure of all new information (which facilitate the appropri-ation of economic returns needed to sustain investment in expanding theknowledge base).16

David is a key player in the policy debate on the current transformationsof the relationships between science, technology and economic perform-ance, and on the various institutional mechanisms to be designed to get abetter protection of the public domain of knowledge from furtherencroachments by the domain of private property rights. A constant argu-ment in his work is that the basic rationale of intellectual property lawdepends on an independent public domain containing a stock of freelyaccessible information. That shared collection of basic knowledge providesthe building blocks for new inventions.17

How will it be possible to maintain and expand this collection of freelyaccessible basic knowledge in the long term? What kind of mechanismsshould be designed to preserve the intellectual commons, given that theymust be ‘incentive compatible’ with the private allocation of resources toinventive and innovative efforts? These are the ‘big questions’ in his explor-ation of the future economic organisation of a well-functioning science andtechnology system.18

A General Insight on this Work

As a rationale for his development of such a rich repertoire of works andstudies, there is in Paul David’s work a sense of inadequacies and erroneousdevelopment in the passage between the understanding that economistshave gained about the very detailed mechanisms of knowledge productionand distribution, and the stylised facts encapsulated in the formalities ofmacroeconomic models. This is perhaps why he continuously probesdeeper in the microanalysis of detailed resource allocation processes invarious areas of the economics of knowledge (open science, open source,


proprietary R&D, knowledge transfer, intellectual property rights), mobil-ising for such purposes not only the microeconomics of innovation but alsohistorical analysis, the sociology of science, legal studies and the manager-ial and organisational literature.

However, the most important feature is that such deep ‘drilling’ into verycomplex structures always leads to the production of very rich stylised factsthat have the destiny to inform the rest of the economic profession inter-ested in the determinants of innovation, productivity and economicgrowth – the foundations of the macroeconomy. Sometimes David himselfuses these stylised facts to enrich his own macroeconomic works on eco-nomic growth. The value of such application is perhaps best illustrated inthe recent set of studies on the economic history of the American macro-economic growth, published with the late Moses Abramovitz.19

The chapters in this volume by Steinmueller, Foray and Hilaire Perez,Mairesse and Turner, Arora, Fosfuri and Gambardella, Cohendet andAmin and Brusoni and Geuna are all contributions to this field in thevarious perspectives marked out by David on the economics of informa-tion and knowledge.

3 PATH DEPENDENCE

During the twentieth century, economists were finally able to combine thework of Vilfredo Pareto and Marie-Ésprit Léon Walras into the theory ofgeneral competitive analysis. The assumptions employed by KennethArrow and Gerard Debreu were numerous, giving rise to both contempor-ary critiques and rejoinders by those with alternative propositions con-cerning the fundamental features of market-based economies and actors.The theory of ‘path dependence’ is often, incorrectly, taken as being amongthese ‘alternative’ approaches to understanding the nature of competitiveequilibrium. It is, instead, an effort to address an issue that general com-petitive analysis leaves unresolved. Namely, what are the consequences ofmoving between different equilibria as new technologies emerge and areintegrated into the market system?

For a scholar of economic history, such as Paul David, answering thisquestion was of significance not because ‘history matters’ for that is tooeasy a target. The more important question is how history might matter.20

David has provided an important answer to this question by developing themodern theory of path dependence, among his most widely cited and leastwell understood contributions to the modern vocabulary of economicanalysis. The theory of path dependence is an attempt to portray the con-sequences that may arise when the processes of growth and accumulation


result in reaching one equilibrium rather than another – to explain in arigorous way how history might matter.

At first glance, it would seem that path dependence contests the basicresult of general competitive analysis – the achievement of a unique equi-librium through Walrasian and other market-clearing mechanisms thataccords with Pareto’s statement of social welfare criteria. The importantpoint is that this equilibrium is at a single point in time.21 It does notaccount for the unanticipated arrival of the new technologies or othershocks and disturbances that cannot be anticipated and therefore incorpor-ated in the trading equilibrium. When moving between different points intime sufficiently distant to provide an opportunity for intervening events,like the arrival of new technology, a new equilibrium will be reached. Thusit is possible to achieve an array of equilibrium positions in the economyover time depending upon the order, nature and timing of interveningarrivals of technology or other shocks.

The foregoing observations only serve to establish the existence of mul-tiple equilibriums that might be reached depending upon the differentorder, nature and timing of intervening variables. The evolution ofeconomies is thus subject to random disturbances and hence fundamen-tally unpredictable. This conclusion may be reached from a number of ana-lytical directions and is, itself, unremarkable. As long as the class of allequilibria remains open for further exploration, the economy is not pathdependent and, from a probability theory viewpoint, the processes of tra-versing the various equilibria is ergodic (without memory). What distin-guishes the class of ‘path dependent’ equilibriums is that they foreclosereaching other classes of equilibriums. This is a much stronger and morecontroversial conclusion than arguing that the path of economic develop-ment is uncertain.22 The nature of the foreclosure is that the costs of switch-ing to another path involve a substantial diminution of welfare for thosethat might benefit from the change. This welfare reduction is so large thatit makes the decision to ‘switch’ irrational.

The key point, however, is that if past movements had led in a differentdirection, a path not taken, a superior equilibrium could have been reached.Recognising that ‘paths not taken’ may have created an entirely different setof economic outcomes is a direct challenge to the interpretation of the suc-cession of competitive equilibriums either as unambiguously progressiveand leading to the ‘best of all achievable worlds’ (subject to the constraintsof resource endowments and technology) or the near equivalent, the ideaof progress as an ergodic process in which better outcomes will come intime.23

The foreclosure of some equilibria amounts to a ‘lock in’ of actors ator near the vicinity of a particular equilibrium. The most straightforward


means for generating such lock-in effects is through local feedback and themost prominent example of such local feedback effects are the networkexternalities that accompany the widespread adoption of a particulartechnology or method of organisation. Network externalities that sub-stantially lower the costs of one technology or method of organisationraise the costs of switching to alternatives. To illustrate this phenomenonDavid has employed a number of examples including the case of theQWERTY keyboard where he argues that the local feedback occurredthrough the decisions of typists to learn ‘touch typing’ and that the ‘lockin’ was reinforced by the ‘quasi-irreversibility’ (high costs of switching) ofthis skill.24

In his 1985 QWERTY article, David chose a cosmological metaphor toconvey the potential significance of path-dependent processes, comparingthese processes to the ‘dark stars’ that influence the ordering of our uni-verse. In the intervening years, the significance of the less obvious con-stituents of our economic universe is receiving more attention.

In particular, the role of adoption externalities in information technolo-gies including software have not only been recognised, but have been incor-porated in strategy. In considering the role of technical compatibilitystandards in accelerating the diffusion of new technologies, David inventednew meanings for the narrow windows of opportunity for intervention inadoption processes, conjured the spectre of angry orphans that would becreated as standards tipped against their (incorrect) choices of the emerg-ing dominant technology, and reanimated the Cyclops ‘blind giant’ – gov-ernment groping to find a path after the ‘white heat’ of technology hadrendered its previous competences obsolete.25

The possibility of ‘premature’ standardisation with the consequence ofsocial welfare losses has become an important component of policy delib-erations and a recognised source of ‘market failure’. David and colleagueshave provided a series of analyses of technological history in which path-dependent processes play a central role. These include the history of elec-trical power,26 data communication standards27 and the mechanisation ofcorn harvesting.28 These contributions as well as his commentaries on thenature of path dependence realise one of David’s central ambitions – toadvance the frontier of understanding how history matters in economicanalysis.

The chapters in this volume by A. Bassanini and G. Dosi, C. Antonelli,F. Malerba and L. Orsenigo, J. Cantwell, P. Swann and R. Cowan not onlypush the frontier of studies of path dependency forward, they also demon-strate the attractive force of the path dependence as a method for economicanalysis.


4 DIFFUSION

A superficial reading of economics would suggest that the process ofexploiting technological and market opportunities is nearly instantaneous –every delay in exploiting opportunity constitutes an opportunity to makegains at the expense of rivals. If a firm cannot reap the first-mover advan-tage, it must strive to be a ‘fast second’ and woe betide those firms or coun-tries that are laggards in the race towards higher levels of prosperity. Thestudy of economic history provides a useful antidote to this febrile frenzy.Through the lens of economic history one can see that even highly promis-ing developments such as the industrial revolution take decades to spreadfrom one country to another, that innovation often requires new people aswell as new machines and that new methods of economic and social organ-isation displace traditional methods slowly through processes of cumula-tive change and adaptation. Thus the systematic study of the rate anddirection of technological change involves understanding how differenteconomic actors take up or adopt changes that become available. PaulDavid has contributed to our understanding of the diffusion of innovationsthrough both theoretical and empirical investigations from the very begin-ning of his career.

While the English language use of the term diffusion in relation toknowledge can be found in the writings of Richard Price (1723–91) andJames Madison (1751–1836), the modern study of the diffusion of innova-tions involves an effort to understand why all who might benefit from it donot instantaneously adopt an innovation. One answer is that the diffusionof information and knowledge is uneven, another is that the value of a newinnovation depends upon the characteristics of the adopter, and still a thirdis that the nature of innovation improves over time, raising the benefit ofadoption. Zvi Griliches’ pioneering exploration of the diffusion of hybridcorn involves all three sources of explanation.29

David’s study of the mechanisation of reaping in the antebellum Midwestmade two very important contributions to this emerging literature.30 First,he positioned the study as an effort to offer an alternative to the existingexplanation in which mechanical reaping was simultaneously the conse-quence of a shortage of labour required to fully exploit the rise in wheatprices during the 1850s, the extension of the planting of wheat in theAmerican Midwest and the growing prosperity of the wheat farmers. AsDavid observes, this explanation falls well short of explaining why an indi-vidual farmer with a particular size farm would choose to replace the morelabour-intensive cradle technology with the mechanical reaper. Second,David undertook a comprehensive review of ‘what changed’ and whatremained static over the period during which reaper use expanded, reducing


the model of innovation diffusion to the question of the threshold farm sizethat would benefit from adopting the new technology. This second contri-bution was an important advance in the study of diffusion. It explicitly setout the comparison between the technology being replaced and the innova-tive technology – establishing a path that remains fruitful in the analysis ofdiffusion of new technologies.

David’s approach to the economic analysis of diffusion is best under-stood by considering the ‘naive’ or ad hoc approach to new technologyadoption processes. Many ad hoc approaches simply dismiss theory andtake the econometric estimation of a logistic diffusion curve as an applica-tion of the ‘well accepted diffusion curve’. Others employ the metaphor ofinformation distribution as contagion, in which a simple specification ofthe percolation of information through a social network provides a basisfor a logistic diffusion curve. Such approaches bypass consideration of themechanisms of adjustment through learning, improvements of competingtechnologies and the effects of network externalities that play importantroles in David’s later works on diffusion.31

The economic theory of diffusion is often seen by other social scientistsas being overly deterministic. Alternative specifications that rely upon cog-nitive limitations and processes of social negotiation such as those involvedin opinion formation are, however, not inherently less deterministic.Theories that rely upon factors that are non-observable before the ‘trial’ ofan actual diffusion process are, in essence, explanatory rather than predict-ive theories. Studies of diffusion based upon these theoretical foundationsare necessarily either retrospective – one can only deduce facts about cog-nitive limits or the percolation of knowledge through social networks in ref-erence of actual experience – or such examinations are speculative – newertechnologies may be taken as ‘similar’ to older ones providing a means tomake a speculative linkage between past and future experience. The advan-tage of economic theories of diffusion is that they direct the attention ofresearchers to the margin, the ordering of individual adopters at the bound-ary between adoption and non-adoption. As time moves on, examining theshifting margin provides the opportunity to sort out whether the factorsresponsible for adoption are essentially structural – that is, whether changesin price, quality or the growing availability of complements bound andorder the underlying heterogeneity of the adopting population – or whetherthe margin, the boundary between non-adoption and adoption, is shapedby dynamic processes of change, such as learning, within the non-adoptingpopulation.

In his famous ‘underground classic’,32 David extended the theory ofdiffusion, reconciling his own work with that of the other students of thediffusion process such as Griliches and Mansfield. The chapters in this


volume by B.H. Hall and M. Trajtenberg, P. Stoneman and O. Toivanenand L. Cabral further extend the basic framework building upon the foun-dations laid by Paul David and other ‘first-generation’ scholars of themodern economic theory of diffusion.

NOTES

1. P.A. David and D. Foray, ‘Accessing and expanding the science and technology know-ledge base’, STI, Review, 16, 1995, 13–68.

2. R. Coase, ‘The lighthouse in economics’, Journal of Law and Economics, 17 (2), 1974,357–76.

3. P.A. David, a reply to ‘Lock and key’, Economic Focus, The Economist, 18 September1999, unpublished draft, 2000.

4. P.A. David, ‘Knowledge, property and the system dynamics of technological change’,Proceedings of the World Bank Annual Conference on Development Economics 1992,Washington, DC: World Bank, 1999.

5. P. Dasgupta and P.A. David, ‘Towards a new economics of science’, Research Policy, 23 (5),1994, 487–521.

6. P.A. David, ‘Common agency contracting and the emergence of open science’, AmericanEconomic Review, 88 (2), 1998, 15–21.

7. P.A. David, ‘Communication norms and the collective cognitive performance of “invis-ible colleges”’, in G. Navaretti, P. Dasgupta, K. Mäler and D. Siniscalco (eds), Creationand Transfer of Knowledge, Heidelberg: Springer, 1998, pp. 115–63.

8. P.A. David, ‘Digital technologies, research collaboration and the extension of protectionfor intellectual property in science: will building “good fences” really make “good neigh-bors”?’, IPR Aspects of Internet Collaborations, Final Report, EUR 19456, EuropeanCommission, 2001.

9. P.A. David, ‘Learning-by-doing and tariff protection: a reconsideration of the case ofthe ante-bellum United States cotton textile industry’, in P.A. David, Technical Choice,Innovation and Economic Growth, Cambridge: Cambridge University Press, 1975,pp. 95–173.

10. P.A. David, ‘Path dependence and varieties of learning in the evolution of technologicalpractice’, in J. Ziman (ed.), Technological Innovation as an Evolutionary Process,Cambridge: Cambridge University Press, 1999, pp. 118–33.

11. P.A. David and W. Sanderson, ‘Making use of treacherous advice: cognitive progress,Bayesian adaptation and the tenacity of unreliable knowledge’, in J. Nye and J. Droback(eds), Frontiers of the New Institutional Economics, Oxford: Academic Press, 1997,pp. 305–66.

12. R. Cowan, P.A. David and D. Foray, ‘The explicit economics of knowledge codificationand tacitness’, Industrial and Corporate Change, 9 (2), 2000, 211–54.

13. M. Abramovitz and P.A. David, ‘American macroeconomic growth in the era ofknowledge-based progress: the long run perspective’, in S. Engerman and R. Gallman(eds), The Cambridge Economic History of the United States, vol. 13, Cambridge:Cambridge University Press, 2000, pp. 1–92.

14. P.A. David, paper delivered at the last TIPIK Conference, Strasbourg, UniversitéPasteur, March 2001.

15. P.A. David and M. Spence, Towards Institutional Infrastructure for E-Science, Oxford:Oxford Internet Institute, 2003.

16. P.A. David, D. Foray and W.E. Steinmueller, ‘The research network and the new eco-nomics of science: from metaphors to organisational behavior’, in A. Gambardella andF. Malerba (eds), The Organisation of Inventive Activity in Europe, Cambridge:Cambridge University Press, 1999, pp. 303–42.


17. P.A. David and B.H. Hall, ‘Heart of darkness: public–private interactions inside theR&D black box’, Research Policy, 29, 2000, 1165–83.

18. P.A. David, ‘The political economy of public science: a contribution to the regulation ofscience and technology’, in L. Smith (eds), The Regulation of Science and Technology,London: Palgrave, 2001, pp. 33–57.

19. M. Abramovitz and P.A. David, ‘American macroeconomic growth in the era of know-ledge-based progress: the long run perspective’, in S. Engerman and R. Gallman (eds),The Cambridge Economic History of the United States, vol. 13, Cambridge: CambridgeUniversity Press, 2000, pp. 1–92.

20. David’s first ruminations on this question occur in the extended introduction ofTechnical Choice, Innovation and Economic Growth: Essays on American and BritishExperience in the Nineteenth Century, Cambridge and New York: Cambridge UniversityPress, 1975. (Second edition, forthcoming in 2004 from Cambridge University Press.)

21. At a point in time does not mean timeless; economic actors may well have expectations(rational or otherwise) about the future. Their expectations are, however, circumscribedby a set of state spaces that are subject to unanticipated change.

22. P.A. David, ‘Path dependence and predictability in dynamic systems with local networkexternalities: a paradigm for historical economics’, in D. Foray and C. Freeman (eds),Technology and the Wealth of Nations, London: Pinter, 1992, pp. 209–31.

23. The critics of the idea of path dependence seem particularly alarmed by this prospect,see P.A. David, ‘Path dependence, its critics and the quest for “historical economics” ’,Evolution and Path Dependence in Economic Ideas: Past and Present, edited byP. Garrouste and S. Ionnides, Cheltenham: Edward Elgar, 2001, pp. 15–40.

24. P.A. David, ‘Clio and the economics of QWERTY’, American Economic Review, 75 (2),May 1985, 332–37.

25. P.A. David, ‘Some new standards for the economics of standardization in the informa-tion age’, in P. Dasgupta and P.L. Stoneman (eds), Economics and TechnologicalPerformance, Cambridge: Cambridge University Press, 1987, pp. 206–39.

26. P.A. David, ‘The hero and the herd in technological history: reflections on ThomasEdison and “The Battle of the Systems” ’, in P. Higgonet, D. Landes and H. Rosovsky(eds), Favorites of Fortune: Technology, Growth, and Economic Development Since theIndustrial Revolution, Cambridge, MA: Harvard University Press, 1991, pp. 72–119;P.A. David and J.A. Bunn, ‘Gateway technologies and the evolutionary dynamics ofnetwork industries: lessons from electricity supply history’, in A. Heertje andM. Perlman (eds), Evolving Technology and Market Structure, Ann Arbor, MI:University of Michigan Press, 1990, pp. 121–56.

27. P.A. David and W.E. Steinmueller, ‘The ISDN bandwagon is coming – but who will bethere to climb aboard? Quandaries in the economics of data communication networks’,Economics of Innovation and New Technology, 1 (1 & 2), Fall 1990, 43–62; P.A. Davidand D. Foray, ‘Dynamics of competitive technology diffusion through local networkstructures: the case of EDI document standards’, in L. Leydesdorff and P. van denBesselaar (eds), Evolutionary Economics and Chaos Theory, London: Pinter, 1994,pp. 63–77.

28. P.A. David, ‘The landscape and the machine: technical interrelatedness, land tenure andthe mechanization of the corn harvest in Victorian Britain’, in D.N. McCloskey (ed.),Essays on a Mature Economy, London: Methuen, pp. 145–205.

29. Z. Griliches, ‘Hybrid corn: an exploration in the economics of technical change’,Econometrica, 25 (4), 1957, 501–22.

30. P.A. David, ‘The mechanization of reaping in the ante-bellum Midwest’, in H. Rosovsky(ed.), Industrialization in Two Systems: Essays in Honor of Alexander Gerschenkron,New York: Wiley and Sons, 1966, pp. 3–39.

31. For example, P.A. David, ‘Technology diffusion, public policy, and industrial competi-tiveness’, in R. Landau and N. Rosenberg (eds), The Positive Sum Strategy: HarnessingTechnology for Economic Growth, Washington, DC: National Academy Press, 1986,pp. 373–91; P.A. David and P.L. Stoneman, ‘Adoption subsidies vs. information provision


as instruments of technology policy’, Economic Journal, 96 (Supplement), March 1986,142–50; P.A. David and T.E. Olsen, ‘Equilibrium dynamics of diffusion when incremen-tal technological innovations are foreseen’, Ricerche Economiche (Special Issue onInnovation Diffusion), 40 (4), October–December 1986, 738–70.

32. P.A. David, ‘A contribution to the theory of diffusion’, Stanford University Center forResearch in Economic Growth Research Memoranda Nos 71, 72, 73, June 1969,mimeograph.


PART II

Path Dependence in Technical Change

2. Competing technologies,technological monopolies and therate of convergence to a stablemarket structure*

Andrea P. Bassanini and Giovanni Dosi

1 INTRODUCTION

In this chapter we address the dynamics of diffusion of different technolo-gies competing for the same market niche.

The stylised fact at the origin of this work is the observation that a stableempirical pattern of market sharing between competing technologies withno overwhelming dominant position rarely occurs in markets with positivefeedbacks.1 For example, even in the case of operating systems, which isoften quoted as a case of market sharing, Apple MacIntosh has never helda market share larger than 1/5 (a partial exception being the submarket ofpersonal computers for educational institutions). This fact has also trig-gered suspicion of market inefficiencies: technological monopolies mayprevail even when the survival of more than one technology may be sociallyoptimal (Katz and Shapiro, 1986; David, 1992). Think for example of thecompetition between Java-based architectures and ActiveX architecturesfor web-based applets: given that with any of the two paradigms the stand-ard tasks that can be performed are different, the general impression ofexperts is that society would benefit from the survival of both.

In turn, from the point of view of interpretation of the processes ofdiffusion of new products and technologies, it is acknowledged that, inmany modern markets, they are characterised by increasing returns toadoption or positive feedbacks. This has partly to do with supply-sidecauses: the cumulation of knowledge and skills through the expansion ofmarkets and production usually reduce the hedonic price of both produc-tion and consumption goods, thus increasing the net benefit for the user ofa particular technology. The Boeing 727, for example, which has been onthe jet aircraft market for years, has undergone constant modification ofthe design and improvement in structural soundness, wing design, payload

23

capacity and engine efficiency as it accumulates airline adoption and hoursof flight (Rosenberg, 1982; Arthur, 1989). Similar observations can bemade for many helicopter designs (Saviotti and Trickett, 1992) as well as forelectric power plants designs (Cowan, 1990; Islas, 1997).

Supply-side causes of this type have received some attention in the eco-nomic literature for quite a while. However, in the last 15 years a great dealof attention has been devoted also to demand-side positive feedbacks,so-called network esternalities or (more neutrally) network effects (Katzand Shapiro, 1994; Liebowitz and Margolis, 1994). For example, telecom-munication devices and networks (for instance, fax machines), as a firstapproximation, tend not to provide any utility per se but only as a functionof the number of adopters of compatible technologies with whom the com-munication is possible (Rohlfs, 1974; Oren and Smith, 1981; Economides,1996). The benefits accruing to a user of a particular hardware systemdepend on the availability of software whose quantity and variety maydepend on the size of the market if there are increasing returns in softwareproduction. This is the case of video cassette recorders (VCRs), micro-processors, hi-fi devices and in general systems made of complementaryproducts which need not be consumed in fixed proportions (Cusumanoet al., 1992; Church and Gandal, 1993; Katz and Shapiro, 1985; 1994). Asimilar story can be told for the provision of post-purchase service fordurable goods. In automobile markets, for example, the diffusion of foreignmodels has often been slow because of consumers’ perception of a thinnerand less experienced network of repair services (Katz and Shapiro, 1985).Standardisation implies also saving out of the cost of investment in com-plementary capital if returns from investment are not completely appropri-able: in software adoption firms can draw from a large pool of experiencedusers if they adopt software belonging to a widespread standard, thus defacto sharing the cost of training (Farrell and Saloner, 1986; Brynjolfssonand Kemerer, 1996). Moreover product information may be more easilyavailable for more popular brands or, finally, there may be conformity orpsychological bandwagon effects (Katz and Shapiro, 1985; Banerjee, 1992;Arthur and Lane, 1993; Bernheim, 1994; Brock and Durlauf, 1995).

Katz and Shapiro (1994) in their review of the literature on systems com-petition and dynamics of adoption under increasing returns distinguishbetween technology adoption decision and product selection decision.

The former refers to the choice of a potential user to place a demand ina particular market. Relevant questions in this case are the conditions foran actual market of positive size, the notional features of a ‘sociallyoptimal’ market size and the conditions allowing penetration of a new(more advanced) technology into the market of an already established one(Rohlfs, 1974; Oren and Smith, 1981; Farrell and Saloner, 1985; 1986; Katz

24 Path dependence in technical change

and Shapiro, 1992). For example, purchasing or not a fax or substituting acompact disc player for an analogical record player are technology adop-tion decisions.

Conversely product selection refers to the choice between different tech-nological solutions which perform (approximately) the same function andare therefore close substitutes. Relevant questions here are whether themarket enhances variety or standardisation, whether the emerging marketstructure is normatively desirable and what is the role of history in the selec-tion of market structure (Arthur, 1983; 1989; Katz and Shapiro, 1985; 1986;David, 1985; Church and Gandal, 1993; Dosi et al., 1994). Choosingbetween VHS or Beta in the VCR market or between Word or Wordperfectin the word-processor market are typical examples of product selectiondecisions.

This work is concerned with the dynamics of product selection. Toexplain the stylised fact recalled above we analyse properties of a fairlygeneral and nowadays rather standard class of models of competing tech-nologies, originally suggested by Arthur (1983) and Arthur et al. (1983) andfurther explored by Arthur (1989), Cowan (1991) and Dosi et al. (1994),among others. This class of models will be presented in details in section 2.

Despite mixed results of some pioneering works on the dynamics ofmarkets with network effects (for example, Katz and Shapiro, 1986),unbounded increasing returns are commonly called for as an explanationof the emergence of technological monopolies. Usually the argument isbased on the results of the model set forth by Arthur (1989). For instance,Robin Cowan summarises it in the following way:

If technologies operate under dynamic increasing returns (often thought of interms of learning-by-doing or learning-by-using), then early use of one tech-nology can create a snowballing effect by which that technology quickly becomespreferred to others and comes to dominate the market.

Following Arthur, consider a market in which two types of consumers adopttechnology sequentially. As a result of dynamic increasing returns arising fromlearning-by-using, the pay-off to adopting a technology is an increasing functionof the number of times it has been adopted in the past. Important with regardto which technology is chosen next is how many times each of the technologieshas been chosen in the past. Arthur shows that if the order of adopters is random(that is, the type of the next adopter is not predictable) then with certainty onetechnology will claim the entire market. (Cowan, 1990: 543, italics added)

It will be shown in the following that this statement does not always hold.Unbounded increasing returns to adoption are neither necessary norsufficient to lead to the emergence of technological monopolies. As provedin the next section, strictly speaking, Arthur’s result applies only whenreturns are linearly increasing and the degree of heterogeneity of agents is, in

Competing technologies, technological monopolies 25

a sense, small. Moreover it cannot be easily generalised further: some mean-ingful counter-examples will be provided. More generally the emergence oftechnological monopolies depends on the nature of increasing returns andtheir relationship with the degree of heterogeneity of the population. Givena sufficiently high heterogeneity amongst economic agents, limit marketsharing may occur even in the presence of unbounded increasing returns.

The bearing of our analysis, in terms of the interpretation of the empir-ical evidence, stems from the results presented in section 3: in essence, wesuggest that the observation of the widespread emergence of monopolies isintimately related to the properties of different rates of convergence (tomonopoly and to market sharing respectively) more than to the propertiesof limit states as such. It will be shown that a market can approach amonopoly with a higher speed than it approaches any feasible limit marketshares where both technologies coexist. Following a line of reasoning putforward by Winter (1986), our argument proceeds by noticing that whenconvergence is too slow the external environment is likely to change beforeany sufficiently small neighbourhood of the limit can be attained. Theresult that we obtain, based on some mathematical properties ofGeneralised Urn schemes,2 is general for this class of models. The empiri-cal implication is that among markets with high rate of technologicalchange and increasing returns to adoption, a prevalence of stable monop-olies over stable market-sharing should be observed.

The applications of Arthur’s result have gone far beyond the dynamicsof competing technologies and typically extended to the role of history inselecting the equilibrium in any situation wherein complementarities arerelevant. The analysis of industry location patterns is a case to the point(for example, Arthur, 1990; Krugman, 1991a; 1991b; Venables, 1996). AsJames Rauch puts it:

In Arthur’s model, firms enter the industry in sequence. Each firm chooses alocation on the basis of how many firms are there at the time of entry and arandom vector that gives the firm’s tastes for each possible location. If agglom-eration economies are unbounded as the number of firms increases, then as theindustry grows large, one location takes all but a finite set of firms with probabil-ity one. (Rauch, 1993: 843–4, italics added)

The implications of our results extend to this domain of analysis as well.The remainder of the chapter is organised as follows. Section 2 reviews

standard models of competing technologies and provides counter-examplesto Arthur’s main result. Section 3 establishes our main results on rate ofconvergence to a stable market structure and builds upon that an alterna-tive explanation for observable patterns of dynamics between competingtechnologies. Section 4 briefly summarises the results.


2 COMPETING TECHNOLOGIES REVISITED:ARE UNBOUNDED INCREASING RETURNSSUFFICIENT FOR THE EMERGENCE OFTECHNOLOGICAL MONOPOLIES?

The class of competing technology dynamics models that we considershares the two basic assumptions that adopters enter the market in asequence which is assumed to be exogenous, and that each adopter makesits adoption choice only once. More than one agent can enter the market ineach period (for example, Katz and Shapiro, 1986) but in order to simplifythe treatment we abstract from this complication. The simple theoreticaltale that underlies these models can be summarised as follows.

Every period a new agent enters the market and chooses the technologywhich is considered best suited to its requirements, given its preferences,information structure and the available technologies. Preferences can beheterogeneous and a distribution of preferences in the population isgiven. Information and preferences determine a vector of pay-off functions(whose dimension is equal to the number of available technologies) forevery type of agent. Because of positive (negative) feedbacks, such asincreasing (decreasing) returns to adoption, these functions depend on thenumber of previous adoptions. When an agent enters the market it com-pares the values of these functions (given its preferences, the available infor-mation, and previous adoptions) and chooses the technology which yieldsthe maximum perceived pay-off. Which ‘type’ of agent enters the market atany given time is a stochastic event whose probability depends on the dis-tribution of types (that is, of preferences) in the population. Because ofpositive (negative) feedbacks, the probability of adoption of a particulartechnology is an increasing (decreasing) function of the number of previ-ous adoptions of that technology.

More formally we can write a general reduced form of pay-off functionsof the following type:

where j � D, D is the set of possible technologies, i �S; S is the set of pos-sible types; n→(t) is a vector denoting number of adoptions for each tech-nology at time t (nj(t)) is the number of adoptions of technology j at timet); a→i represents the network-independent components of agent i’s prefer-ences ( identifies a baseline pay-off for agents of type i from technology j),and hi(�) is an increasing (decreasing) function that can differ across agentscapturing increasing (decreasing) returns to adoption. Information andexpectations are incorporated in hi(�). If, at time t, an agent of type

a ji

ij( n→(t)) hi(a

ij, n→(t)),


i comes to the market, it compares the pay-off functions, choosing A if andonly if:3

(2.1)

Equation (2.1) can be seen as describing an equilibrium reaction func-tion. Consequently, strategic behaviours (including sponsoring activitiesfrom the suppliers of technologies) are not ruled out by the foregoing for-malisation.

In the remainder of this chapter we assume that the order of agents enter-ing the market is random, hence i(t) can be considered as an iid sequence ofrandom variables whose distribution depends on the distribution of thepopulation of potential adopters. Under this assumption, the dynamics ofthe foregoing model can be analysed in terms of generalised urn schemes.

Consider the simplest case where two technologies, say A and B, competefor a market. Let us denote A’s market share with X(t). Given the relation-ships between (a) total number of adoptions of both technologies n(t)t � 1�nA(0)�nB(0), (b) the current market share X(t) of A, and (c) numberof adoptions of one specific technology, ni(t), iA, B, that is, nA(t)n(t)X(t), the dynamics of X(t) is given by the recursive identity

Here t (x), t � 1 are random variables independent in t such that

and t(�) is a function of market shares dependent on the feedbacks inadoption. f(t, x) equals the probability that (2.1) is true when X(t)x andis sometimes called urn function. Denoting t(x)�E( t(x)) t(x)�f(t, x)with �t(x) we have

(2.2)

Provided that there exist a limit urn function f (�) (defined as that func-tion f (�) such that f(t,.) tends to it as t tends to �) and the following condi-tion is satisfied

(2.3)�t�1

t�1 sup x� |0,1|�R(0,1)

| f (t, x) � f (x)| � �,

X(t � 1) X(t) �[ f (t, X(t)) � X(t)] � �t(X(t))

t � nA(0) � nB(0).

t(x) �1 with probability f (t, x)0 with probability 1 � f (t, x)

,

X(t � 1) X(t) � t(X(t)) � X(t)

t � nA(0) � nB(0).

iA( n→(t)) arg max

j�D {ij ( n→)}


where R(0,1) is the set of rational numbers in (0,1), limit market sharesattainable with positive probability can be found by analysing the proper-ties of the function

Particularly, treating g(x) in the same way as the right-hand side of anordinary differential equation, it is possible to show that the process (2.2)converges almost surely to the set of stable zeros.4 The foregoing formalrepresentation is employed for every result of the present chapter.

In some cases, equation (2.1) can be expressed directly in terms of sharesrather than total numbers: in this case f(.,.) is independent of t and (2.3) iseasily verified.

The foregoing formal model can be better visualized by looking at somewell-known example. Consider for instance the celebrated example of theVCR market. JVC’s VHS and Sony’s Beta were commercialised approxi-mately at the same time. According to many studies (see Cusumano et al.,1992; Liebowitz and Margolis, 1994), none of the two standards has everbeen perceived as unambiguously better and, despite their incompatibil-ity, their features were more or less the same, due to the common deriv-ation from the U-matic design. The relevant decisions were likely to besequential. First, a consumer chooses whether or not to adopt a VCR –technology adoption decision in Katz and Shapiro’s terminology. Then,once the adoption decision has been made, it devotes its mind to choosewhich type of VCR to purchase – product selection decision – (in generalit can be expected that most of the consumers buy one single itemand not both). Network effects in this market come mainly from increas-ing returns in design specialisation and production of VCR models (sothat historically all firms specialised just in one single standard) on thesupply side, and from increasing returns externalities and consequentavailability of home video rental services on the demand side (Cusumanoet al., 1992). Despite technical similarities between the two standards,preferences are likely to be strongly heterogeneous, due mainly to a brand-name-loyalty in consumer behaviour, which was in fact exploited (espe-cially by JVC) through original equipment manufacturers’ (OEM)agreements with firms with well-established market shares in electronicdurable goods.

The size of VCR market is sufficiently large (hundreds of millions ofsold units) to make it approximable by the abstract concept of an infinitecapacity market. Therefore, the asymptotic dynamics of this market maybe meaningfully analysed through the asymptotics of generalised urnschemes.

g(x) f (x) � x lim t→�

f (t, x) � x.


Many other markets display somewhat similar characteristics (forinstance, spreadsheets, wordprocessors, computer keyboards, Pc hardware,automobiles and so on). In particular, in many markets product selectioncan be assumed to sequentially follow technology adoption decisions.5 Thefact that decisions are sequential suggests that product selection decisionsmight be dependent on market shares rather than on the absolute size ofthe network. In this case the urn scheme would be even more simplified,with the urn function independent of t.

Arthur (1983; 1989) considers a pay-off function of the following type:

where jA, B, i�S, S is the set of possible types [in the simplest case,considered also in the above quotation from Cowan (1990), S{1,2}],and r is an increasing function (common for every agent) capturingincreasing returns to adoption. If, at time t, an agent of type i comesto the market, it compares the two pay-off functions choosing A if andonly if:

that is

(2.4)

Suppose that the selection of which type of agent enters the market attime t is the realisation of an iid random variable i(t). Thus (2.2) impliesthat the agent coming to the market chooses A with probability

where F�(�) denotes the distribution function of �(t)aiB(t)�ai

A(t).From these considerations Arthur’s main theorem was derived:

Theorem 1 (Arthur (1989), Theorem 3)

If the improvement function r increases at least at rate ��0 as nj increases,the adoption process converges to the dominance of a single technology, withprobability one.

The proof of the theorem offered by Arthur is based on theorem 3.1 ofArthur et al. (1986). In fact it is easy to check that in this case, whatever thedistribution of aj is, the limit urn function f(�) is a step function defined inthe following way:

P(A(t)) F0(r(nA(t) � r (nB(t))),

aiA � r(nA(t)) � ai

B � r (nB(t)).

iA( n→(t)) � i

B( n→(t)),

ij( n→(t)) ai

j � r(n j(t)),


(2.5)

A generaliszed urn scheme characterised by an urn function such as (2.5)converges to {0,1} with probability 1.6

However theorem 3.1 of Arthur et al. (1986) is not applicable herebecause condition (2.3) does not hold in this case. Actually the urn func-tions are defined by:

ft(x)F�(r (x (t�nw�nb))�r((l�x)(t�nw�nb))).

Moreover for t�K�0, t even, they are such that ft(0)0, ft(l/2)F�(0),ft(l)1 and they are continuous in a left neighbourhood (which dependson t) of 1/2; therefore

which is constant with respect to t.7

Even though Arthur’s proof is wrong, the theorem is right and an adhoc proof can be constructed by showing that nA(t)�nB(t) is a time-homogeneous Markov chain with two absorbing barriers (Bassanini, 1997,proposition 2.1, provides a complete proof along these lines). However thisresult strictly depends on the fact that the function r(�) is asymptoticallylinear or more than linear. Arthur’s result is not generalisable to any typeof unbounded increasing returns. Both in the case of increasing returnsthat are diminishing at the margin and in the case of heterogeneous increas-ing returns it is possible to find simple examples where convergence to tech-nological monopolies is not an event with probability 1.

Let us illustrate all this by means of two straightforward counter-examples.

Example 1Let us assume that increasing returns have the common sense property thatthe marginal contribution to social benefit of, say, the 100th adopters islarger than that of, say, the 100 000th and that this contribution tendsasymptotically to zero; formally this means that

(this class of functions has been considered byKatz and Shapiro, 1985).

Focusing on the case set forth by Robin Cowan in the above quotation, letus assume that there are only two types of agents (i1, 2) and two tech-nologies. Recall Arthur’s pay-off functions (2.4),and assume that r(·)s log(·) is a function (which is common for every agent:

ij ( n→(t)) ai

j � r (n j(t)),

and limn j→�

ddn j f (n

j ) 0

ddn

j f (n j) � 0, d2

dn j2 f (n j ) � 0

supx�[0,1]�R(0,1)

| ft(x) � f(x)| � min{F�(0), 1 � F�(0)}

f (x) �1 if x � 1�2F�(0) if x 1�20 if x � 1�2


s is a constant) that formalises unbounded increasing returns to adoption.Agent i chooses technology A if and only if . Bytaking the exponential on both sides and rearranging we have:

(2.6)

The function of the attributes of agent’s type which is on the right-handside can be considered a random variable because, as discussed above, i(t)is a random variable. Moreover such agent characteristics are iid becausei(t) is iid. Denoting the random variables on the right-hand side with �(t),from (2.6) we have that the adoption process can be seen as a generalisedurn scheme with urn function given by:

f (x)F�,(x/(l�x)), (2.7)

where F�(·) is the distribution function of �(t). Because i(t) takes just twovalues (1, 2), also �(t) takes just two values:

where we have assumed without loss of generality that

Thus F� is by construction a step function with two steps:

Therefore, taking into account (2.7), we have that the urn function has twosteps and is defined in the following way:

If the following condition is satisfied

e1s (a1

B�a1A)

1 � e1s (a1

B�a1A) � � �

e1s (a2

B�a2A)

1 � e1s (a2

B�a2A)

f (x) �0 if x �

e1s(a1

B�a1A)

1 � e1s (a1

B�a1A)

� if e

1s (a1

B�a1A)

1 � e1s (a1

B�a1A) � x �

e1s (a

2B�a

2A)

1 � e1s (a

2B�a

2A).

1 if x �e

1s (a2

B�a 2A)

1 � e1s (a2

B�a 2A)

F�(y) �0�

1

if if if

y � e1s(a1

B�a1A)

e1s(a1

B�a1A) � y � e

1s(a2

B�a2A)

y � e1s(a2

B�a2A)

a1B � a1

A � a2B � a2

A.

�(t) �e1s (a1

B�a1A) with probability �

e1s (a2

B�a2A) with probability 1 � �

X(t)1 � X(t)

� e1s (ai

B(t) � aiA(t)).

iA( n→(t)) � i

B � ( n→(t))


the urn function has five fixed points, three of which are down-crossing,therefore there is a set of initial conditions (that imply giving both tech-nologies a chance to be chosen ‘at the beginning of history’) for whichmarket sharing is asymptotically attainable with positive probability.8 Theabove condition imply that the ratios are sufficiently differentbetween the two types. In other words there might be sufficient heterogeneityamong agents to counterbalance the effect of increasing returns to adoption.9

Example 2Consider now pay-off functions of this type:

where rj, aj, jA, B, are bounded random variables which admit density.Such functions allow agents to be heterogeneous also in terms of the degreeof increasing returns which they experience. By applying (2.1), dividingpay-off functions by total number of adoptions, and rearranging we havethat A is chosen if and only if:

(2.8)

Denoting the random variables on the right-hand side with �(t), from(2.8) we have that the adoption process can be seen as a generalised urnscheme with urn function f(t, x)F�(t)(x), where F�(t)(·) is the distributionfunction of �(t). Now suppose that rA and rB are highly correlated and bothhave bimodal distributions very concentrated around the two modes, insuch a way that the distribution of rA/rB is also bimodal and very concen-trated around the two modes too. Furthermore suppose that the two modesare far away from each other. To fix the ideas say that for a percentage � ofthe population rA/rB, is uniformly distributed on the interval [1/(1�b),1/(1�a)], while for a percentage 1�� of the population rA/rB is uniformlydistributed on the interval [1/(1�d), 1/(1�c)], with 0�a�b�c�d. First,let us consider the case of aj0, jA, B. F� is by construction independentof t, implying the following urn function:

f(x) F�(t)(x) �0�� if y � a�

1b � a

(y � a)�� if a � y � b�� if b � y � c� � (1 � �)[ 1

d � c (y � c)] if c � y � d

1�� if y � d

X(t) �rB

rA � rB�

aB � aA

(t � nA(0) � nB(0))(rA � rB).

j( n→(t)) aj � rj n

j,

(e ai

A)1�r�(e a

iB)1�r


If b��c, then there are three stable fixed point of f(x) and, as statedabove, it can be shown that there is a set of initial conditions (that implygiving both technologies a chance to be chosen ‘at the beginning ofhistory’) for which market sharing is asymptotically attainable with posi-tive probability. If aj�0 but has bounded support and admits density,then condition (2.3) applies and the same argument holds: in fact,relying on the fact that rj, aj are bounded it is easy to show that

where K�0 is a constant. The essentialingredient of this example is that the distribution of rA/rB is bimodal andvery concentrated around the two modes. The argument has nothing to dowith the particular (and extreme) distributional form assumed above: fol-lowing the same constructive procedure adopted here it is easy to buildexamples with any other distributional form. The only requirement is thatthe two modes are sufficiently distant. In other words the only requirementis a sufficient degree of heterogeneity in the population to counterbalancethe pro-standardisation effects of increasing returns to adoption.

The two examples above show that the degree of increasing returnsneeds to be compared to the degree of heterogeneity. Unbounded increas-ing returns that are diminishing fast at the margin are not sufficient to gen-erate asymptotic survival of only one technology, provided that agents arenot completely homogeneous (see also Farrell and Saloner, 1985; 1986, forearly models with homogeneous agents that lead to the survival of onlyone technology). Even more interesting, when heterogeneity is so widethat there are agent-specific increasing returns, the emergence of techno-logical monopolies is not guaranteed even with returns that are linearlyincreasing.

To summarise, the foregoing examples show that if preferences aresufficiently heterogeneous and/or increasing returns to adoption are lessthan asymptotically linear, then Arthur’s result cannot be generalised andvariety in the asymptotic distribution of technologies can be an outcomewith positive probability.

From the point of view of empirical predictions, at first look, the forego-ing results might sound, if anything, as a further pessimistic note on ‘inde-terminacy’. That is, not only ‘history matters’ in the sense that initial smallevents might determine which of the notional, technologically attainable,asymptotic states the system might ‘choose’: more troubling, the argumentso far suggests that, further, the very distribution of the fine characteristicsand preferences of the population of agents might determine the very natureof the attainable asymptotic states themselves. Short of empirically con-vincing restrictions on the distribution of agents (normally unobservable)characteristics, what we propose is instead an interpretation of the generaloccurrence of technological monopolies (cum increasing returns of some

supx�[0,1]�R(0,1)| f (t, x) � f (x)| � K �t,


kind) grounded on the relative speed of convergence to the underlying (butunobservable) limit states.

3 RATE OF CONVERGENCE INONE-DIMENSIONAL MODELS OF COMPETINGTECHNOLOGIES

In the example of the VCR market, as well as for many other markets, thepossibility of predicting limit market shares depends on the feasibility offormalising the structure of the market in question in terms of a specificurn function. Heterogeneity of preferences, the degrees of increasingreturns, the type of expectations, price-policies of producers, all affectthe functional form of the urn function. As mentioned before, the goal ofthis chapter in general and of this section in particular is to providesome general asymptotic results that can be used as guidance for theinterpretation of the empirical evidence on emergence of dominantdesigns.

Propositions 2, and 4 suffice for the task. Together they imply the rele-vant statements on the rate of convergence to technological monopoly orto a limit market share where both technologies coexist.10 Furthermore theanalysis that follows applies even in the absence of a clear pattern ofincreasing returns to adoption. In essence, in the presence of constantreturns to adoption, the urn function would be completely constant but thefollowing theorems would still hold.

As above, denote the urn function with f(·,·); the following propositiongives a first result on the rate of convergence to 0 and 1.

Proposition 1Let ��0 and c�1 be such that eventually

f (t,x)�cx for x�(0, �)

( f (t,x))�1�c(1�x) for x�(1��,1)). (2.9)

Then for any ��(0,1�c) and ��0

where X(·) stands for the random process given by (2.2).The proposition is proved in the appendix (section 5).

(limt→�P{t1�c��[1 � X(t)] � �|X(t) → 1 1),

limt→�P{t1�c��X(t) � �|X(t) → 0} 1


A similar result can be expressed in terms of variances (L2�convergence):

Proposition 2Let ��0 and c�1 be such that eventually (2.9) holds. Then for any��(0,1�c)

The proposition is proved in the appendix (section 5).Notice that proposition 2 states that the rate of convergence of the mean

square distance from the limit market share is of the order of 1/t as t → �,conditional to the fact that the process is actually converging to 0 or 1.Roughly speaking it defines the rate of convergence of mean square errorswhen the process converges to a technological monopoly. If the set of limitmarket shares that the process can reach with positive probability containsonly these two points, proposition 2 implies a similar statement in termsof the unconditional mean square distance from the limit market share.

One would like to derive a counterpart of proposition 1 and 2 for the caseof market sharing, whenever this can be attained with positive probability.For a differentiable f(·) at 0(1), (2.9) holds with c arbitrarily close to

.We can easily derive a similar result for a differentiable f(·)independent of t from the following conditional limit theorem for the gen-eralised urn scheme.

Theorem 2 (Arthur et al., 1987)

Let � � (0,1) be a stable root of f(x)�x0 and f(·) is differentiable at � withThen for every y �(��, �)

where �(·) stands for the Gaussian distribution function having zero mean andvariance 1.

From this theorem, we can give an even better characterisation of thelowest possible convergence rate for a limit market share where both tech-nology coexist that can be attained with positive probability. Indeed, thenext proposition follows immediately:

limt→�

P�√t 1 � 2 d

dx f (�)2�(1 � �)

[X(t) � �] � y|X(t) → �� (y),

ddx f (�) � 1�2.

ddx f (0) ( d

dx f (1))

(limt→�t2(1�c)��P{X(t) → 1} �{X(t)→1}

(X(t) � 1)2dP 0

limt→�t2(1�c)��P{X(t) → 0} �{X(t)→0}

X(t)2dP 0


Proposition 3Let � � (0,1) be a stable root and let

[ f(x)��](x��)�k(x��)2 for x�(��, � ��), x��

take place for some ��0 and k�1/2. Then for every �, ��0

Differentiability of the urn function at the limit point is a highly demand-ing restriction, as well as the fact that the urn function has to be indepen-dent of t. As said before, several actual markets can present overwhelmingproblems of formalisation. Consequently, it may be impossible to checkthese conditions, albeit intuitively there is no reason why differentiabilityshould matter. Conversely, we can obtain a general result in terms of L2�convergence that suffices to the task:

Proposition 4Let � � (0,1) be such that

(2.10)

takes place eventually for some ��0 and k�1. Then for every ��0

The proposition is proved in the appendix (section 5).Propositions 2 and 4 show that convergence to 0 and 1 can be much

faster (almost of order 1/t as t → �) than to an interior limit (which canbe almost of order only).11 Here t stands for the number of adoptionsto the urn. That is, we are talking about relative rates (the ideal time whichis considered here is the time of product selection choices). This result is,however, stronger than it may seem at first glance. In fact it has also impli-cations for the patterns of product selection in ‘real’ (empirical) timewhere plausibly the speed of the market share trajectory depends alsoon technology adoption decisions. There is much qualitative evidenceand some econometric results [for example, Koski and Nijkamp, 1997)showing that technology adoption is at the very least independent ofmarket shares if not enhanced by increasing asymmetry in their distribu-tion. Thus a fortiori we can conclude that there is a natural tendency ofthis class of processes to converge faster to 0 or 1 rather than to an inter-ior limit. The explanation is that the variance of �t(x), which characterises

1� √t

limt→�

t 2min{1�k,˛1�2}� ˛�P{X(t) → �} �

{X(t)→�}(X(t) � �)2dP � 0.

[ f (t, x) � �](x � �) � k(x � �)2 for x�(� � �, � � �), x˛ � ˛�

limt→�

P{t1�2 � �|X(t) � �| � �|X(t) → 0} 0.


the level of random disturbances in the process (2.2), is f(t, x)(1�f(t, x)).Under condition (2.9) this value vanishes at 0 and 1 but it does not vanishat ��(0,1), being equal to �(1��), under condition (2.10). Notice alsothat in example 1 c0 and in example 2 c � 0.

As shown in the previous section, the urn function can have any shape andthere is no reason to believe that problems characterised by 0 and 1 as theonly stable points are the only ones that we can expect. Therefore, in princi-ple, an asymptotic outcome where both technologies survive should beobservable with positive frequency in real markets. As discussed in the pre-vious section the tendency to converge to market sharing or technologicalmonopolies is an outcome induced by the relative impact of heterogeneityof preferences and increasing returns to adoption. What tendency is realiseddepends on which of the two prevails. Notice, however, that the prevalenceof one of the two factors is not always predictable ex ante even for a nearlyomniscient agent fully aware of all fundamentals of the economy: in theexamples of the previous section both type of outcomes are possible, butwhich one is realised depends on the actual sequence of historical eventsthat lead to it. In this type of models, in general, when multiple asymptoticequilibria are attainable, history plays a major role in the selection of theactual one.12

If asymptotic patterns were observable, the results of the previoussection would imply that we should observe both stable market sharingand technological monopolies. However, for the interpretation of empir-ical stylised facts, the point where the process eventually would convergemay be irrelevant. Indeed, the rate of change of the technological and eco-nomic environment can be sufficiently high that one can always observediffusion dynamics well short of any meaningful neighbourhood of thelimit it would have attained under forever constant external conditions. Sowhile it is true that a convergent process should generate a long-lastingstable pattern, the time required to generate it may be too long to actuallyobserve it: the world is likely to change well before convergence is actuallyattained. In a sense these changes can be viewed as resetting the game toits starting point.

On the basis of the propositions of this section we notice that conver-gence to technological monopolies tends to be much faster (in probabilisticterms) than to any stable market sharing where both technologies coexist,because of the intrinsic variability that market sharing carries over. Thusthe empirical prediction of these results can be stated as follows: in marketswith increasing returns to adoption and a high rate of technological changewe expect to observe a prevalence of both unstable market sharing (persis-tent fluctuations in the market shares) and stable technological quasi-monopolies as compared to stable patterns of market sharing. The reason


for this is that technological monopolies can be easily attained in a reason-ably short time, that is, sufficiently before any significative change in theunderlying basic technological paradigms takes place (Dosi, 1982).

Finally note that the observation of the frequent emergence of differentmonopolies in different related markets (for example, different geograph-ical areas) does not contradict our empirical predictions. Of course, it istrivially true that, with mutually independent markets, different trajectoriescould emerge in different markets as if they were different realisations ofthe same experiment. In a related paper (Bassanini and Dosi, 1999a) weshow that the foregoing results can be extended also to the case whenmarkets are interdependent: not contrary to the intuition, it is the balancebetween local and global feedbacks which determines whether the systemconverges to the same or different monopolies in every market. However,even though at high level of aggregation a system of different local mono-polies looks like a stable market sharing, it is shown there that it has thesame rate-of-convergence properties of a ‘univariate’ system converging toa monopoly.

4 CONCLUSIONS

This Chapter has reassessed the empirical evidence on prevalence of tech-nological monopolies over market sharing in the dynamics of competingtechnologies. First, we have argued that the dominant explanation in theliterature, namely that unbounded increasing returns can be identified asthe factor responsible for this pattern, does not always hold. BrianArthur’s results – we have shown – hold only when increasing returns toadoption are linear or more than linear and the degree of heterogeneityof agents is small. The presented counter-examples suggest that asymp-totic patterns of the dynamics of competing technologies depend on therelative impact of (unbounded) increasing returns and the degree of het-erogeneity of the population of adopters. Second, given all this, wepropose, however, that in a market with high technological dynamism, nointeresting predictions can be made by simply looking at theoreticalasymptotic patterns. If convergence is too slow the environment maychange before the limit can be actually approached. Conversely, develop-ing upon some mathematical properties of Polya urns, we show that con-vergence to technological monopolies tends to be (in probabilistic terms)much faster than to a limit where both technologies coexist, the empiricalimplication being that in markets with high turnover of basic techno-logies, a prevalence of technological monopolies over stable marketsharing is likely to be observed.


APPENDIX

For the purpose of exposition, to keep the notation simple, all proofs areexposed for generalised urn scheme involving time-independent urn func-tions. They can be easily repeated for the general case.

Proof of Proposition 1

Consider only the first case – convergence to 0. Without loss of generalitywe can assume that . Indeed the theorem, being a statementabout the convergence rate to 0, does not make any sense if X(·) does notconverge to 0.

Let Z(·) be a conventional urn process with cx as the urn-function andthe same initial numbers of balls nnw�nb. Then

and consequently

where ot(l) → 0 as t → �. Hence from Chebychev’s inequality

(2.11)

for every � � (0,1�c) and ��0.For arbitrary � � (0, �) and v�0 there is N depending on these variables

such that

where . Also since Z(t) → 0 with probability 1 as t → �,we can choose this N so large that

(2.12)

To prove the theorem it is enough to show that

P{t1�c��X(t) � �, X(t) → 0} → 0,

P{{X(t) → 0} � {X(s) � �, Z(s) � �, s � N}} � v.

A˛�˛B (A\B)�(B\A)

P{{X(t) → 0} � {X(s) � �, s � N}} � v

P{t1�c��Z(t) � �} → 0 as t → �

� EZ(1)ec˛� ˛1 �n�t�1

n�1

1x dx

EZ(1)( n � 1n � t � 1)1 � c EZ(1)tc�1[1 � ot(1)],

EZ(t) � EZ(1)t�1j1(1 � 1 � c

j � n) � EZ(1)ec˛� ˛1

�

t�1

j1

1(n�j)

EZ(t � 1) � �1 �1 � ct � n�EZ(t), t � 1,

P{X(t) → 0} � 0


or, taking into account that v in (2.12) can be arbitrary small, that

(2.13)

However

(2.14)

where is the set of values that X(t) can attain(not necessarily with positive probability). Due to lemma 2.2 of Hill et al.(1980), there exists a probability space such that Z(·) dominates X(·) on theevent Z(t)��, t�N, providing that these processes start from the samepoint. Therefore, for any

However for every

as t →�by (2.11). Thus (2.14) is a sum of a finite number – namely N – ofterms each converging to zero. This completes the proof.


As before, consider only the first case – convergence to 0.Let Z(·) be a conventional urn process with cx as the urn-function and

the same initial numbers of balls nnw�nb. Then

� 1(n � t)2c(1 � c)2Z(t)2

� 2n � t

(c � 1)Z(t)2 � 1(n � t)2

(c � 1) 2Z(t)2

E [Z(t � 1)2 |Z(t)] Z(t)2

� P{t1�c��Z(t) � �} → 0

P{t1�c��Z(t) � �, X(s) � �, Z(s) � �, s � N}

y � SXN

� P{t1�c��Z(t) � �, X(s) � �, Z(s) � �, s � N˛|˛X(N) y}.

P{t1�c��X(t) � �, X(s) � �, Z(s) � �, s � N˛|˛X(N) y}

y � SXN

SXt {X(1)(n � i)

n � t � 1 , 0 � i � t � 1}

P{X(N) y},

�y�SX

N

P�t1�c��X(t) � �, X(s) � �,Z(s) � �, s � N

|X(N ) y�P{t1�c��X(t) � �, X(s) � �, Z(s) � �, s � N

P{t1�c��X(t) � �, X(s) � �, Z(s) � �, s � N} → 0.


and

and consequently

where tsot(1/ts) → 0 as t → �. Hence:

(2.15)

for every � � (0,1�c).For arbitrary � � (0, �) and v�0 there is N depending on these variables

such that

where . Also since Z(t) → 0 with probability 1 ast → �, we can choose this N so large that

(2.16)

To prove the theorem it is enough to show that

or, taking into account that v in (2.16) can be arbitrary small, that

(2.17)

However,

(2.18) �y �SX

N

t 2(1�c��) �

{X(s) � �, Z(s) � �, s � N, X(N)y}X(t)2dP

t2(1�c��) �{X(s) � �, Z(s) � �, s � N}

X(t)2dP

t2(1�c��) �{X(s) � �, Z(s) � �, s � N}

X(t)2dP → 0,

t2(1�c��) �{X(t)→0}

X(t)2dP → 0,

P{{X(t) → 0} � {X(s) � �, Z(s) � �, s � N}} � v.

A˛�˛B (A\B)�(B \A)

P{{X(t) → 0} � {X(s) � �, s � N}} � v,

t2(1�c��)E [Z(t)2] → 0 as t → �

E [Z(1)2]( n � 1n � t � 1)2(1�c) E [Z(1)2[t2(c�1)]1 � ot(1)],

� E [Z(1)2]e 2(c�1) �

t�1

j1

1(n�j)

� E [Z(1)2]e 2(c�1) �

n�t�1

n�1 1x dx

E [Z(t)2] E [Z(1)2]t�1j1 [1 � 2(1 � c)

j � n � oj(1� ˛j)]

E [Z(t � 1)2] �1 �2(1 � c)

t � n�

1 � c(t � n)2�E [Z(t)2], t � 1,


where is the set of values that X{t) can attain(not necessarily with positive probability). Due to lemma 2.2 of Hill et al.(1980), there exists a probability space such that Z(·) dominates X(·) on theevent Z(t)��, t�N, providing that these processes start from the samepoint. Therefore, for any t�N and

However, for every

as t →�by (2.15). Thus (2.18) is a sum of a finite number – namely N – ofterms each converging to zero. This completes the proof.


The proof is based on the following lemmas:13

Lemma 1Let f(.) be the urn function of the process Z(t) such that

[ f(x)��](x��)�k(x��)2

for some k�1 and ��(0,1) and f(x)[ l�f (x)]��0. Then limt→� dtlimt→� E(Z(t)��)2, where

where K is a constant term.

dt � K �(n � t)�1�� if 2(1 � k) � 1 � 0

(n � t)�1 log(n � t) if 2(1 � k) � 1 0(n � t)�2(1�k)��

if 2(1 � k) � 1 � 0

� t2(1�c��)E [Z(t)2] → 0

t2(1�c��) �{X(s) � �, Z(s) � �, s � N, X(N)y}

Z(t)2dP

y �SXN

� t2(1�c��) �{X(s) � �, Z(s) � �, s � N, X(N)y}

Z(t)2dP

t2(1�c��) �{X(s) � �, Z(s) � �, s � N, X(N)y}

X(t)2dP

y �SXN

SXt {X(1)(n � i)

n � t � 1 , 0 � i � t � 1}


ProofConsider the process (2.2) and write nnw�nb, then:

Setting �tE(Z(t)��)2, from the assumptions of the lemma, taking intoaccount that f(Z(t))[l�f(Z(t))]��, and that

we have

Thus

Since

where �it and �it are small terms (ot(1)) not necessarily non-negative,14 then

Since, in terms of asymptotic behaviour,

�t

i1(n � i)�2�2(1�k) � �

12(1 � k) � 1

(n � t) 2(1�k)�1 if 2(1 � k) � 1 � 0

log(n � t)�� if 2(1 � k) � 1 0

� �(n � t)�2(1�k)�t

i1(n � i)�2�2(1�k)(1 � �it).

�t�1 � �t�n � 1n � t �

2(1�k)(1 � �1t)

� e�2(1�k)[log(n�t�1) � log(n�i)](1 � �it) �n � in � t�

2(1�k)(1 � �it),

t

ji�1�1 �

2(1 � k)n � i e�2(1�k) �

t

ji�1

1(n�j )

(1 � �it)

�t�1 � �t t

i1�1 �

2(1 � k)n � i � �

t

i1

�

(n � i)2 t

ji�1�1 �

2(1 � k)n � j

�t�1 � �t�1 �2(1 � k)

n � t ��

(n � t)2

[ f (Z(t)) � Z(t)](Z(t) � �) [ f (Z(t)) � �](Z(t) � �) � (Z(t) � �)2,

� 1(n � t)2 f (Z(t))[1 � f (Z(t))].

� 2n � t

[ f (Z(t)) � Z(t)](Z(t) � �) � 1(n � t)2 [ f (Z(t)) � Z(t)]2

E [(Z(t � 1) � �)2]X(t)] (Z(t) � �)2


we have that

which implies the statement of the lemma.Set

Then, with probability Also by lemma 1,

(2.19)

As for proposition 2, we can ignore the case when X(t) does not convergewith positive probability.

We have to show that as t →�

(2.20)

For every ��0 there is t(�) such that

Since � can be arbitrarily small, (2.20) holds if and only if

However,

t2min{1�k1˛�˛2}�� y�St(�)

� {|X(s) � �|� �, s � t(�), X(t(�))y}

(X(t) � �)2dP,

t2min{1�k,1˛�˛2}�� {|X(s) � �|� �, s � t(�)}

(X(t) � �)2dP

t2min{1�k,1�2}�� {|X(s)��|�� , s�t(�)}

(X(t) � �)2dP → 0,

P{{X(s) → �} � {|X(t) � �| � � , t � t(�)}} � �.

t2min{1�k,1�2}�� {X(t)→�}

(X(t) � �)2 dP → 0,

t2min{1�k,1�2}��E [(X (t) � �)2] → 0 as t → �.

1, X (t) → � as t → �.

X (1) X(1).

X (t � 1) X (t) �j

t(X (t)) �X (t)

t � n ,

f (x) �

f (� � � )� if x � � � �

f (x)�� if � � � � x � � � �

f (� � � )� if x � � � �

,

�t�1 � K�1

2(1 � k) � 1 (n � t)�1�� if 2(1 � k) � 1 � 0

(n � t)�1 log(n � t)�� if 2(1 � k) � 1 0�1(n � t)�2(1�k)�� if 2(1 � k) � 1 � 0

,


where is the set of values that X(t) andX(t) can attain (not necessarily with positive probability). Notice that forany t�t(�) and y �St(�)

This follows from the fact that f(x) and f(x) are the same for x [��,��].However,

as t →�by (2.19). This completes the proof.

NOTES

* The views expressed here cannot be ascribed to the Organisation for EconomicCo-operation and Development (OECD) Secretariat or its Member Countries. We areindebted to Yuri Kaniovski for very helpful suggestions. We thank also Brian Arthur,Robin Cowan, Klaus Desmet, Judith Gebauer, Michael Horvath, Andrea Prat, AldoRustichini, Valter Sorana, and participants to the 3rd workshop on Economics withHeterogeneous Interacting Agents, Ancona, Italy, May 1998, and to the Conference onEconomic Models of Evolutionary Dynamics and Interacting Agents, Trieste, Italy,September 1998, for their comments. Financial support from International Institute forApplied Systems Analysis (IIASA), Banca Nazionale del Lavoro (BNL), ItalianResearch Council (CNR), and Italian Ministry of University and Scientific Research(MURST) is gratefully acknowledged. All errors are ours.

1. See, for example, the empirical literature on dominant designs (for a survey, cf. Tushmanand Murmann, 1998).

2. Throughout this chapter we label the generalisation of Polya urn schemes set forth byHill et al. (1980) as generalised urn scheme. That generalisation is the most popular ineconomics but obviously it is not the only possible one (see, for example, Walker andMuliere, 1997).

3. We assume that, if there is a tie, agents choose technology A. Qualitatively, breaking thetie in a different way would not make any difference.

4. A convenient review of analytical results on generalised urn schemes can be found inDosi et al. (1994). The reader is referred to that volume for the results that are not provedin this chapter. Particularly, X(�) converges almost surely, as t tends to infinity, to the setof appropriately defined zeros of the function g(x)f (x)�x. However, since we are not

t2min{1�k,1�2}��E [(X (t) � �)2] → 0,

t2min{1�k,1�}2�� {X (t)→�}

(X (t) � �)2dP

t2min{1�k,1˛�˛2}�� y �St(�)

� {|X(s) � �|� �, s � t(�), X (t(�))y}

(X (t) � �)2dP

�{|X (s) � �|� �, s � t (�), X (t(�))y}

(X (t) � �)2dP

�{|X(s) � �|� �, s � t (�), X(t(�))y}

(X(t) � �)2dP

St St {X(1)(n � i)n � t � 1 , 0 � i � t � 1}


going to restrict ourselves to the case when g(�) is a continuous function, we need somestandard definition concerning equations with discontinuous functions. For a functiong(�) given on R(0,1) and a point x � [0,1] set

where {yk} is an arbitrary sequence converging to x. Then the set of zeros A(g) of g(�)on [0,1] is defined by the following relation

Note that for a continuous g(·) this definition gives the roots of the equation g(x)0 inthe conventional meaning. One particular class of attainable singleton components com-prises the downcrossing or stable ones, that is, the points where f (x)�x changes its signfrom plus to minus. More precisely, ��R(0,1) is said to be stable if there exists ��0 suchthat for every ��(0, �)

(*)

If ��R(0,1) is stable then X(·) converges to � with positive probability for some initialcombination . If in addition to (*)

f (x)�(0,1) for all x�R(0,1),

then it converges with positive probability to � for any initial combination . Finallyif the urn function does not have touchpoints and the set A(g) with g(x)f (x)�x is com-posed only of singleton components then almost surely the process converges to the setof stable components.

5. For instance, in the data set of Computer Intelligence InfoCorp employed by Breuhan(1996), more than 80 per cent of the firms in the sample report using a single word-processing package.

6. See note 4 above, or Dosi et al. (1994), theorems 1 and 3.7. To be precise Arthur (1989) quotes also Arthur et al. (1983), though there the properties

are stated only as yet-to-be-proved good sense conjectures.8. See note 4 above, or Dosi et al. (1994), theorem 2.9. Cowan and Cowan (1998) acknowledge this role of heterogeneity, although only for

models where interactions are local. They suggest that many models from other scien-tific disciplines can be adapted to show market-sharing survival as a result of localinteraction effects, and they provide one such example, although restricted to linearreturns.

10. Notice that, provided that inequalities (2.9) and (2.10) are eventually satisfied for any t,propositions 1, 2 and 4 hold even if the inequality (2.3) does not hold as may happenwhen agents are assumed to be forward looking.

11. If returns are constant, the result of proposition 3 and 4 simply becomes the well-knowntextbook result on rate of convergence of the sample mean and its variance. Bassaniniand Dosi (1999b) shows that 2min{l�k, 1/2} is also an upper bound to the rate of con-vergence to an interior limit, therefore proposition 4 could be written in an even strongerway, although not necessary for the task of the present chapter.

12. For a general discussion on this point see also Dosi (1997).13. We are indebted to Yuri Kaniovski for suggesting us the line of the following proof.14. The line of reasoning here is the same as for the proof of proposition 1 and 2.

n→

(0)

n→

(0)

inf�� |x��|�c

[ f(x) � x](x � �) � 0.

A(g) {x � [0,1]: [a(x, g), a(x, g) � 0}.

a(x, g) sup{yb}� R(0,1)lim supk→� g(yk).

a(x, g) inf{yk}� R(0,1)lim infk→� g(yk),


REFERENCES

Arthur, W.B. (1983), ‘On competing technologies and historical small events: thedynamics of choice under increasing returns’, IIASA Working Paper WP-83-90(reprinted in W.B. Arthur (1994) Increasing Returns and Path-Dependence in theEconomy, Ann Arbor, MI: University of Michigan Press, 1983).

Arthur, W.B. (1989), ‘Competing technologies, increasing returns and lock-in byhistorical events’, Economics Journal, 99, 116–31.

Arthur, W.B. (1990), ‘ “Silicon Valley” locational clusters: when do increasingreturns imply monopoly?’, Mathematical Social Sciences, 19, 235–51.

Arthur, W.B. and D. Lane (1993), ‘Information contagion’, Structural Change andEconomic Dynamics, 4, 81–104.

Arthur, W.B. (1983), Y. Ermoliev and Y. Kaniovski, ‘Generalised urn problem andits applications’, Cybernetics, 19, 61–71.

Arthur, W.B., Y. Ermoliev and Y. Kaniovski (1986), ‘Strong laws for a class of path-dependent urn processes’, Proceedings of the International Conference onStochastic Optimization, Lecture Notes on Control and Information Sciences, 81,187–300.

Banerjee, A. (1992), ‘A simple model of herd behaviour’, Quarterly Journal ofEconomics, 107, 797–817.

Bassanini, A.P. (1997), ‘Localized technological change and path-dependentgrowth’, IIASA Interim Report IR-97-086.

Bassanini, A.P. and G. Dosi (1999a), ‘Heterogeneous agents, complementarities,and diffusion of technologies: do increasing returns imply convergence to inter-national monopolies?’, in D. Delli Gatti, M. Gallegati and A. Kirman (eds),Market Structure, Aggregation and Heterogeneity, Berlin, Springer.

Bassanini, A.P. and G. Dosi (1999b), ‘Competing technologies, technologicalmonopolies and the rate of convergence to a stable market structure’, Laboratoryof Economics and Management Working Paper no. 3, Sant ‘Anna School ofAdvanced Studies.

Bernheim, B.D. (1994), ‘A theory of conformity’, Journal of Political Economy, 102,841–77.

Breuhan, A.L. (1996), ‘Innovation and the persistence of technological lock-in’,unpublished manuscript.

Brock, W.A. and S.N. Durlauf (1995), ‘Discrete choice with social interactions I:theory’, NBER Working Paper no. 5291.

Brynjolfsson, E. and C.F. Kemerer (1996), ‘Network externalities in microcomputersoftware: an econometric analysis of the spreadsheet market’, ManagementScience, 42, 1627–47.

Church, J. and N. Gandal (1993), ‘Complementary network externalities and tech-nological adoption’, International Journal of Industrial Organization, 11 239–60.

Cowan, R. (1990), ‘Nuclear power reactors: a study in technological lock-in’,Journal of Economic History, 50, 541–67.

Cowan, R. (1991), ‘Tortoises and hares: choice among technologies of unknownmerit’, Economic Journal, 101, 801–14.

Cowan, R., and W. Cowan (1998), ‘Technological standardization with and withoutborders in an interacting agents model’, unpublished paper.

Cusumano, M.A., Y. Milonadis and R.S. Rosenbloom (1992), ‘Strategic maneu-vering and mass-market dynamics: the triumph of VHS over Beta’, BusinessHistory Review, 66, 51–94.


David, P. (1985), ‘Clio and the economics of QWERTY’, AEA Papers andProceedings, 75, 332–7.

David, P. (1992), ‘Heroes, herds and hysteresis in technological theory: ThomasEdison and the battle of systems reconsidered’, Industrial and Corporate Change,1, 129–80.

Dosi, G. (1982), ‘Technological paradigms and technological trajectories’, ResearchPolicy, 11, 142–67.

Dosi, G. (1997), ‘Opportunities, incentives and the collective patterns of techno-logical change’, Economics Journal, 107, 1530–47.

Dosi, G., Y. Ermoliev and Y. Kaniovski (1994), ‘Generalized urn schemes and tech-nological dynamics’, Journal of Mathematical Economics, 23, 1–19.

Economides, N. (1996), ‘The economics of networks’, International Journal ofIndustrial Organization, 14, 673–99.

Farrell, J. and G. Saloner (1985), ‘Standardization, compatibility and innovation’,Rand Journal of Economics, 16, 70–83.

Farrell, J. and G. Saloner (1986), ‘Installed base and compatibility: innovation,product preannouncements, and predation’, American Economic Review, 76,940–55.

Hill, B.M., D. Lane and W. Sudderth (1980), ‘A strong law for some Generalizedurn processes’, Annals of Probability, 8, 214–26.

Islas, J. (1997), ‘Getting round the lock-in in electricity generating systems: theexample of the gas turbine’, Research Policy, 26, 49–66.

Katz, M.L. and C. Shapiro (1985), ‘Network externalities, competition, and com-patibility’, American Economic Review, 75, 424–40.

Katz, M.L. and C. Shapiro (1986), ‘Technology adoption in the presence ofnetwork externalities’, Journal of Political Economy, 94, 822–41.

Katz, M.L. and C. Shapiro (1992), ‘Product introduction with network externali-ties’, Journal of Industrial Economics, 40, 55–84.

Katz, M.L. and C. Shapiro (1994), ‘Systems competition and network effects’,Journal of Economic Perspectives, 8, 93–115.

Koski, H. and P. Nijkamp (1997), ‘The installed base effect: some empirical evi-dence from the microcomputer market’, unpublished manuscript.

Krugman, P. (1991a), ‘History vs. expectations’, Quarterly Journal of Economics,106, 651–67.

Krugman, P. (1991b), Geography and Trade, Cambridge, MA: MIT Press.Liebowitz, S.J. and S.E. Margolis (1994), ‘Network externality: an uncommon

tragedy’, Journal of Economic Perspectives, 8, 133–150.Oren, S. and S. Smith (1981), ‘Critical mass and tariff structure in electronic com-

munications markets’, Bell Journal of Economics, 12, 467–87.Rauch, J.E. (1993), ‘Does history matter only when it matters little? The case of city-

industry location’, Quarterly Journal of Economics, 108, 843–67.Rohlfs, J. (1974), ‘A theory of interdependent demand for a communication service’,

Bell Journal of Economics, 5, 16–37.Rosenberg, N. (1982), Inside the Black Box, Cambridge: Cambridge University

Press.Saviotti, P.P. and A. Trickett (1992), ‘The evolution of helicopter technology’,

1940–1986, Economics of Innovation and New Technology, 2, 111–30.Tushman, M.L. and J.P. Murmann (1998), ‘Dominant designs, technology cycles,

and organizational outcomes’, in B. Staw and L.L. Cummings (eds), Research inOrganizational Behaviour, 20, Greenwich, CT: JAI Press.


Venables, A.J. (1996), ‘Localization of industry and trade performance’, OxfordReview of Economic Policy, 12 (3), 52–60.

Walker, S. and P. Muliere (1997), ‘Beta-Stacy processes and a generalisation of thePolya-urn scheme’, Annals of Statistics, 25, 1762–80.

Winter, S.G. (1986), ‘Comments on Arrow and Lucas’, Journal of Business, 59,S427–34.


3. Path dependence, localisedtechnological change and the questfor dynamic efficiency1

Cristiano Antonelli

1 INTRODUCTION

The quest for the conditions of dynamic efficiency can be considered one ofthe key aims and scopes of much contemporary work in economic theory.

Neoclassical economics has provided an elaborate and sophisticatedframework to understand the conditions for static efficiency. In thatcontext, growth and development are the consequences of exogenouschanges in the shapes of the utility functions, in the characteristics of thetechnology and in the actual conditions of the demography and of thenatural resources.

The theory of economic growth elaborated in that context does notaddress the actual causes of growth. It is limited to analysing the comple-mentary conditions in terms of rates of growth in the supply of labour andsavings, that make it possible for exogenous growth to take place.

The notion of path dependence elaborated by Paul David provides oneof the most articulated and comprehensive frameworks to move towardsthe analysis of the conditions that make it possible for an economic systemto generate and exploit endogenous growth. Path dependence is an essen-tial tool to move from the analysis of static efficiency and enter into theanalysis of the conditions for dynamic efficiency.

The notion of path dependence proves to be especially attractive forEuropean economists raised in a tradition that considers growth andchange rather than equilibrium as the relevant object of analysis and,hence, values historic time and philological investigations as basic tools tostudy the dynamics of social events. The pages that follow not only try andarticulate the wealth of the contribution of Paul David, but also are anattempt show how fertile and stimulating his framework is.

The rest of the Chapter is structured as follows. Section 2 explores indetails the notion of path dependence and identifies its basic ingredients.

51

Section 3 shows how the different combinations of the basic ingredientslead different types of path dependence, especially when the analysis isapplied to understanding the introduction and diffusion and technologicalinnovations. Section 4 outlines the implications for the analysis of the con-ditions of dynamic efficiency. The conclusions summarise the analysis.

2 THE INGREDIENTS OF PATH DEPENDENCE

According to Paul David, path dependence is an attribute of a special classof dynamic processes. A process is path dependent when it is non-ergodicand yet it is subject to multiple attractors: ‘systems possessing this propertycannot shake off the effects of past events, and do not have a limiting,invariant probability distribution that is continuos over the entire statespace’ (David, 1992a: 1).

As a matter of fact, historic analysis and much empirical evidence in eco-nomic growth, and specifically in the economics of innovation and newtechnologies, confirm that these characteristics apply and are most relevantto understanding the laws of change and growth of complex systems.

Path dependence provides a unique and fertile analytical framework ableto explain and assess the ever-changing outcomes of the combination andinterplay between factors of continuity and discontinuity, growth anddevelopment, hysteresis and creativity, routines and ‘free will’, that charac-terise economic action in a dynamic perspective which is able to appreciatethe role of historic time.

The notion of ergodicity (Figure 3.1) is quite complex and deservescareful examination. A process is ergodic when its initial conditions haveno influence on its development and eventual outcomes. The general equi-librium framework of analysis is typically ergodic, although the analysis ofthe competition process and its building blocks such as the theory of costsand of the firm are based upon short-term conditions where some costs arefixed and their irreversibility bears major consequences on the outcome ofthe interaction among firms in the marketplace.

When a process is non-ergodic, its initial conditions have an effect on itsdevelopment and on the final outcome. Past dependence is an extreme formof non-ergodicity. Historic, as well as social and technological, determin-ism fully belongs to past dependence. Here the characteristics of theprocesses that are analysed and their results are considered to be fully deter-mined and contained in their initial conditions. In the economics of innov-ation, past dependence has often been practised: the epidemic models ofdiffusion of innovations and the notion of technological trajectory aretypical examples of technological and social determinism. As such, these


53

Non

-erg

odic

pro

cess

esE

rgod

icpr

oces

ses

Pat

h de

pend

ence

Sto

chas

tical

lyco

nver

gent

Det

erm

inis

tical

lyco

nver

gent

and

glob

ally

stab

le

Pas

tde

pend

ence

Fig

ure

3.1

Typ

olog

y of

dyna

mic

pro

cess

es

models are non-ergodic and fully past dependent. The process takes placewithin a single corridor, defined at the outset, and external attractorscannot divert its route, nor can the dynamics of the process be altered byinternal factors.

Path dependence differs from past dependence in that irreversibility playsa role together with the initial conditions of a process. Its development andthe final result, however, are shaped by the influence of local externalitiesand especially local and internal feedbacks. Path dependence builds uponthe mix of the notions of irreversibility, local externalities and feedbacks.In so doing, path dependence can be considered at the border between fullyergodic processes and fully non-ergodic processes. In the former, historydoes not matter; in the latter, only history matters. Path dependence differsfrom deterministic past dependence in that irreversibility arises from eventsalong the path, and not only the initial conditions may play a role in select-ing from among the multiplicity of possible outcomes. The analysis of apath-dependent stochastic system is based upon the concepts of transientor ‘permanent micro-level’ irreversibilities, creativity and positive feed-backs. The latter self-reinforcing processes may work through the pricesystem, or they may operate through non-pecuniary externalities. The con-ceptualisation of stochastic path dependence can be considered to occupythe border region between the view of the world in which history enters onlyto establish the initial conditions after which the dynamics unfolds deter-ministically, and the conceptualisation of historical dynamics in which one‘accident’ follows another relentlessly and unpredictably. Path dependenceprovides economists with the tools to include historical forces in theiranalysis without succumbing to naive historical determinism. When pathdependence applies, history matters, but together with other factors.

The sequence of steps becomes a relevant issue in path dependence. Ateach step, in fact, the direction of the process can be changed because ofthe influence of new forces and attractors.

The full understanding of a path-dependent process requires a detailedanalysis of the sequence of the steps that have been made, and of theinteractions between the effects of irreversibility, local externalities andfeedbacks. Irreversibility pushes towards a trajectory where the initial con-ditions are replicated and command the direction. Local externalities andfeedbacks may exert a diverting effect. The dynamic interplay betweenthese elements shapes the actual characteristics of the process, its directionat each stage, and bears an influence on the following stages. At each stagethe balance between such dynamic forces may differ and, hence, so may thedirection of the process.

In path dependence the sequential interplay between the past-dependenteffects of irreversibility and the diverting effects of local externalities and


feedbacks plays a key role and qualifies the conditions of path depen-dence itself as an interface between ergodic and non-ergodic approachesto economics.

The careful identification of the stages of each dynamic process and ofthe changes in their alignment, and therefore in the interplay betweenergodic and non-ergodic forces, and the consequent localisation andexplanation of the points of inflection, bifurcation or change in directionand intensity, becomes a major task and a fertile area for investigation. Letus analyse briefly the three main ingredients of path dependence: irre-versibility, feedbacks and local externalities.

Irreversibility clearly is the prime ingredient and causal factor of thenon-ergodic dynamics. Irreversibility consists in the lack of dynamic mal-leability of production factors as well as tastes and preferences, reputationand routines. When irreversibility applies it is difficult and costly to changea given set of conditions; irreversibility engenders specific costs – the costsof switching from one condition to another.

The induced, and hence endogenous, change of both production andutility functions provides here a major contribution to understanding thedynamic feature of path dependence. Technological knowledge and prefer-ences are endogenous to the economic system as they are the result of thecreative reaction of agents to a mismatch between the actual state of prod-ucts and factors markets and their original plans and expectations. Buildingupon such expectations, agents embedded in historical time have madeplans and commitments upon which irreversible decisions have been taken.

The mismatch between the products and factors markets, conditions asplanned and built into the irreversible decisions which have been taken, andtheir actual conditions, is the basic inducement factor and focusing devicethat leads to creative reactions. The induced technological change traditionof analysis provides the glue between irreversibility and creativity (Arrow,2000; Ruttan, 1997; 2001).

Feedbacks are the most important ingredient of path dependence.Creativity and reactivity are key factors in understanding path dependence.Path dependence assumes that agents are able to react to the changing con-ditions of their environment, not only by adjusting prices to quantities, andvice versa, but also, and mainly, by changing their technology as well astheir preferences and tastes. The contribution of the prospect theoryaccording to which the more agents are exposed to frustration the less riskadverse they are, is most relevant in this context (Kahneman and Tversky,1979; Witt, 1998).

Path dependence assumes that knowledge and preferences are the resultof a dynamic process which is influenced by the initial conditions and yetis open to a variety of local attractors that shape the characteristics of the

Path dependence and the quest for dynamic efficiency 55

learning process and its eventual outcome. Learning is possible mainly,although not exclusively, when it is based upon repeated actions in a well-defined contextual set of techniques and relationships in doing and inusing. Hence, learning is mainly local in that it is limited by the scope andthe perimeter of the expertise acquired within a given technology, product-ive conditions and organisational sets. The cumulated effects of localisedlearning by doing and learning by using consist in higher levels of compe-tence and, eventually, the opportunity to generate new knowledge. Allchanges to the given set of techniques, production conditions and relation-ships are likely to engender relevant opportunity costs, in terms of missedopportunities for learning and hence for acquiring higher levels of compe-tence and generating new technological knowledge (David, 1993; 1994).

While irreversibility exerts a past-dependent effect that keeps the processwithin well-defined corridors, local externalities and local feedbacks exerta path-dependent effect which may diverge and push the process away fromthe initial direction (Table 3.1).

The locality of both externalities and feedbacks is clearly central in pathdependence. The understanding of the local span of influence of external-ities and feedbacks, as opposed to a general or global one, paves the way toappreciate the role of the variety of local contexts and, therefore, to under-standing the existence of multiple equilibria and hence multiple directionsand intensities of the process. The same forces at play at the outset of theprocess can lead to different paths and different outcomes when applied indifferent contexts and in different conditions.

Much systematic evidence and stylised classifications find here a clearexplanatory framework: the same technology applied in different regionscan prove to be more or less effective. Relative factor prices may exert astrong diverging effect that pushes each firm away from the initial trajec-tory. The path of followers cannot replicate that of leaders. The stages ofeconomic growth for each player differ according to the global context inwhich they take place.

The specific characteristic of technological knowledge as a collectiveactivity has major implications here, because of its low levels of appropri-ability, excludability and divisibility, articulated in complexity, cumulability


Table 3.1 Converging and diverging forces in past and path dependence

Diverging forces in path dependence Converging forces in past dependence

Collective learning Internal learningChanges in factors’ prices CreativityLocal externalities and feedbacks Irreversibility

and fungibility, in its production and in its use. Technological knowledgecannot be produced by any single agent in isolation. Technological know-ledge is itself both an input and an output; as such it is the product of thecontinual recombination and internal and external knowledge inputs. Inturn, knowledge output spills within well-circumscribed ‘commons’ andthis has major effects on the conduct of the firms able to access the techno-logical commons. Relative prices of production factors act as a powerfulfactor in selection and discrimination favouring some technologies andactions against others. Proximity among firms in regional, product andtechnological spaces is most important to access technological spillovers.

When the local conditions change, an issue of co-evolution and covari-ance between the internal dynamic conditions dictated by static anddynamic irreversibilities and the external irreversibilities emerges. Theco-evolution of the local environment and the dynamic characteristics ofthe processes in place can determine major discontinuities and drasticchanges in the path (Gould, 2002).

3 TYPES OF PATH DEPENDENCE IN THEECONOMICS OF INNOVATION

The identification of the basic ingredients of path dependence and theappreciation of their diverse and complementary roles in shaping thedynamics of path-dependent processes make it possible to classify differentforms of path dependence. The evidence of the economics of innovation isespecially rich in this context.

Two basic types of factors in path dependence can be identified whenattention is focused on the location of the engine for growth: whether inter-nal or external to each agent, and when the framework is applied to under-standing the introduction of new technologies or, rather, their diffusion.

Internal factors play a major role in past dependence. The latter takesplace when the path along which the process takes place is mainly deter-mined by the interplay between the irreversibility of production factorsand the conditions for localised learning, internal to each firm. The intro-duction of a new technology is determined here by the induced creativityof agents facing unexpected and possibly adverse emerging conditionsmainly determined by the lack of flexibility engendered by irreversibility.Creativity, however, is stimulated so as to compensate for the mismatch.Positive feedbacks are at work. Creativity builds upon localised learningand, hence, is better able to implement the techniques originally in place.The notion of path dependence elaborated by Paul David (1975) belongsto this case: firms are induced to introduce a new technology by their


internal characteristics in terms of irreversibility, and follow a path oftechnological change shaped by their internal characteristics in terms oflearning processes and acquired competence.

The analysis of path-dependent innovation identifies the conditions thatexplain the localised character of technological change, such as the mix ofirreversibility, induced innovation, local externalities and local endow-ments. Here the notion of localised technological change, first introducedby Atkinson and Stiglitz (1969) and subsequently elaborated by Paul David(1975), provides important complementarities. The mismatch between theirreversibility of the tangible and intangible stock of sunken inputs andthe actual conditions of both factors and products markets, affected by thecontinual introduction of unexpected innovations in the system, is theprime engine. This mismatch induces the creative reaction and the eventualintroduction of new localised technologies.

Myopic, but creative, agents introduce technological changes that arelocalised by their knowledge base built upon localised learning processes,the switching costs stemming from irreversible production factors and theexternal conditions of factors markets. While the rate of introduction ofnew technologies is induced by the mismatch between the irreversibility ofproduction factors and the actual conditions of the markets, the directionof the new technologies is induced by the relative prices of productionfactors. The access conditions to technological spillovers and externalknowledge available in the local pools of collective knowledge help explain-ing the localised direction of introduction of new technologies along a well-designed technological, technical, institutional and industrial path. Suchaccess conditions, in fact, affect in depth the actual results of the inducedinnovation activity of each agent and, hence, the incentives to innovate.

External factors play a role when path dependence is determined by con-ditions that are external to firms, but internal to the system. Firms areinduced to innovate and to follow a well-defined technological path by con-ditions that are found in the markets for products and factors, as well as inthe behaviours of consumers and competitors. A clear case for externalpath dependence emerges when the role of external knowledge is taken intoaccount and its contribution to the actual span of technological alterna-tives is appreciated. The understanding of the role of technological exter-nalities in the generation of new technological knowledge and in theintroduction of new technologies makes it possible to stress the role of sys-temic path dependence (Antonelli, 1999; 2001). Firms searching for a pos-sible reaction to unexpected events help each other with localised spilloversand reciprocal knowledge transfer that build upon a local knowledgecommon based on the competence and experience acquired in learning bydoing and learning by using. The direction of the process of introduction


of new technologies of each agent here is seen as the result of the collectiveknowledge available locally. External factors are also at work in pathdependence when local factors markets affecting the introduction andselection of new technologies are appreciated. The productivity and cost-effectivity of a new technology is influenced by composition effects(Antonelli, 2003).

Composition effects are the outcome of the sensitivity of output to therelative scale of each single factor, rather than to the scale of the bundle ofproduction factors. They are positive when the relative scale of most pro-ductive factors is augmented and that of the least productive factors isreduced. In general, because of composition effects, the larger the produc-tivity of the factor which is more widely used and the lower the productiv-ity of the factor which is less used, the larger are the effects of any changesin the relative levels of factor costs.

When the most productive factor is cheaper and hence its use is moreintensive, and the least productive factor is most expensive and hence its useis least intensive, production costs are lowest. The higher the growth oftotal factor productivity stemming from the introduction of a given tech-nology, the higher is the output elasticity of the productive factor whichlocally is most abundant.

Different agents, rooted in different regions, with different endowmentsand hence different conditions of their local factors markets may react withsimilar levels of creativity to similar changes in their current conditions,introducing new technologies with marked differences in terms of factorsintensity not only because of the effects of internal localised learning andthe access conditions to the local pools of collective knowledge, but alsobecause of the powerful consequences of composition effects. Here com-position effects act as an inducement factor that explain the direction of theintroduction of new technologies rather than their diffusion.

Irreversibility and the consequent switching costs matter also in a quitedifferent context. Irreversibility, in fact, exerts important effects on theselective diffusion of rival innovations. New technologies are sorted out notonly by their absolute levels of efficiency, but also with respect to their com-plementarity and compatibility with the installed stocks of fixed and irre-versible production factors. Firms with important stocks of fixed capitaland irreversible competence of their employees, which attach great value totheir customer base and to the relationships with their providers of inter-mediary inputs, will select the new technology which is not only more pro-ductive but also more compatible and more easily integrated within theexisting production process and within the network of relations in place.External factors such as relative prices in the factor markets and levels ofcompatibility and interoperativity between different products sold by


different firms play an important role in shaping the choice among rivaltechnologies, and in so doing play a role in path-dependent diffusion.

The identification of path-dependent diffusion, as distinct from path-dependent innovation, becomes relevant in this context and can be found inthe path-breaking analysis of Paul David (1975). Path-dependent innov-ation can be defined as the set of explanations that make it possible tounderstand why firms are actually better able to innovate within a limitedset of techniques. Path-dependent diffusion is the class of explanatoryanalysis that makes it possible to understand why firms adopt some tech-nologies, possibly in the proximity of existing technologies, both internaland external to each firm, rather than any possible technological innova-tions. More specifically, path-dependent diffusion assumes that technologi-cal innovations have been already introduced and that some rivalry andsubstitutability exists between them and the existing ones, as well as amongthem. Path-dependent innovation instead explains how and why firms inno-vate. The matrix of Table 3.2 provides a synthesis of the basic argument.

Path dependence has contributed substantially to understanding thediffusion of innovations, often bounded by the analysis of the adoptionof single innovations, isolated with respect to their own historic context ofintroduction. In the analysis of the determinants of path-dependentdiffusion, the distinction between internal and external factors is againrelevant.

The notion of path dependence elaborated by Brian Arthur (1989) andPaul David (1985) clearly contributes the analysis of diffusion: new tech-nologies are sorted out mainly by the effects of increasing returns to adop-tion at the system level. In their analysis, the selection and eventualdiffusion of new technologies is path dependent in that it is influencedby the timing of their sequential introduction, which in turn affects theirrelative profitability of adoption as shaped by the powerful conse-quences of positive feedbacks consisting of the interplay between network


Table 3.2 Types of path dependence

Path-dependent innovation Path-dependent diffusion

Factors internal Switching costs Complementarityto firms Localised internal learning Increasing returns

to productionFactors external Localised knowledge commons Network externalities

to firms Composition effects and on supply and demandrelative factors’ prices Composition effects and

relative factors’ prices

externalities on the demand side and increasing returns in production.Here the choice of the new technology is shaped by the factor and prod-ucts markets’ conditions.

The diffusion of new and better technologies can be delayed or barred bythe lack of compatibility with the internal and irreversible characteristicsof potential adopters. Here internal factors play the key role in explainingpath-dependent diffusion. The durability and irreversibility of internalfactors such as the capital stock, but also the skills of human capital, thelocation in a given space and the relationships, will determine adoption ofnew technologies with customers and suppliers.

For a given supply of new and rival technologies, composition effectsact as powerful selection devices and the diffusion of technologies will beinfluenced by the local conditions of factors markets. Labour-intensivetechnologies will diffuse faster in labour-abundant countries, and capital-intensive technologies will be adopted in capital-abundant technologies.The adoption of new technologies that are characterised by high levels ofoutput elasticity of labour, but small shift effects, might be delayed forever in capital-intensive countries.

Too much emphasis has been put on the effects of path-dependentdiffusion in terms of ‘lock in’. Technological ‘lock in’ is a possible outcomeof path-dependent diffusion, although new waves of better technologies,possibly introduced by other competitors may eventually break the tech-nological resilience.

The real key point is in fact not the ‘lock-in’ effects but rather the ‘lock-out’ effects: in path dependence firms are induced to change their currentstate of affairs by some unexpected events they cannot cope with – bymeans of traditional price-quantity adjustments – because of irreversibil-ities and constraints, on the one hand, and the opportunities for the intro-duction of new technologies, on the other. Such dynamics are fuelled byirreversibility and shaped by the changing effects of local externalities andfeedbacks, within a path.

4 THE ENGINE OF GROWTH: PATH DEPENDENCEAND LOCALISED TECHNOLOGICAL CHANGE

The different forms of path dependence that have been identified so far canbe considered complementary components of a broader dynamic frame-work which make it possible to understand the engine of growth and theconditions for dynamic efficiency.

This can help in understanding that the basic aim of path dependenceis to provide a first, and yet quite elaborate, framework to address the key


issues of the analysis of the conditions for dynamic efficiency. In so doing,path dependence provides a framework which is not necessarily in conflictwith static analysis and in general with the quest for static efficiency. Pathdependence addresses a different set of problems and a different set ofconditions.

In the standard static framework growth is exogenous and ergodic.Growth takes place when and if exogenous changes concerning the tech-nology, the preferences and the distribution of natural resources anddemography take place. The direction of change can be both positiveand negative; the system is able to adjust automatically to any new set ofexogenous conditions.

This context does not change, even when assumptions about the varietyand heterogeneity of agents are taken into account. Low levels of irration-ality, at least on the supply side, are also compatible. Irrational conductof firms will lead to suboptimal performances and, eventually, to forcedexit. Rational conduct will lead to survival in the marketplace. The notionof Brownian movement has been successfully borrowed from physics tocharacterise such a micro-dynamic context.

Entry and exit of firms, as well as changes in the levels of output and pos-sibly the exclusion of suboptimal conducts, are the prime engine of such aBrownian movement. New firms enter the market attracted by the gapbetween prices and costs margins, and firms leave the market when pricesfall below costs. Firms are not supposed to be able to change their tech-nology and, hence, their production functions. At best, these firms are ableto influence the position and slope of supply curves as a consequence oftheir entry and exit.

In the approach which builds upon the path-breaking contributions ofPaul David, and further elaborated so far, growth is endogenous and pathdependent. Growth is primarily the result of the endogenous changes intechnologies and tastes, hence in production and utility functions, whichtake place because of the creativity and reactivity of agents. Agents arecharacterised both by high levels of irreversibility and by high levels of cre-ative capability. Irreversibility exposes agents to substantial losses andrigidities when their plans are not fulfilled and their expectations are notrealised.

Creativity, however, makes it possible to consider, next to entry and exitand changes in output levels, the introduction of innovations as the otherpossible reaction to the mismatch between expectations and the actual con-ditions of factors and products markets.

Growth is the possible outcome of a system exposed to the continualmismatch between expectations and actual events, and yet is able toorganise and structure its creativity so as to change the technology and the


psychology. When the creative reaction is not appropriate and consistent,the system is kept within the standard conditions of the static equilibrium.

In this context it is clear that innovation feeds innovation. When and ifthe creative reaction to any mismatch is conducive to the actual introduc-tion of new products and new processes, a self-sustaining process of growthand innovation takes place. The introduction of innovations, in fact, is itselfthe factor that changes the conditions of equilibrium of both products andfactors markets.

The stochastic character of the growth process becomes clear in thiscontext. The actual levels of creativity engendered by the mismatch and theactual levels of technological change introduced depend upon a number ofcomplementary conditions. The mismatch is a necessary but non-sufficientcondition to induce the successful introduction of new technologies. Onlyin special circumstances do all the conditions apply and are conducive to astrong and positive creative reaction in terms of fast rates of introductionof new and highly productive technologies.

Even hyper-rational agents who have access to Olympic rationalitycannot be expected to foresee the outcomes of the innovative process. Wheninnovation is taken into account, the mismatch between expectations andactual factors and products market conditions is bound to take place withvarying levels of intensity and different gaps. At each round the mismatchcan lead to the generation of new technologies or decay into an equilibriumadjustment process leading to the conditions of static efficiency.

The conditions that are conducive to the actual introduction of innov-ations become, clearly, the central focus of the quest for the conditions ofdynamic efficiency.

The understanding of the relationship between the amount of entropywithin the system, that is, the amount and the distribution of the mismatchbetween expectations and the actual conditions of product and factors’markets, and the amount of creativity, is a first and central area of concern.Low levels of mismatch can be easily absorbed by firms that can simplyadjust prices to quantities and vice versa with low levels of attrition. Highlevels of entropy, however, are also likely to endanger the actual capabilityof firms to react properly and to be actually able to introduce successfulinnovations. Empirical analysis in this area is still lacking and can providethe basic ground upon which theoretical investigations can subsequently beelaborated.

The conditions that affect the levels of creativity and reactivity ofagents, for given levels of entropy, play a major role in assessing thedynamic efficiency of an economic system. The organisation of firms interms of hierarchical structure and decision-making is most importantboth with respect to their effects in terms of accumulation of competence


and technological knowledge, to their capability to actually convert tacitknowledge into timely innovations and, most important, to their ability toorganise a timely and successful reaction to the changing conditions oftheir business environment.

The debate about the effects of centralised versus decentralised decision-making, in other words, finds a new context of understanding which valuesthe dynamic characteristics of firms and the timing of their reactions.Decentralised decision-making favours the levels of creativity and reactiv-ity for agents facing unexpected changes in products and factors markets.The size of firms and the structure of hierarchical decision-making,however, favour the accumulation of knowledge and its timely conversioninto technological innovations.

The conditions for entry and mobility in the products markets are keyfactors in this context. The conditions of entry and the general conditionsfor entrepreneurship are most important. Barriers to mobility acrossmarkets also play a role as far as they limit the capability of incumbents totake advantage of their localised knowledge and the scope for techno-logical fungibility and technological convergence. The notion of competi-tion as a discovery process here is consistent with the quest for theconditions for dynamic efficiency.

The conditions for the effectiveness of the creativity are the third layer ofanalysis. The provision of knowledge externalities and the general condi-tions for the generation of new technological knowledge here becomecentral. The access to knowledge in general is the key factor that make itpossible for creativity to be valorised and to lead to actual solutions interms of higher productivity. The institutions of intellectual property rightsand quality of the public knowledge infrastructure are major factors in thiscontext.

The quality and the density of communication channels among learningagents and the access conditions to the flow of technological spillovers andtechnological interactions are also essential. The organisation of financialmarkets, the access to credit and financial resources at large are mostimportant in that they make it possible for new firms to identify the newmarkets and take advantage of the opportunities for generating localisedtechnological change. The institutions of labour markets also play a majorrole. Seniority systems favour the accumulation of competence and makeit easier to generate new technological knowledge. Mobility of qualifiedexperts embodying high levels of tacit knowledge and technological com-petence across firms, however, is conducive to better dissemination and cir-culation of technological knowledge.

Finally, the conditions that favour the working of the path-dependentengine of growth come into question. In this approach the variety of actors


and interacting markets matters (Metcalfe, 1997). Firms are induced toinnovate when their myopic expectations do not match the actual markets.Firms, in other words, react to all changes to their myopic expectations, notonly with changes in the price–output mix, but also with the introduction ofnew technologies. The larger the variety of firms, the larger are likely the mis-matches and, especially, the variety of alternatives that are elaborated andtested by the firms in the marketplaces. Hence, the larger are the chances thattechnological innovations that are actually superior and able significantly toincrease the general efficiency of the system are likely to be introduced.

Moreover, it is clear that the new technologies reflect the specific contextof action. Such a context includes firms active in a broad industrial struc-ture, which includes regions and countries with significant heterogeneity inboth technologies and endowments. The larger the more differentiated thecompetitive arena, the larger the incentives to introduce innovations and tosort the most productive ones: globalisation is likely to be the cause ratherthan the consequence of accelerated rates of introduction of technologicalchange.

In a population of heterogeneous agents, rooted in a heterogeneous eco-nomic space, with different local factors markets, different needs and pref-erences of consumers, and different pools of collective knowledge, alldiscrepancy between the expectations upon which irreversible action istaken and the actual conditions of the factors and products markets islikely to engender the localised introduction of new technologies. The cap-ability of agents actually to generate successful innovations will varyaccording to the stock of localised learning they can mobilise and theaccess conditions to the external pools of collective knowledge.

The diffusion of such technologies will vary according to their charac-teristics in terms of output elasticity and shift effects. Increasing returns toadoption on the demand side and the rates of adaptation of consumers tothe new products will play a major role in shaping the diffusion of newproduct innovations.

The introduction and diffusion of innovations, however, is likely to affectthe conditions of the product and factors markets upon which the expect-ations of other agents had been built. New discrepancies arise and new feed-backs are likely to take place with an endless process of creative reaction.

It is clear that the wider the heterogeneity within the system, the largerthe chances that expectations do not match the actual conditions and,hence, the faster the rates of introduction of new technologies and, pos-sibly, the faster the rates of growth.

The dynamic complementarity and interdependence among the creativeefforts of each agent is likely to play a major role in assessing the eventualoutcome. When the innovative efforts of agents happen to be co-ordinated


so as to add complementary bits of new technological knowledge, newtechnological systems can emerge. New gales of innovation are introduced,and growth, at the system level, can take momentum.

5 CONCLUSIONS

The notion of path dependence provides a unique framework within whichto analyse the conditions for dynamic efficiency. Path dependence provides,in fact, an extraordinary and articulated set of interrelated notions, whichmake it possible to go beyond the analysis of the static conditions forgeneral equilibrium and efficiency to apply. Path dependence builds upon afew simple basic elements: irreversibility and historic time, innovationviewed as a creative reaction, and local externalities and feedbacks. Thesefour elements are put in place and integrated into basic economics, so as toprovide a basis upon which a fully articulated post-Walrasian approach canbe elaborated – one where the central role of the markets as mechanismsfor the creation and distribution of incentives is recognised and empha-sised. In such an approach, however, the welfare attributes of equilibriumare no longer valid. Equilibrium itself is questioned. Sequences of possibleequilibria can be identified and traced.

The investigation of the conditions that make it possible for firms toconvert the entropy of the systems, as determined by the continual mis-match between expected and actual market conditions, into new and bettertechnologies, and hence to feed a rapid and effective growth, is the basicobject of investigation and problematic core of an approach that buildsupon the notion of path dependence.

Path dependence provides a framework, which makes it possible to gobeyond the deterministic and static world of the Walrasian equilibrium.A new landscape springs up, one where the dynamic outcomes of the inter-actions of agents in the market places cannot be fully anticipated. A varietyof possible outcomes can be predicted, as well as their sequence and theirdynamic relationship. Walrasian equilibrium is but one of the many possi-ble outcomes, as well as growth. Growth is the positive, and stochastic,result of a system of interactions where agents are able to react to the short-comings of the mismatch between expectations and the actual conditionsof the products and factors markets changing the equilibrium conditionsof the system. A new scope for economic policy is now available. The goalis clearly the build-up and maintenance of a social, institutional and eco-nomic environment, which is conducive to fully appreciating, valorisingand making effective the innovative reactions of myopic but creative andlearning agents.


From the viewpoint of economic analysis, path dependence proves to befar less hostile and conflicting with neoclassical analysis than is oftenassumed. As a matter of fact, path dependence seems even to provideimportant opportunities to rescue relevant portions of the received trad-ition, especially when the distinction between short-term and long-termanalysis is appreciated. The definition of the interfaces of compatibilityand incompatibility between the conditions for static efficiency and theconditions for dynamic efficiency becomes an interesting area for theoret-ical analysis.

The consequences of the interactions of agents in markets, however, nowinclude the generation and use of new and better technologies, rather thanthe single gravitation around a stable and unique attractor: the notion ofcomplex dynamics seems to provide a suitable framework to pursue suchanalysis.

NOTE

1. This work is the result of systematic redrafting, polishing and refining, since its first pre-sentation at the Conference in Honour of Paul David at the Accademia delle Scienze inTorino, in May 2000. The comments of many, including Paul David’s, are acknowledged.It makes explicit many intellectual debts and tentative recombinations: in so doing it pro-vides evidence on the attempt to introduce incremental knowledge along a well-definedpath sufficient to become a clear example of a deliberate effort to stand on the shouldersof a giant. The present version of the paper is also the text of the ‘Laudatio’ for the awardof the Laurea Honoris Causa in Communication Studies to Paul David for his contribu-tions to the economics of innovation, held at the Università degli Studi di Torino, 12 May2003.

BIBLIOGRAPHY

Antonelli, Cristiano (1995), The Economics of Localized Technological Change andIndustrial Dynamics, Boston, MA, and Dordrecht, Kluwer.

Antonelli, Cristiano (1999), The Microdynamics of Technological Change, London,Routledge.

Antonelli, Cristiano (2001), The Microeconomics of Technological Systems, Oxford,Oxford University Press.

Antonelli, Cristiano (2003), The Economics of Innovation, New Technologies andStructural Change, London, Routledge.

Arrow, Kenneth J. (2000), ‘Increasing returns: historiographic issues and pathdependence’, European Journal of History of Economic Thought, 7, 171–80.

Arthur, Brian (1989), ‘Competing technologies increasing returns and lock-in bysmall historical events’, Economic Journal, 99, 116–31.

Arthur, Brian (1994), Increasing Returns and Path Dependence in the Economy, AnnArbor, MI, Michigan University Press.


Atkinson, Anthony B. and Stiglitz, Joseph E. (1969), ‘A new view of technologicalchange’, Economic Journal, 79, 573–8.

David, Paul A. (1975), Technical Choice Innovation and Economic Growth,Cambridge, Cambridge University Press.

David, Paul A. (1985), ‘Clio and the economics of QWERTY’, American EconomicReview, 75, 332–7.

David, Paul A. (1987), ‘Some new standards for the economics of standardizationin the information age’, in Dasgupta, Paul and Stoneman, Paul (eds), EconomicPolicy and Technological Performance, Cambridge, Cambridge University Press.

David Paul A. (1988), ‘Path dependence: putting the past into the future of eco-nomics’, mimeo, Department of Economics, Stanford University.

David, Paul A. (1990), ‘The dynamo and the computer: a historical perspective onthe productivity paradox’, American Economic Review, (P&P) 80, 355–61.

David, Paul (1992a), Path dependence in Economic Processes: Implications for PolicyAnalysis in Dynamical System Contexts, Torino, Fondazione Rosselli.

David, Paul A. (1992b), ‘Heroes herds and hysteresis in technological history’,Industrial and Corporate Change, 1, 129–79.

David, Paul A. (1993), ‘Knowledge property and the system dynamics of techno-logical change’, Proceedings of the World Bank Annual Conference onDevelopment Economics, Washington, DC, World Bank.

David, Paul A. (1994), ‘Positive feedbacks and research productivity in science:reopening another black box’, in Granstrand, Owe (ed.), Economics andTechnology, Amsterdam, Elsevier.

David, Paul A. (1997a), ‘From market magic to calypso science policy: a review ofTerence Kealey’s “The economic laws of scientific research” ’, Research Policy,26, 229–55.

David, Paul A. (1997b), Path Dependence and the Quest for Historical Economics:One More Chorus of the Ballad of QWERTY, University of Oxford DiscussionPapers in Economic and Social History, Number 20.

David, Paul A. (1998), ‘Communication norms and the collective cognitiveperformance of “Invisible Colleges” ’, in Barba Navaretti, Giorgio et al. (eds),Creation and the Transfer of Knowledge: Institutions and Incentives, Berlin,Springer-Verlag.

David, Paul A. (2000a), ‘Path dependent learning, and the evolution of beliefs andbehaviours’, in Pagano, Ugo and Nicita, Antonio (eds), The Evolution ofEconomic Diversity, London, Routledge.

David, Paul A. (2000b), ‘Path dependence and varieties of learning in the evolutionof technological practice’, in Ziman, John (ed.), Technological Innovation as anEvolutionary Process, Cambridge, Cambridge University Press, ch. 10.

David, Paul A. (2001), ‘Path dependence, its critics, and the quest for “historical eco-nomics” ’, in Garrouste, Pierre and Ioannidis, Stravos (eds), Evolution and PathDependence in Economic Ideas: Past and Present, Cheltenham, Edward Elgar.

Gould, Steven J. (2002), The Structure of Evolutionary Theory, Cambridge, MA,Harvard University Press.

Kahneman, Daniel and Tversky, Amos (1979), ‘Prospect theory: an analysis ondecision under risk’, Econometrica, 47, 263–91.

Metcalfe, John S. (1997), Evolutionary Economics and Creative Destruction,London, Routledge.

Pfenninger, Karl H. and Shubik, Valerie R. (eds) (2001), The Origins of Creativity,Oxford, Oxford University Press.


Ruttan, Vernon W. (1997), ‘Induced innovation evolutionary theory and pathdependence: sources of technical change’, Economic Journal, 107, 1520–29.

Ruttan, Vernon W. (2001), Technology Growth and Development: An InducedInnovation Perspective, Oxford, Oxford University Press.

Witt, Ulrich, (1998), ‘Imagination and leadership: the neglected dimension of anevolutionary theory of the firm’, Journal of Economic Behavior and Organization,35, 161–77.


4. A history-friendly model ofinnovation, market structure andregulation in the age of randomscreening of the pharmaceuticalindustry1

Franco Malerba and Luigi Orsenigo2

1 INTRODUCTION

In this chapter we present a discussion of the long-term history of the evo-lution of the pharmaceutical industry and a related ‘history-friendly’model of the random screening period. This is in line with Paul David’smajor interests (two of many) of linking history and modelling.

Pharmaceuticals constitutes an ideal subject for history-friendly analysis.Pharmaceuticals are traditionally a highly research and development(R&D) intensive sector, which has undergone a series of radical technolog-ical and institutional ‘shocks’. However, the core of leading innovativefirms and countries has remained quite small and stable for a very longperiod of time, but the degree of concentration has been consistently low,whatever level of aggregation is considered. In addition, these patterns ofindustrial dynamics are intimately linked to two main factors: the nature ofthe processes of drug discovery and the fragmented nature of the relevantmarkets. Specifically, innovation processes have been characterised for avery long time by a low degree of cumulativeness and by ‘quasi-random’procedures of search (random screening). Thus, innovation in one market(a therapeutic category) does not entail higher probabilities of success inanother one. Moreover, pharmaceuticals represents a case where competi-tion is less dissimilar to the model of patent races. Understanding if theseintuitive factors can indeed explain the observed patterns of industrialdynamics and articulating the mechanisms through which they exert theirimpact are in themselves interesting challenges; the more so, if this modelis compared with the analysis of the computer industry. The comparisonmight allow for some generalisations about the determinants of the relevant

70

similarities and differences in the patterns of industrial evolution acrossindustries.

In addition, the pharmaceutical industry, ever since its inception, has beendeeplyaffectedbya largevarietyof institutional factorsandpolicies, rangingfrompatents,different formsof regulation(procedures forproductapproval,price controls, and so on), organisation of the public research systems, andso on. From this perspective, pharmaceuticals constitute an ideal case forstudying the differential impact and the working of alternative policies.

The chapter is organised as follows. Section 2 provides a historicalaccount of the evolution of the pharmaceuticals. Section 3 introduces themain theoretical issues that are raised by the previous historical accountand presents the model. Section 4 examines some alternative runs andsection 5 concludes.

2 INNOVATION AND THE EVOLUTION OFMARKET STRUCTURE IN THEPHARMACEUTICAL INDUSTRY: AN OVERVIEW

The patterns of development of the pharmaceutical industry have beenanalysed extensively by several scholars. In what follows, we rely especiallyon the work by Chandler (1990; 1999), Galambos and Lamoreaux (1997),Galambos and Sewell (1996), Galambos and Sturchio (1996), Gambardella(1995), Henderson et al. (1999), Orsenigo (1989) and Schwartzman (1976).But actually, this account of the history is largely drawn from Hendersonet al. (1999) and Pisano (1996).

In very general terms, the history of the pharmaceutical industry can beanalysed as an evolutionary process of adaptation to major technologicaland institutional ‘shocks’. It can be usefully divided into three majorepochs. The first, corresponding roughly to the period 1850–1945, was onein which little new drug development occurred, and in which the minimalresearch that was conducted was based on relatively primitive methods.The large-scale development of penicillin during the Second World Warmarked the emergence of the second period of the industry’s evolution.This period was characterised by the institution of formalised in-houseR&D programmes and relatively rapid rates of new drug introduction.During the early part of the period the industry relied largely on so called‘random’ screening as a method for finding new drugs, but in the 1970s theindustry began a transition to ‘guided’ drug discovery or ‘drug develop-ment by design’ a research methodology that drew heavily on advances inmolecular biochemistry, pharmacology and enzymology. The third epochof the industry has its roots in the 1970s but did not come to full flower

A history-friendly model of innovation 71

until quite recently as the use of the tools of genetic engineering in the pro-duction and discovery of new drugs has come to be more widely diffused.

2.1 Early History

The birth of the modern pharmaceutical industry can be traced to the mid-nineteenth century with the emergence of the synthetic dye industry inGermany and Switzerland. During the 1880s, the medicinal effects (such asantiception) of dyestuffs and other organic chemicals were discovered. Itwas thus initially Swiss and German chemical companies such as Ciba,Sandoz, Bayer and Hoechst, leveraging their technical competencies inorganic chemistry and dyestuffs, who began to manufacture drugs (usuallybased on synthetic dies) later in nineteenth century. For example, salicylicacid (aspirin) was first produced in 1883 by the German company, Bayer.

In the US and the UK, mass production of pharmaceuticals also beganin the later part of the nineteenth century. However, the pattern of devel-opment in the English-speaking world was quite different from that ofGermany and Switzerland. Whereas Swiss and German pharmaceuticalactivities tended to emerge within larger chemical producing enterprises,the USA and the UK witnessed the birth of specialised pharmaceuticalproducers such as Wyeth (later American Home Products) Eli Lilly, Pfizer,Warner-Lambert and Burroughs-Wellcome. Up until the First World WarGerman companies dominated the industry, producing approximately80 per cent of the world’s pharmaceutical output.

In the early years the pharmaceutical industry was not tightly linked toformal science. Until the 1930s, when sulfonamide was discovered, drugcompanies undertook little formal research. Most new drugs were based onexisting organic chemicals or were derived from natural sources (forexample, herbs) and little formal testing was done to ensure either safety orefficacy.

2.2 The ‘Random Screening’ Period

World War II and wartime needs for antibiotics marked the drug industry’stransition to an R&D-intensive business. With the outbreak of the SecondWorld War, the US government organised a massive research and produc-tion effort that focused on commercial production techniques and chemicalstructure analysis. More than 20 companies, several universities, and theDepartment of Agriculture took part. The commercialisation of penicillinmarked a watershed in the industry’s development. Due partially to thetechnical experience and organisational capabilities accumulated throughthe intense wartime effort to develop penicillin, as well as to the recognition


that drug development could be highly profitable, pharmaceutical compa-nies embarked on a period of massive investment in R&D and built large-scale internal R&D capabilities. At the same time there was a verysignificant shift in the institutional structure surrounding the industry.Whereas, before the war, public support for health-related research hadbeen quite modest, after the war it boomed to unprecedented levels, helpingto set the stage for a period of great prosperity.

This period was a golden age for the pharmaceutical industry. Researchand development spending literally exploded, and with this came a steadyflow of new drugs. Drug innovation was a highly profitable activity duringmost of this period. During the early 1980s, double-digit rates of growth inearnings and return-on-equity were the norm for most pharmaceuticalcompanies and the industry as a whole ranked among the most profitablein the USA and in Europe.

A number of structural factors supported the industry’s high averagelevel of innovation and economic performance. One was the sheer magni-tude of both the research opportunities and the unmet needs. In the earlypost-war years, there were many physical ailments and diseases for whichno drugs existed. In every major therapeutic category – from painkillersand anti-inflammatories to cardiovascular and central nervous systemproducts – pharmaceutical companies faced an almost completely openfield (before the discovery of penicillin, very few drugs effectively cureddiseases).

Faced with such a ‘target rich’ environment but very little detailed know-ledge of the biological underpinnings of specific diseases, pharmaceuticalcompanies invented an approach to research now referred to as ‘randomscreening’. Under this approach, natural and chemically derived com-pounds are randomly screened in test tube experiments and laboratoryanimals for potential therapeutic activity. Pharmaceutical companiesmaintained enormous ‘libraries’ of chemical compounds, and added totheir collections by searching for new compounds in places such as swamps,streams and soil samples. Thousands of compounds might be subjected tomultiple screens before researchers honed in on a promising substance.Serendipity played a key role since in general the ‘mechanism of action’ ofmost drugs were not well understood. Researchers were generally forced torely on the use of animal models as screens. Under this regime it was notuncommon for companies to discover a drug to treat one disease whilesearching for a treatment for another. Since even the most productivechemist might find it difficult to synthesise more than a few compoundsover the course of a week, researchers tended to focus their attention onsynthesising variants of compounds that had already shown promisingeffects in a screen, but that might not be ideally suited to be a drug. Any


given compound might have unacceptable side effects, for example, or bevery difficult to administer. The ‘design’ of new compounds was a slow,painstaking process that drew heavily on skills in analytic and medicinalchemistry. Several important classes of drugs were discovered in this way,including most of the important diuretics, many of the most widely usedpsychoactive drugs and several powerful antibiotics. While chemistsworking within this regime often had some intuitive sense of the linksbetween any given chemical structure and its therapeutic effect, little of thisknowledge was codified, so that new compound ‘design’ was driven asmuch by the skills of individual chemists as it was by a basis of systematicscience.

Random screening worked extremely well for many years. Severalhundred new chemical entities (NCEs) were introduced in the 1950s and1960s, and several important classes of drug were discovered in this way.However, the successful introduction of NCEs has to be considered as aquite rare event. Indeed, estimates suggest that, out of all new compoundsthat were discovered only one in 5000 reached the market. So, the rate ofintroduction has been of the order of a couple of dozens per year, and con-centrated in some fast-growing areas such as the central nervous system,cardiac therapy, anti-infectives and cytostatics. Innovative new drugsarrived quite rarely but after the arrival they experienced extremely highrates of market growth. In turn, this entailed a highly skewed distribution ofthe returns on innovation and of product market sizes as well as of the intra-firm distribution of sales across products. So a few ‘blockbusters’ dominatethe product range of all major firms (Matraves, 1999; Sutton, 1998).

As is well known, however, new products do not ensure profits. Rents frominnovation can be lost through competition unless ‘isolating mechanisms’are in place to inhibit imitators and new entrants. Indeed, for most of thepost-war period, pharmaceutical companies (particularly those operatingin the USA) had a number of isolating mechanisms working in their favour.

Several of these mechanisms, including the strength of intellectual prop-erty protection and the nature of the regulatory regime for pharmaceuticalproducts, were institutional in origin and differed significantly acrossnational boundaries. We discuss these types of mechanisms in more detailbelow. However, it is important to note that the organisational capabilitiesdeveloped by the larger pharmaceutical firms may also have acted as iso-lating mechanisms. Consider, for example, the process of random screen-ing itself. As an organisational process, random screening was anything butrandom. Over time, early entrants into the pharmaceutical industry devel-oped highly disciplined processes for carrying out mass screening pro-grammes. Because random screening capabilities were based on internalorganisational processes and tacit skills, they were difficult for potential


entrants to imitate and thus became a source of first-mover advantage. Inaddition, in the case of random screening, spillovers of knowledge betweenfirms was relatively small since, when firms essentially rely on the law oflarge numbers, there is little to be learned from the competition.

Moreover, entirely new products (NCEs) only capture a part of innovativeactivities. ‘Inventing around’ existing molecules, the introduction of newcombinations among them or new ways of delivering them, and so on, con-stituted a major component of firms’ innovative activities broadly defined.

Thus, competition centred not only on new product introductions butalso on incremental advances over time, as well as on imitation and genericcompetition after patent expiration (allowing a large ‘fringe’ of firms tothrive). Processes of generation of new markets and of diversificationacross product groups was followed by processes of incremental innova-tion, development of therapeutic analogues, imitation and licensing. Fast-expanding markets allowed for the steady growth of both the first-comerand other early innovators.

The successful exploitation of the economic benefits stemming frominnovation also required the control of other important complementaryassets, particularly, competencies in the management of large-scale clinicaltrials, the process of gaining regulatory approval, and marketing and dis-tribution, which also have acted as powerful barriers to entry into theindustry.

As a consequence, throughout its history, the industry has been charac-terised by a significant heterogeneity in terms of firms’ strategic orienta-tions and innovative capabilities.

Indeed, ever since its inception, other firms specialised not in R&D andinnovation, but in the imitation/inventing around, production and mar-keting of products often invented elsewhere and sold over the counter. Thisgroup of firms included companies like Bristol-Myers, Warner-Lambert,Plough, American Home Products as well as almost all the firms in coun-tries like France, Italy, Spain and Japan. Conversely, the ‘oligopolistic core’of the industry has been composed of the early innovative entrants, joinedafter World War II by a few American and British firms, which maintainedover time an innovation-oriented strategy. The isolating mechanisms dis-cussed previously, combined with the presence of scale economies inpharmaceutical research, and marketing, may help to explain the dearth ofnew entries prior to the mid-1970s. Indeed, many of the leading firmsduring this period – companies like Roche, Ciba, Hoechst, Merck, Pfizer,and Lilly – had their origins in the ‘pre-R&D’ era of the industry. At thesame time, until the mid-1970s only a small number of new firms enteredthe industry, and even less entered its ‘core’. At the same time, the indus-try was characterised by quite low levels of concentration, at the aggregate


level (the pharmaceutical industry) but also in the individual sub-marketslike, for example, cardiovascular, diuretics, tranquillizers, and so on.

2.3 The Advent of Molecular Biology

Beginning in the early 1970s, the industry also began to benefit moredirectly from the explosion in public funding for health-related researchthat followed the Second World War. From the mid-1970s on, however, sub-stantial advances in physiology, pharmacology, enzymology and cellbiology – the vast majority stemming from publicly funded research – ledto enormous progress in the ability to understand the mechanism of actionof some existing drugs and the biochemical and molecular roots of manydiseases. This new knowledge had a profound impact on the process of dis-covery of new drugs. First, these advances offered researchers a signifi-cantly more effective way to screen compounds. In the place of the request‘find me something that will lower blood pressure in rats’ pharmacologistscould make the request ‘find me something that inhibits the action of theangiotensin II converting enzyme in a test tube’. In turn, the more sensitivescreens made it possible to screen a wider range of compounds, triggeringa ‘virtuous cycle’ in that the availability of drugs whose mechanisms ofaction was well known made possible significant advances in the medicalunderstanding of the natural history of a number of key diseases, advanceswhich in turn opened up new targets and opportunities for drug therapy.

These techniques of ‘guided search’ made use of the knowledge that aparticular chemical pathway was fundamental to a particular physiologicalmechanism. But until quite recently the new knowledge was not used in thedesign of new compounds that could be tested in such screens. The tech-niques of ‘rational drug design’ are the result of applying the new biologicalknowledge to the design of new compounds, as well as to the ways in whichthey are screened. If, to use one common analogy, the action of a drug ona receptor in the body is similar to that of a key fitting into a lock, advancesin scientific knowledge in the 1970s and 1980s greatly increased knowledgeof which ‘locks’ might be important, thus making the screening processmuch more precise. However, organic chemists were still forced to rely onrandom screening or on the elaboration of existing compounds in theirsearch for new drugs since they had no guidance as to what appropriate‘keys’ might look like. More recently, an improved understanding of mole-cular kinetics, of the physical structure of molecular receptors, and of therelationship between chemical structure and a particular compound’smechanism of action has greatly increased knowledge of what suitable‘keys’ might look like. Chemists are now beginning to be able to ‘design’compounds that might have particular therapeutic effects.


These techniques were not uniformly adopted across the industry. Forany particular firm, the shift in the technology of drug research from‘random screening’ to one of ‘guided’ discovery or ‘drug discovery bydesign’ was critically dependent on the ability to take advantage of publiclygenerated knowledge (Gambardella, 1995; Henderson and Cockburn,1996) and of economies of scope within the firm (Henderson andCockburn, 1996). Smaller firms, those farther from the centres of publicresearch and those that were most successful with the older techniques ofrational drug discovery, appear to have been much slower to adopt thenew techniques than their rivals (Gambardella, 1995; Henderson andCockburn, 1996). There was also significant geographical variation inadoption. While the larger firms in the USA, the UK and Switzerland wereamongst the pioneers of the new technology, other European and Japanesefirms appear to have been slow to respond to the opportunities afforded bythe new science. These differences had significant implications for theindustry’s later response to the revolution in molecular biology.

This transition was in mid-course when molecular genetics and recombin-ant DNA technology opened an entirely new frontier for pharmaceuticalinnovation. The application of these advances initially followed two rela-tively distinct technical trajectories. One trajectory was rooted in the use ofgenetic engineering as a process technology to manufacture proteins whoseexisting therapeutic qualities were already quite well understood in largeenough quantities to permit their development as therapeutic agents. Thesecond trajectory used advances in genetics and molecular biology as toolstoenhancetheproductivityof thediscoveryof conventional ‘smallmolecule’synthetic chemical drugs. More recently, as the industry has gained experi-ence with the new technologies, these two trajectories have converged.

The advent of ‘biotechnology’ had a significant impact both on theorganisational competencies required to be a successful player in the phar-maceutical industry through their impact on the competencies required todiscover ‘conventional’, small molecular weight drugs and on industrystructure in general.

In the USA, biotechnology was the motivating force behind the firstlarge-scale entry into the pharmaceutical industry since the earlypost-Second World War period. The first new biotechnology start-up,Genentech, was founded in 1976 by Herbert Boyer (one of the scientistswho developed the recombinant DNA technique) and Robert Swanson,a venture capitalist. Genentech constituted the model for most of the newfirms. They were primarily university spin-offs and they were usuallyformed through collaboration between scientists and professional man-agers, backed by venture capital. Their specific skills resided in the know-ledge of the new techniques and in the research capabilities in that area.


Their aim consisted in applying the new scientific discoveries to commer-cial drug development, focusing on two main directions: diagnostics, on thebasis of monoclonal antibodies, and therapeutics.

Genentech was quickly followed by a large number of new entrants.Entry rates soared in 1980 and remained at a very high level thereafter,favoured also by the availability of venture capital and by the gradual estab-lishment of a very favourable climate concerning patenting.

Patents constituted a major problem. Particularly in the early stages, therelevant knowledge was to a large extent of a generic nature and could inprinciple be codified. Since the product of new biotechnology firms (NBFs)was essentially scientific results, patents were therefore crucial requisites forthe private appropriation of the profits generated by innovations. Yet con-siderable confusion surrounded the conditions under which patents couldbe obtained. These hurdles were gradually overcome, in the direction ofgranting ample concessions to industry. In particular, in 1980 Congresspassed the Patent and Trademark Amendments Act (Bayh-Dole Act),which in effects liberalised and actually encouraged the pursuit of patentprotection for inventions funded by government agencies. On the otherhand, again in 1980, the US Supreme Court ruled in favour of grantingpatent protection to living things (Diamond v. Chakrabarty) and in subse-quent years, a number of patents were granted establishing the right forvery broad claims (Merges and Nelson 1994).

Despite the high rates of entry, it took several years before the biotech-nology industry started to have an impact on the pharmaceutical market.The first biotechnology product, human insulin, was approved in 1982, andbetween 1982 and 1992, 16 biotechnology drugs were approved for the USmarket. As is the case for small molecular weight drugs, the distribution ofsales of biotechnology products is highly skewed. Three products weremajor commercial successes: insulin (Genentech and Eli Lilly), tPA(Genentech in 1987) and erythropoietin (Amgen and Ortho in 1989). By1991 there were over 100 biotechnology drugs in clinical development and21 biotechnology drugs with submitted applications to the Food and DrugAdministration (FDA) (Grabowski and Vernon, 1994; PharmaceuticalManufacturers Association, 1991): this was roughly one-third of all drugsin clinical trials (Bienz-Tadmor et al., 1992). Sales of biotechnology-derived therapeutic drugs and vaccines had reached $2 billion, and two newbiotechnology firms (Genentech and Amgen) have entered the club of thetop eight major pharmaceutical innovators (Grabowski and Vernon, 1994).

However, the large majority of these new companies never managed tobecome a fully integrated drug producer. The growth of NBFs as pharma-ceutical companies was constrained by the need to develop competencies indifferent crucial areas.


First, it was necessary to understand better the biological processesinvolved with proteins and to identify the specific therapeutic effects ofsuch proteins. Companies, in fact, turned immediately to produce thoseproteins (for example, insulin and the growth hormone) which weresufficiently well known. The subsequent progress of individual firms and ofthe industry as a whole was, however, predicated on the hope of being ableto develop a much deeper knowledge of the working of other proteins inrelation to specific diseases. Yet, progress along this line proved moredifficult than expected. Second, these companies lacked competencies inother different crucial aspects of the innovative process: in particular,knowledge and experience of clinical testing and other procedures relatedto product approval, on the one hand, and marketing, on the other. Thus,they exploited their essential competence and acted primarily as researchcompanies and specialised suppliers of high-technology intermediate prod-ucts, performing contract research for and in collaboration with establishedpharmaceutical corporations.

Collaboration allowed NBFs to survive and – in some cases – to pave theway for subsequent growth in many respects. First, clearly, it provided thefinancial resources necessary to fund R&D. Second, it provided the accessto organisational capabilities in product development and marketing.Established companies faced the opposite problem. While they needed toexplore, acquire and develop the new knowledge, they had the experienceand the structures necessary to control testing, production and marketing.

Indeed, large established firms approached the new scientific develop-ments mainly from a different perspective, that is, as tools to enhance theproductivity of the discovery of conventional ‘small molecule’ syntheticchemical drugs. There was enormous variation across firms in the speedwith which the new techniques were adopted. The adoption of biotechno-logy was much less difficult for those firms who had not made the transi-tion from ‘random’ to ‘guided’ drug discovery. For them, the tools ofgenetic engineering were initially employed as another source of ‘screens’with which to search for new drugs. Their use in this manner required a verysubstantial extension of the range of scientific skills employed by the firm;a scientific workforce that was tightly connected to the larger scientific com-munity and an organisational structure that supported a rich and rapidexchange of scientific knowledge across the firm (Gambardella, 1995;Henderson and Cockburn, 1996). The new techniques also significantlyincreased returns to the scope of the research effort (Henderson andCockburn, 1996).

In general, the larger organisations who had indulged a ‘taste’ for scienceunder the old regime were at a considerable advantage in adopting the newtechniques compared with smaller firms. On the contrary, firms that had


been particularly successful in the older regime and firms that were muchless connected to the publicly funded research community were muchslower to follow their lead. The embodiment of the new knowledge was inany case a slow and difficult process, because it implied a radical change inresearch procedures, a redefinition of the disciplinary boundaries withinlaboratories and, in some cases, in the divisional structure of the companyas well. Collaborative research with the NBFs and with universities allowedthese companies, in any case, to get access to the new technology and toexperiment in alternative directions. The advantages stemming from theseinteractions could be fully exploited, however, only through the contextualdevelopment of in-house capabilities, which made it possible to absorb andcomplement the knowledge supplied by external sources (Arora andGambardella, 1992). Collaboration with universities, NBFs and internalresearch were, indeed, strongly complementary.

Thus, a dense network of collaborative relations emerged, with the start-up firms positioned as upstream suppliers of technology and R&D services,and established firms positioned as downstream buyers who could providecapital as well as access to complementary assets. Networking was facili-tated by the partly ‘scientific’, that is, abstract and codified nature of theknowledge generated by NBFs (Gambardella, 1995), which made it possi-ble, in principle, to separate the innovative process in different verticalstages: the production of new scientific knowledge, the development of thisknowledge in applied knowledge, and the use of the latter for the produc-tion and marketing of new products. In this context, different types of insti-tutions specialised in the stage of the innovative process in which they wererelatively more efficient: universities in the first stage, NBFs in the secondstage and large firms in the third. A network of collaboration between theseactors provided the necessary co-ordination of the innovative process. Thenew firms acted as intermediaries in the transfer of technology betweenuniversities – which lacked the capability to develop or market the newtechnology – and established pharmaceutical firms that lacked technicalexpertise in the new realm of genetic engineering but that had the down-stream capabilities needed for commercialisation.

However, substantial costs remained in transferring knowledge acrossdifferent organisations, especially for the tacit and specific component ofknowledge. Moreover, the innovative process still involved the effectiveintegration of a wide range of pieces of knowledge and activities, whichwere not ordered in a linear way and might not easily be separated(Orsenigo, 1989). Thus, the processes of drug discovery and drug develop-ment still required the integration of different disciplines, techniques,search and experimental procedures and routines, which were not generallyseparable and codified.


Hence, firms were encouraged to pursue higher degrees of vertical inte-gration. Thus, some NBFs tried to vertically integrate downstream to pro-duction and marketing, becoming fully-fledged pharmaceutical companiesand directly challenging incumbents. Some of the latter tried to verticallyintegrate upstream, gaining full mastery of the new knowledge. In recentyears, moreover, a significant process of consolidation has begun, withboth mergers between biotechnology firms (for example, Cetus andChiron), as well as the acquisition of biotechnology firms by pharmaceuti-cal corporations (for example, Genentech and Hoffman LaRoche). Ernstand Young (1990) reported that the large majority (71 per cent of the com-panies surveyed) of the NBFs expected to be acquired by a large firm or tomerge with another NBF.

However, since knowledge is still fragmented and dispersed, and sincethe rate of technological change is still very high, no single institution isable to develop internally in the short run all the necessary ingredients forbringing new products to the marketplace. Each NBF, in effects, representsa possible alternative approach to drug discovery and a particular instanti-ation of the opportunities offered by the progresses of science. New gener-ations of NBFs have been created which adopt different approaches to theuse of biotechnology in the pharmaceutical industry. Large established cor-porations continue, therefore, to explore these new developments throughcollaborative agreements.

2.4 Institutional Environments

The proliferation of NBFs was essentially an American (and partly British)phenomenon. The development of the biotechnology segment in Europeand Japan lagged considerably behind the USA and rested on the activitiesof large established companies. The British and the Swiss companiesmoved earlier and more decisively in the direction pioneered by the largeUS firms in collaborating or acquiring American startups. But those firmsthat had smaller research organisations, were more local in scope or weremore orientated towards the exploitation of well-established research tra-jectories – in short, those firms that had not adopted the techniques of‘rational’ or ‘guided’ drug discovery – have found the transition moredifficult (Gambardella, 1995; Henderson and Cockburn, 1996): almost allthe established French, Italian and Japanese companies – but also theGerman giants – have been slow to adopt the tools of biotechnology as anintegral part of their drug research efforts.

More generally, ever since the mid-1970s the American, British andSwiss companies appear to have gained significant competitive advantagesvis-à-vis European firms, including the Germans. And traditionally the


continental European (except Germany and Switzerland) and Japaneseindustries have been much less orientated towards innovation than tostrategies based on imitation, production and marketing mainly for thedomestic market.

While the reasons for these differentiated patterns of evolution are stillcontroversial, institutional factors seem to have played a decisive role.Indeed, from its inception, the evolution of the pharmaceutical industryhas been tightly linked to the structure of national institutions. The phar-maceutical industry emerged in Switzerland and Germany, in part, becauseof strong university research and training in the relevant scientific areas.Organic chemistry was literally invented in Germany by Professor JustusLeibig, and German universities in the nineteenth century were leaders inorganic chemistry. Basel, the centre of the Swiss pharmaceutical industry,was the home of the country’s oldest university, long a centre for medicinaland chemical study. In the USA the government’s massive wartime invest-ment in the development of penicillin, as we discussed above, profoundlyaltered the evolution of American industry. In the post-war era, the insti-tutional arrangements surrounding the public support of basic research,intellectual property protection, procedures for product testing andapproval, and pricing and reimbursement policies have all strongly anddirectly influenced both the process of innovation and the economic returns(and thus incentives) for undertaking such innovation. We now turn to abrief review of these four key areas.

Public support for health-related researchNearly every government in the developed world supports publicly fundedhealth-related research, but there are very significant differences acrosscountries in both the level of support offered and in the ways in which it isspent. In the USA, public spending on health-related research took off afterthe Second World War and it is now the second largest item in the federalresearch budget after defence. Most of this funding is administered throughthe National Institutes of Health (NIH), although a significant fractiongoes to universities. Both qualitative and quantitative evidence suggeststhat this spending has had a significant effect on the productivity of thoselarge US firms that were able to take advantage of it (Henderson andCockburn, 1996; Maxwell and Eckhardt, 1990).

Public funding of biomedical research also increased dramatically inEurope in the post-war period, although total spending did not approachAmerican levels. Moreover, the institutional structure of biomedicalresearch evolved quite different in continental Europe as opposed to theUSA and the UK. For example, in continental Europe biomedical researchwas mainly concentrated in national laboratories rather than in medical


schools as happened in the USA and the UK. These differences in the levelsand sources of funds, along with a number of other institutional factors,have interacted in continental Europe to create an environment which, ingeneral, not only produces less science of generally lower quality, but alsoone in which science is far less integrated with medical practice.

Intellectual property protectionPharmaceuticals has historically been one of the few industries wherepatents provide solid protection against imitation (Klevorick et al., 1987).Because small variants in a molecule’s structure can drastically alter itspharmacological properties, potential imitators often find it hard to workaround the patent. Although other firms might undertake research in thesame therapeutic class as an innovator, the probability of them findinganother compound with the same therapeutic properties that did notinfringe on the original patent could be quite small. However, the scope andefficacy of patent protection has varied significantly across countries.

Both the USA and the majority of the European countries have providedrelatively strong patent protection in pharmaceuticals. In contrast, in Japanand in Italy, until 1976 and 1978 (respectively), patent law did not offer pro-tection for pharmaceutical products; only process technologies could bepatented. As a result, Japanese and Italian firms tended to avoid productR&D and to concentrate instead on finding novel processes for makingexisting molecules.

Procedures for product approvalPharmaceuticals are regulated products. Procedures for approval have aprofound impact on both the cost of innovating and on firms’ ability tosustain market positions once their products have been approved. As in thecase of patents, there are substantial differences in product approvalprocesses across countries.

Since the early 1960s most countries have steadily increased the strin-gency of their approval processes. However, it was the USA, with theKefauver-Harris Amendment Act in 1962, and the UK, with the MedicineAct in 1971, that took by far the most stringent stance among industrialisedcountries. Germany and especially France, Japan and Italy have historicallybeen much less demanding.

In the USA, the 1962 Amendment Act introduced a proof-of-efficacyrequirement for approval of new drugs and established regulatory controlsover the clinical (human) testing of new drug candidates. Specifically, theAmendments required firms to provide substantial evidence of a new drug’sefficacy based on ‘adequate and well controlled trials.’ As a result, after1962 the FDA shifted from a role as essentially an evaluator of evidence


and research findings at the end of the R&D process, to an active partici-pant in the process itself (Grabowski and Vernon, 1983).

The effects of the Amendments on innovative activities and marketstructure have been the subject of considerable debate (see, for instance,Chien, 1979, and Peltzman, 1974). They certainly led to large increases inthe resources necessary to obtain approval of a new drug application(NDA), and they probably caused sharp increases in both R&D costs andin the gestation times for NCEs, along with large declines in the annual rateof NCE introduction for the industry and a lag in the introduction of sig-nificant new drugs therapies in the USA when compared with Germanyand the UK. However, the creation of a stringent drug approval process inthe USA may also have helped create an isolating mechanism for innov-ative rents. Although the process of development and approval increasedcosts, it significantly increased barriers to imitation, even after patentsexpired.3

The institutional environment surrounding drug approval in the UK wasquite similar to that in the USA. As in the USA, the introduction of atougher regulatory environment in the UK was followed by a sharp fall inthe number of new drugs launched into Britain and a shakeout of theindustry. A number of smaller weaker firms exited the market and the pro-portion of minor local products launched into the British market shrunksignificantly. The strongest British firms gradually reoriented their R&Dactivities towards the development of more ambitious, global products(Thomas, 1994).

In other European countries, procedures for products approval were lessstringent. This allowed the survival of smaller firms specialised in the com-mercialisation of minor domestic products.

The structure of the health-care system and systems of reimbursementPerhaps the biggest differences in institutional environments across coun-tries was in the structure of the various health-care systems. In the USA,pharmaceutical companies’ rents from product innovation were furtherprotected by the fragmented structure of health-care markets and by theconsequent low bargaining power of buyers. Moreover, unlike mostEuropean countries (with the exception of Germany and the Netherlands)and Japan, drug prices in the USA are unregulated by government inter-vention. Until the mid-1980s the overwhelming majority of drugs weremarketed directly to physicians who largely made the key purchasing deci-sions by deciding which drug to prescribe.

The ultimate customers – patients – had little bargaining power, even inthose instances where multiple drugs were available for the same condition.Because insurance companies generally did not cover prescription drugs


(in 1960, only 4 per cent of prescription drug expenditures were funded bythird-party payers), they did not provide a major source of pricing lever-age. Pharmaceutical companies were afforded a relatively high degree ofpricing flexibility. This pricing flexibility, in turn, contributed to the pro-fitability of investments in drug R&D.

Drug prices were also relatively high in other countries that did not havestrong government intervention in prices, such as Germany and theNetherlands. In the UK, price regulation left companies to set their ownprices, but a global profit margin with each firm was negotiated which wasdesigned to assure each of them an appropriate return on capital invest-ment including research. The allowed rate of return was negotiated directlyand was set higher for export-oriented firms. In general, this scheme tendedto favour both British and foreign R&D-intensive companies which oper-ated directly in the UK. Conversely, it tended to penalise weak, imitativefirms as well as those foreign competitors (primarily the Germans) tryingto enter the British market without direct innovative effort in the UK(Burstall, 1985; Thomas, 1994).

On the contrary, in Japan, France and Italy price regulation was organ-ised in such a way to protect the domestic industry from foreign competi-tion and offered little incentive to ambitious innovative strategies(Henderson et al. 1999; Thomas, 1994).

In more recent times, the introduction of cost containment policies inalmost all countries has led to profound changes in these systems andintense debates about the efficiency of alternative systems in resolving thetrade-off between lower prices and incentives for innovation.

3 THE MODEL

3.1 Challenges for a History-friendly Model

As was discussed in section 2, there are several important conceptual issuesthat are raised by an analysis of the evolution of the pharmaceutical indus-try. In particular, we mentioned three of them: the relationships betweenthe properties of the regimes of search, the nature of markets, the patternsof competition and the evolution of market structure; the relationshipsbetween science and innovation; and the role and the impact of alternativeforms of public policy and regulation. In this chapter, we will restrict theanalysis to the era of random screening and we will address a subset ofthese issues.

Here, the thrust of the story can be summarised as follows. Firmscompete to discover, develop and market new drugs for a large variety of


diseases. They face a large space of – at the beginning – unexplored oppor-tunities. However, the search for new promising compounds is essentiallyrandom, because the knowledge of why a certain molecule can ‘cure’ a par-ticular disease and of where that particular molecule can be found islimited. That is to say, the role of ‘science’ here is modest. Thus, firmsexplore randomly the ‘space of molecules’ until they find one which mightbecome a useful drug and on which they obtain a patent. The patent pro-vides protection from imitation for a certain amount of time (patent dura-tion) and over a given range of ‘similar’ molecules (width of patents). Afterpatenting, firms engage in the development of the drug, without knowinghow difficult, time-consuming and costly the process will be and what thequality of the new drug will be. Then, the drug is sold on the market, whosesize is defined by the number of potential patients. Marketing expendituresallow firms to increase the number of patients they can access. At the begin-ning, the new drug is the only product available on that particular thera-peutic class. But other firms can discover competing drugs or imitate.Indeed, firms are characterised by different propensities towards innov-ation, on the one hand, and imitation and marketing, on the other.Innovators will therefore experience a burst of growth following the intro-duction of the new drug, but later its revenues and market shares will beeroded away by competitors and imitators.

Since discovery of a drug in a particular therapeutic class does not entailany advantage in the discovery of another drug in a different class(market) – except for the volume of profits they can reinvest in research anddevelopment – firms will start searching randomly again for a new producteverywhere in the space of molecules. Firms’ growth will then depend onthe number of drugs they have discovered (that is, in diversification intodifferent therapeutic categories), on the size of the markets they are in, onthe number of competitors, and on the relative quality and price of theirdrug vis-à-vis competitors. Given the large number of therapeutic cate-gories and the absence of any form of cumulativeness in the search anddevelopment process, no firm can hope to be able to win a large share in theoverall market, but – if anything – only in specific therapeutic categories fora limited period of time. As a result, the degree of concentration in thewhole market for pharmaceuticals will be low. However, a few firms willgrow and become large, thanks essentially to diversification.

Market structure is also likely to be heavily affected by institutional vari-ables. Essentially, in the previously recounted history and in the model thisis due to the patenting regime, to the strictness of the procedures forproduct approval and to the forms of price regulation. Here we look onlyat the former two variables, leaving the analysis of price controls to futureexercises.


Patents constitute a fundamental means for the appropriability of theeconomic benefits of innovation. Weak patenting regimes reduce the via-bility of innovative strategies vis-à-vis imitative, marketing-based strate-gies. Thus, the strengthening (weakening) of the degree of protectionoffered by patents should increase the rate of innovation and generatehigher (and more persistent) degrees of concentration. However, the rela-tionship between the tightness of the patenting regime and the rate ofinnovation is not necessarily a linear one. Beyond (below) certain maximum(minimum) levels, further increases (decreases) of the protection providedby patents may have little effect.

The introduction of more stringent procedures for product approval hasoften been indicated as leading to a reduction of the number of new drugs,to an increase of their quality, to the exit of smaller firms and to the growthof more innovative companies. Countries characterised by a more lenientapproach in this respect are reported to have been losing competitivenessin the long run, especially as far their ability to innovate is concerned.

3.2 The Topography of the Model

In this section we describe the basic structure of the model. The techno-logical and market environment in which pharmaceutical firms are active iscomposed by several therapeutic categories (TCs). Each TC has a differenteconomic dimension according to the number of potential customers. Thiseconomic size is expressed by the total potential sales (VTC) and it is exoge-nously given in the model. In our model there are n therapeutic categoriesTC, each of which has a specific VTC. VTC is set at the beginning of eachsimulation, it is a random number drawn from a normal distribution[VTCN(�V, �V)]. VTC grows in every period at a certain rate, ranging ran-domly between 0 and 2 per cent. Firms active in a certain TC get a share ofVTC equal to their market share.

Within each TC there are a certain number M of molecules, which firmsaim to discover and which are at the base of pharmaceutical products thatlater are introduced in the market. Each molecule has a certain quality Qthat could be visualised as the ‘height’ of a certain molecule (see Figure 4.1).In most of the cases (70 per cent), Q has a value equal to zero. In the other30 per cent of cases, it has a positive value, drawn from a normal distribu-tion [Q N(�Q, �Q)]. Figure 4.1 depicts the ‘landscape’ in term of thera-peutic areas and molecules that firms face.

Firms do not know the ‘height’ Q of a molecule. Once they engage in asearch process in a specific therapeutic category, they may ‘discover’ a mol-ecule or not. In case they do, firms start a research process (see below): theyobtain a patent only if the molecule has a positive quality. Molecules whose



quality does not pass the FDA quality check will not give an origin to theircorrelated products. In the standard run, this quality threshold (FDA) is setat fixed value.

A patent has a specific duration (pd) and width (w). Once patent dura-tion expires, the molecule becomes free for all the firms. A patent gives afirm the right to extent the protection also on the molecules situated in the‘neighbourhood’ – as defined by the width – of the molecule that has beenpatented. The protection in the neighbourhood of the existing patent by theinnovator has major consequences on the search process of competingfirms. Competing firms are in fact blocked in the developments of poten-tial molecules near the patented one.

Once the patent has been granted, the firm can start the development ofthe product based on that molecule. If product development is successful,the product gets an economic value PQ. The value of the product PQi is afunction of the value of the molecule Qi. That is:

PQi(1��)Qi (4.1)

where i1, 2 . . . 150 for each TC and ��U [�.25,� .25].Each product gives a certain level of utility to consumers (see section 2.3

for a discussion of demand). The value of the product influences the con-sumers demand for such drug.

3.3 The Firms

3.3.1 The basic features of firmsThe industry is populated by f firms. Each firm has a budget B, initiallyequal for all firms. Firms are characterised by three activities – search,research and marketing – but with different intensity in each activity. Some

Q TC1 TC2 TCn

Figure 4.1 Therapeutic categories and molecules

firms, in fact, may want to spend relatively more on research and less onmarketing; other firms the contrary. In the model, the marketing propen-sity of the firms, �, is characterised by a share of the budget randomly setin the interval [0.2–0.8]. Relatedly the search–research propensity by firmsis characterised by a share of the budget that complements the share of thepropensity to marketing, which is (1��).

Thus, the firms’ budget B is divided among search, research and market-ing activities according to the specific propensities discussed above:

Resources for search (BS): (1��)!BResources for research (BR): (1��) (1�!)B

Resources for marketing (BM): �B

where ! (randomly drawn from a uniform distribution ranging between0.05 and 0.15) is invariant and firm specific.

3.3.2 Innovators and imitatorsFirms are heterogeneous also in another respect: they can be innovators orimitators. The propensity to research determines whether a firm is an innov-ator or an imitator. If a firm has a propensity to research (1��) greaterthan r (a random number from a uniform distribution ranging from 0 to 1in each period), then the firm is an innovator. If the firm has a propensityto research (1��) lower than r, then it will be an imitator. In this way,innovative or imitative nature is a time dependent variable of the specificfirm: while the propensity to research is given initially and does not changeover time, r does.

3.3.2.1 Innovative search activities If firms are innovators, they look fornew molecules. The amount of money invested in search activities, BS, deter-mines the number (approximated to the nearest integer) X of therapeuticcategories TCs which are explored by a firm during its current project:

XBS/Drawcost (4.2)

where Drawcost is a parameter fixed in our simulations. This simple speci-fication implies that the number of TCs explored is linear with that fractionof budget BS that any one firm allocates for search processes. If X is lowerthan 1, then the firm is assumed to be able to explore only one TC.

3.3.2.2 Imitative search activities If a firm follows an imitation strategy,after having drawn a certain number of TCs (as defined in equation (4.3)),it looks for an existing molecule which is free (and thus not protected by


a patent any more). Imitators rate molecules according to their quality Qiwhich is, however, only imperfectly known. They select the molecule withthe highest ‘perceived’ quality, Ri. R is also a measure of the probability ofchoosing a molecule, that is a function of Q:

Ri(1��)Qi (4.3)

where i1, 2 . . . 150 for each TC and ��U[�.25,� .25].Hence, high-quality molecules will be more frequently picked up by imi-

tators, generating a congestion effect.

3.3.2.3 Research activities By ‘research activities’ we mean productdevelopment. Both innovators and imitators do research. If the molecule ispotentially interesting (that is, it has a quality Q greater than zero), the firmstarts a development project, using the budget BR. Remember that firms donot know the quality of the drug. Over time, given the research budget BRthe firm progresses towards the full development of the drug (attaining thevalue Q of the drug). That is to say, firms have to ‘climb’ Q steps in orderto develop a drug having a quality Q.

Each step implies a unitary cost CS. Thus, the total cost of developing adrug having a quality Q is equal to CS �Q. In each period a firm pays CS.However, firms differ in the speed of their development process: higherspeed implies higher costs. In each period, the progress that a firm makestowards Q (SP) is randomly drawn, and ranges from 1 to 5. Firms thatmove ahead faster in their research per period, pay more for each unitarystep the unitary cost CS of each step increases as SP increases according tothe following relationship:

(4.4)

where Cur is the cost of a single step for a firm that has a SP equal to 1 (thatis, it progresses with 1 step each period: only in this case CSCur). CS forimitative firms is set ¼ of the CS of innovating firms. In our simulations,Cur is equal to 15 and is fixed for every firm.

With its research resources, a firm may be able to reach Qi. In this case itstarts the commercialisation of the product. Otherwise, if Q is too ‘high’for the resources BR of the firm, the project fails with the firm. Moreover,a product must have a minimum quality, already defined as FDA, to beallowed to be sold in the marketplace. Below this value the drug cannot be

CS

(Cur �iSP

i1i)

SP


commercialised and the project fails. Only when research activities are over(that is, after a product has been created), will the firm start another processof search.

3.3.2.4 Marketing activities As previously mentioned, if a firm reachesQi, it has to launch a ‘new born’ product on the market. The firm has abudget BM available for that purpose. BM is divided by the firm in two parts,with shares h and 1�h (equal for all firms).

BM �h defines the marketing investment AjTL for the product j, where TLrefers to the launch of the product. It is spent only once, at the moment ofthe launch of the product. Marketing expenditures yield a certain level of‘product image’ for the consumers. This level of the ‘image’ is eroded withtime at a rate equal to eA in each period. In addition, the firm will profitfrom a marketing spillover � from its previous products k� j. The level ofthe ‘image’ Ajt in period t is given by:

(4.5)

BM �(1�h) defines the total yearly marketing expenditures that will bespent over TA periods. This factor captures the firms’ attempt of keepingthe level the ‘image’ over time.

Yearly marketing expenditures YAt will therefore be:

(4.6)

In our simulations, h is equal to 0.5, erosion (eA)0.01 and TA20.

3.3.3 Utility, demand and market shareDecisions to buy a specific drug depend on several factors, which togetheryield a specific ‘merit’ to each drug j. The value of this ‘merit’, Uj, is givenby:

(4.7)

PQj is the economic value of the drug as in equation (4.1). mup is thedesired rate of return that each firm wants to obtain from its drug. Ajt isthe product ‘image’ derived by the marketing investment for that productand YAjt is the yearly marketing expenditure for the product j, as already

Ujt PQaj · (1�mup)b · Ac

jt · YAdjt

YAt (1 � h)BM

TA t TL, . . ., TL � TA

Ajt Ajt�1·(1 � eA) � ��k�j

Ak�for t TL � 1, . . .,100

Ajt AjTL � ��k�j

Ak�� for t TL


defined. Exponents a, b, c are drawn from uniform distributions and arespecific to each therapeutic category; on the contrary d is equal in allmarkets. Finally mup is double for innovative products (mupinn) comparedwith that applied to imitative products (mupimi).

The market share MSfj of the firm f for the product j in each TC is thenproportional to its relative merit as compared to other competing drugs inthe same TC and it is given by:

(4.8)

Please keep in mind that a firm may have more than one product in a TC.Thus its market share in a TC is the sum of the market shares for all itsproducts.

3.3.4 Budget and accountingRevenues of firm f for product j are fj; because firm f may have more thanone product, total revenues (tot) are the sum of revenues obtained from allthe products of the firm, whatever TC has been explored:

(4.9)

In each period, the excess gross profits, that is, the difference between rev-enues, the current costs search, development and launch of the new products(in the periods when these activities take place) and the yearly expenditureson marketing, accumulate in an account that is used as a budget to financesearch, research and marketing investment when a new project is started.The division of the budget among the different activities follows the firm-specific parameters (� and !) already defined in the section 3.3.1.

Firms exit the market when their budget falls to zero.

4 THE SIMULATION RUNS

4.1 The Dynamics of Market Structure and Innovation

In the standard case (Standard), 30 firms with the same budget start theirinnovative and imitative activities. They differ in their propensity toresearch and marketing, and in their rate of progress. These firms search ina space composed of 50 therapeutic areas; each of them has 150 molecules(see the Appendix for the full set of parameters).

tot �fj

ij �fjk

(MSfj · VTCk)� f 1, . . .,30; k 1, . . .,50

MSfj Ufj

�TC

UTC


The dynamics of innovation and market structure in this industry is asfollows. Innovative firms start doing their innovative activities by search-ing over the therapeutic categories. Some of them will succeed to find apositive-quality molecule; they patent it and start doing research on thatmolecule. If they succeed in completing their research, they start doingmarketing activities for the support of their sales in the market. Demandfor the product will go to those firms that have higher product quality,higher initial marketing investments and higher marketing yearly expendi-tures, with different importance depending on the specific therapeutic cat-egory. Firms with larger demand will obtain higher market shares andhigher profits. Once the patent expires, the molecule becomes available onthe market for imitative activities.

Imitators do not have to do any search for new molecules. Rather, theylook around among existing available molecules whose patents haveexpired. They choose a molecule whose quality is higher than a given value.Then they start doing their research on that molecule, paying lower costs,and they proceed as in the case of innovators (with the exception that, oncetheir research is over, they do not obtain a patent). Then they start doingtheir marketing activities.

4.2 The Standard Run

The dynamics of market structure and innovation has been examined for100 periods and the values shown in the figures are averages over 100 runs.

The standard runs (Standard) show that in each therapeutic categoryconcentration (in terms of the Herfindahl index) is quite high and a firmmonopolizes the market (Figures 4.2 and 4.3). However, the overall marketconcentration is very high at the beginning, and then drastically declines asmore firms discover more molecules in various therapeutic areas(Figure 4.4). Selection is intense and 10 out of the 30 initial firms exit themarket (Figure 4.5). In the long run, almost all the therapeutic areas arediscovered (Figure 4.6). In each therapeutic area there are an increasingnumber of products (Figure 4.7a) and firms (Figure 4.7b). At the beginningthere are only innovative products, and then also imitative products (Figure4.8a). With the passage of time, the share of innovative products on thetotal products in the market (what we have called the Innovation Index)declines from 1 to approximately 0.6 at the end of the run (Figure 4.9). Onaverage each firm is increasingly present in more than one therapeutic cat-egory. They reach around 17 at the end of the run (see Figure 4.10). Finally,the value of the molecules discovered by firms over the total potential valueof the market (called Performance Index) increases with time and reaches11 percent of the total potential value of the whole market (Figure 4.11).


94

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.91

14

710

1316

1922

2528

3134

3740

4346

4952

5558

6164

6770

7376

7982

8588

9194

9710

0T

ime

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.2

Her

finda

hl in

dex

in e

ach

TC

95

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.91

18

1522

2936

4350

5764

7178

8592

99T

ime

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.3

Sha

re o

fle

ader

in e

ach

TC

96

0

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

0.55

0.60

16

1116

2126

3136

4146

5156

6166

7176

8186

9196

Tim

e

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.4

Her

finda

hl in

dex

– to

tal m

arke

t

97

01020

30

40

50

607080

90

100

15

913

1721

2533

3741

4549

5357

6165

6973

7781

8589

93

Tim

e

97

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.5

Sur

vivi

ng fi

rms

98

015304560

14

710

1316

1922

2528

3134

3740

4346

4952

5558

6164

6770

7376

7982

8588

9194

9710

0T

ime

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.6

Num

ber

ofdi

scov

ered

TC

s

99

0510152025303540

14

710

1316

1922

2528

3134

3740

4346

4952

5558

6164

6770

7376

7982

8588

9194

9710

0T

ime

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.7a

Ave

rage

num

ber

ofpr

oduc

ts in

eac

h T

C

100

036912151821242730

14

710

1316

1922

2528

3134

3740

4346

4952

5558

6164

6770

7376

7982

8588

9194

9710

0T

ime

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.7b

Ave

rage

num

ber

offir

ms

in e

ach

TC

101

050100

150

200

250

300

17

413

1019

1625

2231

2837

3443

4049

4655

5261

5867

6473

7079

7685

8291

8897

9410

0T

ime

Inno

vativ

e S

tand

ard

Imita

tive

Sta

ndar

d

Fig

ure

4.8a

Tot

al n

umbe

r of

inno

vati

ve a

nd im

itat

ive

prod

ucts

102

0

200

400

600

800

1000

1200

1400

Tim

e

Inno

vativ

e S

im 2

Imita

tive

Sim

2

17

413

1019

1625

2231

2837

3443

4049

4655

5261

5867

6473

7079

7685

8291

8897

9410

0

Fig

ure

4.8b

Tot

al n

umbe

r of

inno

vati

ve a

nd im

itat

ive

prod

ucts

103

Inno

vativ

e S

im 3

Imita

tive

Sim

3

050100

150

200

250

300

Tim

e1

74

1310

1916

2522

3128

3734

4340

4946

5552

6158

6764

7370

7976

8582

9188

9794

100

Fig

ure

4.8c

Tot

al n

umbe

r of

inno

vati

ve a

nd im

itat

ive

prod

ucts

104

020406080100

120

140

160

180

Tim

e

Inno

vativ

e S

im 4

Imita

tive

Sim

4

17

413

1019

1625

2231

2837

3443

4049

4655

5261

5867

6473

7079

7685

8291

8897

9410

0

Fig

ure

4.8d

Tot

al n

umbe

r of

inno

vati

ve a

nd im

itat

ive

prod

ucts

105

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

16

1116

2126

3136

4146

5156

6166

7176

8186

9196

Tim

e

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.9

Inno

vati

on I

ndex

(sh

are

ofin

nova

tive

pro

duct

s on

tot

al p

rodu

cts)

106

0510152025

Tim

e

Sta

ndar

dS

im 2

Sim

3S

im 4

17

413

1019

1625

2231

2837

3443

4049

4655

5261

5867

6473

7079

7685

8291

8897

9410

0

Fig

ure

4.10

Num

ber

ofT

Cs

in w

hich

firm

s ar

e ac

tive

107

0

0.05

0.10

0.15

0.20

0.25

0.30

0.35

14

710

1316

1922

2528

3134

3740

4346

4952

5558

6164

6770

7376

7982

8588

9194

9710

0T

ime

Sta

ndar

dS

im 2

Sim

3S

im 4

Fig

ure

4.11

Per

form

ance

Ind

ex (

aggr

egat

ed q

ualit

y of

disc

over

ed m

olec

ules

ove

r to

tal a

vaila

ble

qual

ity)

These results are the outcome of very different microdynamics in eachtherapeutic class. Results (not reported here) show that in few therapeuticareas are there no firms at all. Interestingly enough, in spite of somedifferences, the Herfindahl index is always rather similar. This shows that aleader soon emerges and that the rest of the firms remain rather marginalin that therapeutic area. Similarly, the number of products (not reportedhere) may range from zero to more than 10. We have reported, on the otherhand, the number of innovative and imitative products in each therapeuticarea: in most of the areas an innovative product corresponds to an imita-tive product (see Figure 4.12a).

4.3 Alternative Runs

The standard set reproduces some stylised facts of the pharmaceuticalindustry related to the level of concentration and the relationship betweeninnovators and imitators. We then did three counterfactual exercisesregarding the role of the initial number of firms, the changes in patent pro-tection and the approval procedures in terms of quality.

4.3.1 Increase in the number of firmsWhat would have happened if the number of firms would have been muchhigher, growing from 30 to 100 (Simulations 2)? As one would expect, con-centration in each therapeutic categories would have been lower, and theoverall concentration even lower. Nearly all the therapeutic areas wouldhave been discovered, with a higher number of firms in the market and ineach therapeutic category. On average, there would have been a greaternumber of innovative and imitative products in each therapeutic category.Moreover the value of the discovered molecules would increase to40 per cent of the total potential value.

4.3.2 The extension of time of patent protectionWhat would have happened if patent protection would be extended from20 periods to 60 periods (Simulations 3)? Interestingly enough, concentra-tion in each therapeutic area would not increase, because of its already highlevel. Only the number of firms in each therapeutic area slightly declines.On the contrary, the number of imitative products drastically decline,which in turns brings up the Innovation Index and reduces the number ofTCs in which firms are active.

4.3.3 Increase in the stringency of approval proceduresWe also did some other runs on the increase in the stringency of approvalprocedures (Simulations 4). We increased the quality check for obtaining a


109

01234567

01

23

45

67

Inno

vativ

e pr

oduc

ts

Imitative productsS

tand

ard

Fig

ure

4.12

aIn

nova

tive

and

imit

ativ

e pr

oduc

ts in

eac

h T

C a

t la

st p

erio

d

110

051015202530

02

46

810

1214

16In

nova

tive

prod

ucts

Imitative productsS

im 2

Fig

ure

4.12

bIn

nova

tive

and

imit

ativ

e pr

oduc

ts in

eac

h T

C a

t la

st p

erio

d

111

0

0.5

1.0

1.5

2.0

2.5

01

23

45

67

Inno

vativ

e pr

oduc

ts

Sim

3Imitative products

Fig

ure

4.12

cIn

nova

tive

and

imit

ativ

e pr

oduc

ts in

eac

h T

C a

t la

st p

erio

d

112

0

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

4.5

5.0

00.

51.

01.

52.

02.

53.

03.

54.

0In

nova

tive

prod

ucts

Imitative productsS

im 4

Fig

ure

4.12

dIn

nova

tive

and

imit

ativ

e pr

oduc

ts in

eac

h T

C a

t la

st p

erio

d

patent from 25 to 50. As expected, compared to the standard simulation,concentration increases and the total number of therapeutic areas discov-ered by firms decreases. Similarly, also the number of therapeutic classes inwhich firms are involved also decreases and the Performance Index is lower.Finally, when the single therapeutic classes are considered (Figure 4.12d),the differences across classes is higher.

5 CONCLUSIONS

In this chapter in honour of Paul David we merged two different types ofanalysis which lie at the core of his interests: a historical analysis regardingthe long-term evolution of the pharmaceutical industry (and focusing on awide range of factors such as the search process of firms, the role of insti-tutions and regulation) and a modelling analysis, attempting to model thedynamics of market structure and innovation in pharmaceuticals in ahistory-friendly way.

The history tried to examine three different periods: the first based onrelatively primitive methods, the second based on ‘random screening’ as amethod for finding new drugs and the third on ‘drug development bydesign’. The model focused on the random screening period. It has beenable to replicate the long-term evolution of the industry in terms of con-centration and the relationship between innovative and imitative productsin the industry. In a companion paper (Malerba-Orsenigo, 2002) we haveexamined also the period of drug development by design.

Some counterfactual exercises show that changes in the number offirms matter negatively for concentration and positively for the discoveryof new products. On the contrary, the increase in the length of patent pro-tection matters positively for the share of innovative products in themarket, and much less for concentration. Finally, an increase in the strin-gency of approval procedures has positive effects on concentration andnegative effects on the overall total quality of the products discovered byfirms.


APPENDIX

Parameter Symbol Value

Mean of normal distribution of positive �Q 50quality molecules

Standard deviation of normal distribution �Q 20of positive quality molecules

FDA quality check threshold FDA 25Mean of normal distribution of TCs value �V 3000Standard deviation of normal �V 500

distribution of TCs valueTotal number of firms (standard run) f 30Total number of TCs n 50Total number of molecules in each TC M 150Initial budget for every firm B 3000Drawcost in search activities [see eq. (4.2)] Drawcost 10 000Standard width of the patents w 10Standard patent duration pd 20Exponent of product quality (PQ) a [1.2–1.4]

(see Utility function [eq.(4.7)])Exponent of inverse of rate of return (1/mup) b [1.0–1.2]

(see Utility function [eq. (4.7)])Exponent of launch marketing expenditures c [0.1–0.2]

(A) (see Utility function [eq. (4.7)])Exponent of yearly marketing expenditures d 0.1

(YA) (see Utility function [eq.(4.7)])Desired rate of return for innovative products mupinn 0.2Desired rate of return for imitative products mupimi 0.1

NOTES

1. This chapter is a revised version of the paper presented at the conference in honour ofPaul David ‘New Frontiers in the Economics of Innovation and New Technologies’,Turin, 20 and 21 May, 2000.

2. We thank Luca Berga, Christian Garavaglia, Marco Gazzola and Nicola Lacetera fortheir invaluable contribution to the development of the model. Wesley Cohen andBronwyn Hall have provided useful and constructive comments. We thank the supportof the Italian CNR and of the 40 percent programme of the Italian Ministry ofUniversity and Research (MIUR).

3. Until the Waxman-Hutch Act was passed in the USA in 1984, generic versions of drugsthat had gone off patent still had to undergo extensive human clinical trials before theycould be sold in the US market, so it might be years before a generic version appearedeven once a key patent had expired. In 1980, generics held only 2 per cent of the US drugmarket.


BIBLIOGRAPHY

Arora, A. and Gambardella, A. (1992), ‘Complementarity and external linkage: thestrategies of the large firms in biotechnology’, Journal of Industrial Economics,37 (4), 361–79.

Arora, A. and Gambardella, A. (1994), ‘The changing technology of technicalchange: general and abstract knowledge and the division of innovative labor’,Research Policy, 23, (5), 523–32.

Arrow, K.J. (1962), ‘Economic welfare and the allocation of resources for inven-tion’, in Nelson, R.R. (ed.), The Rate and Direction of Inventive Activity.Economic and Social Factors, Princeton, NJ, Princeton University Press,pp. 609–25.

Bienz–Tadmor, B., Di Cerbo, P., Tadmor, G. and Lasagna, L. (1992), ‘Biopharma-ceuticals and conventional drugs: clinical success rates’ Bio/Technology, 10,521–5.

Bottazzi, G., Dosi, G., Lippi, M., Pammolli, F. and Riccaboni, M. (2001),‘Innovation and corporate growth in the evolution of the drug industry’,International Journal of Industrial Organization, 19 (7), 1161–87.

Breschi, S., Malerba, F. and Orsenigo, L. (2000), ‘Technological regimes andSchumpeterian patterns of innovation’, Economic Journal, 110, 388–410.

Burstall, M.L. (1985), The Community Pharmaceutical Industry, Luxembourg:Office for the Official Publications of the European Communities.

Chandler, A.D. (1990), Scale and Scope: The Dynamics of Modern Capitalism,Cambridge, MA, Bellknap Press of Harvard University Press.

Chandler, A.D. (1999), ‘Paths of learning: the evolution of high technology indus-tries’, manuscript, forthcoming.

Chien, R.I. (1979), Issues in Pharmaceutical Economics, Lexington, MA, LexingtonBooks.

Comanor, W.S. (1986), ‘The political economy of the pharmaceutical industry’,Journal of Economic Literature, 24, 1178–217.

David, P.A., Mowery, D.C. and Steinmueller, W.E. (1992), ‘Analyzing the economicpayoffs from basic research’, Economic of Innovation and New Technology, 2 (4),73–90.

Ernst and Young (1990), Biotech 91: A Changing Environment, San Francisco: Ernstand Young.

Galambos, L. and Lamoreaux, N.R. (1997), ‘Understanding innovation in thepharmaceutical industry’, paper presented at conference on ‘UnderstandingInnovation’, Baltimore, MD, June.

Galambos, L. and Sewell J.E. (1996), Network of Innovators. Vaccine Developmentat Merck, Sharp & Dohme and Mulfor, 1895–1995, Cambridge, CambridgeUniversity Press,

Galambos, L. and Sturchio, J. (1996), ‘The pharmaceutical industry in the twenti-eth century: a reappraisal of the sources of innovation’, History and Technology,13 (2), 83–100.

Galambos, L. and Sturchio, J. (1998), ‘Pharmaceutical firms and the transition tobiotechnology: a study in strategic innovation’, Business History Review, 72,250–78.

Gambardella, A. (1995), Science and Innovation in the US Pharmaceutical Industry,Cambridge, Cambridge University Press.


Gambardella, A., Orsenigo, L. and Pammolli, F. (2000), ‘Global competitivenessin pharmaceuticals: a European perspective’, report prepared for theDirectorate General Enterprise of the European Commission, EPRIS WorkingPaper n. 3.

Garavaglia, C., Lacetera, N., Malerba, F. and Orsenigo, L. (2002), ‘Project diversi-fication in pharmaceuticals in the random screening paradigm: a simulationmodel’, working paper.

Grabowski, H. And Vernon, J. (1983), The Regulation of Pharmaceuticals:Balancing the Benefits and the Risks, Washington, DC, AEI.

Grabowski, H. and Vernon, J. (1994), ‘Innovation and structural change in phar-maceuticals and biotechnology’, Industrial and Corporate Change, 3, (2),435–50.

Henderson, R. (1994), ‘The evolution of integrative competence: innovation incardiovascular drug discovery’ Industrial and Corporate Change, 3 (3), 607–30.

Henderson, R. and Cockburn I. (1996), ‘Scale, scope and spillovers: the determin-ants of research productivity in drug discovery’, Rand Journal of Economics,27 (1), 32–59.

Henderson, R., Orsenigo, L. and Pisano, G.P. (1999), ‘The pharmaceutical industryand the revolution in molecular biology: exploring the interactions between scien-tific, institutional and organizational change’, in Mowery, D.C. and Nelson, R.R.(eds), The Sources of Industrial Leadership, Cambridge, Cambridge UniversityPress.

Klepper, S. (2002), ‘Firm survival and the evolution of oligopoly’, Rand Journal ofEconomics, 33 (1), 37–61.

Klevorick, A., Levin, R., Nelson, R. and Winter, S., (1987), ‘Appropriating thereturns from industrial research and development’, Brookings Papers onEconomic Activity, 3, 783–820.

Malerba, F. and Orsenigo, L. (2001), ‘Towards a history-friendly model of innov-ation, market structure and regulation in the dynamics of the pharmaceuticalindustry: the age of random screening’, CESPRI Working Paper.

Malerba, F. and Orsenigo, L. (2002), ‘Innovation and market structure in thedynamics of the pharmaceutical industry and biotechnology: towards a history-friendly model’, Industrial and Corporate Change, 11 (4), 667–704.

Malerba, F., Nelson, R., Orsenigo, L. and Winter, S. (1999), ‘History-friendlymodels of industry evolution: the computer industry’, Industrial and CorporateChange, 8 (1), 3–40.

Malerba, F., Nelson, R., Orsenigo, L. and Winter, S. (2001), ‘Competition andindustrial policies in a “history-friendly” model of the evolution of the computerindustry’, International Journal of Industrial Organization, 19, 635–64.

Matraves, C., (1999), ‘Market structure, R&D, and advertising in the pharmaceuti-cal industry’, Journal of Industrial Economics, 48 (2), 169–94.

Maxwell, R.A. and Eckhardt, S.B. (1990), Drug Discovery: A Case Book andAnalysis, Clifton, NJ, Humana Press.

McKelvey, M. (1995), Evolutionary Innovation: The Business of Biotechnology,New York, Oxford University Press.

Merges, R. and Nelson, R.R. (1994), ‘On limiting or encouraging rivalry in techni-cal progress: the effect of patent scope decisions’, Journal of Economic Behaviorand Organization, 25, 1–24.

Nelson, R.R. (1959), ‘The simple economics of basic scientific research’, Journal ofPolitical Economy, 67 (2), 297–306.


Nelson, R. and Winter, S. (1982), An Evolutionary Theory of Economic Change,Cambridge MA, Belknapp Press of Harvard University Press.

Orsenigo, L. (1989), The Emergence of Biotechnology, London, Pinter.Orsenigo, L., Pammolli, F. and Riccaboni, M. (2001), ‘Technological change and

network dynamics’, Research Policy. 30 (5), 485–508.Peltzman, S. (1974), Regulation of Pharmaceutical Innovation: The 1962

Amendments, Washington, DC, Washington American Enterprise Institute forPublic Policy.

Pharmaceutical Manufacturers Association (1991), 1989–1991 Statistical Report,Washington, DC: Pharmaceutical Manufacturers Association.

Pisano, G. (1991), ‘The governance of innovation: vertical integration and collabo-rative arrangements in the biotechnology industry’, Research Policy, 20, 237–49.

Pisano, G. (1996), The Development Factory: Unlocking the Potential of ProcessInnovation, Cambridge, MA, Harvard Business School Press.

Scherer, F.M. (2000), ‘The pharmaceutical industry’, in Culyer, A.J. andNewhouse, J.P. (eds), Handbook of Health Economics, vol. 1, Amsterdam,Elsevier, pp. 1297–336.

Schwartzman, D. (1976), Innovation in the Pharmaceutical Industry, Baltimore,MD, Johns Hopkins University Press.

Sutton, J. (1998), Technology and Market Structure: Theory and History,Cambridge, MA, MIT Press.

Thomas, L.G. (1994), ‘Implicit industrial policy: the triumph of Britain and thefailure of France in global pharmaceuticals’, Industrial and Corporate Change,3 (2), 451–89.


5. Path dependence and diversificationin corporate technological historiesJohn Cantwell1

1 INTRODUCTION

David (1985; 1991; 1993) has applied the notion of path dependence in thecontext of a continuing impact of individual micro decisions and actionsundertaken at certain key historical junctures that represent ‘windows’ ofstructural change, on the subsequent emergence and evolution of a widernetwork of actors. Path dependence is hence defined as non-ergodic, in thesense that the particular historical events that occur at these stages of struc-tural transformation, and the precise sequencing of those events, has alasting effect on the asymptotic probability distribution of outcomes (seealso Arthur, 1994). This formulation of path dependency confers greaterprecision on the phrase that ‘history matters’. This chapter applies David’sconcept to the large firm viewed as a network of actors and activities, andin particular as a network of various types of technological effort thatpromote innovation and learning across the different parts of the firm.

The chapter also follows David’s (1994) suggestion that human organ-isations and institutions act as the ‘carriers of history’, owing to the factthat within the context of the form and functioning of large firms, pathdependence plays an especially crucial role. A similar argument is advancedby Cantwell and Fai (1999), supported by empirical historical evidence thatthe composition of the products or markets of the largest firms tends tohave shifted more markedly over time than have their profiles of corporatetechnological specialisation. Thereby large firms represent repositories ofcompetence or expertise (Winter, 1988), and so provide some historicalcontinuity to the composition of productive and technological capabilitiesof a society, amidst the sometimes more radical changes that are observedthrough innovation at the product level. Thus, the argument can beextended to the path dependency observed among national groups of largefirms when the profile of their capabilities is considered collectively, or inthe enduring character of specific alternative national systems of innov-ation (Cantwell, 2000).

118

However, as David (2001) has contended, there is a productive tensionwithin evolution (whether of a biological or a social and cultural kind)between the path dependency that is associated with the constraints ondevelopment imposed by the structure carried forward from past events (orthe current contents of the gene pool), and the teleological principle ofnatural selection according to inclusive fitness, which may lead to depart-ures from the existing structure in accordance with the newer requirementsof the current environment. This argument suggests that there may be bothstrong and weak versions of path dependency, in terms of differences in thelongevity of the effects that are engendered by past events. In the strongkind of path dependency positive feedback effects from some initial eventsdominate, providing a ‘lock-in’ phenomenon as in the QWERTY illustra-tion (David, 1985), while in the weaker kind negative feedbacks (or newstructural shocks) from the selection environment play a greater role inmoving a system away from its starting point rather more easily or quickly.The implication is that there may be a variety of individual paths, andindeed with stochastic processes the strength of the observed path depend-ency is likely to include a chance element. In this chapter corporate tech-nological trajectories are treated as being generally path-dependent, butwith a continual drift.

Cantwell and Fai (1999) had observed path dependency in the profiles ofcorporate technological specialisation of 30 large firms over the 60-yearperiod from 1930 to 1990, in that these profiles tended to persist over time,even if they were subject to gradual or incremental change. However, thatstudy also found some evolution and diversification in the patterns of cor-porate technological capabilities over time. This chapter looks in greaterdepth at the specific historical paths followed by four of these large firms intheir corporate technological trajectories, and over the longer period1890–1995. In doing so more can be said about the precise nature of theevolution that occurred, and about the character of changes in corporatetechnological diversification. What is more, the question can be addressedof whether, if profiles of corporate technological capabilities tend to persistover periods of 60 years (such that the specificities of a firm’s primary tech-nological origins can still be identified as being present 60 years later), doesthe same hold for periods as long as 100 years or more?

The four firms chosen were two of the world leaders in each of the chem-ical and electrical equipment industries respectively. The origins of most ofthe present leaders in the chemical industry can be traced back to the endof the nineteenth century and even further back in the case of the Germancompanies (Beer, 1959). The leading German firms, which were the inter-national pioneers of in-house corporate research and development (R&D),enjoyed great early success, none more so than Bayer (Haber, 1971). The

Path dependence and diversification 119

most prominent companies, including Bayer, merged their operations intoIG Farben between 1925 and 1945, before reconstituting themselves indi-vidually again in the post-Second World War era (as Hoechst, BASF andBayer). Their continuing presence in an international leader capacity over100 years later constitutes a remarkable demonstration of the perseverenceof technological prowess. Only sightly less remarkably, Du Pont had beenthe first USA company to open in-house research facilities in the chemicalindustry, and having risen to a leading world position in the inter-warperiod, also has a record of success that traces back over 100 years.Likewise, in the electrical equipment industry, the largest firms in the USAand Europe trace their origins back to the nineteenth century. In terms ofthe longevity and the significance of their technological contributions,perhaps the two best known of these are General Electric and AmericanTelephone and Telegraph (AT&T), whose research histories have been welldocumented (Reich, 1985), but both of which continue to be world leadersto the present day.

The focus of attention here is on the long-term paths of technologicaldevelopment of these companies, based on the proposition that historymatters, in the sense that the technological characteristics of such largecompanies (during the period under consideration) were heavily influ-enced – and constrained – by the type of technological activities that theyor their predecessors had carried out in the past. This notion of organisa-tional continuity can be supported with reference to David’s (1994) expla-nation of the role of historical experience in forming mutually consistentexpectations that facilitate co-ordination without the need to rely perpetu-ally on centralised direction, and the role of the interrelatedness that tendsto develop among the constituent elements of complex human organisa-tions, as well as by the earlier concept of the central place of organisationalroutines as representing embedded experience in the course of evolution-ary social learning (Nelson and Winter, 1982). In order to compare the evo-lutionary paths in the sectoral composition of the innovative activity offirms over time we require a quantitative measure of their technologicalactivities. This chapter uses patents granted in the USA to Du Pont, IGFarben (and later Bayer), General Electric and AT&T as a measure of theextent and the spread of the technological achievements of these compa-nies. We contend that patents may be used with relatively good confidenceas a proxy measure of the rate and direction of the technological change ofthese companies, active as they all are in science-based industries.

The chapter is divided into five sections. The next section introduces thedata to be used in the analysis, discusses the suitability of using patent sta-tistics as a measure of corporate technological activities and briefly reviewsthe methodology adopted. Section 3 looks at the evolution of technological


capabilities at Du Pont and IG Farben (and later Bayer) through the use ofcorporate patent data, and section 4 conducts an equivalent analysis ofGeneral Electric and AT&T. In the final section some summary cross-firmcomparative measures of corporate technological diversification are pre-sented and assessed, and conclusions are drawn with respect to the distinc-tive technological paths followed by each company.

2 THE DATA AND METHODOLOGY

The Companies Selected

The research presented in this chapter, based on a study of four firms, ispart of a wider project on long-term patterns of technological change (overa period of more than a century) of the largest US and European indus-trial companies. For the purposes of the wider project, two types of infor-mation have been collected manually from the US Index of Patents and theUS Patent Office Gazette. First, all patents were recorded that wereassigned to a selection of large US-owned and European-owned firmsbetween 1890 and 1968. From 1969 onwards equivalent information hasbeen computerised by the US Patent and Trademark Office (USPTO). Thefirms selected for the historical patent search were identified in one of threeways. The first group consisted of those firms which have accounted for thehighest levels of US patenting after 1969; the second group comprisedother US, German or British firms which were historically among thelargest 200 industrial corporations in each of these countries (derived fromlists in Chandler, 1990); and the third group was made up of other compa-nies which featured prominently in the US patent records of earlier years(a method that proved most significant for a number of French firms thathad not been identified from other sources).

In each case, patents were counted as belonging to a common corporategroup where they were assigned to affiliates of a parent company. Affiliatenames were normally taken from individual company histories. In all, theUS patenting of 857 companies or affiliates was traced historically;together these comprise 284 corporate groups. Owing to historical changesin ownership, 17 of the affiliates were allocated to more than one corporategroup over the period as a whole. Where patents have been assigned tofirms, the inventor is normally an employee of the company or is directlyassociated with it in some other way, but occasionally independent indi-vidual inventors do choose to assign their patents to firms (Schmookler,1966). Assignments by independent individuals were more common in thenineteenth century but, at least from the inter-war years onwards, the


typical assignor was a prominent member of a corporate research labora-tory, or some other similar in-house company facility. Although it is nor-mally difficult to trace these named individuals in secondary sources on thefirms concerned (as they are not usually also senior managers), the locationof assignors can be checked against business history sources on the inter-national location of activity in particular firms. Such checks on a selectionof large firms have confirmed that whenever a location has been responsiblefor significant numbers of patents being assigned to a company, that firmdid indeed have some in-house facility in the location in question at the rel-evant time. Companies checked in this fashion include various US firmsactive abroad and European companies in the USA (Stocking and Watkins,1946; Beaton, 1957; Wilkins, 1974; 1989; Chandler, 1990) including IGFarben and its predecessors (Plumpe, 1990), Du Pont and ICI (Hounshelland Smith, 1988), Courtaulds and British Celanese (Coleman, 1969), andAT&T, General Electric and the British GEC (Jones and Marriot, 1971;Reich, 1985).

The six firms granted the largest volume of US patents historically were,in descending order, General Electric, AT&T, Westinghouse Electric, IGFarben, RCA and Du Pont (for the years 1890–1947, the details for whichare given in Tables 5.3 and 5.6 below). So as to be able to compare the long-run trends in technological specialisation of the two leading firms in eachof two broadly defined industries – electrical equipment and chemicals – inthis chapter GE, AT&T, IG and Du Pont are the companies selected forcloser study. For the purposes of data continuity in the case of the earlierhistorical years, the founders of IG – namely, Bayer, BASF, Hoechst andAgfa – are treated together collectively prior to the formation of IG Farbenin 1925. With the break-up of IG Farben after 1945, for the purposes of thepost-war period attention is directed instead to the leading member of whathad been the IG group, namely, Bayer.

To construct a measure of technological specialisation of firms the firststep is to devise a classification of fields of technological activity, which isderived from the USPTO patent class system. Fortunately, as these classeschange, the USPTO reclassifies all earlier patents accordingly, so the classi-fication is historically consistent. This study uses the classification schemethat was in operation at the end of 1995, which is then applied backwardsin time. Every patent was classified by the USPTO under at least one suchclass and sub-class. Although patents can be assigned to more than onefield, the primary classification was used in all cases. Various broad cat-egories of technological activity were derived by allocating classes or sub-classes to common groups of activity. Patents granted to the companiesincluded in the study were classified in this manner to a total of 23 techno-logical sectors for each industrial group, representing the principal areas of


development in each of these industries respectively. For the wider project,patents have been allocated to one of the 56 fields of technological activityset out in Table 5.1. However, not all these fields are important for the firmsof a given industry, so for this study some of the less significant fields aregrouped together in each industry, such that the sectoral composition across23 areas is specifically designed to suit the analysis of patterns of special-isation in the chemical and electrical equipment industries separately. Theparticular disaggregation chosen for each industry is shown in Table 5.2.

Patent Statistics as a Measure of Technological Activities

Patent statistics present a potentially very rich source of empirical evidenceon questions related to technological change (see Scherer et al., 1959;Schmookler, 1966; Pavitt, 1985; 1988; Griliches, 1990). The learningprocess which generates accumulated capability in companies relies oninputs of new knowledge and inventions, and so long as the pattern ofknowledge requirements thus reflects the underlying distribution of tech-nological competence across firms, corporate patents may be used as aproxy for the underlying pattern of technological change, and not merelyas a direct measure of inventions. It is argued that US patent data providethe most useful basis for international comparisons, given the commonscreening procedures imposed by the US Patent Office (Pavitt and Soete,1980; Soete, 1987; Pavitt, 1988). Additionally, as the USA is the world’slargest single market, it is likely that firms (especially large ones), will regis-ter for a patent there after patenting in their home countries. It is also rea-sonable to assume that such foreign patents registered in the USA are likelyto be on average of higher quality or significance. United States patentsreveal to which firm each patent was granted, and with which type of tech-nological activity the patent is associated.

Looking within the innovating firms themselves, the hypothesis here ofpath dependence and persistence in the profiles of corporate technologicalspecialisation comes not so much from the characteristics of the knowledgegeneration process (R&D) itself, but from the structure of downstreamlearning and problem-solving in and around production, which calls for thecreation of specialised knowledge inputs in specific fields. Thus, our use ofpatent statistics regards them as a measure of inputs (into innovation, thecreation of new commerical products and processes) and not outputs (fromR&D); that is, codified knowledge inputs into the processes of problem-solving and learning in production, through which technological compe-tence is created. Of course, this does imply that there may still be potentialproblems with an input-based classification scheme derived from the patentclass system, given the way in which technologies from different disciplinary



Table 5.1 The classification of 56 fields of technological activity

1 Food and tobacco products2 Distillation processes3 Inorganic chemicals4 Agricultural chemicals5 Chemical processes6 Photographic chemistry7 Cleaning agents and other compositions8 Disinfecting and preserving9 Synthetic resins and fibres

10 Bleaching and dyeing11 Other organic compounds12 Pharmaceuticals and biotechnology13 Metallurgical processes14 Miscellaneous metal products15 Food, drink and tobacco equipment16 Chemical and allied equipment17 Metalworking equipment18 Paper-making apparatus19 Building material processing equipment20 Assembly and material handling equipment21 Agricultural equipment22 Other construction and excavating equipment23 Mining equipment24 Electrical lamp manufacturing25 Textile and clothing machinery26 Printing and publishing machinery27 Woodworking tools and machinery28 Other specialised machinery29 Other general industrial equipment30 Mechanical calculators and typewriters31 Power plants32 Nuclear reactors33 Telecommunications34 Other electrical communication systems35 Special radio systems36 Image and sound equipment37 Illumination devices38 Electrical devices and systems39 Other general electrical equipment40 Semiconductors41 Office equipment and data processing systems42 Internal combustion engines43 Motor vehicles44 Aircraft


Table 5.1 (continued)

45 Ships and marine propulsion46 Railways and railway equipment47 Other transport equipment48 Textiles, clothing and leather49 Rubber and plastic products50 Non-metallic mineral products51 Coal and petroleum products52 Photographic equipment53 Other instruments and controls54 Wood products55 Explosive compositions and charges56 Other manufacturing and non-industrial

Table 5.2 The relationship of the 23 fields of technological activity usedfor each industry, to the original 56 sector classification

Field Chemical industry Electrical equipment industry56 sector codes included 56 sector codes included

1 2 2–12, 552 3 133 4 144 5 245 6 15–23, 25–306 7 337 8 348 9 359 10 36

10 11 3711 12 3812 13, 14 3913 16 4014 15, 17–30 4115 33–41 42, 4316 42–47 44–4717 48 4818 49 4919 50, 54 50, 5320 51 5121 52, 53 5222 55 5323 1, 31, 32, 56 1, 31, 32, 56

foundations may be integrated, and given some arbitrariness in the divisionbetween certain patent classes. We have tried to alleviate this difficulty bydevising a classification scheme that groups together patent classes that arethe most technologically related, as described above.

So while Schmookler (1966) used patents as a direct measure of inven-tion as such and others (Scherer, 1983; Bound et al., 1984) have since usedthem as an indirect (output) measure of R&D inputs, the patents grantedto the largest industrial firms are used here instead as an indirect (input)measure of the pattern of technological change in these companies. In thissense patents represent knowledge inputs into the corporate learningprocesses that give rise to changes in production methods, the creation ofwhich knowledge has generally been tailored to the problem-solvingagenda of such learning in production. This is a valid inference so long asthe knowledge requirements of the learning processes by which firms gen-erate accumulated capabilities reflect the profile of those resultant techno-logical competences across types of innovative activity. Just as the locationof inventors that assigned patents to each firm has been checked against theknown location of corporate research facilities as mentioned above, so toothe sectoral distribution of corporate patenting has been checked againstthe more qualitative or archival evidence of business history sources on theequivalent firms. Again, we have found an approximate matching betweenthe quantitative patterns of patenting and the qualitative accounts of theprimary fields of R&D and productive expertise of the same firms(as described in Reich, 1985; Hounshell and Smith, 1988; Plumpe, 1995).In one respect what is done here is to provide a greater formalisation ofpropositions on the evolution of the composition of technological special-isation and the degree of technological diversification that can already befound descriptively in business history stories.

Without fully reviewing the literature on the use of patent statistics, itmay be worth mentioning two of the problems that have been raised. First,the fact that companies do not patent all their inventions and, therefore,any comparison may be biased in favour of those firms which rely more onpatenting relative to secrecy; and, second, the fact that those inventionswhich do get patented differ in their economic and commercial significance.With regard to the first problem, there is evidence which suggests thatdifferences in the propensity to patent are more significant when compar-ing firms from different industries (Scherer, 1983). We may assume thatcompanies in the chemical industry (or alternatively in the electrical equip-ment industry) have a similar attitude towards patenting, allowing for acomparison of their technological capabilities based on their patentingactivity. However, we have to consider that even within a company thepropensity to patent varies between technological fields. Therefore, for the


comparative analysis here an indicator of relative specialisation is usedrather than absolute numbers of patents, as explained below.

The second question regarding the difference in quality of patentedinventions may not necessarily be a significant obstacle when assessing thebreadth of a company’s innovative capabilities. Since patents are not onlyissued on the most significant inventions but also on other related discov-eries, they are useful in the analysis of general trends in the sectoral com-position of technological activities (which may cover a wider spectrum thanimportant invention alone). For instance, a major shift in the level of acompany’s patenting in a particular sector, relative to the aggregate level,is likely to indicate a shift in the focus of its technological efforts.Consequently we may conclude that despite differences in the ‘quality’ ofindividual patents, comparisons between rival companies across the entiredistribution of their respective patenting can shed light on the compositionor spread of the technological expertise of firms. Other evidence confirmsthe suitability of patent data as a measure of corporate technological effort,particularly in inter-firm comparisons within an industry (Griliches, 1990).They are available, they go back over 100 years and allow for a technolog-ical classification at a greater level of detail than any other measure of tech-nological activity (such as R&D statistics).

The business histories of the firms studied here can also be cited to showhow patenting mattered to them as part of their strategy, and with respectto the construction of technology exchange arrangements with other largefirms in their respective industries (see Cantwell and Barrera, 1998). Reich(1977) discusses the importance of patents to the struggle to control radiothat included both GE and AT&T. Plumpe (1990) shows that the level of IGFarben’s patent applications in Germany was consistent with its researchand development (R&D) activities, which gives us an indication of thecompany’s reliance on patents. In addition, IG Farben’s negotiations withStandard Oil, which led to the establishment of a joint venture in the USA,together with the activities of the US subsidiary of IG, General Aniline andFilm, confirm how these patents had to be extended to the US market.However, since we will be using data on patents granted to these companiesin the USA, it should be allowed that Du Pont, being a US company, wasmore prone to patent in its home country than would have been IG Farben.This is another reason for using relative rather than absolute numbers.

The Indicators of Corporate Technological Specialisation Derived fromPatenting

Arguments such as those just considered, and other issues that have beenwell documented, demonstrate the need for caution in the use of patent


statistics (Basberg, 1987; Pavitt, 1988; Griliches, 1990; Archibugi, 1992;Patel and Pavitt, 1997; 1998). However, a number of the difficulties in theuse of patent data have been avoided in our approach through relevant dis-aggregation and the construction of an appropriate index. Inter-industrydifferences in patenting propensity are reduced as the chapter deals withintra-industry comparisons only, although admittedly the industrialgroups are defined very broadly (but so, too, is the span of activity ofthe very large companies under consideration). It is recognised that inter-sectoral (across technological fields) or inter-firm differences in the propen-sity to patent may arise, but these are controlled for here by the use of theRevealed Technological Advantage (RTA) index (Cantwell, 1989; 1993;Cantwell and Andersen, 1996; Patel and Pavitt, 1997; 1998). The RTA isan indicator of a firm’s technological specialisation across a spectrum oftechnological activity relative to that of other firms in the same industry.The RTA of a firm in a particular field of technological activity is given bythe firm’s share in that field of US patents granted to all companies in thesame industrial group, relative to the firm’s overall share of all US patentsassigned to all firms in the industry in question. If Pij denotes the numberof US patents granted in a particular industry to firm i in technologicalactivity j, the RTA index is defined as:

The index varies around unity, such that a value in excess of one showsthat the firm is specialised in that field of activity in relation to other firmsin its industrial group. In this manner inter-sectoral differences in thepropensity to patent are normalised in the numerator of the RTA index,and inter-firm differences are normalised in the denominator. There stillremains the possibility of intra-firm and intra-sectoral differences in thepropensity to patent, but it is likely that the respective variances of thesetwo factors are systematically lower than the inter-firm and inter-sectoraldifferences.

The degree of technological diversification of the firm is measured by theinverse of the coefficient of variation of the RTA index, CVi, across all therelevant fields for the firm. Therefore, for firm i in each period considered,the proxy DIVi for technological diversification will be the reciprocal of theCVi. In particular:2

where is the standard deviation and is the mean value of theRTA distribution for the firm i.

�RTAi�RTAi

DIVi 1�CVi �RTAi��RTAi

RTAij (Pij ��iPij �(�jPij ��ijPij))


3 TECHNOLOGICAL DEVELOPMENT ATIG FARBEN AND DU PONT

A central proposition here is that, to a great extent, the research traditionsof the companies which joined in the formation of IG Farben in 1925regulated the subsequent technological development of the new company,and are likely to be reflected as well in the path of its descendants. Themajor German chemical companies such as Bayer, BASF and Hoeschsthad their origins in the dyestuffs industry, one of the most dynamic chem-ical sectors at the turn of the century. These companies managed to bridgethe gap between industrial and academic chemistry by incorporating andorganising research within the dye workplaces. Consequently, they hadaccomplished great innovative and technical achievements prior to theiramalgamation into IG Farben. Certainly in the first part of the twentiethcentury, the technological supremacy of these companies was still unchal-lenged (Beer, 1959). These German companies had traditionally dom-inated the world dyestuffs market, including in the USA. Furthermorethey had established barriers to entry which proved very difficult to cir-cumvent. These seemed to be the case when, with the outbreak of theFirst World War, the US market was cut off from German imports.Lacking a research tradition in this area, American companies found italmost impossible to replicate the broad range of dyes which the Germanshad introduced.

It seems that after the First World War, dyestuffs had nevertheless lostrelative importance, and the German chemical companies expanded intoother areas of activity, such as nitrogenous fertilisers, plastics, photo-graphic products and synthetic materials (Haber, 1971). However, the 1926figures for IG Farben’s research staff show that as much as 76 per cent ofthe labour employed at the company’s laboratories was still concentrated indyestuffs and dyeing processes research; with about 9 per cent of theresearch staff engaged on the newer fields of pharmaceuticals, syntheticfibres and photographic chemicals (Plumpe, 1990). Other technologiesgaining ground on dyestuffs included cleaning agents and other composi-tions (coatings and plastics in particular), and coal and petroleum research(the oil from coal hydrogenation process).3 Yet the fastest-growing tech-nologies from the 1930s were those grouped under synthetic resins andfibres. By the 1930s the chemical industry seems to have built upon organicchemistry to move into synthetic materials, photographic chemistry andpetrochemicals.

By contrast, Du Pont’s origins had been in the explosives business. Itsfirst research laboratory (which opened in 1902) had been established todeal with the problems inherent in the manufacturing process. However,


a threat to the company’s monopolistic position in the smokeless powdermarket seems to have prompted the need for diversification. This move wasinfluenced by the company’s excess capacity following the expansion ofproduction during the First World War, and for which alternative useswould have to be found in peacetime. With cash available after the war, andbased on its experience with cellulose technology (used in the manufactur-ing of explosives), the company expanded into related areas mainlythrough the acquisition and further refinement of firms and technologies(Hounshell and Smith, 1988). Following this strategy, by the time of IGFarben’s merger, Du Pont had broadened its areas of activity. In additionto explosives, the company had diversified into paints (Duco being a majorinnovation), silk, leather and cellulose. As with explosives, these productswere based on nitrocellulose technology. The company had bought its wayinto the production of cellophane and developed moisture-proof cello-phane, a successful new product. In addition, Du Pont had finally accom-plished the production of dyestuffs following a major struggle tocircumvent the barriers established by the German companies (whichinvolved the government’s establishment of a higher import tariff in theyears following the war). Research and development expenditures in 1926give us an idea about the extent of this diversification – roughly 15 per centof research spending was into explosives, 25 per cent in general chemicals,27 per cent in paints and related chemicals, 19 per cent in dyestuffs and6 per cent in plastics (Hounshell and Smith, 1988).

Comparing the evolution of the German companies with that ofDu Pont prior to 1926, it appears that their technological strategies hadbeen different (Dornseifer, 1989; 1995). Whereas the German firms hadrelied more on internal generation of innovation (through their strong in-house research), Du Pont had relied more on external sources, that is,acquisitions. Lacking a research tradition in areas outside its core technol-ogy, Du Pont took advantage of the strengths of its organisation to incor-porate those acquired technologies and develop them further (Chandler,1990). By contrast, IG Farben’s predecessors had continued to rely on theinternal dynamism of the research organisations that they had built.

Table 5.3 presents evidence on US patents granted to the largest chem-ical firms. In absolute terms (see last row) US patenting activity in thechemical sector experienced sustained growth throughout 1890 to 1947,with the partial exception of the period 1920–24, when there was a fallattributable to the effects of the First World War. The position of theGerman dyestuffs companies (here amalgamated under IG Farben) is evenmore remarkable than the high shares tell us, given that, as mentionedabove, US companies had a higher propensity to patent their innovationsin the USA. Indeed, Du Pont lagged behind in second place until the late


131

Tab

le 5

.3P

erce

ntag

e sh

are

ofU

S p

aten

ting

of

larg

e ch

emic

al a

nd p

etro

chem

ical

com

pani

es

1890

–191

919

20–2

419

25–2

919

30–3

419

35–3

919

40–4

7A

vera

ge19

91–9

5A

vera

ge18

90–1

947

1890

–199

5

Che

mic

al fi

rms

IG F

arbe

n56

.22

11.3

728

.86

30.7

518

.74

5.58

19.8

1n/

a2.

64H

oech

stn/

an/

an/

an/

an/

an/

an/

a8.

474.

01B

ayer

n/a

n/a

n/a

n/a

n/a

n/a

n/a

8.39

5.57

BA

SFn/

an/

an/

an/

an/

an/

an/

a6.

022.

79D

u Po

nt7.

1715

.19

8.99

11.6

119

.13

17.4

615

.27

7.95

9.17

Dow

Che

mic

al1.

265.

032.

375.

123.

474.

944.

097.

106.

04U

nion

Car

bide

7.53

15.6

87.

583.

664.

734.

165.

051.

413.

59A

llied

Che

mic

al4.

7613

.81

6.69

6.87

4.34

2.35

4.44

0.08

1.50

ICI

1.82

1.54

1.86

4.36

4.07

2.61

3.13

2.66

2.87

Bri

tish

Cel

anes

e0.

402.

113.

086.

378.

103.

894.

93n/

a0.

87C

elan

ese

Cor

pn/

an/

an/

an/

an/

an/

an/

a0.

340.

86Sw

iss

IG5.

805.

045.

203.

383.

332.

893.

56n/

an/

aC

iba

Gei

gyn/

an/

an/

an/

an/

an/

an/

a5.

064.

25Sa

ndoz

n/a

n/a

n/a

n/a

n/a

n/a

n/a

0.72

0.88

Tota

l che

mic

als

%96

.75

82.0

577

.60

81.8

478

.12

66.7

876

.26

80.7

474

.19

132

Tab

le 5

.3(c

onti

nued

)

1890

–191

919

20–2

419

25–2

919

30–3

419

35–3

919

40–4

7A

vera

ge19

91–9

5A

vera

ge18

90–1

947

1890

–199

5

Oil

firm

s

Stan

dard

Oil

NJ

0.77

4.14

5.91

5.68

4.90

9.45

6.45

n/a

n/a

Exx

onn/

an/

an/

an/

an/

an/

an/

a3.

325.

20Sh

ell

0.13

0.24

0.63

1.47

4.00

5.38

3.42

3.62

3.52

Tota

l oil

%3.

2517

.95

22.4

018

.16

21.8

833

.22

23.7

419

.26

25.8

2

Tota

l num

ber

451

81

231

269

28

079

1139

817

906

4582

437

033

343

703

Not

e:P

rior

to

1926

IG

Far

ben’

s an

d IC

I’s fi

gure

s co

rres

pond

to

the

aggr

egat

e of

thei

r pr

edec

esso

r co

mpa

nies

;a s

imila

r pr

oced

ure

appl

ies

toSw

iss

IG.n

/a

not

appl

icab

le.

Sou

rce:

US

Inde

x of

Pat

ents

,Pat

ent

Gaz

ette

,and

com

pute

rise

d da

ta f

rom

the

US

Pat

ent

and

Tra

dem

ark

Offi

ce.

1930s. This overwhelmingly commanding position of IG Farben (in whatfollows this name also applies to its predecessors prior to 1925) was par-ticularly apparent in the period prior to the First World War. However, thefigures for the 1920–24 and 1940–47 periods may be somewhat underesti-mated, since after the two wars the Alien Property Custodian confiscatedthousands of German patents, which had not yet been assigned to the rele-vant companies. Nevertheless, note that, although IG Farben was dis-solved in 1945, the relevant period allows for the inclusion of twoadditional years (to 1947) during which the confiscated American sub-sidiaries of IG Farben were still patenting.

The Evolution of Technological Capabilities in IG Farben and Bayer

Table 5.4 presents the cross-field RTA index over time for the IG group andlatterly for Bayer. As explained earlier, the RTA index shows the firm’sshare of patenting in a given technological field (among patents granted toall large firms in the chemical industry), relative to its equivalent share oftotal chemical industry patenting in all fields for each period considered. Itcan be seen that, in spite of the fact that IG Farben’s research remainedheavily concentrated in organic chemistry (as that of its predecessor com-panies had been) as shown by an RTA value consistently above 1, changesbegan to occur to its profile of technological specialisation in other partsof the cross-sectoral distribution. There was even some decline in relatedbleaching and dyeing processes after the merger that constituted the IGgroup in 1925.

Looking back to the turn of the twentieth century, it can be seen thatTable 5.4 provides evidence that is entirely consistent with what is alreadyknown about the early development of the chemical industry, and of thelarge German firms in the chemical industry in particular. That their his-tories were essentially coincident at that stage reflects the dominant pos-ition in the industry of the German leaders. The industry began with thedevelopment of artificial dyestuffs in the 1870s, from which it moved intophotographic chemicals and synthetic fibres in the early years of the twen-tieth century. The predecessors of IG Farben led the way, as is shown bythe RTA values above 1 in 1890–1919 in Table 5.4. In the 1920s and 1930sthe IG group diversified further into fertilisers (listed here under agricul-tural chemicals), and pharmaceuticals. Following the formation of IG in1925 synthetic resins and fibres steadily gained in importance in the group’sprofile of technological effort, the RTA rising from 0.45 in 1925–29 to 1.14in 1940–47. The latter may partly reflect the company’s research into syn-thetic rubber which led to the development of PVC (Freeman, 1982). Inaddition, intensification of research in photochemicals (from 1.92 to 2.65


134

Tab

le 5

.4E

volu

tion

of

patt

erns

of

tech

nolo

gica

l spe

cial

isat

ion

at I

G F

arbe

n,18

90 t

o 19

95

Tec

hnol

ogic

al s

ecto

rsIG

Far

ben

Bay

er

1890

–19

20–

1925

–19

30–

1935

–19

40–

Cum

ul.

1991

–C

umul

.19

1924

2934

3947

1890

–95

1947

–19

4795

Dis

tilla

tion

pro

cess

es0.

340.

801.

010.

360.

600.

580.

421.

330.

74In

orga

nic

chem

ical

s0.

580.

840.

961.

160.

730.

620.

970.

810.

48A

gric

ultu

ral c

hem

ical

s0.

472.

400.

901.

300.

990.

220.

921.

961.

53C

hem

ical

pro

cess

es0.

500.

420.

880.

630.

680.

610.

590.

560.

61P

hoto

grap

hic

chem

istr

y1.

457.

211.

922.

403.

002.

651.

952.

483.

26C

lean

ing

agen

ts a

nd o

ther

0.64

0.14

0.59

0.81

0.63

0.52

0.56

0.57

0.51

com

posi

tion

sD

isin

fect

ing

and

pres

ervi

ng0.

000.

000.

000.

000.

001.

710.

170.

330.

41Sy

nthe

tic

resi

ns a

nd fi

bres

1.58

0.34

0.45

0.96

0.90

1.14

0.70

1.41

1.41

Ble

achi

ng a

nd d

yein

g pr

oc.

1.53

4.38

1.44

0.82

0.94

1.27

1.16

1.01

0.88

Oth

er o

rgan

ic c

ompo

unds

1.33

1.96

1.30

1.33

1.35

1.33

1.48

1.23

0.91

(inc

ludi

ng d

yest

uffs)

Pha

rmac

euti

cals

0.92

1.55

1.14

0.93

1.28

0.97

0.96

0.86

1.31

Met

als

0.49

0.72

0.57

0.60

0.46

0.36

0.45

0.40

0.48

Che

mic

al a

nd a

llied

equ

ip.

0.38

0.18

0.64

0.53

0.44

0.39

0.47

1.11

0.90

Mec

hani

cal e

ngin

eeri

ng n

es.

0.13

0.29

0.52

0.36

0.25

0.20

0.24

0.38

0.53

Ele

ctri

cal e

quip

men

t ne

s.0.

070.

000.

070.

330.

610.

930.

280.

370.

38T

rans

port

equ

ipm

ent

0.34

0.00

0.34

0.53

0.78

0.00

0.39

0.37

0.51

135

Tex

tile

s,cl

othi

ng,l

eath

er0.

000.

000.

000.

000.

000.

000.

000.

000.

42R

ubbe

r an

d pl

asti

c pr

oduc

ts0.

110.

420.

290.

490.

440.

310.

290.

760.

78N

on-m

etal

lic m

iner

al a

nd0.

260.

390.

090.

430.

500.

310.

320.

580.

56w

ood

prod

ucts

Coa

l and

pet

role

um p

rodu

cts

1.17

0.00

1.08

0.95

1.58

2.72

1.30

0.38

0.29

Pro

fess

iona

l and

sci

enti

fic0.

060.

481.

561.

622.

621.

881.

460.

982.

39in

stru

men

tsE

xplo

sive

com

posi

tion

s an

d0.

070.

000.

000.

000.

060.

000.

050.

000.

01ch

arge

sO

ther

man

ufac

turi

ng0.

100.

000.

000.

170.

170.

890.

210.

260.

20

Sou

rce:

as fo

r T

able

5.3

.

over the same period) seems to have been associated with developments inrelated product areas, namely an involvement in photographic equipment(here included under the other scientific instruments category, the RTA forwhich was well above 1 throughout 1925–47). In this case technologicaldiversification was linked to product diversification into the photographicsector, as represented by the Agfa branch of the group, this and other partsof the IG group having developed the complementary technological expert-ise in photographic chemistry.

One noteworthy diversification feature of the interwar period is that coaland petroleum research reached an all-time high in the period following themerger (particularly from the late 1930s). This confirms other evidenceabout IG Farben’s renewed interests in coal liquefaction, which had beeninitiated at one of the company’s predecessors (Beer, 1959). While this mayhave originally been commercially inspired in the 1920s and early 1930swhen other chemical firms followed suit in investing in the apparentlypromising field of oil from coal, most other large firms that had experi-mented with this possibility divested from the area during the 1930s as theearly promise was disappointed. This helps to explain how IG’s RTA in thefield rose from 1.08 in 1925–29 to 2.72 in 1940–47. The reason seems fairlyclear – by this stage IG Farben had become obliged to assist in the Nazi wareffort, and as is well known the principal German military weakness wasthat it lacked its own oil supply.

Perhaps the most remarkable part of the IG story comes from updat-ing the account to consider the subsequent profile of technological spe-cialisation of its largest successor company, Bayer. In this context we focuson a comparison of the pattern of technological specialisation of the IGFarben group as a whole over 1890–1947 with that of Bayer over 1947–95,but with an eye on the position of Bayer in the most recent sub-period,1991–95, to see whether there are any substantive shifts in the pattern ofspecialisation that appear to have emerged in the latest years. A rationalefor comparing two very long-term periods of 1890–1947 and 1947–95 isitself the belief that the knowledge base of firms is cumulative and incre-mental (Nelson and Winter, 1982; Rosenberg, 1982; Cantwell, 1989).What comes out of this comparison is indeed a picture of the mostremarkable continuity. The oldest strength in dyestuffs has held up and,indeed, even reasserted itself most recently (an RTA of 1.23 in 1991–95),while the new developments that had emerged early in the twentiethcentury, and had been either reinforced or established as a strength in IGFarben in the inter-war period, remain evident in Bayer today. Thus, con-sider the RTA values in 1947–95 in synthetic resins and fibres (1.41), pho-tographic chemistry (3.26) and instruments (2.39); pharmaceuticals alsostands at 1.31, but this advantage has been eroded recently (0.86 in


1991–95). In contrast, the inter-war specialisation in oil-related chemicalshas disappeared (an RTA in the Bayer years of 0.29), driven as it was moreby external military demands than by the specificities of internal capabil-ities and commercial logic. In contrast, what had been for IG in the 1930sthe newly emergent strength in fertilisers and agricultural chemicals hassuccessfully reappeared in Bayer (an RTA of 1.53 in 1947–95, and 1.96 in1991–95).

The Evolution of Technological Capabilities at Du Pont

Looking at the evolution of Du Pont’s technological specialisation(Table 5.5), through to 1947 the company preserved its founding strengthin the field of explosives. However, given its incredibly high degree of focuson explosives (an RTA of 11.16 in 1890–1919), it was almost inevitable thatgrowth would take the form of to some extent moving away from its origi-nal research activities in and around explosives and allied technologies(chemical and distillation processes, cellulose, silk and leather with textileapplications – with RTAs in 1890–1919 of 1.76, 5.40 and 13.49 respect-ively), initially into rubber and plastics (2.70 in 1890–1919), and then intosynthetic resins and fibres (3.05 in 1925–29). The emergence and subse-quent rise of the development of synthetic materials in the 1920s and 1930sseems to be for Du Pont the most striking phenomenon of the inter-warperiod. This confirms other evidence regarding the company’s research intopolymers which eventually led to the discovery of nylon, the first syntheticfibre and Du Pont’s most successful product.4

By contrast with IG Farben, organic chemistry was never an area ofcomparative strength at Du Pont, and it emerged to respectability in thisfield (an RTA of 0.99 in 1935–39) only after the First World War. Since thiswas not a traditional area of research for the company, its moderate catch-ing up partly reflects the company’s co-operative agreement with the ICI(after 1929), a central objective of which was to enable these partners tobetter match the industry leader, IG Farben (see Cantwell and Barrera,1998). A sustained position in rubber and plastic technologies through theinter-war years reflects in part the rise in Du Pont’s research in polymersleading to the discovery of neoprene (Hounshell and Smith, 1988). Aninteresting feature hidden by aggregation in Table 5.5 is the growth of thedevelopment of textile and clothing machinery (within mechanical engin-eering), which may have spun off from the interest in synthetic fibres, andparticularly nylon.

What is most striking from a comparison with IG Farben is that whileDu Pont began with a much more focused spectrum of technological spe-cialisation early in the twentieth century than that of IG, reflecting its later


138

Tab

le 5

.5E

volu

tion

of

patt

erns

of

tech

nolo

gica

l spe

cial

isat

ion

at D

u P

ont,

1890

to

1995

Tec

hnol

ogic

al s

ecto

rs18

90–

1920

–19

25–

1930

–19

35–

1940

–C

umul

.19

91–

Cum

ul.

1919

2429

3439

4718

90–

9519

47–

4795

Dis

tilla

tion

pro

cess

es5.

401.

803.

241.

080.

870.

821.

082.

801.

48In

orga

nic

chem

ical

s0.

490.

630.

710.

851.

161.

160.

870.

911.

35A

gric

ultu

ral c

hem

ical

s0.

000.

000.

001.

030.

491.

040.

780.

571.

18C

hem

ical

pro

cess

es1.

761.

250.

810.

761.

261.

201.

201.

261.

15P

hoto

grap

hic

chem

istr

y0.

000.

002.

470.

520.

501.

060.

971.

821.

50C

lean

ing

agen

ts a

nd o

ther

1.98

1.51

2.48

1.71

1.52

1.33

1.61

0.90

1.21

com

posi

tion

sD

isin

fect

ing

and

pres

ervi

ng0.

000.

000.

002.

820.

001.

641.

091.

050.

62Sy

nthe

tic

resi

ns a

nd fi

bres

0.00

0.77

3.05

2.29

1.39

1.53

1.81

1.06

1.19

Ble

achi

ng a

nd d

yein

g pr

oc.

0.00

0.39

0.32

0.75

0.73

0.85

0.73

0.49

0.82

Oth

er o

rgan

ic c

ompo

unds

0.18

0.63

0.85

0.82

0.99

0.87

0.80

0.70

0.78

(inc

ludi

ng d

yest

uffs)

Pha

rmac

euti

cals

0.00

0.77

1.32

1.43

0.73

0.72

0.85

0.27

0.29

Met

als

0.96

0.72

0.92

0.81

0.68

0.74

0.78

1.10

1.03

Che

mic

al a

nd a

llied

equ

ip.

3.16

1.98

1.40

0.87

0.99

0.82

1.03

0.86

1.06

Mec

hani

cal e

ngin

eeri

ng n

es.

3.04

1.15

0.38

0.75

0.59

0.61

0.73

1.57

1.23

Ele

ctri

cal e

quip

men

t ne

s.0.

140.

050.

080.

190.

450.

640.

292.

561.

05T

rans

port

equ

ipm

ent

6.75

3.60

2.16

0.00

0.51

0.38

1.00

0.78

1.69

139

Tex

tile

s,cl

othi

ng,l

eath

er13

.49

3.60

0.00

0.00

1.48

0.38

1.15

6.25

1.26

Rub

ber

and

plas

tic

prod

ucts

2.70

0.95

1.23

0.97

0.97

1.47

1.43

2.79

1.70

Non

-met

allic

min

eral

and

1.04

1.75

1.39

2.14

1.39

1.64

1.77

1.88

1.68

woo

d pr

oduc

tsC

oal a

nd p

etro

leum

pro

duct

s0.

540.

950.

430.

690.

970.

750.

781.

741.

56P

rofe

ssio

nal a

nd s

cien

tific

1.35

0.36

0.20

0.24

0.25

0.32

0.33

1.00

0.71

inst

rum

ents

Exp

losi

ve c

ompo

siti

ons

and

11.1

64.

025.

344.

501.

871.

783.

140.

562.

70ch

arge

sO

ther

man

ufac

turi

ng8.

343.

473.

841.

431.

290.

881.

690.

951.

09

Sou

rce:

as fo

r T

able

5.3

.

start and smaller size, it had become notably more diversified than IG bythe 1940s. This again can be related to the strategy of acquisitions ratherthan internal growth. Over 1890–1947 as a whole, Du Pont held a techno-logical specialisation among others in cleaning agents and compositions(an RTA of 1.61), chemical processes (1.20), synthetic resins and fibres(1.81), disinfecting and preserving (1.09), textiles and leather (1.15), non-metallic mineral products (1.77), rubber and plastic products (1.43), distil-lation processes (1.08) and, of course, in explosives technologies (3.14).Within chemical processes, the most important technologies were coatingprocesses, adhesive bonding and chemistry, and electrical and wave energy.Cleaning agents and other compositions included paints and lacquers, with‘Duco’ (the trademark product) being its most important innovation in thisfield (Hounshell and Smith, 1988).

However, despite a path that emphasised technological diversification atDu Pont compared to greater consolidation among the IG group, DuPont’s corporate technological trajectory from 1900 to 1947 was in manyrespects even more coherent in its direction than that followed by IG. Theearly moves away from explosives mainly represented the development ofnitro-cellulose technologies – cellulose (used in the manufacturing processof explosives), silk and leather, chemical processes (including coating andbonding) and paints. From here came the development of cellophane.Then, having worked with rayon and cellulose acetate textile fibres, thecompany was well placed to diversify into (and to lead) the synthetic fibrerevolution (Hounshell and Smith, 1988).

Much of this structure was then preserved in the post-war period. In1947–95 (and in 1991–95) Du Pont retained an RTA greater than one indistillation processes, chemical processes, synthetic resins and fibres, tex-tiles, clothing and leather, rubber and plastic products. However, althoughthis remained true also of explosives in 1947–95 as a whole (an RTA of2.70), it is striking that by 1991–95 this original primary source ofstrength had finally lapsed (the RTA value standing at 0.56). Conversely,to a greater extent than in the post-war experience of Bayer, new strengthshave emerged in Du Pont in 1947–95 in mechanical processes (an RTA of1.23 in 1947–95, and 1.57 in 1991–95), electrical equipment (1.05 in1947–95 and 2.56 in 1991–95) and coal and petroleum products (1.56 in1947–95 and 1.74 in 1991–95). The latter emergent strength in oil-relatedchemicals may be somewhat ironic given Bayer’s post-war retreat fromthat field, but in Du Pont it can be traced back to the development ofpolymer intermediates that derived from the long tradition in explosives(Hounshell, 1995). So perhaps some residue of Du Pont’s path-dependenthistory stemming from its beginnings in explosives remains through to thepresent day after all.


4 TECHNOLOGICAL DEVELOPMENT ATGE AND AT&T

As has been seen already in the case of chemicals, the science-based indus-tries that began towards the end of the nineteenth century were charac-terised by growth through horizontal diversification into technologicallyrelated fields. However, vertical diversification into related mechanicalfields mattered too, as in the case of IG’s move into photographic equip-ment, and Du Pont’s move into textile machinery. Nevertheless, horizontalscience-related diversification was the essential theme of corporate techno-logical trajectories in the chemical industry. In comparison, the develop-ment of vertically integrated systems was relatively more important in theelectrical equipment industry, broadly defined. The electrical equipmentindustry focused from the outset on the design of complex and interrelatedtechnological systems, of which certain components might lie either ahead(salients) or behind (reverse salients) the general front of development atany point in time, whereby progress in other parts of the overall system iseither facilitated or constrained (Hughes, 1983; 1989; Aitken, 1985).

While in the chemical industry the largest German companies were theworld leaders, in the electrical industry that role was taken up by the largestUS firms. The firms examined here were associated with the founding of thetwo central planks of the industry – namely, electrical lighting, power, trac-tion and related machinery (in the person of Edison, whose role is discussedby David, 1991), and the telephone (in the person of Bell). General Electricwas formed in 1892 from the merger of the Edison General ElectricCompany and the Thomson-Houston Electric Company, while AT&T wasoriginally a subsidiary of American Bell set up in 1885, which with finan-cial reorganisation in 1899 became the holding company for the entire Bellgroup (Reich, 1985). The original overlap between these two branches ofthe industry lay in electrical devices and systems, and in some generalmachinery. This overlap illustrates how the leading firms in this industrywere concerned to establish integrated systems, and not the (perhaps inter-connected range of) more narrowly defined products that were typical inthe chemical industry. The connection between the two segments of theindustry became much sharper from the inter-war years onwards with thedevelopment of the radio, and subsequently the television. The radio wasthe primary focus of growth in the electrical equipment industry in theinter-war period, and both parts of the industry made critical contributionsto this new area of development.

Table 5.6 shows the comparative patenting records of the leading firmsin the electrical equipment industry. It shows how AT&T (Bell) was theleading research organisation and the dominant corporate patenter in the


142

Tab

le 5

.6P

erce

ntag

e sh

are

ofU

S p

aten

ting

of

larg

e el

ectr

ical

equ

ipm

ent

com

pani

es

1890

–19

20–

1925

–19

30–

1935

–19

40–

Ave

rage

1991

–A

vera

ge19

1924

2934

3947

1890

–95

1890

–19

4795

Gen

eral

Ele

ctri

c49

.71

21.5

225

.21

28.5

327

.21

26.9

730

.90

23.5

825

.26

AT

&T

20.1

734

.65

34.2

625

.96

20.3

219

.81

24.0

213

.58

16.9

4W

esti

ngho

use

Ele

ct.

23.7

237

.64

29.8

521

.44

15.2

419

.09

22.4

58.

0015

.20

RC

A1.

661.

722.

548.

8619

.37

19.5

010

.78

0.30

9.71

ITT

0.05

0.43

0.78

2.07

1.39

5.95

2.31

2.61

5.11

Siem

ens

2.43

1.52

2.45

4.12

3.62

1.39

2.59

14.4

07.

26A

EG

0.45

0.43

1.57

2.87

6.38

2.36

2.54

0.26

1.27

Tota

l Num

ber

1305

74

883

952

412

003

1200

518

503

7002

519

184

241

069

Sou

rce:

as fo

r T

able

5.3

.

1920s (which made it the highest ranked patenting firm in any industry),but that at other times – before the First World War and once again fromthe 1930s onwards – this role fell to GE. The early 1920s saw a substantialgrowth and diversification of research at AT&T, which gave rise to theestablishment of Bell Telephone Laboratories in 1925, the formation ofwhich gave a further impulse to R&D in the company (Maclaurin, 1949;Nobel, 1979; Reich, 1985). Westinghouse Electric was not far behind thebig two, while RCA was set up by GE in 1919 to take over the assets it hadacquired from American Marconi in the nascent radio industry. By 1921Westinghouse Electric and AT&T had also entered into a partnership withRCA in the radio industry, which therefore, of course, was also a partner-ship with GE in the development of radio systems (Reich, 1977). Duringthe post-war period while GE has remained the industry leader the othershave gradually declined (and in the case of AT&T have been broken up),and large firms from other countries have caught up. The German firm,Siemens, is a prime example, but there are also the newer Japanese compa-nies that have not been considered here.

The Evolution of Technological Capabilities at GE

Given the complex systems nature especially of that part of the electricalbusiness in which GE was active, and given what is known about thebreadth of interest of GE in particular from the beginning across the rangeof electric lighting, power and transport components, it is hardly surpris-ing that the evidence on the spread of GE’s technological specialisation setout in Table 5.7 shows that it was far more diversified in its range of expert-ise in 1890–1919 than were our other large companies. Although this highinitial span of corporate technological diversification is, indeed, partlyattributable to the systems nature especially of this segment of the electricalequipment industry from the start, it may also have something to do withrelatively wide inventive interests of Edison, and his comparative advan-tage as an inventor in the design and construction of electromechanicalsystems (Reich, 1985). This is akin to David’s own (1991) assessment of thesignificance of the individual personality of Edison as an influence on his-torical paths, in this case helping to establish GE as an innovator across abroad front, which has remained as a feature of GE’s capabilities (com-pared to its major competitors) through to the present day.

This having been said, GE embarked on a further technological diversi-fication in the final decade of the 1890–1919 period (Reich, 1985), and thisis encompassed in the aggregation of the years 1890–1919 into a combinedperiod for the purposes of comparison with other companies (for whichchange was concentrated in the inter-war years). The earliest strengths in


144

Tab

le 5

.7E

volu

tion

of

patt

erns

of

tech

nolo

gica

l spe

cial

isat

ion

at G

ener

al E

lect

ric,

1890

to

1995

Tec

hnol

ogic

al s

ecto

rs18

90–

1920

–19

25–

1930

–19

35–

1940

–C

umul

.19

91–

Cum

ul.

1919

2429

3439

4718

90–

9519

47–

1947

95

Che

mic

als

and

phar

mac

euti

cals

1.14

1.09

0.93

1.04

1.19

2.09

1.40

2.48

1.82

Met

allu

rgic

al p

roce

sses

1.21

2.14

1.15

1.12

1.39

1.21

1.21

1.89

1.35

Mis

cella

neou

s m

etal

pro

d.0.

891.

431.

030.

981.

241.

121.

030.

971.

36E

lect

rica

l lam

p m

anuf

.equ

ip.

1.61

1.69

1.27

1.42

1.75

1.54

1.46

0.45

1.06

Oth

er m

ach.

and

ind.

equi

p.1.

411.

250.

890.

931.

411.

151.

221.

601.

26T

elec

omm

unic

atio

ns0.

040.

160.

130.

110.

110.

170.

110.

170.

16O

ther

ele

ctri

cal

0.33

0.46

0.84

1.11

0.80

0.62

0.64

0.54

0.56

com

mun

icat

ion

syst

ems

Spec

ial r

adio

sys

tem

s0.

001.

600.

600.

240.

280.

200.

240.

640.

31Im

age

and

soun

d eq

uipm

ent

0.19

0.07

0.37

0.43

0.11

0.16

0.23

0.28

0.30

Illu

min

atio

n de

vice

s1.

301.

551.

621.

511.

371.

021.

280.

901.

12E

lect

rica

l dev

ices

and

sys

tem

s1.

011.

371.

331.

261.

081.

201.

150.

591.

06O

ther

gen

eral

ele

ct.e

quip

.1.

141.

211.

681.

521.

561.

451.

440.

881.

43Se

mic

ondu

ctor

s0.

000.

000.

791.

280.

640.

480.

610.

330.

46O

ffice

equ

ipm

ent,

com

pute

rs,

0.13

0.20

0.75

0.61

0.18

0.18

0.30

0.47

0.44

and

othe

r da

ta p

roce

ssin

g

145

Mot

or v

ehic

les

and

engi

nes

1.73

1.55

1.26

1.11

1.38

1.69

1.92

0.75

1.50

Oth

er t

rans

port

equ

ipm

ent

1.10

0.77

0.97

1.07

1.33

1.05

1.34

2.30

1.44

Tex

tile

s,cl

othi

ng a

nd le

athe

r2.

010.

003.

970.

001.

051.

481.

030.

000.

96R

ubbe

r an

d pl

asti

c pr

oduc

ts1.

010.

620.

910.

501.

020.

550.

762.

071.

22B

uild

ing

mat

eria

ls1.

161.

391.

720.

951.

791.

581.

352.

271.

55C

oal a

nd p

etro

leum

pro

duct

s2.

010.

003.

972.

041.

471.

011.

662.

831.

91P

hoto

grap

hic

equi

pmen

t1.

440.

000.

570.

980.

460.

560.

720.

200.

41O

ther

inst

rum

ents

and

con

trol

s1.

251.

281.

041.

131.

141.

101.

160.

880.

90O

ther

man

uf.a

nd n

on-i

nd.

1.37

0.91

0.79

1.20

1.58

0.85

1.13

2.08

1.33

Sou

rce:

as fo

r T

able

5.3

.

electric lighting, power and traction are here reflected in RTA values greaterthan 1 for 1890–1919 in illumination devices, electrical lamp manufactur-ing equipment, general electric equipment, other machinery and industrialequipment, vehicle (components) and other transport equipment. Otherareas of GE technological advantage are best related to developments thatcame in the early part of the twentieth century. An RTA greater than unityin metallurgy can be related to the development of ductile tungsten fila-ments for incandescent electric lighting, while that in chemicals may relateto advances in heat insulation and refrigeration (Reich, 1985).

What is most striking about GE’s subsequent corporate technologicaltrajectory is its high degree of path dependency and persistence, even acrossthe very wide front of capabilities over which it operated. Almost all theprimary fields of advantage in 1890–1919 were also areas of advantage forGE in 1940–47 (including illumination devices, electric lamp manufacture,chemicals and pharmaceuticals, metallurgical processes, other machineryand industrial equipment, other general electrical equipment, motorvehicles and engines, other transport equipment, building materials andcoal and petroleum products). Perhaps even more remarkably, GE contin-ued to enjoy an RTA greater than 1 in all these fields in 1947–95! However,in its original core field of lighting it had declined by 1991–95. Thus, by thislast period, GE’s RTA in illumination devices had fallen to 0.90, while inelectric lamp manufacturing equipment it had dropped as low as 0.45. Nodoubt lighting is not as central to new development in the electrical equip-ment industry now as it once was, but given the overall forcefulness of GE’stechnological path dependency over a wide range of activities, this retreatfrom its historical origins is noteworthy, even though it has taken the bestpart of a century to reach that turning point away from the past.

The Evolution of Technological Capabilities at AT&T

AT&T (Bell Telephone as it was) began with a much sharper focus in its tech-nological specialisation than that at GE, concentrating its efforts on the tele-phone, and on related technologies. This is readily apparent from Table 5.8,and from a comparison of Table 5.8 with Table 5.7. In 1890–1919 thecompany was primarily focused on its origins in telecommunications (anRTA value of 4.50), together with the closely allied fields of other electricalcommunication systems (3.68) and image and sound equipment (3.12).Secondarily, it had established a related base in metal product technologies(an RTA of 1.22), no doubt given the need to work on the detailed develop-ment of telephone receivers, transmitters, cables and the like. Between them,these four fields of technological endeavour were the only ones out of the23 areas of activity under consideration in which AT&T held RTA values


147

Tab

le 5

.8E

volu

tion

of

patt

erns

of

tech

nolo

gica

l spe

cial

isat

ion

at A

T&

T,18

90 t

o 19

95

Tec

hnol

ogic

al s

ecto

rs18

90–

1920

–19

25–

1930

–19

35–

1940

–C

umul

.19

91–

Cum

ul.

1919

2429

3439

4718

90–

9519

47–

1947

95

Che

mic

als

and

phar

mac

euti

cals

0.34

0.61

0.85

0.87

0.87

0.73

0.71

0.82

0.72

Met

allu

rgic

al p

roce

sses

0.41

0.48

1.08

1.54

1.38

1.80

1.27

0.86

1.20

Mis

cella

neou

s m

etal

pro

d.1.

220.

700.

691.

501.

821.

221.

210.

300.

85E

lect

rica

l lam

p m

anuf

.equ

ip.

0.28

1.57

0.47

0.18

0.82

0.36

0.47

0.16

0.37

Oth

er m

ach.

and

ind.

equi

p.0.

370.

530.

751.

041.

021.

400.

900.

420.

95T

elec

omm

unic

atio

ns4.

502.

532.

322.

261.

952.

152.

652.

772.

77O

ther

ele

ctri

cal

3.68

2.21

1.28

1.09

1.83

2.04

1.94

1.12

1.48

com

mun

icat

ion

syst

ems

Spec

ial r

adio

sys

tem

s0.

350.

701.

020.

930.

470.

480.

580.

200.

63Im

age

and

soun

d eq

uipm

ent

3.12

2.48

1.77

1.43

1.50

1.05

1.59

0.86

1.00

Illu

min

atio

n de

vice

s0.

420.

590.

560.

350.

390.

580.

440.

080.

39E

lect

rica

l dev

ices

and

sys

tem

s0.

730.

640.

700.

710.

910.

730.

740.

990.

92O

ther

gen

eral

ele

ct.e

quip

.0.

150.

150.

240.

350.

340.

400.

260.

950.

45Se

mic

ondu

ctor

s0.

002.

890.

000.

390.

620.

510.

451.

191.

04O

ffice

equ

ipm

ent,

com

pute

rs,

0.31

0.75

1.40

1.07

1.12

0.98

0.95

0.99

1.13

and

othe

r da

ta p

roce

ssin

g

148

Tab

le 5

.8(c

onti

nued

)

Tec

hnol

ogic

al s

ecto

rs18

90–

1920

–19

25–

1930

–19

35–

1940

–C

umul

.19

91–

Cum

ul.

1919

2429

3439

4718

90–

9519

47–

1947

95

Mot

or v

ehic

les

and

engi

nes

0.18

0.43

0.53

0.09

0.12

0.00

0.21

0.00

0.17

Oth

er t

rans

port

equ

ipm

ent

0.12

0.15

0.07

0.32

0.15

0.29

0.14

0.00

0.20

Tex

tile

s,cl

othi

ng a

nd le

athe

r0.

002.

890.

000.

641.

410.

000.

760.

000.

80R

ubbe

r an

d pl

asti

c pr

oduc

ts0.

170.

381.

332.

111.

091.

931.

410.

691.

19B

uild

ing

mat

eria

ls0.

510.

510.

430.

890.

991.

130.

800.

550.

73C

oal a

nd p

etro

leum

pro

duct

s0.

000.

000.

000.

320.

000.

000.

120.

000.

28P

hoto

grap

hic

equi

pmen

t0.

710.

001.

671.

230.

820.

000.

830.

000.

63O

ther

inst

rum

ents

and

con

trol

s0.

320.

610.

580.

870.

770.

690.

651.

180.

86O

ther

man

uf.a

nd n

on-i

nd.

0.65

1.09

1.34

1.00

1.27

1.14

1.10

0.26

0.43

Sou

rce:

as fo

r T

able

5.3

.

greater than 1 in 1890–1919. In contrast, GE had an RTA above unity in noless than 16 out of 23 fields over the same period (see Table 5.7).

From the 1920s onwards, and especially after the formation of BellLaboratories in 1925, AT&T was engaged in a substantial technologicaldiversification largely in support of (rather than away from) its continuingcore interests in telecommunications (Reich, 1985). This theme is again wellillustrated in Table 5.8. In 1925–29, AT&T had added to the fields in whichit held an RTA value above 1 (among others) the three areas of metallurgi-cal processes (1.08), rubber and plastic products (1.33) and office equip-ment and other data processing (1.40). These three new fields of strengthare likely to have been related to respectively the firm’s development in the1920s of metallic contacts for telephone switching apparatus and the prop-erties of magnetic materials, enamel and phenol-fibre insulation, and tele-phone and telegraph transmission parameters (Reich, 1985). In 1930–34 anRTA above one was further attained in other machinery and industrialequipment (then 1.04), representing again a recognition of the role of ver-tical systems integration in the direction of diversification in the electricalequipment industry.

By 1940–47 all these new fields just referred to continued to hold an RTAgreater than unity in AT&T – metallurgical processes at 1.80, machineryand industrial equipment at 1.40 and rubber and plastic products at 1.93 –but with the partial exception of office equipment and data processing at0.98. Needless to say, the established advantage was retained through to1940–47 in telecommunications (2.15), other electrical communicationsystems (2.04), image and sound equipment (1.05) and metal products(1.22). In other words, over a combined period of well over 50 years AT&Tpreserved its core technological competence, but effectively built around it.As has been commented upon already, AT&T was also involved as a majorcontributor to the development of the inter-war radio industry. However,while it had moved substantively into special radio systems (an RTA of 1.02in 1925–29) and various instrument technologies (0.87 in 1930–34), thesenever became its relative strength compared with its major rivals (and mostnotably compared with RCA, of course). Even in image and sound equip-ment there was a significant relative decline in the 1940s (from 1.50 in1935–39 to 1.05 in 1940–47), following the advent of commercial televisionin the USA and the new focus of experimentation that this provided(Abramson, 1995).

In essence, AT&T’s technological profile persisted also into the post-warperiod. In the years 1947–95 considered together, AT&T retained its indi-cator of positive corporate technological specialisation in telecommunica-tions (2.77), other electrical communication systems (1.48), image andsound equipment (1.00), as well as in metallurgical processes (1.20) and


rubber and plastic products (1.19), but had lost its position in metal prod-ucts (0.85). Key strengths persist all the way through to the latest period of1991–95 in its core fields of telecommunications (2.77) and other electricalcommunication systems (1.12). However, by 1991–95 the company had wit-nessed some decline of its relative capabilities in image and sound equip-ment (0.86) and metallurgical processes (0.86), and a clear fall in metalproducts (0.30) and rubber and plastic products (0.69). Replacing theseareas, AT&T now has advantages in instruments (1.18) and semiconduct-ors (1.19), for which the antecedents had been laid in the inter-war years,but on which a relative focus of development attention emerged onlyrecently. Yet despite some shift in the overall composition of its technolog-ical trajectory, AT&T’s continuing strong concentration on its telecommu-nications origins (2.77) is evident.

5 DIVERSIFICATION REVISITED AND SOMECONCLUSIONS

The evidence seems to confirm that all the large firms considered here fol-lowed specific and path-dependent corporate technological trajectories, inthat the distinctive characteristics of their early years exercised an influenceon the composition and breadth of their subsequent technological cap-abilities, and the direction in which they evolved. Despite the fact that theselarge companies broadened their areas of research, until at least the post-war period they remained specialised in those fields which had been theiroriginal stronghold. However, Du Pont and GE seem to have experienceda more radical departure from their original core technology than did IGFarben (and later Bayer) and AT&T. Nevertheless, even the extension ofBayer’s and AT&T’s research activities also led to a gradual diversificationinto some other related areas of strength. It is noticeable that the intensifi-cation of research in certain key areas appears to have spun-off alliedinnovations in other fields. This underlines the importance of interrelated-ness of technology, whereby a major technological breakthrough tends togenerate further innovations in connected fields, a feature of particular sig-nificance for firms in the science-based industries. It also helps to show theusefulness of patent statistics as a means of tracing the historical paths fol-lowed by large firms when considered across their entire distribution, asimportant patents are unlikely to be isolated while unimportant ones maybe. While the patterns described here are entirely consistent with the qual-itative evidence of business histories, they add some precision by facilitat-ing clearer comparisons between the positions and paths of firms operatingin similar industries.



Having commented on the extent of corporate technological diversifica-tion and how it evolved historically in each of our individual companydescriptions of paths of specialisation, it is time to summarise these trendsthrough an examination of our more formal indicator of diversification,DIV, the reciprocal of the coefficient of each company’s RTA distributionacross fields (as explained above). The values of this indicator are set out inTable 5.9. The first row of the table, concerning 1890–1919, affirms whathas already become clear from the discussion of Tables 5.5 to 5.8. That is,corporate technological diversification was much more pronounced fromthe outset in general electrical systems (as represented by GE, with a DIVvalue of 1.64 in 1890–1919) than it was in chemicals (1.04 in what becamethe IG group combined, and a still more concentrated 0.71 in Du Pont), buttelecommunications was more like the latter than the former (with a DIVas low as 0.61 in AT&T for the equivalent period).

By the end of the inter-war period both Du Pont and AT&T had largelycaught up with the span of diversified technological development at GE.Thus, in 1935–39 Du Pont’s DIV value had risen as high as 2.00, and thatof AT&T to 1.71, as against a value of 2.01 in the case of GE. IG Farben

Table 5.9 Evolution of corporate technological diversification, 1890to 1995

Period IG Farben Bayer Du Pont GE AT&T

1890–1919 1.04 n/a 0.71 1.64 0.661920–24 0.57 n/a 1.06 1.35 1.071925–29 1.23 n/a 1.00 1.32 1.331930–34 1.27 n/a 1.14 2.03 1.591935–39 1.10 n/a 2.00 2.01 1.711940–59 n/a 0.79 2.68 1.87 1.651960–64 n/a 0.83 1.90 1.42 1.371965–68 n/a 0.91 2.66 1.68 1.091969–72 n/a 0.91 2.63 1.87 1.161973–77 n/a 0.95 1.98 1.41 1.151978–82 n/a 1.34 1.58 1.31 1.191983–86 n/a 1.34 1.50 1.36 1.321987–90 n/a 1.50 1.46 1.20 0.891991–95 n/a 1.31 1.17 1.28 0.98

1890–95 0.89 1.15 2.56 2.04 1.52

Note: n/a not applicable.

Source: As for Table 5.3.

was the firm out of step in this respect, in that while it had been graduallydiversifying through to 1930–34, at which stage it remained more diversi-fied than Du Pont (a DIV value of 1.27 as against 1.14 for Du Pont), in thelater 1930s what appeared to be its natural commercial path became dis-torted by the need to be attuned to the military objectives of the newGerman government. The fall in IG’s DIV value in 1935–39 (to 1.10) wasparticularly associated with its further move into oil-related chemicals,exactly at the when Du Pont was diversifying heavily into a range of newchemical processes following (and associated with) its successful transitioninto synthetic resins and fibres.

Bringing the story closer to the present day, there tends now to be lessdifference between large firms in the scope of their technological diversifi-cation, compared to the inter-company variety of diversification that wasoften observed in the past. By 1991–95 the DIV measure seems to have con-verged among firms to a range of around 1.0 through to 1.3. Compared tothe long-term historical average value of DIV (for 1890–1995 as a whole)this represents a rise for Bayer (to 1.31 in 1991–95, as opposed to 1.15 in itsown post-war history, or even 0.89 for the IG group in earlier times), a fallfor AT&T (from 1.52 as a long run average through to 1.32 in 1983–86, andthen 0.98 in 1991–95, although the sharpness of this recent structural shiftreflects the break-up of AT&T and the greater focusing of its remainingbusiness), but a significant decline for Du Pont (1.17 in 1991–95 v. 2.56 in1890–1995) and GE (1.28 in 1991–95 v. 2.04 in 1890–1995).

Looking across companies and over time the general trend that isobserved is of a steady initial increase in technological diversification his-torically, followed by a renewed concentration in more recent times. Thethree of our four companies with the greatest continuity of historical iden-tity (Du Pont, GE and AT&T) describe this pattern most clearly, in that theDIV values with which they began in 1890–1919, and those with which theyfinished in 1991–95, were both well below their respective long-term aver-ages for DIV in 1890–1995 as a whole. If one allows for the specificities ofthe experience of IG Farben in its later years (its contribution to theGerman war effort), and for the smaller size of Bayer which was only partof the original group and the recovery of the German chemical industry inthe early post-war period, one can also see an historical trend towardsdiversification, from a DIV value of 1.04 in 1890–1919 through to 1.27 in1930–34 and then to 1.34 in Bayer in 1978–82. Although Bayer’s DIV valuedid not decline after 1978–82, it has seen no further sustained increase sincethat time (standing at 1.31 in 1991–95).

While in the early post-war years there was some continuation of theinter-war diversification trend (into, for example, photographic chemistry)in both Du Pont and Bayer, from around 1970 Du Pont has refocused its


technological efforts (with a fall in DIV from 2.63 in 1969–72 to 1.17 in1991–95). In comparison the reversal of the diversifying trend beganearlier in the post-war period in the electrical equipment industry, withsome moderate refocusing of technological efforts in GE until around 1970(the DIV indicator fell from 2.01 in 1935–39 to 1.87 in 1969–72), and inAT&T until the mid-1980s (a drop from 1.71 in 1935–39 to 1.32 in 1983–86,having been as low as 1.16 in 1969–72). Since then there has been a clearerrefocusing upon a more closely related set of technological activities in GEfrom around 1970 (from 1.87 in 1969–72 to 1.28 in 1991–95) and in AT&Tfollowing its break-up (from 1.32 in 1983–86 to 0.98 in 1991–95).

Some explanation can be offered for the apparent switch in the long-term direction of corporate technological diversification in at least thisgroup of the largest firms (in terms of their patent volume), away frompro-diversifying change and towards an increasing focus of effort. In thefirst phase of the growth of large industrial companies, from around theend of the nineteenth century through to the Second World War, productdiversification and technological diversification were much more closelyconnected to one another, through attempts to realise the joint economiesof scale and scope (as documented in depth by Chandler, 1990). In thesecond phase of such growth since 1945, and especially since around 1970,corporate technological diversification has acquired a new motive apartfrom the simple support of product or market diversification. That is, inmore recent times the primary motive has become the potential rewardsfrom rising technological interrelatedness between formerly largely sepa-rate and discrete branches of innovative activity. These are new and moredynamic economies of scope, associated with continuous knowledgespillovers between allied fields of learning, and with the creation of newand more complex technological combinations.

While for many smaller companies this shift of motives has meant a newimpulse towards greater technological diversification than in the past toincorporate what have become the most closely related areas to their owncore business, in some giant firms the greater potential inner benefits ofinterrelatedness has meant, instead, a refocusing of efforts around that com-bination of their established areas which have become most closely related(Cantwell and Santangelo, 2000). Thus, the drivers of corporate techno-logical diversification have shifted from the coverage of related productsand markets in the first phase of large company growth, to the relatednessto be found in innovative activity itself, and in the construction of new tech-nological combinations in the second phase in the growth of large firms. Sotechnological diversification associated with a steady movement outwardsinto new markets and technologically related products has been graduallyreplaced by often more focused combinations of technological activity


to capture the fruits of interrelatedness in the competence creation processitself.

Meanwhile, path dependency has prevailed throughout the last hundredyears or so of the corporate technological trajectories of these large firms,but it is a path dependency accompanied with the gradual drift that is asso-ciated with most stochastic processes. Bayer has continued with IG Farben’straditions in dyestuffs, photographic chemistry and fibres to the presentday, but Du Pont has eventually moved away from its historic origins inexplosives. AT&T continues to hold its primary position in electrical com-munication systems, but GE has eventually moved away from illuminationdevices and lamp manufacture. Taken together these large company tech-nological histories show how corporate technological trajectories have thetypical property of path dependency with some continual drift, and not thestrong kind of ‘lock-in’ configuration of the QWERTY kind.

NOTES

1. The author is grateful for the support of the UK Economic and Social Research Council,who funded the project on long-term technological change in the largest US andEuropean firms on which this chapter draws, and to Pilar Barrera, who worked with himon that project. He is also grateful for the help of Jane Myers and Jim Hirabayashi at theUS Patent and Trademark Office.

2. The CV measure has often been used as well in the analysis of business concentrationacross firms within an industry, as opposed to concentration or dispersion across sectorswithin a firm (see Hart and Prais, 1956). It is worth noticing that alternative measurescould be used (for example, the Herfindhal index) but that for a given number of firmsor sectors (N), there is a strict relationship between the Herfindahl index (H) and thecoefficient of variation (CV) (Hart, 1971). The relationship is H (CV2 � 1)/N.

3. There is evidence that during the interwar period there was a close overlap between oiland petrochemical research in the chemical industry (Freeman, 1982).

4. ‘Nylon became far and away the biggest money-maker in the history of the Du PontCompany’ (Hounshell and Smith, 1988: 273).

REFERENCES

Abramson, A. (1995), Zworykin: Pioneer of Television, Urbana, IL: University ofIllinois Press.

Aitken, H.G.J. (1985), The Continuous Wave: Technology and American Radio,1900–1932, Princeton, NJ, Princeton University Press.

Archibugi, D. (1992), ‘Patenting as an indicator of technological innovation: areview’, Science and Public Policy, 19, 357–68.

Arthur, W.B. (1994), Increasing Returns and Path Dependence in the Economy, AnnArbor, MI, University of Michigan Press.

Basberg, B.L. (1987), ‘Patents and the measurement of technological change: asurvey of the literature’, Research Policy, 16, 131–41.


Beaton, K. (1957), Enterprise in Oil: A History of Shell in the United States, NewYork: Appleton-Century-Crofts.

Beer, J.J. (1959), The Emergence of the German Dye Industry, Urbana, IL:University of Illinois Press.

Bound, J., Cummins, C., Griliches, Z., Hall, B.H. and Jaffe, A.B. (1984), ‘Who doesR & D and who patents?’, in Z. Griliches (ed.), R & D, Patents and Productivity,Chicago, IL, University of Chicago Press.

Cantwell, J.A. (1989), Technological Innovation and Multinational Corporations,Oxford, Basil Blackwell.

Cantwell, J.A. (1993), ‘Corporate technological specialisation in internationalindustries’, in M.C. Casson and J. Creedy (eds), Industrial Concentrationand Economic Inequality: Essays in Honour of Peter Hart, Aldershot, EdwardElgar.

Cantwell, J.A. (2000), ‘Technological lock-in of large firms since the inter-warperiod’, European Review of Economic History, 4, 147–74.

Cantwell, J.A. and Andersen, H.B. (1996), ‘A statistical analysis of corporate tech-nological leadership historically’, Economics of Innovation and New Technology,4, 211–34.

Cantwell, J.A. and Barrera, M.P. (1998), ‘The localisation of corporate techno-logical trajectories in the interwar cartels: cooperative learning versus anexchange of knowledge’, Economics of Innovation and New Technology, 6, 257–90.

Cantwell, J.A. and Fai, F.M. (1999), ‘Firms as the source of innovation and growth:the evolution of technological competence’, Journal of Evolutionary Economics,9, 331–66.

Cantwell, J.A. and Santangelo, G.D. (2000), ‘Capitalism, profits and innovation inthe new techno-economic paradigm’, Journal of Evolutionary Economics, 10,131–57.

Chandler, A.D. (1990), Scale and Scope: The Dynamics of Industrial Capitalism,Cambridge, MA, Harvard University Press.

Coleman, D.C. (1969), Courtaulds: An Economic and Social History (Volume II:Rayon), Oxford and New York: Oxford University Press.

David, P.A. (1985), ‘Clio and the economics of QWERTY’, American EconomicReview, 75, 332–7.

David, P.A. (1991), ‘The hero and the herd in technological history: reflections onThomas Edison and the battle of the systems’, in P. Higonnet, D.S. Landes andH. Rosovsky (eds), Favorites of Fortune: Technology, Growth and EconomicDevelopment Since the Industrial Revolution, Cambridge, MA, HarvardUniversity Press.

David, P.A. (1993), ‘Historical economics in the long run: some implications of pathdependence’, in G.D. Snooks (ed.), Historical Analysis in Economics, Londonand New York, Routledge.

David, P.A. (1994), ‘Why are institutions the “carriers of history”? Path dependenceand the evolution of conventions, organisations and institutions’, StructuralChange and Economic Dynamics, 4, 205–20.

David, P.A. (2001), ‘Path dependence, its critics and the quest for “historical eco-nomics” ’, in P. Garrouste and S. Ioannides (eds), Evolution and Path Dependencein Economic Ideas: Past and Present, Cheltenham, Edward Elgar.

Dornseifer, B. (1989), ‘Research, innovation and corporate structure: Du Pont andIG Farben in comparative perspective’, Harvard University Graduate School ofBusiness Administration Working Paper, Spring.


Dornseifer, B. (1995), ‘Strategy, technological capability and innovation: Germanenterprises in comparative perspective’, in F. Caron, P. Erker and W. Fischer(eds), Innovations in the European Economy Between the Wars, Berlin: de Gruyter.

Freeman, C. (1982), The Economics of Industrial Innovation, London, FrancesPinter.

Griliches, Z. (1990), ‘Patent statistics as economic indicators: a survey’, Journal ofEconomic Literature, 28, 1661–707.

Haber, L.F. (1971), The Chemical Industry: 1900–1930, Oxford, Oxford UniversityPress.

Hart, P.E. (1971), ‘Entropy and other measures of concentration’, Journal of theRoyal Statistical Society, Series A, 134, 73–85.

Hart, P.E. and Prais, S.J. (1956), ‘The analysis of business concentration: a statis-tical approach’, Journal of the Royal Statistical Society, Series A, 119, 150–91.

Hounshell, D.A. (1995), ‘Strategies of growth and innovation in the decentralizedDu Pont company 1920–1950’, in F. Caron, P. Erker and W. Fischer (eds),Innovations in the European Economy Between the Wars, Berlin, de Gruyter.

Hounshell, D.A. and Smith, J.K. (1988), Science and Corporate Strategy: Du PontR&D, 1902–1980, Cambridge, Cambridge University Press.

Hughes, T.P. (1983), Networks of Power: Electrification in Western Society,1880–1930, Baltimore MD, Johns Hopkins University Press.

Hughes, T.P. (1989), American Genesis: A Century of Invention and TechnologicalEnthusiasm, New York, Viking.

Jones, R. and Marriot, O. (1971), Anatomy of a Merger: A History of GEC, AEI andEnglish Electric, London: Cape.

Maclaurin, W.R. (1949), Invention and Innovation in the Radio Industry, Londonand New York, Macmillan.

Nelson, R.R. and Winter, S.G. (1982), An Evolutionary Theory of Economic Change,Cambridge, MA, Harvard University Press.

Nobel, D.F. (1979), America by Design: Science, Technology and the Rise ofCorporate Capitalism, New York, Alfred A. Knopf.

Patel, P. and Pavitt, K.L.R. (1997), ‘The technological competencies of the world’slargest firms: complex and path-dependent, but not much variety’, ResearchPolicy, 26, 141–56.

Patel, P. and Pavitt, K.L.R. (1998), ‘The wide (and increasing) spread of techno-logical competencies in the world’s largest firms: a challenge to conventionalwisdom’, in A.D. Chandler, P. Hagström and Ö. Sölvell (eds), The Dynamic Firm:The Role of Technology, Strategy, Organisation, and Regions, Oxford andNew York, Oxford University Press.

Pavitt, K.L.R. (1985), ‘Patent statistics as indicators of innovative activities: possi-bilities and problems’, Scientometrics, 7 (1–2), 77–99.

Pavitt, K.L.R. (1988), ‘Uses and abuses of patent statistics’, in A. van Raan (ed.),Handbook of Quantitative Studies of Science Policy, Amsterdam, North Holland.

Pavitt, K.L.R. and Soete, L.L.G. (1980), ‘Innovative activities and export shares:some comparisons between industries and countries’, in K.L.R. Pavitt (ed.),Technical Innovation and British Performance, London: Macmillan, 38–66.

Plumpe, G. (1990), Die IG Farbenindustrie AG. Wirtschaft, Technik und Politik1904–1945, Berlin, Duncker and Humblot.

Plumpe, G. (1995), ‘Innovation and the structure of the IG Farben’, in F. Caron,P. Erker and W. Fischer (eds), Innovations in the European Economy Between theWars, Berlin, de Gruyter.


Reich, L.S. (1977), ‘Research, patents and the struggle to control radio: a study ofbig business and the uses of industrial research’, Business History Review, 51,208–35.

Reich, L.S. (1985), The Making of American Industrial Research: Science andBusiness at GE and Bell, 1876–1926, Cambridge and New York, CambridgeUniversity Press.

Rosenberg, N. (1982), Inside the Black Box: Technology and Economics, Cambridgeand New York, Cambridge University Press.

Scherer, F. (1983), ‘The propensity to patent’, International Journal of IndustrialOrganisation, 1, 107–28.

Scherer, F.M., Herzstein, S.E., Dreyfoos, A.W., Whitney, W.G., Bachmann,O.J., Pesek, C.P., Scott, C.J., Kelly, T.G. and Galvin, J.J. (1959), Patents and theCorporation: A Report on Industrial Technology Under Changing Public Policy,Boston, MA, J.J. Galvin (privately published).

Schmookler, J. (1966), Invention and Economic Growth, Cambridge, MA, HarvardUniversity Press.

Soete, L.L.G. (1987), ‘The impact of technological innovation on internationaltrade patterns: the evidence reconsidered’, Research Policy, 16, 101–30.

Stocking, G.W and Watkins M.W. (1946), Cartels in Action, New York: TwentyCentury Fund.

Wilkins, M. (1974), The Maturing of Multinational Enterprise: American BusinessAbroad from 1914 to 1970, Cambridge, MA: Harvard University Press.

Wilkins, M. (1989), The History of Foreign Investment in the United States to 1914,Cambridge, MA: Harvard University Press.

Winter, S.G. (1988), ‘On Coase, competence and the corporation’, Journal of Law,Economics and Organization, 4, 163–80.


6. Is the world flat or round? Mappingchanges in the taste for artG.M. Peter Swann1

1 INTRODUCTION

Public taste, I believe, as far as it is the encourager and supporter of art, has beenthe same in all ages; a fitful and vacillating current of vague impression, perpet-ually liable to change, subject to epidemic desires, and agitated by infectiouspassion, the slave of fashion, and the fool of fancy. (Ruskin 1843 [1996], vol. 3:617–18)

The main lesson imparted by the test of time is the fickleness of taste whosemeanderings defy prediction. (Baumol 1986: 14)

This chapter is a preliminary attempt to map the changing tastes for worksof art as manifested in the prices of paintings sold at auction. There are twomain goals in this work: first, to describe a space in which we can representthe work of different artists; and second, to describe how ‘cultivated taste’moves around that space.

In pursuing the first goal, moreover, we have to confront a furtherquandary: what is the appropriate shape of the space suitable for thispurpose? Is the appropriate world ‘round’ or ‘flat’? This is an empirical ques-tion, which can be assessed by reference to measures of ‘goodness of fit’. Butthere is also an interesting theoretical interpretation. We shall argue that theanswer depends on whether one takes a historicist or a modernist perspec-tive on the development of art. When we turn to the second goal, the quotefrom Ruskin illustrates why our task will be a difficult one: it is not easy tomap a ‘fitful and vacillating current’. Even more, we shall argue that move-ment in taste around the space of painters is a path-dependent process.

That provides the link between this chapter and the work of Paul David.One of the many areas in which his work has been very influential is in theeconomics of path dependence – see for example David (1985; 1986; 1987;1988; 1992; 1993; 1994; 1997), David and Foray (1994), David et al. (1998).Antonelli (1997) introduces a special issue of the International Journal ofIndustrial Organisation on path dependence, inspired by Paul David’s work.

158

In a very loose sense, path dependence is the idea that ‘history matters’when we try to understand how markets evolve today. But Paul David hasprovided us with much more precise definitions of path dependence, andwe shall turn to this later (section 4). The path dependence we observe inthis case is a little different from that observed in much of his work. Forhere, as art historians and theorists have recognised (see section 3), there isboth positive and negative feedback: positive feedback from the elite to theaspirants but negative feedback from aspirants to the elite.

The rest of the chapter is in eight sections. Section 2 briefly describes theeconomic models of taste change, which guide what follows – though theseare discussed at greater length in the companion chapter by Cowan in thisbook (Chapter 7). Section 3 looks at a few central themes in art theory andhistory that illuminate our study. Section 4 provides a more precise defin-ition of path dependence. Section 5 compares two leading techniques formapping products in a common space: the characteristics approach andmultidimensional scaling (MDS). Section 6 then examines (in the light ofsection 3) whether an appropriate space for mapping should be flat or round:precisely, should we work with a (two-dimensional) plane or the surface ofa (three-dimensional) sphere? Section 7 states precise results for mappingfrom price correlations to points on a plane or on the surface of a sphere.Section 8 presents some preliminary calculations, which illustrate in a roughway how tastes have changed over a 100-year period. Section 9 concludes.

2 SOCIAL THEORIES AND ECONOMIC MODELSOF TASTE

This chapter aims only to map changes in taste, and not to explain why theycome about. The companion chapter by Robin Cowan (Chapter 7) offers amodel of changing tastes towards works of art. However, to help motivatethe chapter, here is a brief sketch of one possible process leading to changesin taste.

For a long time, we have known that the character of demand for pres-tige goods, including works of art, is rather different from the elementaryneoclassical picture of demand. Veblen (1899) described how the newlywealthy indulged in conspicuous consumption – that is, the visible con-sumption of things that other people do not have. So long as these items ofconspicuous consumption are not owned by others from whom one wishesto distinguish oneself (Bourdieu, 1984), then they serve their purpose well.But when others who aspire to share the consumption activities of the elitestart to catch up, then it is time for the elite to move on to other forms ofconspicuous consumption.

Mapping changes in the taste for art 159

Cowan et al. (1997; 2004) have developed a model of demand driven bythe conflicting desires of association, distinction and aspiration. The con-sumer may seek to associate in consumption with some groups, may seekdistinction in consumption from some other groups, and all the while isaspiring to share some consumption activities with an elite. In this model,demand exhibits waves: to begin with, a particular good may be in demandfrom the elite, but then demand shifts downmarket, and the elite desert thegood. Under some circumstances, these waves can repeat themselves. Thismodel generates rich and complex patterns in the consumption of particu-lar goods.

The companion chapter by Cowan (Chapter 7) shows how models of thissort may be applied to understanding changing tastes towards works of art.In this last study, cycles in relative popularity are a natural outcome whenthe products are arrayed around a circle.

Although some other social scientists believe that economics has rela-tively little to say about such historical patterns in demand, the opposite istrue. There is indeed a large economics literature on association, distinctionand aspiration in demand. We do not attempt to review that literature here,but have done so elsewhere (Swann, 1999; 2002). See also McPherson(1987) for a very useful overview. Some of the most important recent con-tributions are by Becker (1996), Bianchi (1998), Dosi et al. (1999) andFrank (1985). Moreover, following an influential paper by Baumol (1986),an interesting and substantial literature has grown up recently on the eco-nomics of the arts and culture, including Cozzi (1998), essays in De Marchiand Goodwin (1999), Frey (1997; 2000), Frey and Pommerehne (1989),essays in Ginsburgh and Menger (1996), Grampp (1989), McCain (1981)and Throsby (2001).

Nor does this chapter try to describe how painters have reacted tochanges in taste. Elsewhere, we have studied how the nature of consumerdemand for distinction goods might influence the product design strategiesof producers (Swann, 2001a). It is certainly true that some of the paintersin our sample painted primarily for the market at the time, and as a resultsome of their works do not appeal today. By contrast, Van Gogh only soldone painting during his lifetime, but his popularity has grown monoton-ically since his death.

3 SOME THEMES IN ART HISTORY AND THEORY

Ruskin’s remark, quoted at the start of the chapter, highlights the volatilityof taste. In view of that, it would not make much sense to assume that tastecan be approximated by a constant! But Ruskin was talking about taste


over all forms of art. Individual episodes are less chaotic but show curiouscycles of popularity. As the art dealer and historian, Maurice Rheims (1959[1961]) shows us, the fluctuating fortunes of art in different times can bequite entertaining. But he notes that it can also be cutting, when he says(Rheims, 1959 [1961]: 133): ‘Fashion is a sorcerer’s charm or talismanchanging the masterpieces of today into the laughing-stock of tomorrow.’

Why do these changes in taste come about? It is unlikely that there couldever be one explanation that accounts for all cases, but some examples aresuggestive. Take the history of the Pre-Raphaelite group of painters.Bell (1984) discusses the fall in popularity of the Pre-Raphaelites in the1920s. Part of their error was to abuse their position as the establishment,and fail to understand innovations elsewhere (Bell, 1984: 14). In turn, the‘Bloomsbury Group’ (notably Bell’s father, Clive Bell, 1927) were influen-tial in turning the tide against the Pre-Raphaelites. As Barnes (1998: 113)puts it: ‘The Bloomsbury set laughed at the Pre-Raphaelites. To themit seemed that “everything of importance in the second half of the19th century had happened in France” .’

In addition, Bell (1984: 16) recognised that part of the reason for theirfall was that the wrong patrons had bought their work: ‘From the very firstthese painters found their market among those whom contemporarieswould have considered an ignorant and philistine clientele, the “self-made”men and manufacturers of the North.’

In that respect, Bell takes a view very similar to that in the economicmodels of distinction and aspiration described in the last section. He is evencloser when he goes on to describe how reproductions of Pre-Raphaelitepaintings became commonplace on middle class, aspirational, walls (Bell,1984: 16):

We all want to exhibit a cultivated taste, we all want to be enlisted in the culturalelite and of course in so doing we deprive the elite of its elitist character; thatwhich had been distinguished becomes in the truest sense vulgar and the publicis ready for something else; it is thus I would suggest that the wheel of fashion ismade to revolve.

While the Pre-Raphaelites became unpopular from the 1920s, and this canbe seen in prices paid at auction (see below), the wheel of fashion cameround again in the latter part of the twentieth century. Barnes (1998: 115)notes that attitudes started to change in the 1960s, and 1970s, and that theTate Gallery exhibition on Pre-Raphaelites in 1984 was one of the mostpopular ever mounted by the gallery.

Art theorists also stress how innovators in art can dispel the unattractiveassociations of the current establishment. Gombrich summarises what he


calls (with a little irony) the ‘treasured legend of the modern movement’(Gombrich, 1963a: 145):

What is called loosely ‘modern art’ sprang indeed from a protest against the liein the soul, a revulsion from false values. When new classes of patrons acquiredunexpected wealth and were bent on ostentation, cheap vulgarity stifled ourcities and choked our drawing rooms. Sentimental trash was taken for Great Art.This sickened the heart of the true artists who went on their lonely and perilousway in the face of public neglect and derision.

Others went further and argued that true modernity was a negation of allthat is past – not just the current establishment. Marcus (1998: 7) argues asfollows: ‘To be modern, design did not just have to be new, it had to be freeof any reference to the decorative styles of the past.’ Some, by contrast,recognised that negation of the present establishment could lead to a redis-covery of past styles. So, for example, while Malraux (1954) describes aconcept of perpetual revolution in art, of the artist as a ‘defiant animal’ andhence a theory of changing styles of (and tastes for) of art, he neverthelessrecognises a potential connection with the distant past. Gombrich sum-marises Malraux’s position very succinctly (Gombrich, 1963b: 83):‘Modern art came into being as a protest against the commercial pseudo-art of prettiness. It is this element of negation that establishes its kinshipwith the religious art of the past . . .’.

Gombrich (1963b: 83) recognises, however, the fundamental paradoxthat today’s revolutionaries in art become tomorrow’s establishment: ‘True,modern art will not be able to “outlive its victory intact”. As an act of defi-ance it will wither away when it becomes dominant.’

Historicism, the practice of borrowing from the more distant past (eventhat untainted by recent associations) was an anathema to true modernists.Marcus (1998: 7) argues:

Ever since the middle of the previous century, reformers had condemned design’sdependence on historicism (and its handmaiden, ornamentalism), and theprogress of modern design could be measured by the extent to which the miningof historic styles was supplanted by the creation of new, anonymous, and uni-versal forms, forms that looked to the future instead of the past.

Nevertheless, it was still widely accepted, especially in the nineteenthcentury, and as McDermott (1992: 120–21) notes, an obsession for the pastlead to the publication of a large number of design source books at thattime.

The history of what was popular and unpopular at particular times inthe past shapes what is popular and unpopular today. Using the term, ‘path


dependence’ in an informal way, we can say that the history of taste is path-dependent. Some of the manifestations of path dependence that havedrawn greatest attention in the literature occur when there is positive feed-back only, leading perhaps to ‘lock in’. Here, in contrast, there is a mix ofpositive and negative feedback.

4 PATH DEPENDENCE

In some parts of the literature, the term ‘path-dependent’ has been used ina rather casual fashion. David (1997: 13) helps us to be much more precise.Path dependence is a dynamic property of allocative processes. It can bedefined with reference to either: (a) the relationship between processdynamics and outcome(s) to which it converges; or (b) the limiting prob-ability distribution of the stochastic process under consideration.

To understand the true meaning of path dependence, it is easiest perhapsto start by defining path independence. A path-independent process is onewhose dynamics guarantee that it will converge to a unique, globally stableequilibrium, or, in the case of stochastic systems, where the outcome hasan invariant stationary asymptotic probability distribution. Stochasticsystems with this latter property are ergodic. That means they are able toshake off the influence of their past.

David (1997:13) offers a ‘negative definition’ of path dependence:‘Processes that are non-ergodic, and thus unable to shake free of theirhistory, are said to yield path dependent outcomes.’ Working on from this,he can provide a ‘positive definition’ (David, 1997: 14): ‘A path dependentstochastic process is one whose asymptotic distribution evolves as a conse-quence (function) of the process’ own history.’

In short, there are three key points about path dependence (David, 1997:18–19):

● Path dependence is a property of stochastic dynamic systems.● It is natural to interpret a path-dependent process as a contingent

branching process.● The definition of path dependence is independent of any issues of

economic efficiency or inefficiency.

From this point of view, there seems little doubt that the evolution of tastesand prices in the market for art is truly a path-dependent process – evenif not all the models that could be invoked to analyse these are strictlypath-dependent.


5 CHARACTERISTICS AND MULTIDIMENSIONALSCALING

We can now return to the primary aim of the chapter: to map the changingtastes for works of art as manifested in the prices of paintings sold atauction.

Whenever economists seek to represent different products in a commonspace, the natural starting point is the characteristics approach, developedin the modern literature by Gorman (1980), Ironmonger (1972) andLancaster (1971). But what characteristics are needed to give an adequaterepresentation of objects so subtle as works of art? Here Bacharach’s(1990) work on commodities and language is helpful. We can judge thenecessary characteristics from the things people say about these objects.But in the case of works of art, people have said a great deal. For example,Wildenstein’s (1996) catalogue of paintings by Monet runs to four volumesand over 1500 pages. It would seem to be a difficult task indeed to captureall this in a set of characteristics scores!

Nevertheless, both economists and art historians have tried to capturethe essence of good art in a few characteristics. As De Marchi (1999: 6)notes, Adam Smith argued that with objects of ‘art’, we derive pleasurefrom four characteristics:

● Form● Colour● Rarity● Ingenuity of design and manufacture.

Equally, the nineteenth-century art and social critic, John Ruskin, said thatthere were four essential characteristics of great art (Ruskin 1856 [1996],vol. 5: 48–63):

● Choice of Noble Subject● Love of Beauty● Sincerity● Invention (Imagination).

And, amongst classical writers on this theme, perhaps the greatest advanceswere made by De Piles, who again identified four characteristics of art(De Piles, 1708, here quoted from De Marchi, 1999: 8–10):

● Design● Colouring


● Composition● Expression.

De Piles, indeed, attempted to rank the work of 56 different painters usinghis own estimates of their scores on these four characteristics. And in themodern literature, the hedonic technique has been applied to works of art(Chanel et al., 1996).

However, in this chapter, we shall take a different approach. We seek toidentify an implicit taste space by reference to observed trends in the pricesof works of art. The basic logic of our approach is this. If the prices ofworks by two painters are closely correlated over time, then we assume thatthese painters are located close together in the taste space. If, on the con-trary, the prices of works by two painters are negatively correlated, then weassume that these painters are located far apart in the taste space.

A variety of techniques exist to enable us to map from an n*n matrix ofcorrelations between objects into a two-dimensional (or higher dimen-sional) representation of those objects. These include principal compon-ents, factor analysis and multidimensional scaling. Such methods give aready way of illustrating the similarities and differences between differententities, whether they are nations, companies, products or, even, consumers.For a wide range of empirical data-sets in economics, and other businessstudies, two components capture a large share of the total variance.

However, these statistical techniques are often used in a rather ad hocway, which does not describe the precise microeconomics of how correl-ations in price map into proximity in taste space. That mapping is set outin detail below.

Accordingly, in what follows, we assume that prices are sufficient statisticsfordescribing theevolutionof tastes in theartmarket. Inreality, theyarenot.The art historian has much to tell us about influences which change tastes,that are not captured in prices. Even stronger, this chapter assumes in effectthat a matrix of correlation coefficients between the prices of different artists’work is a sufficient set of statistics. This is somewhat stronger, because it doesnot make use of the more detailed trends in popularity of different artists.

Are these assumptions too strong? In the second edition of ThePrinciples of Economics, Marshall provided an interesting insight into whyprices might not give a wholly accurate measure of taste (here quoted fromWhite, 1999: 79):2

And therefore the price at which such a thing is sold will depend very much onwhether rich persons with a fancy for that particular thing happen to be presentat its sale. If not, it will probably be bought by dealers who reckon on being ableto sell it at a profit; and the variations in the price for which the same picture sells


at successive auctions, great as they are, would be greater still if it were not forthe steadying influence of professional and semi-professional purchasers. The‘equilibrium price’ for such sales is very much a matter of accident.

Indeed, since our data source (Reitlinger, 1961; 1970) simply records theauction prices of works (and also some prices relating to private sales), wedo not know how many potential buyers were ‘in the market’ for a particu-lar work at any time.

However, Ruskin believed that price was a reasonable measure of theartist’s rank (Ruskin 1843 [1996], vol. 3: 617–18): ‘Of course a thousandmodifying circumstances interfere with the action of the general rule; but,taking one case with another, we shall very constantly find the price whichthe picture commands in the market a pretty fair standard of the artist’srank of intellect.’

The technique developed below refers to the ‘height of taste’. We meanby this that location on our space of tastes and artists where demand isstrongest. We shall assume throughout that at any time the artist whosework comes closest to the ‘height of taste’ will ‘enjoy’ the highest prices.This is an important, if strong, assumption in what follows.

6 FLAT OR ROUND? MODERNISM ORHISTORICISM?

Before starting to construct our map of painters, there is one further issuethat needs attention. Should the space be flat or round? It is conventionalin multidimensional scaling and principal components to use a linear orplanar representation of the data. But there is an implicit but rarelyexplored assumption behind this procedure, which could be described asthe ‘transitivity of distance’. This is an inherent property of planar repre-sentations but would not be found, for example, in a spherical projection.3

But it is possible that goods exhibiting cycles in popularity may be betterrepresented by circular or spherical projections.

Ultimately, this is an empirical question. And in what follows we shallcreate both planar and spherical representations. However, the ideas fromart history and art theory summarised in section 3 also give us an import-ant lead here. To achieve the modernist goal that design has to be differentfrom any designs of the past, it is necessary to locate painters in anunbounded plane. If there are any bounds to this space, it is impossible fordesign to be innovative and continually to evade any reference to the past.Sooner or later the artist will be forced back into a part of the space thathas been occupied by earlier artists.


On the other hand, if one accepts that a degree of historicism is aninevitability, then a bounded space will suffice. In that case, we may wish todistinguish between two types of historicism:

● Distant past historicism: it is legitimate to use the styles of the distantpast, but not those of the recent past.

● Recent past historicism: it is legitimate to use the styles of the recentpast, but not those of the distant past.

Assume that the movement of the ‘height of taste’ around the space ofartists is continuous and smooth. If artists and tastes are represented on areal line, then when the ‘height of taste’ hits the boundary of the space,tastes must move in the reverse direction, and this will entail recent past his-toricism. On the other hand, if artists and tastes are represented on a circle,then the ‘height of taste’ can continue to cycle in one direction (say clock-wise) without reversal: this will entail distant past historicism.

When we move from lines and circles up to planes and spheres, somemore subtle possibilities emerge. But if distant past historicism is the morecommon than recent past historicism (and the references cited in section 3suggest it is), then a spherical projection may be more useful for the pur-poses of this chapter.

7 TECHNICAL DETAILS

This section presents techniques for mapping from a correlation matrix rep-resenting the similarities and differences between products to a planarspherical representation of the positions of those products. The reason wechoose to represent products on the surface of a sphere and not on theperimeter of a circle is simply so that two degrees of freedom (or two com-ponents) are obtained, rather than one. A result of considerable power andgreat simplicity is derived: for two products located on the surface of asphere, the correlation between their prices is equal to the cosine of theangle between them – as measured from the centre of the sphere.

A variety of standard data reduction methods (principal components,multidimensional scaling) take a matrix of distances (or similarities)between entities and project these entities onto a plane. Points closetogether imply that the entities are similar, while points far apart imply thatthe entities are dissimilar.

As indicated above, however, there are some disadvantages from project-ing onto a plane. In particular, if the ‘height of taste’ moves around in acontinuous fashion, it implies a particular and perhaps restricted pattern of



fashion cycles. By contrast, there are some attractions in locating points onthe surface of a sphere for in that case a rather different pattern of cycles isavailable.

In this section we show how to map from a matrix of correlations intoplanar and spherical projections. But in each case, we start with the easiercase of a line and a circle – because these help us to grasp the intuition ofwhat is happening.

Line

Assume that each artist is located at some point along a line (0,xmax). The‘height of taste’ is defined as x*, and the strength of demand at any otherpoint (xi) along the line is defined as:

S(xi, x*)1� |x*�xi| (6.1)

As Figure 6.1 shows, the signal is at a peak when x*xi, and drops awayon either side. We can use this simple model to compute the correlationbetween demand prices for artists located at different points along the line.To see this, consider Figure 6.2. Swann (2001b) shows a very simple andconvenient result. Under some assumptions, we can define the correlationbetween these demand prices as follows:

(6.2)"12

[(x1 � (x2 � x1)] � (xmax � x2)

xmax 1 �

2(x2 � x1)xmax

x1 xmax

Demand

0

1

0

Figure 6.1 Signal strength along a line


The basic intuition of this result is as follows. We can split Figure 6.2into three areas. From 0 to x1, the correlation between demand prices is 1;from x1 to x2 that correlation is �1; and from x2 to xmax that correlationis again 1.

In short, there is a natural mapping from positions to correlations: whenx1 and x2 are located close together, the correlation between their demandprices is close to 1; when they are located far apart, at either end of thepainter spectrum, then, the correlation between their demand pricesapproaches �1.

Plane

Extending the previous result to a plane is messy rather than difficult. Tomake it as simple as possible, it is helpful to use a grid measure of distance(Figure 6.3) rather than an Euclidean measure. Thus, the distance betweenany two points (x1,y1) and (x2,y2) is defined as:

L |x1�x2|� |y1�y2| (6.3)

And as above, we assume that when the ‘height of taste’ is (x*,y*), then thedemand at point (x1,y1) is given by:

S1�L1� |x1�x*|� |y1�y*| (6.4)

Figure 6.2 Correlation in demand prices

xmax

Demand

0

1

x10 x2

D1 D2


Swann (2001b) shows that the correlation between demand for 1 anddemand for 2 can be calculated by the following procedure. Split theplane into nine zones as shown in Figure 6.4. (Note that this can bedone wherever the two points lie. If y1y2 then the central horizontalband W–C–E disappears; if y20 then the lower horizontal band disap-pears; and so on). Swann (2001b) shows that under some conditions,the overall correlation between 1 and 2 over the entire plane can be com-puted as:

(6.5)

where a(i) is the area of zone i and "12 is the correlation between demandfor 1 and demand for 2 in zone i. These zone areas and zone correlationsare given in Table 6.1.: Once again, if artists are clustered together in a par-ticular part of the plane, then there is a strong positive correlation in theirdemand prices. By contrast, if they are located at opposite corners of theplane, then this correlation will be strong and negative.

corr12 �9

1a(i)"12(i)

y2

y1

x1 x2

ymax

0 xmax

Figure 6.3 A grid measure of distance


Circle

While superficially it may look harder to project points onto a circle than aline (and on the surface of a sphere than on a plane), in fact it is in somerespects easier. We shall see that a very simple result obtains: the correlationbetween the demand for and price of work by two different artists is givenby the angle between those two artists as located on the circle and viewedfrom the centre of the circle. Consider Figure 6.5. Suppose that a particu-lar artist is located at point x1 on the circle, and that at a particular timeand date, the most popular artist or ‘height of taste’ is at point x*. As in the

Figure 6.4 Nine zones to describe correlation in demand

y2

y1

x1 x2

C

NW N NE

W E

SW S SE

ymax

0 xmax

Table 6.1 Zone areas and zone correlations

Zone Area Correlation12i a (i) "12(i)

NW (ymax�y2)·x1 �1N (ymax�y2)·(x2�x1) [(ymax�y2)�(x2�x1)]/[(ymax�y2)�(x2�x1)]NE (ymax�y2)·(xmax�x2) �1W (y2�y1)·x1 [x1�(y2�y1)]/[x1�(y2�y1)]C (y2�y1)·(x2�x1) �1E (y2�y1)·(xmax�x2) [(xmax�x2)�(y2�y1)]/[(xmax�x2)�(y2�y1)]SW y1·x1 �1S y1·(x2�x1) [y1�(x2�x1)]/[y1�(x2�x1)]SE y1·(xmax�x2) �1


case of the line or plane, we need a function relating demand or demandprice for x1 to x*.

Here we adopt a slightly different convention to that used in the case ofthe line or plane. We liken the question to that of computing the brightnessof daylight at different places on the globe. Suppose that demand is a signaltransmitted from the dotted line at the top of Figure 6.5 (which is tangentto the circle at x* and which passes through 1 on the vertical axis). Theintensity of demand felt at any other point on the circle depends on the ver-tical distance from the transmitter line to that point on the circle.

From the diagram, it is readily apparent that this vertical distancedepends on the angle between x1 and x*. As drawn, the perpendicular dis-tance from transmitter line to x1 is 1�cos(#). Hence, if we measure thestrength of the signal at x1 on a scale from �1 to 1, then that strength ofsignal is given by cos(#). This is reasonable. When the angle between x1 andx* is small, so that our chosen artist is close to the ‘height of taste’, thendemand is very strong: as # → 0, cos(#) → 1. By contrast, when the angle# is large, suggesting our chosen artist is far from the ‘height of taste’, thendemand is small.

Now, using this framework, we can obtain a remarkably simple andpowerful result about the correlation between demand prices for differentartists. Take two artists, 1 and 2, located at points x1 and x2 on the cir-cumference of the circle. Suppose that the angle between each artist and

x1

.

cos(#)

�1

0

1

#

*x

Figure 6.5 Signal strength around a circle

the vertical (as drawn in Figure 6.5) is denoted by #1 and #2 respectively.Suppose also that the angle between the location of the ‘height of taste’at any date and the vertical is given by #*. (Note that in the diagram,#* 0.) Then, the strength of demand for the work of painters 1 and 2is simply defined by cos(#1�#*) and cos(#2�#*), respectively.

From this we can obtain the following expression for the covariance ofdemand prices for 1 and 2. Assume that the ‘height of taste’ might in thefullness of time occur above any point on the circle, and with equal prob-ability. This means that we can assume that the density of #* is constantover the range �$ to $. We obtain the covariance by integrating from �$to $ as shown:

(6.6)

In fact, this expression can be simplified considerably by using two results.First, when k is an integer (positive or negative) and because sin (a�k$)sin(a�k$):

(6.7)

Second, Ayers (1987: 143, eqn 9) shows that:

(6.8)

Applying the first result to equation (6) we see that the second row of thatexpression is simply zero. Moreover, using the second result, we see that thefirst line of equation (6.6) simplifies to:

(6.9)

Now, we can use equation (6.9) to derive the variance of demand for 1(or 2):

(6.10)var1 cos(#1 � #1)

2

12

cos(#1 � #2)

2

cov12 1

4$ �$

�$

cos (#1 � #2)d#* �1

4$ �$

�$

cos (#1 � #2 � 2#*)d#*

cos (x) cos (y) 12

[cos (x � y) � cos (x � y)]

�$

�$

cos (a � k#*)d#* 1k

[sin(a � k$) � sin (a � k$)] 0

�1

2$ �$

�$

cos (#1 � #*)d#* · 1

2$ �$

�$

cos (#2 � #*)d#*

cov12 1

2$ �$

�$

cos (#1 � #*)cos (#2 � #*)d#*



From which it is clear that the simple correlation between the demandprices of 1 and 2 is given by:

(6.11)

In short, the formula describing the demand intensity for one artist whenanother is the ‘height of taste’ also defines the correlation between demandprices for the work of different artists. This simple but powerful result playsa central role in what follows.

Sphere

An equivalent result obtains in the case of the sphere, but the proof is rathermore cumbersome. It is also helpful to use a slightly different notation.Suppose that we follow cartographical principles and identify any point ona sphere by its latitude (a) and longitude (n).

Start by assuming that the ‘height of taste’ is located at the ‘North Pole’,that is, where latitude90%N (or �90%, or $/2). In this case the result fromthe previous section carries over in a straightforward way. To see this,examine Figure 6.6. As in the case of the circle, a radiating plane is tangentto the sphere at the point (n0,a0), where a090%, which defines the ‘heightof taste’. The strength of demand signal at another point is defined by thevertical distance between the radiating plane and that other point. Now, as

"12 cos(#1 � #2)�2

√(1�2) · (1�2) cos(#1 � #2)

Figure 6.6 Signal strength on a sphere

(n0, a1)

(n0, a0)

(n1, a1)

S(n0, a1)= cos(a0 � a1)


is clear from the diagram, when the ‘height of taste’ is at the North Pole,then the strength of demand anywhere else depends only on the latitude atthat other point. The diagram shows two points (n0,a1) and (n1,a1) on thesame latitude a1, and it is readily apparent that the demand strength is thesame at both: cos(a0�a1).

This convenient result does not apply when the ‘height of taste’ is not atthe North Pole. Then a more complex formula applies. This asymmetrydoes not imply any imperfection in the sphere, rather it results from a fun-damental asymmetry in the treatment of longitude and latitude in cartog-raphy: latitude is defines between – 90% and �90% while longitude is definedbetween �180% and �180%. Figure 6.7 shows the more general case. Herethe ‘height of taste’ is at (n0,a0) and we wish to compute the strength ofdemand at another point (n1,a1). It is easiest to do this in two stages. First,define another point (n0,a1) which is located on the same longitude as the‘height of taste’ and the same latitude as the other artist. Compute thestrength of the demand signal at (n0,a1) by the perpendicular distance

(n0, a1)

(n0, a0)

(n1, a1)

S(n0, a1)R

Figure 6.7 Signal strength on a sphere

between the ‘radiating plane’ tangent to the sphere at (n0,a0) and this inter-mediate point (n0,a1). The strength of demand at (n0,a1) is given by S(n0,a1).Then, second, compute the perpendicular distance between this intermedi-ate point and the final point (n1,a1). This second distance is defined by R inthe diagram. The strength of the signal at (n1,a1) can then be computed asS(n0,a1)�R.

The first stage is easy. Because (n0,a0) and (n0,a1) are on the same longi-tude, then we can use the result of the previous section to compute thestrength of signal at (n0,a1). It is simply given by cos(a0�a1).

The second stage is a bit harder. First, compute the horizontal distancefrom (n0,a1) and (n1,a1) along the horizontal (and dotted) latitude line. Thiscan be calculated as follows. The horizontal girth of the sphere at latitudea1 is given by 2cos(a1). At the equator (a10), the horizontal girth is at itsmaximum (equal to the diameter of the sphere, which is 2). Away from theequator, the girth is smaller. The horizontal distance between (n0,a1) and(n1,a1) however, is only a proportion of this girth, equal to:

(6.12)

Moreover, this horizontal distance between (n0,a1) and (n1,a1) overstatesthe distance when measured perpendicular to the ‘radiating plane’ (R). Toobtain R, we need to multiply the horizontal distance by cos(a0).

Hence the strength of demand at (n1,a1) when the ‘height of taste’ is at(n0,a0) is defined by:

(6.13)

By putting a090%, we can see that this formula handles the special casewhere the ‘height of taste’ is at the North Pole, and where the strength ofdemand is simply cos(a0�a1).

Once again, it can be shown that this formula for the strength of demandcan also be used to compute the correlation between demand prices of twodifferent artists. The proof is very cumbersome but the basic idea is asfollows. Once again, define two different artists 1 and 2 by the longitude/latitude coordinates: (n1,a1) and (n2,a2). Again, assume that the ‘height oftaste’ could be found with equal probability above any point on the surfaceof the sphere. Then we can show (Swann, 2001b) that the correlationbetween the demand prices of 1 and 2 is given by:

(6.14)corr12 cos (a1 � a2) � cos (a1) cos(a2) · [1 � cos (n1 � n2)]

S(n1, a1) cos(a0 � a1) � cos(a0) cos(a1) · [1 � cos(n0 � n1)]

2cos(a1) · [1 � cos (n0 � n1)]

2 cos (a1) · [1 � cos (n0 � n1)]


Finally, we can – as a special case of the sphere – locate points on a hemi-sphere. This is done simply by limiting the longitude to the range{�$/2,�$/2}.

8 SOME ILLUSTRATIVE RESULTS

The main data for this study come from Reitlinger’s (1961; 1970) studies ofthe auction prices of works of art. In addition, we have generated anapproximate price deflator using data from Mitchell (1980) and Mitchelland Deane (1971). In particular, we have focused on the auction prices ofoil paintings by some of the major artists of the seventeenth, eighteenth,nineteenth and twentieth centuries. We should add, of course, that some ofthe artists commanding high prices in the past have fallen right out offavour in the twentieth century. Indeed, that is one of the phenomena thatthis study has set out to explore. Accordingly, the criterion for inclusion isthat the artist commanded relatively high prices (in real terms) at somepoint of the evolution of the art market between 1800–1970, even if his/herwork is not highly priced now.

Guerzoni (1995) has discussed some of the shortcomings of these data.The raw data are not in an ideal form for econometric analysis. This is noreflection on the accuracy of the data. Indeed, since these are for the mostpart, auction prices, where the agreed sums are recorded, then the data arevery accurate. A few prices are estimated by Reitlinger, but this is not aserious source of error.

Rather, the problems with these data reflect two main factors. The firstis the non-homogeneity of works of art. Clearly, different works areunlikely to be of equal merit. Some are small paintings, others are largecanvasses. While art historians have documented these paintings in greatdetails, we did not (as discussed above) think it practical to attempt toturn these qualitative descriptions into a list of characteristics or to con-struct a hedonic analysis of art prices. At most, we have tried to reducethe degree of variance in art prices by restricting our attention to oil paint-ings, neglecting most watercolours or prints, and neglecting the smallerworks or studies. Also, where two or three paintings were sold as a lot wehave attributed the total price paid between the constituent parts of thebundle.

Second, while the volume of art traded increased markedly in the post-war period, and in particular during the 1960s, this is still a fairly thinmarket. Amongst the great old masters, indeed, few works of significancecame up for auction in recent years. As a result, it has not been practical toinclude a number of old masters in our sample – there is just too little data,


and what sales there are do not represent their great works. Moreover, salesare infrequent and unevenly spaced.

While some econometricians might wish to delve into the peculiar time-series properties of these data, we have chosen instead to perform a verysimple analysis of these data, in three steps. First, we have constructed anaverage price for each artist in each year as follows. We take a 20-yearmoving average covering all the items sold in the last 20 years. Naturally,this smoothes the raw data considerably.

Second, we have deflated these prices by a general price index derivedfrom Mitchell (1980) and Mitchell and Deane (1971). A little splicing wasrequired to make the index continuous back to 1840, and while it is far fromperfect, it is adequate for our immediate purposes. Some have argued thatit would be even more interesting to deflate by an art price index, so thatone can look simply at relativities within the art market, and abstract fromthe secular upwards trend in art prices during (for example) the 1960s.Eventually, we may be able to do that, but it would require a more com-prehensive set of prices that we have analysed so far.

Third, we have computed a simple correlation coefficient between theprice series for each painter. In practice, it is unlikely that a simple correla-tion coefficient is a sufficient statistic for all the analysis we might want todo here. For example, some artists have shown price cycles with relativelylow periodicity, while others have exhibited perhaps one major cycle in100 years or more. Figure 6.8 illustrates this for three artists, chosen moreor less at random, but who do happen to show some very different time-series properties. However, for the preliminary analysis in this chapter, weshall work just with these correlation coefficients using data on the pricesof works by 20 artists (listed in Appendix 1) for up to 130 years.

Econometric Methods

With the correlation matrices computed as described in the previoussection, we have applied the methods of Section 7 to create some prelim-inary artist maps. Of course, as noted above, this is by no means a new pro-cedure since there is a very well-established tradition of using principalcomponents (or multidimensional scaling) to construct such maps.However, the analysis of Section 7 helps to bring out more precisely the wayin which product locations in a characteristics space map into demand pricecorrelations. And it is only with that precision that we can hope to dis-criminate between planar and spherical projections as a means of repre-senting these data.

The basic procedure is straightforward. For any set of coordinates forall the artists in our sample, we can compute the matrix of implied


179

10100

1000

1000

0

100

000 18

4018

6018

8019

0019

2019

4019

6019

80

£

Céz

anne

Can

alet

to

Alm

a-T

adem

a

Sou

rce:

Aut

hor’s

cal

cula

tion

s ba

sed

on d

ata

in R

eitl

inge

r (1

961;

1970

),M

itch

ell a

nd D

eane

(19

71)

and

Mit

chel

l (19

80).

Fig

ure

6.8

Pri

ce o

foi

l pai

ntin

gs –

197

0 pr

ices

,20-

year

mov

ing

aver

age

correlations between them. We can then compare that to the actual andcompute a matrix of residuals. We then add up the absolute values ofthese residuals across each element in the correlation matrix to obtain asum of absolute residuals. The aim of our optimisation algorithm, then,is to choose the set of coordinates which minimises this sum of absoluteresiduals.

The minimand is of course a highly non-linear function of the coord-inates. But we can apply the Lasdon–Waren GRG2 (Generalised ReducedGradient) nonlinear optimisation routine to solve this (Lasdon andWaren, 1978; 1981; Lasdon et al., 1978). The problem can be seen as anexercise in mathematical programming: we have to minimise the sum ofabsolute residuals with upper and lower limits on all the coordinates. Forthe planar representation, we assume that all artist coordinates arebetween 0 and 1 on both axes. For the spherical representation, the lati-tude must be between �$/2 and $/2 while the longitude must be between�$ and $.

Results

The sum of absolute residuals for the 400 elements of the correlationmatrix comes to 23.5 for the planar representation and 17.4 for the spher-ical representation. This could suggest that the spherical representation isslightly preferable, but we are cautious about making such a claim. Theresults described here are preliminary and we cannot be certain that we havereached the global optimum in each case. These statistics correspond to amean absolute error of 4 percentage points in the spherical representationand 6 percentage points in the planar representation. These figures are quiteacceptable.

Figures 6.9 and 6.10 show, respectively, the coordinates obtained for theplanar and spherical representations. They are actually strikingly similar.In Figure 6.9 we have superimposed a dotted circle, within the plane, andit is striking how most artists cluster around this circle. Indeed, this sug-gests that were we to constrain the artists to lie on a circle and not on asphere then that constraint would not do too much damage to the data.By contrast, it is clearly inappropriate to constrain the artists to a (one-dimensional) line. We can also use these charts to interpret the maintrends in the prices of these artists during the period 1840–1970. In thenineteenth century and at the start of the twentieth century, the artists(Landseer, Collins, Meissonier, Alma-Tadema) in the top left-hand part ofthese charts were at the ‘height of taste’. During the early and middle partof the twentieth century, the ‘height of taste’ was moving in an anti-clockwise fashion, with a strong revival in the prices of Nattier and Hals.



The first of the impressionists in the bottom right (Renoir) was also thefirst to enjoy a rapid growth in prices in the early twentieth century, whilethe prices of artists further to the right took off later (van Gogh andBonnard).

In short, the ‘height of taste’ has, over 100 years or so, completed rathermore than half an anticlockwise circuit from top left to bottom right.Where next? The results here probably do not bring out the full advantageof the spherical projection over the planar. Indeed, in the best solutionobtained to date, no painter is located on the back of the sphere: the fronthemisphere fits everybody in comfort. This work is continuing, and whenwe complete computations for a much larger number of artists, it is likelythat the algorithm will need to spread out the artists over a wider area, andsome will locate on the back of the sphere.

0

0.5

1

0 0.5 1

CollinsLandseer

MeissonierAlma-Tadema

Boucher

ConstableNattier

Rembrandt

Hals RenoirDegas

Manet

PissarroCézanne

MonetSisley

van Gogh

BonnardCanaletto

Claude

Figure 6.9 Planar map of 20 artists


9 CONCLUSIONS

The aim of this work is to construct a map of artists and to illustrate howtastes evolve within that map. We have seen that there is an interesting simi-larity between some of the economic models of evolving tastes and whatart historians and art theorists have written about the evolution of art.Moreover, we have seen that the evolution of tastes is clearly a path-dependent process, in the sense of the term defined by Paul David. Ratherthan constructing a characteristics space for works of art, the chapter con-structs an implicit space of painters derived from the correlations betweenprices of different painters’ work. We discuss whether a planar representa-tion or a spherical representation would be preferable. We suggest that themodernist ideal, that new art be divorced from any reference to thepast, requires an unbounded planar representation. We also distinguishbetween two types of historicism: distant past historicism and recent past

Figure 6.10 Spherical map of 20 artists

�$/2

$/2

�$/2 0

$/2

CollinsLandseer

Meissonier

Alma-Tadema

Boucher

Constable

Nattier

Rembrandt

Hals

RenoirDegas

Manet

PissarroCézanne

MonetSisley van Gogh

BonnardCanaletto

Claude

historicism. The former fits more comfortably with a spherical representa-tion while the latter fits best with a planar representation.

The main theoretical results of the chapter are to show exact methodsfor deriving planar and spherical representations of artists from a correl-ation matrix of prices. The spherical representation offers a result of strik-ing power and similarity. For two artists located on a circle, the correlationbetween the prices of their work is shown to equal the cosine of the anglebetween them’ and an analogous result applies to the spherical case. Thesemethods are applied to a representative sample of data on the prices ofwork by 20 artists, taken from the period 1840–1970. The maps generatedshow these painters arrayed over three-quarters of a hemisphere, and the‘height of taste’ has, over 100 years or so, moved top left to bottom right.

The results presented here are incomplete, but the techniques describedin this chapter can help us to understand the path-dependent nature oftastes in art, and the associated waves in popularity of different artists.

APPENDIX 1 LIST OF ARTISTS ANALYSED IN THISSTUDY

Alma-Tadema, Lawrence: 1836–1913Bonnard, Pierre: 1867–1947Boucher, François: 1704–70Canaletto, Antonio: 1697–1768Cézanne, Paul: 1839–1906Claude Le Lorrain (or Claude Gellée): 1600–82Collins, William: 1788–1847Constable, John: 1776–1837Degas, Edgar: 1839–1917Hals, Frans: 1584–1666Landseer, Edward: 1802–73Manet, Edouard: 1832–83Meissonier, Ernest: 1815–91Monet, Claude: 1840–1926Nattier, Jean-Marc: 1685–1766Pissarro, Camille: 1831–1903Rembrandt van Ryn: 1606–69Renoir, Pierre Auguste: 1841–1919Sisley, Alfred: 1840–99Van Gogh, Vincent: 1853–90


NOTES

1. Nottingham University Business School, Jubilee Campus, Wollaton Road, Nottingham,NG8 1BB, UK. Email: [email protected]. I am grateful to participants inthe conference on New Frontiers in the Economics of Innovation and New Technology,held in honour of Paul David, Turin, May 2000, at a seminar at London Business School,and at the International Schumpeter Society Conference in Manchester, June 2000, andalso to Catherine Beaudry, John Cantwell, Robin Cowan, Stuart MacDonald and JennySwann for helpful discussion about this chapter, but none of these are responsible for anyerrors and idiosyncrasies. Funding from the Social Sciences and Humanities ResearchCouncil of Canada (with R.A. Cowan) is gratefully acknowledged.

2. As White (1999) notes, this paragraph was abbreviated in later editions.3. To explain this, consider the travel analogy. Suppose we take a train journey from London

to Edinburgh via York. Since York is north-west of London and Edinburgh is north-westof York then the distance from London to Edinburgh exceeds the distance from Londonto York. This transitivity works on a local scale, but it does not work on a global scale.Thus, for example, to fly from London to Tokyo is a journey of 9 600 km to the east, andto fly from Tokyo to New York is a further journey of 10 800 km to the east. But this doesnot imply that the distance from London to New York is greater than the distance fromLondon to Tokyo. Indeed, at 5600 km that is the smallest distance of the three. So on thisglobal sphere, distance rankings are not transitive.

REFERENCES

Antonelli, C. (1997), ‘The economics of path dependence in industrial organisa-tion’, International Journal of Industrial Organisation, 15, 643–52.

Ayers, F. (1987), Calculus, 2nd edition, New York: Schaum/McGraw-Hill.Bacharach, M. (1990), ‘Commodities, language and desire’, Journal of Philosophy,

87 (7), 346–68.Barnes, R. (1998), The Pre-Raphaelites and Their World, London, Tate Gallery.Baumol, W.J. (1986), ‘Unnatural value: or art investment as floating crap game’,

American Economic Review, 76 (2), 10–14.Becker, G.S. (1996), Accounting for Tastes, Cambridge, MA, Harvard University

Press.Bell, C. (1927), Landmarks in Nineteenth-Century Painting, London, Chatto and

Windus.Bell, Q. (1984), ‘The Pre-Raphaelites and their critics’, in L. Parris (ed.), Pre-

Raphaelite Papers, London: Tate Gallery, pp. 11–22.Bianchi, M. (ed.) (1998), The Active Consumer: Novelty and Surprise in Consumer

Choice, London and New York: Routledge.Bourdieu, P. (1984), Distinction: A Social Critique of the Judgement of Taste,

London, Routledge and Kegan Paul.Chanel, O., L-A. Gerard-Varet and V. Ginsburgh (1996), ‘The relevance of

hedonic price indices: the case of paintings’, Journal of Cultural Economics,20, 1–24.

Cowan, R.A., W. Cowan and G.M.P. Swann (1997), ‘A model of demand with inter-action between consumers’, International Journal of Industrial Organisation, 15,711–32.

Cowan, R.A., W. Cowan and G.M.P. Swann (2004), ‘Waves in consumption withinterdependence between consumers’, Canadian Journal of Economics, 37 (1).


Cozzi, G. (1998), ‘Culture as a bubble’, Journal of Political Economy, 106, 376–94.David, P.A. (1985), ‘CLIO and the economics of QWERTY’, American Economic

Review, 75, 332–36.David, P.A. (1986), ‘Understanding the economics of QWERTY: the necessity of

history’, in W.N. Parker (ed.), Economic History and the Modern Economist,London, Basil Blackwell.

David, P.A. (1987), ‘Some new standards for the economics of standardization inthe information age’, in P. Dasgupta and P.L. Stoneman (eds), The Economics ofTechnology Policy, London: Cambridge University Press.

David, P.A. (1988), Path-Dependence: Putting the Past into the Future of Economics,Technical Report 533, Institute for Mathematical Studies in the Social Sciences,Stanford University.

David, P.A. (1992), ‘Heros, herds and hysteresis in technological history: ThomasEdison and the “Battle of the systems” reconsidered’, Industrial and CorporateChange, 1 (1), 129–80.

David, P.A. (1993), ‘Path dependence and predictability in dynamic systems withlocal network externalities: a paradigm for historical economics’, in D. Foray andC. Freeman (eds), Technology and The Wealth of Nations: The Dynamics ofConstructed Advantage, London, Pinter.

David, P.A. (1994), ‘Why are institutions the “carriers of history”? Path-dependenceand the evolution of conventions, organizations and institutions’, StructuralChange and Economic Dynamics, 5 (2), 205–20.

David, P.A. (1997), Path Dependence and the Quest for Historical Economics. OneMore Chorus of the Ballad of QWERTY, Discussion Papers in Economic andSocial History, Number 20, University of Oxford.

David, P.A. and D. Foray (1994), ‘Percolation structures, Markov random fields andthe economics of EDI standards diffusion’, in G. Pogorel (ed.), GlobalTelecommunications Strategies and Technological Change, Amsterdam, ElsevierScience.

David, P.A., D. Foray and J.-M. Dalle (1998), ‘Marshallian externalities and theemergence and spatial stability of technological enclaves’, Economics ofInnovation and New Technology, 6 (2/3), 147–82.

De Marchi, N. (1999), ‘Introduction’, in N. De Marchi and C.D.W. Goodwin (eds),Economic Engagements with Art, Durham, NC, and London, Duke UniversityPress, pp. 1–30.

De Marchi, N. and C.D.W. Goodwin (eds) (1999), Economic Engagements with Art,Durham, NC, and London: Duke University Press.

De Piles, R. (1708), Cours de peinture par principes, reprinted as English edition(1743), The Principles of Painting, London.

Dosi, G., R. Aversi, G. Fagiolo, M. Meacci, and C. Olivetti (1999), ‘Demanddynamics with socially evolving preferences’, in S.C. Dow and P.E. Earl (eds),Economic Organisation and Economic Knowledge: Essays in Honour of BrianLoasby, Cheltenham, UK and Lyme, USA: Edward Elgar.

Frank, R.H. (1985), Choosing the Right Pond: Human Behaviour and the Quest forStatus, New York, Oxford University Press.

Frey, B.S. (1997), ‘Art markets and economics: introduction’, Journal of CulturalEconomics, 21, 165–73.

Frey, B.S. (2000), Art and Economics, Heidelberg: Springer-Verlag.Frey, B.S. and W.W. Pommerehne (1989), Muses and Markets: Explorations in the

Economics of the Arts, Oxford, Basil Blackwell.


Ginsburgh, V.A. and P.-M. Menger (eds) (1996), Economics of the Arts: SelectedEssays, Amsterdam, North-Holland.

Gombrich, E.H. (1963a), ‘The vogue of abstract art’, in E.H. Gombrich (ed.),Meditations on a Hobby Morse and Other Essays on the Theory of Art, Londonand New York, Phaidon Press.

Gombrich, E.H. (1963b), ‘André Malraux and the crisis of expressionism’, inE.H. Gombrich (ed.), Meditations on a Hobby Morse and Other Essays on theTheory of Art, London and New York, Phaidon Press.

Gorman, W.M. (1980), ‘A possible procedure for analysing quality differentials inthe egg market’, Review of Economic Studies, 47 (5), 843–56.

Grampp, W.D. (1989), Pricing the Priceless: Art, Artists and Economics, New York,Basic Books.

Guerzoni, G. (1995), ‘Reflections on historical series of art prices: Reitlinger’s datarevisited’, Journal of Cultural Economics, 19, 251–60.

Ironmonger, D.S. (1972), New Commodities and Consumer Behaviour, Cambridge,Cambridge University Press.

Lancaster, K.J. (1971), Consumer Demand: A New Approach, New York, ColumbiaUniversity Press.

Lasdon, L.S. and A.D. Waren (1978), ‘Generalized reduced gradient software forlinearly and nonlinearly constrained problems’, in H.J. Greenberg (ed.), Designand Implementation of Optimization Software, Leiden, the Netherlands, Sitjhoffand Noordhoff.

Lasdon, L.S. and A.D. Waren (1981), ‘GRG2 – an all FORTRAN general purposenonlinear optimizer’, ACM SIGMAP Bulletin, 30, 10–11.

Lasdon, L.S., A.D. Waren, A. Jain and M. Ratner (1978), ‘Design and testing of ageneralized reduced gradient code for nonlinear programming’, ACMTransactions on Mathematical Software, 4 (1), 34–50.

Malraux, A. (1954), The Voices of Silence, London, Secker and Warburg.Marcus, G.H. (1998), Design in the Fifties, Munich and New York, Prestel-Verlag.McCain, R.A. (1981), ‘Cultivation of taste, catastrophe theory, and the demand for

works of art’, American Economic Review, 71, 332–4.McDermott, C. (1992), Essential Design, London, Bloomsbury.McPherson, M.S. (1987), ‘Changes in tastes’, in J. Eatwell, M. Milgate and

P. Newman (eds), The New Palgrave: A Dictionary of Economics, London,Macmillan, pp. 401–03.

Mitchell, B.R. (1980), European Historical Statistics 1750–1975, second revisededition, London: Macmillan.

Mitchell, B.R. with P. Deane (1971), Abstract of British Historical Statistics,Department of Applied Economics Monograph No. 17, Cambridge, CambridgeUniversity Press.

Reitlinger, G. (1961), The Economics of Taste: The Rise and Fall of Picture Prices,1760–1960, London, Barrie and Rockliff.

Reitlinger, G. (1970), The Economics of Taste, Volume III: The Art Market in the1960s, London, Barrie and Jenkins.

Rheims, M. (1959), La Vie Etrange des Objets, published in English translationby D. Pryce-Jones (1961), Art on the Market, London, Weidenfeld andNicholson.

Ruskin, J. (1843), Modern Painters: Volume I, reprinted in E.T. Cook andA. Wedderburn (1996), The Works of John Ruskin, Library edition on CD-ROM,Cambridge, Cambridge University Press.


Ruskin, J. (1856), Modern Painters: Volume III, reprinted in E.T. Cook andA. Wedderburn (1996), The Works of John Ruskin, Library edition on CD-ROM,Cambridge, Cambridge University Press.

Swann, G.M.P. (1999), ‘Marshall’s consumer as an innovator’, in S. Dow and P. Earl(eds), Economic Organisation and Economic Knowledge: Essays in Honour ofBrian Loasby, Cheltenham, UK and Lyme, USA, Edward Elgar.

Swann, G.M.P. (2001a), ‘The demand for distinction and the evolution of the pres-tige car’, Journal of Evolutionary Economics, 11 (1), 59–75.

Swann, G.M.P. (2001b), ‘From correlations to implicit spaces: some results for planesand spheres’, unpublished paper, Manchester Business School.

Swann, G.M.P. (2002), ‘There’s more to the economics of consumption than utilitymaximization’, in A. McMeekin, K. Green, M. Tomlinson and V. Walsh (eds),Innovation by Demand: Interdisciplinary Approaches to the Study of Demand andits Role in Innovation, Manchester, Manchester University Press.

Throsby, D. (2001), Economics and Culture, Cambridge, Cambridge UniversityPress.

Veblen, T. (1899), The Theory of the Leisure Class: An Economic Study ofInstitutions, New York: Macmillan.

White, M.V. (1999), ‘Obscure objects of desire? Nineteenth-century British econo-mists and the price(s) of “rare art” ’, in N. De Marchi and C.D.W. Goodwin (eds),Economic Engagements with Art, Durham, NC, and London, Duke UniversityPress, pp. 57–84.

Wildenstein, D. (1996), Monet Volumes I–IV, Köln: Taschen Verlag.


7. Waves and cycles: explorations inthe pure theory of price for fine art*

Robin Cowan

1 INTRODUCTION

The popularity of painters rises and falls. The Impressionists were scornedin the mid-1800s, became the mode a few decades later, largely disappearedfrom view early in the twentieth century and stormed back to popularity atthe end of that century. Their history illustrates two interesting phenom-ena. First is the rise and fall in popularity of artists. Artists become themode; their prices rise, and galleries, critics and the public alike praise themas expressing the spirit of the times. But later they fade from view and arereplaced by others. Second is that artists often come and go as a group.While there may be the ultimate Impressionist (some say that Manet was‘the finest painter of all the Impressionists’ [Janson and Janson, 1997: 722),as his popularity rises and falls so does that of other ‘similar’ artists.Whatever is driving the popularity of Manet, drives also the popularity ofMonet and Renoir. This grouping together of painters acts as a type of con-formity effect in art fashion – if Van Gogh is popular, artists whose work issimilar to his in the appropriate respects will also be popular, whereaspainters whose work is different from his will tend to be unpopular. If onecan imagine painters located in a space, with their popularity representedas a distribution over that space, then the modal painter will be surroundedby similarly popular painters. The mode, though, moves through the spaceover time.

There is another striking feature of the market for fine art. This is thepresence of ‘avant-garde’ consumers. These are the fashion-setters who areunwilling to have yesterday’s heroes on their walls. This need not soundquite so snobbish. One explanation for the behaviour of the avant-garde isthat their search for new painters is driven by their need to express newideas and concerns. Currently popular art will express today’s (or yester-day’s) concerns and ideas, but will in all likelihood not express tomorrow’s.Avant-garde consumers perform the valuable role of finding the modes ofexpression for emerging concerns. While the conformity effect of painting

188

‘schools’ creates inertia in fashion, the presence of avant-garde consumerscreates motion. These consumers will create popularity for painters situ-ated far from the modal painter in painting space.

These two features of the art market suggest the possibility of waves inpopularity. Groups of painters become popular together, though one maystand out, but their popularity fades as avant-garde consumers look fornew painters. The choices of these fashion leaders are thereafter taken upby less forward-looking consumers, which contributes to a shift in overalldemand within the market. This is the phenomenon at the centre of thischapter. Our concern here is to model it in a way that sheds light on the pat-terns that can emerge from this interplay of conformity and distinction.

The world of fine art is not the only one in which waves of fashion comeand go. While the consumers of textbook theory have preferences inde-pendent of each other, this seems restrictive when considering the widevariety of both behaviour and type within the population of agents referredto as consumers. And indeed, a more general view, namely, that there areexternalities in consumption or that the utility of an agent will depend inpart on what other particular agents are consuming (whether there arephysical spillovers such as pollution or not), is not new in economics. Itstretches back to Smith, who claimed that the ‘the chief enjoyment ofriches consists in the parade of riches’. Recently, the idea has been taken upby Becker (1996) and Akerlof (1997) each in his own way.

One key aspect of externality in consumption, and the one emphasisedby sociologists, is distinction.1 The idea here is that individuals gain utilityfrom what they perceive to be their relative status in some hierarchy, andthat one way to express, or even to change that status is through consump-tion. Consumption of some things will raise our relative status, and otherswill lower it. Frank (1985) has described some of the economic effects ofthe desire for distinction. Desire for distinction may be a powerful motiva-tor but, nonetheless, for most agents it is important to function within apeer group – after all, from whom would we like to distinguish ourselvesbut our ‘former peers’ – so it is important that some activities create con-formity with at least some part of the population around us. Thus withinany utility function we would expect to see two (possibly occasionally con-flicting) forms of non-independence: desire for distinction and desire forconformity.

The effects of conformity and distinction have been addressed in ageneral way by Cowan et al. (1997). That model forms the background formuch of the work presented in this chapter.

Curiously, given the occasion for which this chapter was written, themodel developed here shows no path dependence. It is a dynamic model,but the dynamics are deterministic, and the final, very long-run outcome is

Explorations in the pure theory of price for fine art 189

predictable from the start. One place it does pick up ideas from Paul David(ignoring the general idea that some things are better than others) is in thenotion of bandwagons, heroes and herds. Bandwagons of taste form,leaving yesterday’s hero a mere member, or at best icon, of today’s herd.How this comes about is through a form of standardisation – ideas andtheir expression become standardised. A new idea or new form of express-ing something important emerges, to the scorn of the majority of the pop-ulation. It is picked up by those who can see the coming thing, and pushedby them, they perhaps acting as translators, or gateways, between the newunusual idea and the old standard. Slowly, the herd come to see its value, itis absorbed (somewhat twisted perhaps to fit) by the mass of the popula-tion and, before you know it, everyone is doing it. But what of the innov-ator? Is he or she doomed to a future of anonymity? Not if he or she is ableto continue producing new, unusual ideas to address emerging issues. Ifhistory is anything to go by, and especially if it matters, this is not some-thing that need concern us on this occasion.

We turn now to a very stylised model of the art world, in which pricesare determined through the utility that paintings (or, more properly, theworks of painters) provide the consumer. This utility is driven in part byexternalities and, if snobbery is an unappealing trait on which to build amodel, externalities can be seen as an expression of the notion that someart forms or genres are better fitted to the concerns of the day. Avant-gardeconsumers are looking at tomorrow’s ideas, and they pull the market withthem in their search for new means of expression or needs to express newideas. From these simple externalities we can derive rich dynamic patternsin which prices rise and fall as painters come into and go out of fashion.

2 A SIMPLE MODEL OF ART PRICE DYNAMICS

In this model we examine demand for painters’ outputs. That is, we con-sider a painter to be a brand name, associated with which there is a fixedsupply of the commodity. We acknowledge the fact that painters producenothing after they have died, but assume also that none of their works dis-appears. We consider the median consumer as representative of the marketdemand. While this is acknowledged to be problematic in some cases (seeKirman, 1992) this problem is alleviated somewhat in the heuristic whereinwe treat the consumer’s demand as distributed over painters. This can beseen as capturing the notion that the median consumer represents a popu-lation of heterogeneous consumers whose demand will be distributed. Themodel is developed in the usual way as a consumer maximisation problem.The assumption of a fixed supply of every good immediately produces the


equilibrium quantities consumed. What is of interest, however, rather thanquantity is price. Prices change to equilibrate the market for each painter.This is not the trivial problem it seems, due to the presence of externalitiesin consumption – the popularity of a painter, indicated by a high price forhis works, will affect the utility gained from consumption of other like, andunlike, painters. In addition, price adjustment is temporal in the sense that,because not every painting is auctioned every period (neither in the modelnor in the actual art market), price adjustments are necessarily partial.

There are two types of goods – paintings and other goods. Other goodsare aggregated into the good Z, whereas paintings remain branded bypainter. Define A as the set of painters: A{a � [1, N ]}. The consumer’sutility function is additively separable, written as U (X1,X2,. . .,Z )&AUa(Xa)�Z where Z is the aggregate bundle of other goods, and Xa rep-resents consumption of the paintings of artist a.

Normalise by setting the price of Z to 1. Utility maximisation under afixed budget yields a first-order condition:

where p(a) is the price of the work of artist a.2

Separating a from the other artists:

(7.1)

The first term is simply the marginal utility of consuming paintings bypainter a; the sum represents the effects of consumption of painter b on theutility gained from consuming painter a. Assume now that d2Ua/dXbdXc0∀c�{a, b}. If the number of painters is large, this permits us to approxi-mate equation (7.1) as

(7.2)

where g(a)dUa/dXa| . We can suppress the Xa argument since by assump-tion there is a fixed number of paintings per painter, X, and the marketclears through price adjustments. The term g(a) can be seen as representingthe inherent value of artist a, regardless of his current (un)popularity. Someartists are just simply better than others.

The integral contains externality effects: consumption of other paintersaffects the marginal utility of the consumption of a. We decompose F(a, b)into a product: 1/�f (b�a)p(b). The first element represents the strength ofthe externality as determined by the distance between the two painters inquestion – made up of the conformity and avant-garde effects; the second

X

p(a) g(a) � �b�A

F(b, a)db,

p(a) dUa �dXa � �

b�adUb

�dXa.

dU�dXa �b�A

dUb �dXa p(a).


is the standing price of the second painter, which is a measure of the extentto which he is consumed, or how popular he is (recalling that supply isfixed). To ease the presentation that follows we do a small violence to thenotation and will treat b as indicating ‘the painter at distance b�a from a’.This allows us to write f (.) as a function of b alone: f (b).

Equation (7.2) describes the equilibrium price vector for the set of artistsA.3 The nature of the art market, however, makes the disequilibriumprocess extremely important. Many out of equilibrium trades take place, inpart because for any painter, only a small proportion of his work is tradedin any period. This is the nature of the art auction market. To capture thiseffect we assume that the work of every artist is trade at the rate �. That is,speaking in discrete time, each period, �'100 per cent of every artist’sworks come up for auction. The price of paintings actually traded is set byequation (7.2). Consider now the ‘standing’ price for work by a particularartist, a. Exposition is more transparent in discrete time:

(7.3)

or,

(7.4)

Writing now in continuous time,

This dynamic structure has been studied in other contexts (see forexample Cowan et al., 1997). A solution is generated by doing a Laplacetransform on the time variable, and a Fourier transform on the a variable(see the appendix in Cowan et al., 1997). Drawing on that work we can statethree propositions:

Proposition 1 If the system is convergent, the steady state is described bythe limit:

where we use the definitions P(k, t)and G(k)

This proposition states that the ‘natural prices’, namely, those that wouldprevail if behaviour were based solely on the inherent worth of the artist

� ��e�ikag(a)da.� �

��e�ikbf(b)db��

e�ikap(a, t) da, F(k)

limt→�

P (k,t) �G(k)

� � F(k)

dp(a)dt

�g(a) � p(a, t) ��b�A

f (b)p(b)db.

p(a, t � 1) � p(a,t) ��p(a, t) � �g (a) ��b�A

f(b)p(b)db.

p(a, t � 1) (1 � �)p(a, t) � �g(a) ��b�A

f(b)p(b)db,


and not on any considerations of externalities, are, in equilibrium,‘distorted’ by any externalities that exist.

We can retrieve the dependence of p on a by the following transforma-tion: p(a, t) In the absence of specified functions,the form of p(a, t) is less easy to interpret than P(k, t). Consider an arbi-trary painter a. P(k, t) measures the extent to which a painter of distance$/k from painter a affects the price (or equivalently the popularity) ofpainter a. Thus large values of k are associated with effects of nearbypainters, small values of k are associated with effects of distant painters.

The next two propositions concern the dynamics of this market, andwhether waves of popularity will be observed.Proposition 2 If f(b) is an even function ( f(�b)f(b)) then the dynamicsare strictly diffusive. That is, the initial state decays and the final equilib-rium state builds up exponentially.Proposition 3 If f(b) is an odd function ( f (�b)�f(b)) then the dynamicbehaviour is captured by travelling waves.

These two propositions follow from basic properties of Fourier transfor-mations. For a discussion of these results see Cowan et al. (1997).

Remark 1 Any well-defined function can be expressed uniquely as the sumof even and odd functions.4 When f(b) contains both even and odd elementsnon-trivially, price dynamics will be a sum of the dynamics described inpropositions 2 and 3. The quantitative features will depend on the detailedforms of the even and odd parts of f(b).

3 PURE FAD

There is almost certainly a strong ‘fad’ component to taste in art. That is,the externalities represented by the sum in the utility function (equation[7.1]) are a major source of utility from art consumption. We can considertwo extreme cases. In the first case, art per se has no inherent value to theconsumer. Formally, dUa/dXag(a)0. In the second case, there is inher-ent value to the consumption of art, but no painter is better or worse thanany other as a direct source of utility: for all a, g(a)c where c is a constant.

Proposition 4 If g(a)0 then if the system is convergent the limiting priceis p(a)0 for all a.

From Proposition 1, the limiting price function is described by the limitof its transform If g(a)0 then G(k)0limt→�P(k, t) �G(k) �(� � F(k)).

� ��1�(2$)eikaP(k, t) dk.


by definition of the transform. Thus the inverse transform of P(k, t) is alsoequal to 0: for all a.

Proposition 5 If g(a)c then if the system is convergent

If g(a)c then G(k)2c$�(k) where �(k) is the Dirac delta function.Substituting into the limit from Proposition 1:

Since �(k)0 this reduces to

By definition, F(k) , and substitutionyields the propostion.

Corollary If g(a)c and then there is a positive price p*such that p(a)p*

This can be shown be checking consistency. Substitute p* for p(a):

Define q1/p*. Then

which has real roots (and therefore at least one positive root), if and only if�/(4c)�

Propositions 4 and 5 give insight into the stability of demand over longperiods. The fact that prices have not gone to zero, even for painters longsince dead, suggests that there is some inherent utility to be gained fromconsuming art. On the other hand, the fact that there is variation in theprices fetched by different painters implies one (or both) of two things:

�F(b)db.

qc � q � 1��F(b)db 0,

1�p* � � p*�F(b)db

�c.

p* �c

� � ��F(b)p*db

.

(a � A.� �(4c) ��F(b)db

��e�ikbf (b)db so F(0) ��

�� f (b)db

limt→�

P(k, t) �2c$

2$

1� � F(0)

.

(k � 0,

limt→�

P(k, t) �2c$�(k) �(� � F(k)).

limt→�

p(a, t) �c

� � ��

� f (b)db

.

limt→� p(a, t) 0


either not all painters are equal in the eyes of the median consumer, or thereare strong ‘fad’ elements in the consumption of art works, and the marketcontinues to exhibit out-of-equilibrium behaviour. We have not analysedthe case in which some painters are ‘better’ than others – analytical resultsare extremely difficult to generate. This is a situation we explore below,however, by simulating the model.

4 CONVERGENCE

One concern that always exists with dynamic systems is whether or not theycoverge. Steady-state results given above were conditional on the systemconverging. In this section we explore the conditions for convergence.

From the solution in the appendix in Cowan et al. (1997), in terms of theconjugate variables k and z the dynamics of the system are defined by

(Ibid.A4)

Thus convergence turns on the final term: . If, for some k,then the system is divergent. Recall that F(k) is the transform of

the externalities in consumption, defined as wheref(b) is the sum of the conformity and avant-garde effects. Convergence isdetermined in a critical way by the functional forms of these externalities.

Assume that the conformity effect is even and the avant-garde effect is odd.We can then write the total, net externality effect as f(b)fc(b)�fa(b) wherefc(�b)fc(b), and fa(�b)�fa(b). In passing we can note that fromRemark 1 the dynamics in this case will exhibit both waves in prices and asecular trend toward the final price distribution. Fourier transforms arelinear, so Because fc(b) is even andfa(b) isodd, this simplifies to

To illustrate, suppose that the two externality effects are each a memberof families of functions: fc(b)C1 f1(b) and fa(b)C2 f2(b). In this case thetransforms become Substituting the functional familydescription of fc and fa we get

The Dirichlet conditions (which have been assumed to hold) ensure that theintegrals are finite. Thus C1 and C2 are scaling parameters that will, jointlywith �, determine the sign of The stronger are the externalities,the faster the system diverges; similarly the smaller is �, that is, the smaller

� � F(k).

F(k) C1��

��

2 cos(bk) f1(b)db � C2��

��

� 2i sin f2(b)db.

C1F1(k) and C2 F2(k).

F(k) � ��2 cos(bk) fc(b)db � � �

��2i sin fa(b)db.F(k) � �

��e�ikbfc(b)db � � ��e�ikbfa(b)db.

F(k) ��e�ikbf (b)db

� � F(k) � 0e�(��F(k))

P(k,t) �G(a)

� � F(k) �

�G(a) � (� � F(k))P(k,0)� � F(k)

e�(��F(k))t.


the proportion of works that come onto the market each period, the fasterthe divergence. If externalities are weak enough, or enough works come onto the market each period, then the system converges.

5 SIMULATION OF THE MODEL

To illustrate some of the results, showing both dynamics and long-runproperties, we simulate the model of art prices developed above. An initialproblem is the space in which painters exist. In a Lancasterian world, paint-ings obviously exist in a very high dimensional space – they have manytypes of qualities that can vary from one painter to another. To what extentthe dimension of the space can be reduced without doing extreme violenceis as yet unknown. Peter Swann explores this issue in his contribution tothese proceedings. He finds that it is indeed possible to reduce the dimen-sion dramatically. Exactly how far this can be done, and the nature of thespace itself (linear, periodic, spherical or toroidal and so on), is somethinghe is currently exploring I believe. Because these details, which are essen-tially empirical matters, are still being explored, we can use this opportun-ity to implement the model in a variety of spaces.

We create a world of 600 painters, each at a fixed location in space.Proximity in this space indicates that painters (or more properly theiroutputs) are ‘like’ each other in the eyes of their beholders. Consumers gainutility from consuming artworks, and there are externalities in that con-sumption. Painters, and schools of painters become fashionable, that is, aconsumer will gain utility from owning the works of a painter who is likeother painters who are widely appreciated. This is the conformity effect,though it could be described as the effect of having concerns and ideas incommon with other consumers. This we model as an even function. On theother hand, there is an avant-garde effect. There are some consumers whoare ahead of the pack in terms of ideas and concerns, and are thus lookingfor ‘new’ painters to express those concerns. In general, this effect is to seek‘unfashionable’ parts of the space in which painters reside. There is,though, a difference between yesterday and tomorrow, so this is modelledas an odd function. It is this effect that gives direction (rather than justchange) to art prices.

5.1 One Dimension

The analytic results developed above were in the context of a one dimen-sional linear space. They apply directly to a (one-dimensional) periodicspace, making allowances for the fact that in this space if a wave travels



forever in one direction, it is actually travelling around a circle. One caveatis that the externality functions must go to zero (in distance) at a distanceless than half the circumference of the circle. This seems entirely reasonablein this case. For completeness we explore both linear and periodic spaces.

5.1.1 Linear spaceSix hundred painters are arrayed along the whole line. Every period 1 percent of each painters paintings comes up for auction. Prices of those paint-ings are set as in equation (7.2). Two types of externalities exist: the con-formity effect is modelled as f1(b)c/|b|. The avant-garde effect is modelledas f2(b)sgn (b)a(1�1/|b|). The dynamic pattern can be seen in Figure 7.1.This figure should be read as a relief map. Darker colours indicate higherprices, and thus higher popularity. What can be observed here is a mainwave of popularity. Initially the painter located at 550 on the horizontal

0

50

100

150

200

250

300

0 100 200 300 400 500 600

Tim

e

Painters

Figure 7.1 Prices of painters over time: linear space

axis is the painter of the day. His popularity fades, though, and popularitymoves to painters to his left. This is driven by the avant-garde effect. Whatis noteworthy, however, is that painters ‘reappear’. There is a second andthird wave – an echo if you will, in which prices rise again, to be followedby a decline. What is interesting here is that this occurs without the intro-duction of new painters, which is clearly one source of new waves in theactual world of art and artists. Here the simple dynamics are alone enoughto produce a main wave and subsequent resurgence of formerly popularpainters.

5.1.2 Periodic spaceThe interpretation of the dimensions of ‘painter-space’ is unclear. Fine arthas a variety of attributes; precisely which of them should appear as majoraxes is unclear. Further, some of these properties seem naturally modelledas extending indefinitely, while others seem naturally periodic (consider acolour wheel for example). In this section we implement the same simula-tion but in a periodic space. Painters are arrayed around a circle rather thanon a line. This is consistent with some of the empirical results of PeterSwann. The periodicity of the space permits a painter to reappear if thereare waves in consumption. A wave travels in one direction, and eventuallyreturns to its original position, travelling round and round. Again, cyclesemerge without the introduction of new painters. Figure 7.2 shows a typicalpattern, using the same representation as Figure 7.1. Darker grey indicateshigher popularity (and prices). Notice that with these parameters prices aresecularly increasing, as the waves shown by the diagonal patches get darkerand darker as time passes.

5.1.3 Inherently good paintersIn the previous illustrations of the model, no painter was any better, inher-ently, than any other. Thus the dynamics were driven purely by external-ities or fashion effects. It may be, though, that some painters or schools areinherently better than others in providing consumers with higher utilityregardless of fashion. The analytic results included this aspect; here weillustrate it. Figure 7.3 shows the same dynamics as Figure 7.2 (a periodicspace with secularly increasing prices) but with the painter located at pos-ition 100 having inherent value. His inherent value spreads a short distanceto include those nearby above and below.5 As can be seen in the figure, thegradual darkening of the graph as time passes indicates general increasesin prices. Repeated waves occur – the diagonal stripes running north-west/south-east. We can see though, that relative to Figure 7.2, those wavesare distorted by the price of painter 100. His inherent, persistent valuecreates a vacuum of low prices to his right, caused by the avant-garde



shunning him and those like him. (Reading horizontally from left to rightjust to the right of 100 the graph becomes lighter ‘than it should be’ rela-tive to the overriding wave pattern). This effect gains in extent in that astime passes more and more painters are affected. This can be seen by thechanging shape of the wave – the stripes are not of uniform width, andhave odd patterns emerging between the waves. Similar effects exist in theone-dimensional linear space.

5.2 Two Dimensions

Reducing the dimension to one is a dramatic simplification of the space inwhich fine art must exist. Nonetheless, it does lend insights into cycles inpricing. A higher dimension space adds realism, but raises an in principleproblem: how does the avant-garde effect operate? In a single dimension

0

200

400

600

800

1000

1200

1400

1600

0 100 200 300 400 500 600

Tim

e

Painters

Figure 7.2 Prices of painters over time: periodic space


this was relatively simple, in that the temporal aspect of avant-garde-ismhad to be equated with a spatial aspect, and this determined the directionof motion. In higher dimensions, however, at any time the avant-gardeeffect operates, it could be pulling the art world in more than one direction,depending on what the avant-garde consumer (or artist) is attempting toexpress. To assume at the outset that this motion always occurs in the samedirection (the avant-garde effect favours painters to the right, as in theone-dimensional case, for example) will simply reproduce the effects of theone-dimensional model, the other dimensions not interacting very muchwith the motion. A more interesting, and probably more realistic, notion isthat from time to time one point, currently ‘ignored’ in the space becomesvery popular. The avant-garde effect would be to pull popularity from thecurrent mode towards and beyond this newly popular artist. This implies

0

200

400

600

800

1000

1200

1400

1600

0 100 200 300 400 500 600

Tim

e

Painters

Figure 7.3 Prices of painters over time: periodic space with one inherentlyvaluable painter

that popularity would follow a more random path around the space, buttending to move from ‘old’ to ‘new’ ideas.

Ignoring the possibility that new painters enter, thereby disturbingexiting market structures and dynamics, the same dynamic patternsobserved in the single-dimension cases are observed in two dimensions,whether the space is a plane or a torus. There are waves of popularity aspainters’ prices rise, fall, and then rise again as the fashion is pulled by theavant-garde away from the currently popular painters. Using the sameframework it is possible to introduce entirely new fashions, by creating asmall island of high prices in a part of the space that is currently out offashion. Doing this in an ad hoc way indicates that the motion of prices isas expected. The island of high prices creates immediate conformity anddistinction effects, and changes the direction in which art fashion wasmoving. It is difficult to perform this experiment in a systematic way (andit is equally difficult to present the results graphically and concisely), so weleave that discussion at this point as part of the agenda for future work.

5.3 Planes and Circles

One of the fascinating results of Peter Swann’s contribution to these pro-ceedings is that even when painters are constrained to exist in a two- orthree-dimensional space, neither the planar nor the spherical projection isfull. In fact, in both cases the painters are (statistically) located around acircle. This suggests that the one-dimensional periodic space may be a goodrepresentation in which to analyse the dynamics of prices or fashion in thefine art world. Swann’s analytic results show that in this representation cor-relations of prices are proportional to the cosine of the angle between thepainters. Thus a test of the suitability of modelling a high dimensionalspace as a circle is whether the dynamics of the model can be parametrisedin such a way that the price correlations satisfy the cosine relationship.Figure 7.4 is a scatter plot of price correlations versus the cosine of theangle between painters on the circle. In this implementation there are 600painters, avant-garde and conformity effects that increase with distance,but are truncated at a distance of 25.6 As is clear in Figure 7.4, this para-metrisation comes very close to satisfying the cosine relationship exactly. Itdeparts from this relationship when the cosine is near �1, that is, betweenpainters who are located on opposite sides of the circle. There are threeexplanatory factors: (1) here the externality effects, which are key to deter-mining price relationships, are weakest; (2) in general the model places nobounds on prices, but for reasons of realism prices have been boundedbelow by zero, which will interfere with the inter-painter price relationships,and will ‘distort’ this relationship most for distant painters; and (3) under



the parameters used in this implementation there is a secular increase inprices which is a source of positive correlation among prices of all painters,which pulls the correlation ‘up’ most where the pattern created by exter-nalities is weakest.

6 DISCUSSION

The rise and fall of painters’ popularity is striking over a period of severalhundred years. This is a phenomenon that has been felt intuitively by manyobservers of modern culture, but is also evident from the hard economicmeasure, namely, prices. It is a common place that the art world goesthrough fads and fashions, so a tempting explanation for the rise and fall of

Figure 7.4 Correlations between prices versus cosine of the angle betweenpainters

�0.8

�0.6

�0.4

�0.2

0

0.2

0.4

0.6

0.8

1

�1 �0.8 �0.6 �0.4 �0.2 0 0.2 0.4 0.6 0.8 1

Pric

e co

rrel

atio

n be

twee

n pa

inte

rs

Cosine between painters on the circle

a painter is something like sunspots. But to explain the rise and fall of groupsof painters implies that there are qualitative similarities between themembers of any group. This suggests that something more stable is goingon (unless one simply invokes sunspots that operate on groups rather thanon individuals, but this begs the question why members of a group fetchdifferent prices from each other). An economist is tempted to point to stablepreferences at least in explaining the grouping phenomenon. But stablepreferences can be used for a more general explanation, provided one iswilling to abandon inter-agent independence in preferences. That has beenthe approach here. Stable preference in a stable world are enough to gener-ate waves in which painters emerge, disappear and emerge again. And a verysimple model has produced a rich set of dynamics which can be treated ana-lytically and numerically to help understand the dynamics of the art market.

One of the observations made by Swann regarding the price time seriesis that there appear to be cycles of different period. This can in principle bereproduced in this simple model, by introducing more complex externalityfunctions. Every function used here was monotonic. But a non-monotonicfunction, with more than one maximum or minimum will produce strongreactions at wherever there is an optimum. If there are optima at severaldistances, this implies that waves of different frequencies will form.

The stability of the world of the model, in the sense that no new paintersare suddenly appearing to upset existing dynamic patterns, is both astrength and a weakness. It shows the power of externalities in creatinginteresting dynamics, but it departs somewhat from reality. Introducingnew painters in a non ad hoc way remains a research challenge.

NOTES

* This chapter was written for the conference New Frontiers in the Economics ofInnovation and New Technology, held in honour of Paul David at the Accademia delleScienze, Torino, 20–21 May 2000. The chapter has benefited from the comments of theparticipants in that conference and especially from ongoing discussions with WilliamCowan and Peter Swann. Funding from the Social Sciences and Humanities ResearchCouncil of Canada is gratefully acknowledged.

1. See most notably Bourdieu (1984).2. Because we are interested in the evolution of prices of all the painters, rather than using

the common pa notation we treat price as a function of the painter, p(a).3. The price on the left-hand side of equation (7.2) is the ‘standing price’, that is, the average

prices most recently paid for the entire oeuvre of an artist. This introduces some historyinto the notion of popularity, which seems appropriate in this context. The market for fineart, at least for the painters that survive the test of time is indeed faddish to some degree,but nonetheless it has a strong sense of value and history. History is gradually eliminated(and replaced) however, as the entire work of an artist comes to auction over time, andmust face competition from other artists and in other eras.


4. Formally, any function satisfying the Dirichlet conditions can be so written. The Dirichletconditions are that the function is square integrable, single-valued, piece-wise continuousand bounded above and below.

5. Specifically there is a scaled normal distribution with variance 10, and mean 100 describ-ing inherent value.

6. Because the circle is translation invariant, many pairs of painters are separated by thesame angle – painters 0 and 10 are separated by the same angle as 1 and 11, and so on.The data in the figure are not averages for an angle but rather are the correlations betweeneach pair of painters plotted against the cosine of the angle between them.

BIBLIOGRAPHY

Akerlof, G. (1997), ‘Social distance and social decisions’, Econometrica, 65 (5),1005–27.

Becker, G.S. (1996), Accounting for Tastes, Cambridge, MA: Harvard UniversityPress.

Bourdieu, P. (1984), Distinction: A Social Critique of the Judgement of Taste,London, Routledge and Kegan Paul.

Cowan, Robin, William Cowan and Peter Swann, (1997), ‘A model of demand withinteraction among consumers’, International Journal of Industrial Organisation,15, 711–32.

David, P.A. (1985), ‘CLIO and the economics of QWERTY’, American EconomicReview, 75, 332–36.

David, P.A. (1992), ‘Heroes, herds and hysteresis in technological history’,Industrial and Corporate Change, 1 (1), 129–80.

David, P.A. (1997), ‘The economics of path dependence in industrial organization’,International Journal of Industrial Organisation, 15, 643–52.

Frank, R. (1985), Choosing the Right Pond: Human Behavior and the Quest forStatus, New York, Oxford University Press.

Janson, H.W. and Anthony F. Janson, (1997), History of Art, 5th edition,New York, Prentice-Hall.

Kirman, A. (1992), ‘Whom or what does the representative individual represent?’,Journal of Economic Perspectives, 6 (2), 117–36.

Smith, A. (1937), The Wealth of Nations, New York, Random House.Swann, G.M.P. (2000), ‘Is the world flat or round? Mapping changes in the taste for

art’, MERIT-Infonomics Research Memorandum series #2001–009.


PART III

The Economics of Knowledge

8. Learning in the knowledge-basedeconomy: the future as viewed fromthe pastW. Edward Steinmueller

The pervasiveness of economies of scale opens up the prospect that past marketconfigurations, which neoclassical theory tempts one to interpret as globallystable equilibria, were in reality unstable positions away from which the systemmoved when disturbed by shifts in demand. The added presence of ‘learning’effects in production (and the implied suggestion that they may have also beenpresent in consumption, in the form of habituation or endogenous taste refor-mation) introduces a degree of irreversibility in the ensuing market adjustmentsof relative costs and prices. As a result of this previous economic configurationsbecome irrevocably lost, and in trying to work backwards by entertaining coun-terfactual variations on the present, one cannot hope to exhibit the workings ofhistorical process. [. . .]

Under such conditions, market divergences between ultimate outcomes mayflow from seemingly negligible differences in remote beginnings. There is noreason to suppose that dynamic processes are ergodic, in the sense of ultimatelyshaking free of hysteresis effects and converging from dispersed initial positionstowards a pre-determined steady state. To understand the process of moderneconomic growth and technological development in such an untidy world neces-sarily calls for the study of history. For, change itself ceases to be mere loco-motion. Economic growth takes on an essentially historical character, and theshape of the future may be presumed to bear the heavy impress of the past.(David, 1975: 15–16)

1 INTRODUCTION

Since this early contribution of Paul A. David, many economists, from avariety of perspectives and using a variety of methodologies, have come toshare the viewpoint that economic change involves processes of organisa-tional and individual learning. This chapter examines the relationshipbetween individual and organisational learning as it bears upon the accu-mulation of knowledge (the result of learning) from a historical perspectiveand the recent opportunities afforded by information and communication

207

technologies for changes in the ‘technology of learning’ (the means forreproducing and exchanging knowledge).

1.1 A Closer Look at Learning

Marking a distinction between individual and organisational learning is anecessary step in explaining one of the most persistent and troubling ofeconomic problems, the uneven and uncertain transfer of technologicalcapability, what is commonly called ‘technology transfer’. The two funda-mental premises underlying technology transfer, that the acquisition ofcapability is related to the acquisition of knowledge and that knowledge isacquired through learning processes, seem simple enough. The problem isthat, between any two individuals, learning processes are likely to producedifferent knowledge outcomes and hence different capabilities. Thesedifferences are magnified by distance (cultural, social, physical) between theindividuals and compounded when a group of receiving individualsattempts to rearticulate the knowledge in an organisation. Within anorganisation, knowledge can only achieve and retain value throughprocesses of exchange that result in the construction of common under-standings which, in turn, provide the basis for consistent and predictableactions.1 The result is that there is rarely a clear map or plan available forspecifying how individual learning processes can be translated into effectiveor usable organisational capabilities.

In other words, in an organisation, individual knowledge is unlikely to‘self-assemble’ into organisational capability. Instead, the construction oforganisational capabilities from individual knowledge is likely to requireprocesses of iteration and interaction that involve constructing a ‘commonground’ between the individuals engaged in purposive activities. Thisrequirement of constructing a ‘common ground’ of knowledge withinorganisations is one of the reasons why the knowledge of organisations isnot the same as that of individuals. Taking account of this requirement,2

along with the costs of transactions and principal-agent incentive prob-lems, offers a basic toolkit for explaining the division of labour within andbetween organisations and a simple or first-order explanation for thedifficulty of the technology transfer problem.

The recognition that knowledge must be articulated within an organisa-tion to become an effective capability leads to a more specific theory of whateconomists refer to as economies of learning.3 In the usual economic treat-ment of learning economies, the practice is to take instrumental variablesreflecting the accumulation of experience, such as the accumulated scale ofoutput or the passage of time, as an explanatory variable for cost reduction.If, instead, we focus on the ‘technology of learning,’ the means by which

208 The economics of knowledge

organisations ‘construct’ knowledge by building a common ground ofunderstanding based on individual learning processes, a new set of ques-tions about the knowledge acquisition and exchange processes emerge.

First, it is possible to ask whether, in a competitive selection environ-ment, the knowledge of two different organisations may differ in significantways. In other words, can different ‘stocks’ of knowledge have the samecompetitive fitness? The answer to this question clearly depends upon atechnological context. The modern-day ‘low-tech’ cement or brick indus-tries seem to offer less opportunity for distinct accumulations of knowledgethan the ‘high-tech’ pharmaceutical, integrated circuit or software indus-tries. It is a common belief that these differences are the results of ‘techno-logical opportunity’; that is, that greater opportunities for product designand discovery are available in high technology industries.

It is also true, however, that these differences arise from the character ofdemand.4 The demand for variety in cement or brick design is far morelimited than the demand for variety in ‘high-technology’ industry products.In essence, knowledge accumulation is more valuable in ‘high-technology’industries because of the interest of customers in differentiated productofferings. If tomorrow, everyone were to wake up with a determination tohave dwellings and workplaces that were truly unique, the relative value ofknowledge accumulation in the brick and cement industries would increaseand, with some delay, so would the pace of innovation in these industries.For the purposes of this chapter, what is significant is that changes inknowledge accumulation initiated by the new characteristics of demandwould likely lead to greater diversity in what is known by cement and brickproducers.

Second, the nature of competitive advantage in an industry is pertinent.For variety in knowledge accumulation between firms to emerge, it is neces-sary that the sources of advantage be pluralistic. Advantages in one areamust be offset by weaknesses in others. For example, if a single parameterdetermines the technological trajectory of a particular industry, the unevenresults of knowledge accumulation (learning) would produce leaders andlaggards. Leading and lagging firms could only coexist to the extent that thetechnological advantage of the leader(s) was not fully translated intomarket advantage due to the possibility of non-technological advantagessuch as marketing capability or market location. Similarly, if the industry’stechnological trajectory is characterised by many technological parame-ters, there may be many coexisting ‘leaders’, none of which has a clear tech-nological advantage over the others. In this case, limitations to the ‘scope’of technological mastery will limit the dominance of individual firms.5

These first two issues suggest that diversity in ‘what’ is known by firmsis a consequence of the characteristics of demand, of technological

Learning in the knowledge-based economy 209

opportunity, and of specific sources of technological advantage that shapethe accumulation of knowledge. Learning provides a basis for the emer-gence of diversity and, perhaps, disparity in firm capabilities. Wheretechnological disparities exist and some firms are at a technological dis-advantage, these firms must find offsetting and non-technological advan-tages in order to remain in the market.

1.2 The Circulation of Knowledge within the Organisation

The third observation to be made, and the focus of this chapter, is on therelation between the sources of diversity (or disparity) of firm capabilitiesand the processes by which knowledge is held and circulated within andbetween firms.6 The possibilities for knowledge circulation may be highlyconstrained. That is, knowledge may only be accumulated ‘locally’ in closeproximity to the work processes and within very tightly bound socialgroupings of individual workers. Alternatively, the knowledge underlying afirm’s capabilities may circulate widely and with few constraints, providingeach of the employees of the firm with access to the knowledge underlyingthe firm’s capabilities. These may be taken as the two extreme points on anaxis along which it is possible, in principle, to locate firms.7

Even if knowledge is highly localised in the sense of not easily circulat-ing within or between organisations, it is not necessarily true that what isknown by various organisations is fundamentally different. In principle,faced with similar environmental conditions and opportunities, individu-als and organisations may independently accumulate similar types ofknowledge. This is especially likely to occur, despite the proclivities ofhumans to be inventive, when the technological and demand conditions inan industry punish rather than reward creativity and invention. Moreover,if knowledge exchange between organisations occurs through processes ofimitation, labour mobility or feedback from capital suppliers and cus-tomers, the knowledge of entire industries may follow convergent paths.Individual firms either learn or die and, if they survive, they eventuallylearn the same things.

The above observations suggest two possible research agendas. The firstis to characterise industries according the sources of their knowledge andthe processes by which knowledge is exchanged and reproduced.8 Thesecond is to examine how changes in market or technological conditionsalter the processes of knowledge accumulation and exchange, and exposeunderlying differences in capabilities among firms. When such changes havesimilar effects on all firms in an industry it is likely that each of the firms isoperating with similar knowledge. If the effects of environmental and tech-nological change have an uneven or irregular effect on individual firms, it


is likely that firms are operating with different knowledge. Thus, market ortechnological changes may help to reveal differences in knowledge ‘hold-ings’ by firms within an industry.

The next section examines market and technological changes in learning(and knowledge holding) within organisations from an historical perspect-ive. This discussion sets the stage for the second half of this chapter, whichexplores some of the structural changes and transformations accompany-ing the application of information and communication technologies to thelearning process within firms and industries.

2 A HISTORICAL PERSPECTIVE ON KNOWLEDGEAND LEARNING

There is no immediately obvious reason to believe that contemporary eco-nomic activities are inherently more knowledge-intensive than those of thepast. Despite the high tide of rhetoric concerning ‘knowledge-based’economies and industries, manufacturing and service activities have alwaysinvolved knowledge inputs from skilled workers, accumulated production-related know-how, the design of products and services, and the configur-ation of production processes to achieve well-defined levels of quality andreliability. Indeed, the embodiment of this knowledge in individual workersin ways that were unarticulated outside a local context is likely to have pro-vided more scope for variety and differentiation in nineteenth-centuryworkplaces than in contemporary ones. Given a time machine, it would betempting to transport and to strand (at least briefly) a few modern know-ledge management experts in a steel, textile or grain mill of 1870.

The survivors of this exercise would likely conclude that the processes theyobserved did encompass bodies of knowledge, albeit ones that would belargely unfamiliar even to those with some knowledge of the modern equiv-alents of these factories. The accumulation of knowledge is, in fact, situatedwithin historical circumstance and it does not immediately follow thatmodern ‘stocks’ of knowledge are larger or more complex than those accu-mulated at previous points in history. A close interrogation of these visitorsto the past would, however, reveal that the structures of knowledge that theysaw being employed, involved a much greater degree of personal expertiseof typical individuals directly engaged in production activities. The creationand sustenance of hierarchies in nineteenth-century industry often reflectedthe accumulation of experience over a lifetime, with individual workersclosely guarding the maintenance and reproduction of this knowledge.

Correspondingly, the transportation into the future of industriallyknowledgeable individuals from 1870 to the present would provoke a


symmetrical observation. They would observe that modern productionoperations were comparatively ‘de-skilled’ in their absence of individualswith specific knowledge of the tools and procedures being employed on thefactory floor. This change in the locus of knowledge ‘holding’, rather thanin the extent of employed knowledge is the fundamental distinctionbetween the role of knowledge in contemporary industry and that of thelate nineteenth century. In the hierarchy of the modern enterprise, pro-duction engineering management and operations are activities conductedbehind the scenes and, for the most part, off the factory floor. While thismode of organisation is sometimes criticised as discarding valuable infor-mation, the trend to relocate these activities largely continues, as does thegrowth in recorded productivity in manufacturing of all industrialisedcountries employing this type of organisation.9 There must, therefore, be areason for the learning activities that are responsible for the accumulationof knowledge having been ‘de-localised’ from direct involvement in factoryoperations. Examining these reasons provides a deeper understanding ofthe learning processes at the level of the organisation and at the level of theindustry.

2.1 The De-localisation of Knowledge Accumulation inNineteenth-Century American Manufacturing

There appear to be two fundamental reasons for the de-localisation ofknowledge in late nineteenth-century American manufacturing.10 The firstis the progressive articulation of knowledge required to produce a large-scale producer goods industry. The expertise required to produce state ofthe art producer goods was, by the late nineteenth century, already movingoutside the factory.11 Producing complex machines such as textile looms orreapers follows the admonition of Smith, Young and Stigler that the divi-sion of labour is limited by the size of the market.12 Widespread marketsfor producer goods assured that the design and manufacture of thesemachines was increasingly remote from the site of production. Factorieswithout machinists who would be capable, if necessary, of rebuilding thefactory’s machinery became increasingly common and, with the end ofthese individuals’ employment, a vast body of specialised knowledge dis-appeared. The new site of knowledge accumulation became the capitalgoods producer and a new form of industrial co-ordination, the pro-ducer–user relationship, came to the fore. This change is important for theensuing discussion.

The second reason for the de-localisation of knowledge in productionwas the rise of ‘systemic’ approaches to the problems of increasing pro-ductivity and standardisation of industrial output. Disentangling the


origins and spread of these ideas has become a monumental quest anda proving ground for historians of technology. One reason for this is that‘systemic’ approaches are tightly bound to understanding the origins of‘modernism’ and for a critical appraisal of its consequences.13 For our pur-poses, however, the main point is established by Rosenberg’s (1976) historyof the emergence of generic and multi-purpose machines to replace the spe-cialised tools developed within specific industries (for example, locomotive,bicycles and sewing machines). These new machines could producemechanical parts for a wide variety of industries and thus could be pro-duced on a larger scale than the industry-specific machine tools. Thesedevelopments produced a degree of interdependence between mechanicalmanufacturing industries, creating the possibility for knowledge spilloversas machinists and mechanics practising in one industry could master skillsthat were applicable for employment in another industry.

The argument that these two forces, scale and interdependency, fosterstandardisation and the need for control is a particularly attractive hypoth-esis in explaining the increasing interest in mechanisation that emerged inthe latter half of the nineteenth century. Several complementary hypothe-ses support this explanation. For example, Beniger (1986) links the originsof systematisation with the control requirements for safely maintaining thespeed of railroads, a critical parameter in reaching higher levels of scaleeconomies in a transportation network.14 Similarly, Hughes (1989) hasargued that solving the ‘reverse salient’ problems introduced by new tech-nologies required the development of a more pervasive ‘systems’ view oftechnology in a variety of sectors, most notably in the network industries.Hounshell (1984) contends that the idea of interchangeable parts wasadopted in view of the impetus it would provide to scale manufacture,rather than that increasing scale led to the rationalisation represented byinterchangeable parts.

Systematisation produced a need for organisational structures withinmodern enterprises that could record, analyse and plan the extension anddevelopment of production. This obligated enterprises to restructure the‘holding’ of knowledge from the shop floor or even from its individual cus-todians within engineering departments to the managerial and researchfunctions of the enterprise. Only by achieving a global overview of pro-duction was it possible to strive for systemic improvements in the organ-isation and conduct of production, to apply ‘scientific methods’ basedupon ordered trial, or quantitative analysis, to the operations of business.

The ‘de-localisation’ of knowledge from the shop floor and its relocationinto the managerial and research functions of the organisation provided thefoundation for the mass production of Henry Ford and perhaps the mostpivotal moment in the history of industrial learning processes. Ford’s vision


of the future of mass production was the creation of standardised goodsthat all could afford. As Edward A. Filene argued, Fordism offered Americaeconomic salvation from the inter-war years growth problems, in support-ing mass consumption by allowing an ‘increasing standardisation of prod-ucts, and an increasing mechanisation of the process of production’.15

The shift in knowledge ‘holding’ in the USA from 1870 until the mid-1920s involved supplanting learning and knowledge at level of the factoryor shop floor with knowledge managed at the level of the organisation. Thetransformation from a ‘craft-based’ to a mass production system obliter-ated this localisation of knowledge and replaced it with large and complexproduction systems with many interdependent components and sub-processes. These systems served as the new locus of knowledge accumula-tion. The ‘canvas’ upon which learning processes were inscribed had grown.

2.2 An Archetypical Case of the Transformation in Knowledge Holdingand Circulation

Providing systematic evidence concerning the rate and direction of thetransformation of knowledge holding and circulation in nineteenth-century manufacturing is beyond the scope of this chapter. It is possible,however, to provide an archetypical narrative of this transformation in oneorganisation, the McCormick Reaper Works.

The following account examines the struggle between Cyrus Hall and hisbrother Leander McCormick over the operation of the McCormickReaper Works following its reconstruction after the Chicago fire of 1871and until the year of Cyrus Hall’s death 1884.16 During these years, thestruggle, often acrimonious, between the two brothers concerned the adop-tion of what Hounshell calls the American system of manufacture, asystem of standardised parts based on the use of jigs and fixtures and acombination of general purpose and specialised machines.

The machines were produced by the emerging New England machinetool industry. The context of this dispute was a continuing disagreementbetween the brothers about the scale of production of the mechanisedreaper for which the company had become famous. Leander held the viewthat there was a significant danger of overproduction in any particularmodel year of the company’s reaper, a view that Cyrus Hall shared formany years. Leander’s view, however, was strengthened by the craftmethods employed in the McCormick factories that were under his direc-tion. These methods involved many parts being individually constructed byskilled workmen following patterns derived from the current year’s modeldesign, but without the techniques necessary to achieve consistent repro-duction of parts and, therefore, the need for additional ‘fitting’ operations


in the production process. In economic terms, the variable costs of pro-duction remained high despite the existence of a relatively ‘standardised’model design.

Weeks before the Chicago fire of that year destroyed their north-sidefactory, Cyrus Hall McCormick decided and announced in 1871 that hethought the company should expand production. To achieve this aim fol-lowing the disaster he dispatched a series of production experts to the newlyconstructed factory in Chicago’s south-west side. Hounshell contends thatin equipping this factory, the nature of Leander’s orders with New Englandmachine companies clearly revealed his inability to exploit the opportun-ities offered by specialised machinery.17

After a protracted struggle with his brother and following the advice ofthe manufacturing experts he had sent to their factory, Cyrus Hall reacheda similar conclusion. He first bought out his brother’s share in the businessand then, with the concurrence of the Board of Directors of therecently incorporated business, dismissed Leander as factory superinten-dent in favour of Lewis Wilkinson, a veteran of several New Englandmanufacturing companies. Working with the young Cyrus McCormickJr, Wilkinson began the process of creating a truly standardised McCormickreaper in the 1880s, a path Cyrus Jr continued after Wilkinson left thecompany in the following year. The resulting changes between 1880 and1884 were, however, not viewed neutrally by the craftworkers who haddevoted their lives to the production methods. In 1885, a strike united all themajor unions for the first time, a development that one author has concludedwas a ‘prelude to Haymarket’.18

There are three lessons from this case. First, the issues of market size andgrowth are relevant to the transformation. Cyrus Hall McCormick’s deci-sion to expand production was taken in the expectation that the marketwould absorb the resulting expansion. If this expectation had not been ful-filled, the McCormick Reaper Works would be a subject for the localhistory of Chicago, and its successor company, International Harvester,would not have been born. Second, to transform the scale of production itwas necessary to replace one set of skills with another, a transformationthat required considerable planning and tooling investment. In the case ofthe McCormick Reaper Works, the shortcomings of craft-based produc-tion made these investments worthwhile, although this need not always bethe case. Third, the process of destroying one set of skills and replacing itwith another had consequences for the livelihood of individual workersand was collectively resisted. The consequence of the change, however, wasthat the new production system, under the direction of Cyrus McCormickJr was able to increase its output fivefold by 1902. The new system involveda different learning process from the old, one based upon the knowledge


required for altering the entire production system to accommodate changesin model designs.

2.3 The Middle Distance: Why Wasn’t Fordism Enough?

The contention that the specific forms in which knowledge is organisedwithin the organisation creates a differentiated advantage that may notpersist with changes in market conditions is well illustrated by the FordMotor Company.

During the first quarter of the twentieth century, the logic of a mass pro-duction system based upon standardised parts was widely deployed. Theprinciples enunciated by Ford of ‘power, accuracy, economy, system, con-tinuity and speed’ became the guiding principles for mass production.19

Ford and his production engineers devised and perfected this logic in theproduction of the Model T, a single basic model for the automobile. TheModel T was produced, with some setbacks, in ever greater numbersand with incremental changes from 1908 through 1927 when the ‘last’ ofthe 15 million Model Ts was driven by Ford to its resting place in Dearborn,Michigan.

The difficulties that Ford and his company had in making the changeoverto the Model A, and the costs to the company whose market share hadslipped to less than 15 per cent in 1927, have become a case study followedby several generations of business school students.20 The central lesson isthat General Motors achieved its stronger position through a combinationof organisational, marketing and production innovations. These innov-ations created ‘flexible mass production’, the ability to re-tool model annualchanges in the automobile and to promote these changes to buyers as animprovement over the standard represented by Ford’s Model T. Thus, asHounshell concludes, ‘Ford had driven the strategy of mass production toits ultimate form thereby into a cul-de-sac’.21 This example of a funda-mental change in the nature of the learning process is no less valid becausethe Ford Motor Company was ultimately able to make the transition to thenew paradigm, for in doing so the original visionary path of Henry Ford,of mass production, had to be abandoned. This path, admired and emu-lated throughout the socialist world, supported the industrialisation of theSoviet Union and China, as well as many smaller countries, even as it wasdisplaced in the USA and other industrialised countries.

Few illustrations of changes in learning processes could be more dra-matic than the contest between the Ford Motor Company and GeneralMotors for the future architecture of industrial production. Along thebranch pioneered by Ford is the indefinite expansion of standardised massproduction, achieving ever-larger scales of production but imposing upon


the buyer a very limited choice. The branch that Alfred Sloan created pro-duces technologies for creating variety and diversity in industrial output,the shifting of skills and employment towards stimulation of new consumerdemands and the creation of new fashions in artefacts. In following thissecond path, the USA was able to discover the new technological frontierrepresented by the application of distributed information technology. Thispossibility might never have emerged had Fordism succeeded in producingwhat Filene claimed was a ‘way out’ of the American economic dilemmasof the inter-war years.

3 APPROACHING TODAY: KNOWLEDGE AND THECOMPUTER REVOLUTION

Early commercial computer manufacturers in the USA such as IBM andBurroughs attempted to produce general-purpose computers that could bereconfigured as information-processing needs shifted. This choice did notreflect technological necessity. Alternative means to build computationaloffice machinery were being demonstrated by the British company, LyonsConfectionary Limited, which had created a special-purpose computerdesigned around many of the most common business processes in accoun-tancy and inventory control.22 It is tempting to conclude that computermanufacturers had learned the lessons of the past century of Americanmanufacturing in producing a flexibly configurable and mass-producedsystem. It is, however, more likely that the initial commercial designs werea response to the uncertainties of commercial computer applications.

Although mass production had resulted in the centralisation of manybusiness operations, those amenable to computerisation might be eithersmaller or larger than anticipated by computer manufacturers. A general-purpose strategy not only simplified the design of the machine, it alsoassured that software would become the principal means for ‘customising’the machine to particular applications. To achieve this aim, it was neces-sary for computer manufacturers to collaborate with users in the designof software systems. Users also exchanged a certain amount of softwarewith each other through user groups sponsored and organised by the com-puter manufacturers, a harbinger of the contemporary ‘open source’ soft-ware movement.

Computer manufacturers abandoned this collaboration in stages.23

During the first stage, customers were provided with ‘general purpose’ lan-guages that could be used to create customised or user-owned applications,and computer manufacturers retreated from the direct implementation ofspecific software applications. A second stage of computer manufacturers’


retreat, beginning in the late 1960s, involved further retrenching of softwareofferings and the blossoming of an independent software vendor market.24

It is difficult to conclude whether these developments would have followeda similar path had Ford’s vision of mass production prevailed over Sloan’s.Until the late 1960s, centralised computation was a necessity because thecosts of equipment and ancillary support were simply too high to be dis-tributed throughout the company. Had Ford’s vision prevailed, however,the longer planning horizon it entails might have encouraged a greaterdegree of hardware and software integration.

By the mid-1960s, the progress in electronic components (anotherexample of a producer goods industry that emerged from the continuingprocess of outsourcing) had provided the means for a new generation of‘minicomputer’ companies. Digital Electronics Corporation and DataGeneral promoted a decentralised model of computation, a model allow-ing the computer to either be integrated as a dedicated part of the produc-tion process or to ‘stand alone’ as a data processing engine for use asneeded. In the framework of this chapter, the minicomputer provided themeans for ‘localising’ the generation of knowledge at the site of applicationand provided a means for escaping from the development and tasking pri-orities established for the central computer facility.25

The trajectory of decentralising information processing applicationscontinued with the development of the personal computer, which, insteadof being an incremental and specialised innovation, created a profoundalteration in the paradigm of information processing. As David has empha-sised in his examination of the parallels between electrification of USindustry and the growth of distributed computational power, the transfor-mational feature of the distributed technology is its capacity to be utilisedat the ‘site’ of application.26 For the personal computer, this ‘site’ was thedesktop of the ordinary business professional as well as the desks ofnumerous clerks, administrators and secretaries. Electrification involvedthe transformation of the physical plant of manufacturing operations,which no longer needed to rely on the economies of the central drive andvertical plant architectures to best utilise central drives.

In the case of the personal computer, the ability to distribute docu-ments (initially by the manual exchange of the floppy disks and soonthereafter through local area networks) offered numerous opportunities torestructure the work flow within organisations. These changes reducedvertical information-processing layers in management. What in the pastneeded to be centrally collected and analysed can now be built into thelocal information processing tools and these can, in turn, be used to con-struct integrated systems in which horizontal as well as vertical commu-nication paths in the organisation can be exploited in the same way as


fractional horsepower electrical motors can be distributed in mechanicaloperations. A new integrated and information-laden system in manufac-turing and services is now characteristic of modern establishments.

In one respect these developments are a direct application of the decen-tralised management and rapid changeover principles of Alfred Sloan.In another, however, they offer potential for more fundamental change.This potential is the consequence of the social issues involved in employingthe new technologies. As in the past, there is resistance to the new tech-nologies.27 It appears in obvious forms such as non-cooperation with theroutines established for data entry or professed incapacity to master thenew tools. One means for addressing these problems is to create a moretransparent information environment in which the individual users of thistechnology receive direct and useful feedback for the performance of theirjobs and in which these same users are encouraged to suggest new tools orimprovements in existing ones. Some have suggested that implementingsuch capacities would support the ‘learning organisation’, an improvedprocess of knowledge accumulation and distribution that would augmentthe designs originally built into such systems. Whether such optimisticassessments of the deconstruction of managerial control and authoritywill, in fact, be either widely accepted or constitute a competitive advan-tage for those organisations adopting such changes remains an unresolvedbut intriguing issue which is explored further in the next section (4).

The central feature of these developments is that they involve the use ofan information processing system to support local learning. The ultimategoal is similar to that pursued within Sloan’s model, facilitating innovationand product variety while simultaneously intensifying the use of fixed tan-gible and intangible capital. The means for doing this, however, involve newlearning processes, both in the central design of such systems and theirlocal application. The remarkable feature of decentralised information pro-cessing systems is the fluidity that they introduce for reconfiguring the lociof learning within the organisation. As information technology use inten-sifies, organisations are free to evolve in patterns that are peculiar to thecomposition of their employees and local circumstances. In this sense, con-temporary learning opportunities involve re-energising the ‘local’ layer thathas been displaced by prior management practices involving larger sys-temic planning and implementation of information systems.

It is possible that these new processes may deliver a more participatoryand, therefore, in certain respects, more democratic work environment. Onthe other hand, it is possible that these new processes will be structured andmanaged using new forms of hierarchy, ones that are perhaps more shallowin layers but that are no less autocratic in defining what procedures indi-vidual workers follow in their everyday work lives. This new environment


offers a new contest between centralised and decentralised modalities oflearning and knowledge accumulation.

In summary, the intensification in the use of computer technology hasbeen accompanied by several shifts in the locus of knowledge generation,from a highly centralised and collaborative activity between computermanufacturers and their customers to develop organisation-wide solutionsto information processing problems, towards a more ‘local’ accumulationof knowledge and capability that is, nonetheless, co-ordinated through theuse of an ‘information processing infrastructure’.

The history of learning processes involving production planning andinformation processing systems in the twentieth century has been mixed.On the one hand, the creation of vast integrated systems managed by verylarge companies have substituted a new organisational-based learning forthe localised and idiosyncratic learning of the shop floor and individualroutines. On the other hand, when these larger organisational structuresprove unable to exploit new opportunities, it appears that the articulationof the producer goods industries supports entry and growth of new indus-trial players.28

What is new, and in particular in need of further analysis, is how thefurther development of information and communication technologies islikely to influence the processes emerging in the ‘learning organisation’ orvariants based on this idea that more strongly emphasise managerialcontrol and authority. The ‘relocation’ of learning activities, in fact, thepotential for the wholesale dispersal of these activities throughout theorganisation, suggests a further profound set of changes in how organisa-tions operate. This change provides a rematch of the contest, in a differentcontext, between centralised and decentralised knowledge accumulation.By reflecting on the similarities with and differences from previous experi-ence, it is possible to hazard a few predictions about the outcomes of thiscontest in the future. This is the principal theme of the next section.

4 FORWARD INTO THE FUTURE: THE NATURE OFTHE KNOWLEDGE-BASED ECONOMY

Mass production supports the convergence of all productive processes thatemploy it towards an ‘information model’ in which the characteristicsusually associated with ‘information goods’, their high initial costs of pro-duction and the relatively lower costs of subsequent copies, become themodel for all mass produced products. There are, of course, significantdifferences in the scale between the ‘first copy’ costs of a General Motorsautomobile and those of a multimedia product, and the variable costs of


tangible artefacts such as automobiles are much higher than the reproduc-tion costs for information goods.29 Nonetheless, many of the same prin-ciples may be employed in product planning such as the ‘reuse’ of designand development efforts, the planning of product ‘versions’ that incorpo-rate upgrades and updates to the initial design and, perhaps most signifi-cantly from an organisational perspective, the granting of substantialautonomy to the product manager.

4.1 The Production of Variety

In terms of economic analysis, a principal decision of the firm is whetherthe relative investment costs of producing more varieties, that is, reinvest-ing in the fixed costs to produce a new variety, have a greater returnthan extending the marketing and production of existing varieties. The‘changeover’ decision is also of direct relevance for issues of learning.A particular path of development may offer greater ‘spillover’ from the pro-duction of one variety to the production of another, reducing the costs ofintroducing new varieties and producing sub-additive costs (economies ofscope) in producing more varieties. Technological change can be seen asopening up new opportunities for learning, some of which will lead to sub-stitute products, eroding the share of existing varieties and perhaps extin-guishing further development along previously developed paths. Learninginfluences not only the supplier’s development path, but also the nature ofdemand (for example, brand or product loyalty or product type familiari-sation), and may either sustain or detract from the continuing productionof a particular set of varieties.30

A complementary analysis can be derived from employing the trad-itional learning curve. If cost reduction is accumulated through repeatingthe same productive process, the relevant economic issue is when to makea change to a new learning curve. The same sorts of elaboration are possi-ble additions to the basic learning curve model. Again, a principal issue isthe timing and nature of changeover processes that alter the variety ofproducts under production.31

These same basic principles can be extended beyond manufacturing to anumber of services, although not to all. When services incorporate repro-ducible elements that are designed once and then replicated in whole or ina significant part of the whole, the ‘information model’ of fixed costs isapplicable. Similarly, when services involve significant learning processes,the same basic model can be applied to services as to manufactured goods.The distinguishing feature of many services, however, is that they arealmost entirely driven by variable costs. For example, the costs of a surgi-cal operation involve commitment of the surgical team’s time and skill on


a one-to-one basis with the patient. Although the costs of the operatingtheatre will be subject to economies of throughput planning and the arte-facts utilised are individually subject to the information model, a majorshare of the costs is for the professional services rendered to a particularpatient.32 Services of this nature are of considerable interest, but are notdealt with in this chapter.

In the ‘information model’ of production, organisations innovate by cre-ating generic products that are capable of reaping economies of scale byreusing the fixed costs of their development. These economies are inherentin the relatively lower fixed costs of creating the ‘first copy’ of the productcompared to the costs of creating succeeding copies. What makes the worldnot only more ‘untidy’, but also more interesting, is the enormous varietyof possibilities for reuse and recombination of knowledge. The resourceallocation decisions governing the number and scope of this reuse andrecombination are of principal importance in determining the rate anddirection of economic growth in the ‘knowledge-based’ economy. Thisraises questions about how to best organise the accumulation of knowledgewithin and between organisations.

As the model of flexible production has developed, capabilities to recon-figure and tailor products to specific markets have improved dramatically.In industrialised nations, the two most desirable strategies are (1) the cre-ation of ‘hit’ products in which the full extent of mass production andconsumption can be engaged or (2) the development of portfolios of spe-cialised products that dominate their market ‘niches.’ Given the uncertain-ties of finding ‘hit’ products, a strategy of risk reduction is to employ thesecond strategy as a means of searching for potential ‘hits’. A less attrac-tive third strategy is to produce ‘commodity’ products in which price com-petition reduces profits towards competitive levels, the ‘normal’ rate ofprofit. This strategy may still be viable because it supports the constructionof a manufacturing base that can be used to introduce imitations of ‘hit’ orspecialised products. If the preceding account is an accurate characterisa-tion, we would expect to observe a continuing growth of product variety.Despite the continuing shortcomings of national income statistics in mea-suring variety (a legacy of the historical concern with recording the per-formance of mass production), evidence can be produced about the extentof this increase. A significant compound annual growth rate of varieties ofconsumer product categories has been experienced in the USA over thepast two decades (see Table 8.1).

Projecting the number of varieties of consumer packaged goods avail-able for another two decades at these rates would yield over 140 000different consumer goods on offer in the year 2018. An individual devotingone minute on average to considering each of them would devote over


2000 hours to the task of becoming a ‘fully informed’ consumer comparedto the 400 hours now required. The problems of competitive analysis andstrategic formulation will, of course, also expand.

If these trends are to continue, the technologies available for conceiving,designing and implementing new product and service varieties must con-tinue to improve. Just as the consumer faces information overload in con-sidering the range of choice available, the companies producing these goodsand services are likely to experience an overload in their design and execu-tion capabilities using the centralised planning and design structure ofSloan’s flexible manufacturing model. Further decentralisation in whichthe term ‘product managers’ achieves a growing independence of actionand initiative seems a likely consequence of these developments.

4.2 Organising Knowledge to Produce Variety

Implementing greater decentralisation within the organisation requires thedevelopment of intra-organisational interfaces between working groupsthat minimise the time required to set up and execute the complex interac-tions needed to conceive, design, and implement new products and services.Doing this requires organisational capabilities that are ‘modular’, they can


Table 8.1 Indicators of increasing variety in the US economy

Product category Number Year Number Year CAGR

Consumer 4 414 1980 24 965 1998 10.1 %packaged goods

Vehicle models 140 1978 260 1998 3.5 %Vehicle styles 654 1978 1 212 1998 3.5 %Breakfast cereals 160 1978 340 1998 4.3 %National soft 20 1978 87 1998 8.5 %

drink brandsNew book titles 40 530 1978 77 446 1998 3.7 %Mouthwashes and 27 1978 130 1998 9.1 %

dental flossesLevi’s jean styles 41 1978 70 1998 3.0 %

Note: The number of consumer packaged goods is estimated from the shelf keeper unitcode registrations, part of the uniform product code standard for standardised point of salescanners. Sources of other figures are available in the source. In order to compute theCAGR (compound annual growth rate) the original sources’ approximation of the years ofobservation ‘late 70s’ and ‘late 90s’ have been fixed at 1978 and 1998 respectively.

Source: Federal Reserve Bank of Dallas (1998).

be reconfigured and recombined to adapt to the growing variety and com-plexity of products and services. Information and communication tech-nologies may aid in this process, but there are many uncertainties aboutwhether the ‘hype’ and promises of ‘work group’ and ‘collaborative’ soft-ware are really meeting the needs of organisations for this flexibility andmodularity.33 At the same time, however, negative ‘systemic’ effects, such asnegative spillovers from one group’s decisions on the performance ofothers, are likely to persist. Limiting these effects requires managerialcontrol.

There are two issues that prominently appear in organisations’ attemptsto grapple with the issue of new forms of organisation associated withvariety and the ‘information model’ of product and service innovation andproduction. The first is the extent to which the ‘centre’ has to intervene inorder to achieve coherent and cohesive outcomes among the decentralised‘work groups’ or ‘production teams’ that are responsible for creating andproducing products and services. Can the centre govern by adopting ‘pro-cedural’ rules for mediating conflicting priorities or contests over commonresources or is it necessary to engage in a more comprehensive planningprocess?34 The second issue is whether new ‘institutional standards’ (that is,norms, rules and practices) can be devised that reduce the need for suchintervention and permit the internal organisation of companies and therelation of companies with each other to exhibit ‘emergent order’, much asmarkets achieve co-ordination between supply and demand without anexplicit co-ordinator or auctioneer.

The first of these issues reflects the culmination of outsourcing and thedevolving of operational decision-making to lower levels within the organ-isation (that is, the elimination of middle management layers that previ-ously processed information and attempted to steer the organisationtowards strategic goals). Management provides not only ‘proceduralauthority’ but also the resources and priorities granted to these groups anda set of rules for governing the interactions among them. The disadvantageof decentralisation, however, is that potential synergies in the accumulationand reuse of knowledge may be ignored, reducing the organisation to aloose confederacy of smaller-sized enterprises operating within a commonfinancial framework.

These potential problems indicate that further co-ordination may haveto be devised to capture the benefits available from operating within asingle organisational framework. One means of doing this is to deviseincentives and institutional standards that are effective in reducing centralmanagement ‘overheads’ (in terms of managerial time and therefore, ulti-mately, cost). How these incentives and institutional standards can bedevised and who will devise them are areas of active experimentation


within contemporary organisations. A vast array of research opportunitiesexist for identifying their emergence in specific industries and therebyaiding in the process of their diffusion and recombination within otherindustries.35

Those who are sceptical about the possibilities for achieving this sort ofdecentralisation emphasise the situated nature of knowledge and indi-vidual cognition in order to deny the transferability of understandingsreached by one social group within an organisation to others.36 This scep-tical position, however, seems to deny that anything has happened sincecraftworkers were the principal repositories of knowledge in manufactur-ing establishments. There can be little doubt that the processes of labourresistance and solidarity have, on occasion, prevented the complete appro-priation of local knowledge through the redesign and centralisation of pro-duction processes and designs. As was noted earlier, however, wherever thisissue has proved to be a significant barrier, the nature of the productionsystem has been changed to make skills and knowledge obsolete. This isdone by a centralised authority devising a set of procedures and routinesand the induction and indoctrination of an (often new) industrial labourforce through a relatively short period of training, which reflects the modestextent of local knowledge required to perform the job.37

The question remains, however, whether standards for localised decision-making and knowledge accumulation can be re-established withoutre-creating the problems of opportunism, co-ordination failure and idio-syncrasies that historically motivated the usurpation of decentralisedcontrol. This amounts to assessing whether Sloan’s idea of divisionalresponsibility can be further extended to permit work groups to negotiate(either internally or with a ‘lean’ and process-oriented centralised manage-ment) the incentives and institutional standards that will govern their work.In some cases this process will not work and central authority will makedirect interventions in the content of incentives and standards. These inter-ventions will forgo the potential advantages of worker participation in thedesign and improvement of their roles, including the more detailed workerknowledge of context and situation.

These developments are not just a source of interesting research ques-tions; they reflect two alternative ‘paths’ for the predominant mode for theorganisation of work in modern enterprises. They will influence the con-ditions of working life for the foreseeable future. There are many possiblepoints of observation within the organisation and between organisationsfor observing the changing locus of learning within the organisation andhow these changes stimulate the development of new incentives and insti-tutional standards. They include human resource policies regardingrecruitment, induction and training, the design of business processes and


structures, and the configuration of information and communication tech-nologies. This chapter focuses only on the last of these areas because of itsdirect and relatively straightforward linkages to institutional stand-ardisation issues.

4.3 Information and Communication Technologies and Knowledge

The design and execution of information and communication technologysystems is a central issue in the development and implementation of incen-tives and institutional standards for meeting the challenges of complexityand variety. The following discussion briefly considers three aspects ofthese technologies that are likely to inform future developments. First, wewill examine the use of interpersonal communication within the organisa-tion and between organisations to enhance the exchange of knowledgenecessary for forming institutional standards. Second, we examine the useof information technologies to model the operations of the company withthe aim of identifying systemic dependencies and then resolving themeither by the decisions of central authority or by introducing incentives tolimit systemic failures and the propagation of negative externalities. Thethird aspect of the use of information and communication technologies isin augmenting collective memory and, ideally, accelerating processes ofcollective learning and adaptation.

The role of interpersonal communicationsInformation and communication technologies provide a means to augmentinterpersonal communication within organisations. The use of email forthe exchange of information allows fluid interpersonal communicationrequired for learning and the negotiation of incentives and institutionalstandards, while reducing the overhead costs of scheduling and conductingmeetings. There is evidence, however, that electronic communication doesrequire ‘real world’ interaction to build interpersonal trust and to resolvenon-routine transactions.38 It is, however, unclear whether reliance on ‘realworld’ interpersonal interactions is a persistent or transitory feature in theuse of electronic communications. Many current users have relied uponother types of communication for most of their careers, and succeedinggenerations with more experience with the medium may exhibit differentbehaviour.

Correspondingly, the tools for facilitating more complex types of com-munication, such as exchanges about work or negotiations in progress, haveonly recently begun to be used on a broad scale.39 Some period of learningmay be required before they can be considered as replacements for existingmethods for negotiating standards about artefacts or procedures and a


continuing critical appraisal of these methods is needed to predict theireventual consequences. Evidence from Europe’s Telematics ApplicationProgramme, a Europe-wide research, development and demonstration pro-gramme, suggests that while some ‘new ways of working’ including emailwere used by a majority of participants, a minority of the participantsutilised more advanced information and communication technologies suchas collaborative software or videoconferencing.40 The relative immaturity inthe use of these technologies suggests limits to their applicability both to theexercise of procedural authority by centralised authority and to theirefficacy in negotiating incentives and institutional norms within decen-tralised work groups.

Recent experience in the operation of open source software communitiesoffers examples of ‘leading edge’ experimentation with new mechanisms forimplementing collaborative work.41 Successful development of a widelyused microcomputer operating system (Linux) and the software that dom-inates the application of ‘serving’ World Wide Web pages on the Internet(Apache) are two recent examples of successful developments undertakenby worldwide collaboration among software developers, many of whomhave only ‘met’ through virtual communications.42 Since the workingprocesses involved in the development of computer software involve manyinstitutional standards issues (and several incentive issues as well), the opensource software development is a likely source of evidence about the possi-bility of more broadly employing electronic communication to supportshared development of artefacts involving the use of incentives and insti-tutional standards within companies.

For example, open-source software developers must develop a number oftools for the smooth interchange of ‘work in progress’ (the source code),which they are collectively developing. One of the features of these tools isinternal documentation of revisions that allows smooth ‘backtracking’ toearlier versions. This capability, in combination with modular code andstandardised interfaces between modules, allows the recombination ofmodules across versions of the system under development. To co-ordinatethese activities, open-source software developers also employ real-timecommunication channels such as ‘chat’ lines (concurrent short messageexchanges) and a hierarchy of procedural authorities (at the level of themain project as well as sub-projects) to achieve closure on versions of thesoftware as it is developed. All these elements involve learning andexchange processes that are likely to be reproduced within other successfulvirtual development efforts.

Interpersonal communications about artefact and procedure standardsare now able to benefit from the very rapid increase in abilities to documentcurrent practice. The Internet technologies for creating web pages as


applied to intranet communication networks substantially reduce the costsof intra-company ‘publications’, the internal dissemination of companyinformation designed to facilitate learning activities. These technologiesalso have the potential to accelerate the timeliness of information publica-tion and to hasten the revision and updating processes.

As in any publication process, the costs of producing the first copy ofinformation remain significant and it is possible that organisations canover-invest in the production of information that does not provide signifi-cant value. Nonetheless, these technologies do appear to offer potentialreductions in the distribution costs and increases in the accessibility ofinformation about artefact and procedure standards. Electronic communi-cation can further augment this process by providing feedback about thevalue and correctness of such publications. Organisational incentives thatsupport the currency and accuracy of publications are necessary foreffective results and provide a means for preventing over-investment byindividual enthusiasts.43

Effective means for organising information about artefact and procedurestandards will also be required to decentralise institutional standards-making processes and devise effective incentives. In medium-sized or largerorganisations it is relatively easy to create information structures whosecomplexity creates a bottleneck to effective use. The development of theskills and procedures for organising this information is likely to become arelatively high priority. The development and application of new metricsfor assessing the complexity of information structures such as web sites isurgently needed.44 Once created, however, the information structure of aparticular organisation is likely to be subject to ‘lock in’ as users invest inlearning how to navigate it to find the information they need.

Modelling business processesThe value of creating a ‘virtual model’, an electronic representation of theprocesses and transactions of the organisation, is the ability to defineorganisational dependencies, overlaps and redundancies that can be fruit-fully resolved through decentralised initiative or centralised intervention.Integrating the analysis and execution of transactions is particularlyimportant because it avoids a separate and often difficult to justify dataentry process. Thus, for example, inventory control systems that rely upona separate data entry procedure are far less reliably maintained than oneswhere placing an item in inventory or removing an item requires a ‘logging’procedure. Moreover, the requirement for data logging provides theimpetus needed to implement complementary technologies such asscanner codes, along with the readers and writers of these codes. In retailestablishments, for example, reorders are often automatically generated by


the integrated processing of received shipments and ‘point of sale’ termi-nal information.

The recording of individual transactions is the basis for an organisation’s‘data warehouse’ and the growth of these warehouses is providing theimpetus for fundamental change in the architecture of enterprise manage-ment software. The goal of software systems devised to exploit data ware-houses is to provide a means for the ex post imposition of an analyticalstructure on data records that may either be tightly or loosely linked to oneanother. Thus, enterprise software may be used to model and analyse theprocesses that originally generated the stored data. They may also be used,however, for entirely different purposes. For example, in retail applicationsdata on weather patterns or television schedules may be correlated with thepurchasing patterns of customers. The data from factories where many ofthe processes are automatically measured and recorded may be analysed forpreviously unrecognised interdependencies and the propagation of localmachine breakdowns on the pattern of work flow and error rates.

The ability of such systems to support many different users may allowsubstantial decentralisation in problem identification, problem solutionand design activities. The particularly intriguing feature of these systems istheir capacity for the dynamic restructuring of interrelationships amongthe data that have been warehoused. It is important to emphasise that thesetools are not costless or automatic; their application requires substantialinvestment and skill. In principle, these tools can support the formation ofinstitutional standards and the specification of incentives that will improveorganisational performances. In the first instance, such systems lower theincremental costs of organisational analysis; ultimately, they may encour-age creative experimentation, learning and critical thought.

Organisational memory and learningThe issues surrounding organisational memory and learning are particu-larly important for assessing the value of localised knowledge accumula-tion. The activities of engineering consulting companies, architects, lawfirms and management consultants involve a constantly shifting pattern ofnovel problem-solving activities and the replication of established know-ledge. Many other organisations, including those that have high numbersof ‘similar’ transactions, could also benefit from improved recovery of pre-vious solutions to related problems. On the one hand, localised problem-solving is stimulated by the immediacy and tangibility of the problem tobe solved. On the other hand, localised problem-solving runs the risk ofproviding idiosyncratic solutions that ignore potentially valuable organ-isational experience. Smaller organisations, particularly those with rela-tively stable employment, can develop effective social referral networks so


that the same individual or group consistently deals with problems of asimilar type.45

Larger organisations, including those involved in the generation ofgreater variety, will face greater problems in creating social referral net-works. They face three challenges. The first is the challenge of identifyingthe salient features of a particular problem that make it ‘like’ some otherproblem the organisation has experienced. If the challenge of ‘likeness’ canbe met, the second problem is identifying the relevant information source.At best, this will involve the identification of a current employee of theorganisation who can service the referral (assuming that proper intra-organisational incentives are in place to do so). Perhaps as likely, part or allof the organisational memory regarding the particular problem has beendisassembled through departures or transfers so that it is not possible tofind the appropriate individual. These three challenges, ‘likeness’, identify-ing relevant referrals and recovering information with no active caretakerare among the most common ‘memory’ problems that medium- and larger-sized organisations face. The ways that these types of problems are resolvedis central to the prospects of decentralisation.

4.4 Knowledge Management: Control versus Empowerment

Eigentlich weiss man nur wenn man wenig weiss; mit dem Wissen wachst desZweifel. (We know accurately only when we know little; with knowledge doubtincreases.) (Goethe, Spruche in Prosa [Proverbs in Prose])

Each of the above instances of the use of information and communicationtechnologies in supporting decentralised formation of institutional stand-ards and incentives intersects with the field of studies that has come to beknown as ‘knowledge management’. The growing recognition of the intan-gible asset value of knowledge as a productive input and, in some cases, asan output of the organisation suggests the need for a management strategyto maximise this asset’s value. The controversy and struggles arising fromknowledge management recapitulate and surmise about the contestbetween centralised and decentralised strategies with regard to learningand knowledge in modern organisations.

Much of this struggle turns on an understanding of the meaning of theterm ‘knowledge’. Throughout this chapter, knowledge has been depicted asdeeply intertwined with learning and transactional or interactive experience.The alternative view is that knowledge can be meaningfully described as a‘stock’ that can be ‘held’ in some repository such as a data warehouse or inthe documentation produced by the organisation about its organisation andoperations. In the first view, knowledge is transitory and provisional, while,


in the second view, knowledge is an equilibrium state reflecting the attain-ment of a ‘best’ solution or understanding. The contention is that the growingcomplexity of modern organisations as well as the variety of their outputsand processes make the first view the most relevant. A version of the secondview, involves developing and employing a ‘best’ set of understanding andsolutions for the circumstances, for ‘coping’ and for using this knowledge toimpose routines and standard operating procedures on the workforce.

Which view of knowledge prevails has implications for the systems usedto gather, store and process information in the organisation as well as forthe applications of these information capacities to solve problems (know-ledge creation) and enable learning (knowledge exchange or reproduction).Within the community of scholars engaged with ‘knowledge management’there are clearly adherents to both viewpoints. As each of the viewpointsis also allied to a position on the issues of decentralisation it would becomforting to believe that competitive selection would test the respectivevalue of the viewpoints in practice. Unfortunately, there are real possibil-ities that the ‘best available’ (also called ‘best practice’) approach willprevail, despite the merits of decentralised options, as the ‘best practice’approach when used as a means for indoctrination serves preferences forpower and dominance.

It is also the case that organisational solutions involving hierarchicalcontrol and dominance are the most familiar. Decentralised solutions needto be examined critically and comparatively. For example, what organisa-tions choose to remember and what they forget should, in principle, be aninvestment decision. This immediately conjures up images of painfulreporting procedures in which individuals are required to maintain recordsof their problem-solving activities categorised and classified in ways theorganisation’s ‘knowledge managers’ believe will eventually facilitate theretrieval of relevant information and thereby a return on investment. Incontemplation of this prospect, it is important to confront this strategy witha decentralised solution based upon incentives and institutional standards.

There are three requirements for such a decentralised knowledge man-agement system. First, there must be rewards (positive incentives) for ‘pub-lication’ (disclosure) of knowledge as information that might be used byothers. Second, these rewards must create greater value for the first disclo-sure of the relevant information (as copying the available information is anopportunistic possibility). Third, the reward system must be funded in a waythat does not require the information user to bear the costs of the reward.Thus, this system cannot employ the common practice of establishing‘internal markets’ for consulting and other services within the organisationin which the user must bear the costs (and uncertainties) of utilising infor-mation produced by others. Such systems are often underutilised because


of the difficulties of assessing the information to be provided in advance.By separating rewards from utilisation the organisation encourages thediffusion of information and is, in effect, investing in the distribution ofinformation within the organisation.46

A single institutional standard is a necessary complement to these fea-tures of the incentive system. A norm of ‘recognition’ in using others ideasmust be established so that it is possible to attribute credit in a decentralisedway. Failing this, it is necessary to monitor the use of information withinthe organisation to ascertain credit.

This incentive system and institutional standard is, of course, virtuallyidentical to the one that Dasgupta and David (1994) propose lies at the rootof the social system known as ‘science’, one of the most powerful systemsdevised by humans for generating and distributing knowledge. By provid-ing incentives for disclosure and identifying ‘priority’ of ‘discovery’ (or con-tribution to the common knowledge), the system encourages others toconsider the current state of knowledge within the organisation and tohasten to add their own contribution lest someone else receive recognitionand the subsequent reward.47 Severing the use of knowledge from a require-ment to ‘pay’ for its use generates further incentives to look for a solutionrather than engage in the risks and incur the costs of trying to create a solu-tion (that is, to reinvent the wheel). The level of reward can be adjusted todiscourage ‘over-investment’ by individuals in disclosure, since the value ofonly a few uses may not justify the costs of submission.

There are, of course, some problems with this scheme as there are withthe social system of science that suggests it. The opportunity costs of theindividual’s time in making disclosures is a further investment in the schemeby the company and the collective opportunity costs plus the costs of therewards may exceed the benefits since, although rewards may be adjustedfrom time to time, they need to be set arbitrarily. The parallel problem inthe case of science is the problem of setting the overall level of funding forscientific endeavour (as, in science, it is the proposed investigation ratherthan the outputs that govern compensation48). There are also practicalproblems in packaging knowledge for disclosure in ways that prevent its‘leakage’ to rivals and in governing rewards for the creation of ‘incremen-tal improvements’ which may be just that or opportunistic attempts to sharethe rewards.49

The purpose of this stylised example is to illustrate that it is possible toview issues of knowledge management involving organisational memoryand learning as a problem of creating the corrective incentive structure andinstitutional standards for knowledge disclosure. It is certainly true thatmemory represents an important asset and learning a major investment inmodern organisations. It does not follow, however, that these assets and


investments can be readily committed to inventory or distributed within theorganisation, as are other corporate assets. Instead, it is necessary to bringproducers and users of different types of knowledge in the organisationtogether through a system of procedural authority (in this example repre-sented by the mechanisms of establishing and operating a ‘disclosure’process). Aligning the incentives of producers and users becomes the prin-cipal incentive design problem, one for which the social institutions ofscience provide an intriguing solution.

5 CONCLUSION

This chapter began with an examination of the relocation of learning andknowledge accumulation within the organisation with the aim of illumin-ating how these processes have changed over time and may be subject tofurther transformation. The principal examples utilised in the first half ofthe chapter stressed the historical processes by which localised learning andknowledge accumulation was displaced by a process of centralisationneeded to support the standard and flexible mass production of Ford andSloan respectively.

In the second half of this chapter, Sloan’s system of flexible mass pro-duction is portrayed as suffering from increasing stress in managing theprocesses of variety creation and problems shared by non-manufacturingorganisations such as services as well. The prospects for avoiding disec-onomies of scale in the management of variety are chosen as the principalfocus of analysis along with the specific contributions of information andcommunication technologies to the standards regarding procedures andartefacts within the organisation.

The information and communication technology examples highlight theconsiderable variability in the existing systems for managing the decentral-isation that is a concomitant of increasing variety. In general these systemscannot be characterised as either mature or well integrated for the purposesof managing decentralised learning or accumulation of knowledge, oravoiding diseconomies as variety continues to expand.

There remains considerable opportunity that has not yet been fullyexploited in such systems for the purposes of standardising artefacts andprocedures within the organisation as well as devising appropriate incen-tive systems that would address the problems of achieving greater productand service variety. Thus, it would be premature to conclude that the Sloanmodel of centralised design authority (at least at a divisional level) is facedwith a viable and currently available substitute. At the same time, it wouldappear that the ‘learning organisation’ represented by a decentralisation of


the processes of making intra-organisational standards has considerablefurther potential than has yet been exploited. This potential is not only rele-vant for further improvements in productivity and economic growth, it alsooffers fundamentally more democratic and participatory processes in theworkplace.

Whether the new ‘learning organisations’ created by these processes willovercome tendencies towards the assertion of dominance and power thatwould indoctrinate workers with ‘best practice’ and lock down the poten-tial of the new technologies remains to be seen. In their infancy, theseorganisations will be vulnerable to market pressure and the undevelopedstate of the information and communication technologies necessary tosupport them.

NOTES

1. A prominent contemporary theory for understanding this process is provided by Laveand Wenger (1991).

2. Such accounting is not usually made in economic theory where it is more straightfor-ward simply to argue that the firm understands its own technological capabilities andmay achieve an understanding of any other ‘technology’ (productive capability) that itdeems useful in the pursuit of profit. Coming to such understanding may represent afixed ‘getting started’ cost worth noting, but the details of the process and its corollariesare not further developed.

3. Arrow (1962) and David (1975).4. Schmookler (1966).5. One type of market structure consistent with the existence of multiple ‘niche’ players.6. This approach is a generalisation of some of the issues raised in David and Foray (1995).7. This is a good example of a theoretical abstraction that would be very difficult to measure.

The argument that follows is only based on the premise that such rankings exist, not thatthey can be empirically ascertained given the large number of co-determinants involved.

8. Among the most influential efforts to undertake this is Pavitt (1984).9. It is often argued that production engineering and operations are more integrated in

Japanese factories than in their Western counterparts and that this is a major source ofthe high rates of Japanese productivity improvement. Two poles defining the spectrumof views on this subject are (Dore, 1973) who reflects on the social consequences ofthis form of organisation and (Schonberger, 1982) who uncritically accepts its inherentsuperiority.

10. In addition, ancillary sociological issues related to the ‘professionalisation’ of certainclasses of employment may have created further impetus to this trend.

11. Rosenberg (1976).12. Smith (1937), Young (1928) and Stigler (1951).13. For example, Mumford’s (1934) monumental effort at identifying the origins and cri-

tiquing the consequences of industrialism.14. Mumford (1934), Giedion (1948), Hounshell (1984) and Beniger (1986).15. Filene (1925), as quoted in Hounshell (1984: 305).16. The following draws heavily upon Hounshell’s (1984) account of these events.17. Hounshell (1984: 174).18. Ozanne (1967).19. Henry Ford, ‘Mass production’, as quoted by Hounshell (1984: 217).


20. Among the texts often used are Nevins and Hill (1957), Abernathy (1978) andChandler Jr (1962).

21. Hounshell (1984: 267).22. Caminer et al. (1996).23. See Steinmueller (1996) for a more complete account of these developments.24. IBM retained a significant position in software production. What distinguished its later

activities in the software field from the earlier period was the lack of sustained commit-ment to producing software in any particular category of application. IBM’s softwareactivities in the 1980s and 1990s focused on what has come to be called ‘enterprise com-puting’, large integrated systems assembled from more generic sub-systems and compon-ents. IBM’s participation in any particular sector was, in this period, based upon thedemand for these services in the use of their computer systems rather than the effort topre-specify bundled packages of hardware and software for particular applications.

25. See Steinmueller (1996) for further details of this history. The emergence of minicom-puters was made possible by shortcomings in the ‘divisibility’ of computing power that,in turn, would have created yet another alternative path for the development of compu-tation. One of the ironies of historical development is that time-sharing related ideas arere-emerging in the use of the Internet. These new forms range from ‘net computers’ (per-sonal computers whose system and application software is regularly upgraded throughnetwork access) to the distributed control of logistic and production systems that useInternet technology and centralised computers.

26. David (1991).27. See Zuboff (1988) and Mansell and Silverstone (1996).28. See Pavitt and Steinmueller (2000) for a complementary analysis reaching a similar con-

clusion.29. A significant qualification to this point of view is the recognition that the relative prices

of tangible inputs may change over time, selecting against particular sets of productvariety in ways that were not initially anticipated. Although this is theoretically animportant issue for all tangible goods, it is often only of practical significance for rela-tively simple products that have many close substitutes and therefore relatively smallprofit margins.

30. Other demand side influences including saturation, fashion, habituation or novelty pref-erence may also involve learning elements.

31. See Gulledge Jr and Womer (1986) for an extended analysis along these lines.32. For example, the margin between the re-sterilisation of instruments and their disposal is

continuously closing.33. See Steinmueller (2000) for an overview of these issues.34. See Cowan et al. (2000).35. Hertog and Huizenga (2000).36. For example, see Ancori et al. (2000).37. Those who believe that the local knowledge of production workers is of particular sig-

nificance need to explain how this belief can be consistent with relatively high turnoverin the labour force.

38. Granovetter (1985) provides a general theory regarding the tradeoffs between interper-sonal trust and market exchanges. The issue of routine and non-routine transactions isexamined empirically in Hart and Estrin (1991). Kraut et al. (1998) examine the issue ofintra-organisational networks as well as providing a thorough conceptual review of theseissues.

39. Steinmueller (2000).40. ASSENT (1998).41. Open-source software involves the publication of the source code that is used to compile

executable computer programs. Typically, source code is the proprietary intellectualproperty of the software author as its analysis would allow the creation of functionallyidentical software by competitors. In the open-source model, alternative business modelsmust be developed to fund the costs of software development and current developments


have relied upon voluntary contributions of programmers. See Mateos-Garcia andSteinmueller (2003).

42. Jeong (1999).43. The design of such incentives is a difficult practical problem since it is desirable to

encourage learning and experimentation with innovative information resources as wellas the adaptation of existing organisational information resources to the electronicmedia.

44. See Steinmueller (1992) for a pre-World Wide Web examination of these issues.45. It is, of course, important to ask whether consistently good answers are found for such

problems.46. See David and Foray (1995) for the implications of this argument for science and tech-

nology investment from a social viewpoint.47. In practice, the efforts required to assign priority are best left primarily to the commun-

ity of users since it is undesirable to exclude incremental improvements to existing solu-tions. Instead, the role of ‘procedural authority’ (management in this case) should beconfined to linking suggestions that appear to be related so that first disclosure isrewarded even if later disclosures are more heavily utilised.

48. This practice is necessary given the often lengthy delay in making use of science and thedifficulty of ‘tracing’ these applications. In the example here, the practicality and instru-mentality of results suggest a shorter delay while the proximity of production and useindicate better ‘tracing’. The latter will ‘of course’ depend on how well the institutionalstandard (or norm) of recognition operates.

49. The parallel issues in science concern the release of interim research results thatmight allow rivals to make a larger claim of discovery or generalisation and the govern-ance system of peer review which helps define the significance of claimed advances inknowledge.

REFERENCES

Abernathy, W.J. (1978), The Productivity Dilemma: Roadblock to Innovation in theAutomobile Industry, Baltimore, MD, Johns Hopkins University Press.

Ancori, A., A. Bureth and P. Cohendet (2000), ‘The economics of knowledge: thedebate about codification and tacit knowledge’, Industrial and Corporate Change,9 (2), 255–88.

Arrow, K.J. (1962), ‘The economic implications of learning by doing’, Review ofEconomic Studies, 29 (3), 155–73.

ASSENT (1998), Assessment of the Telematics Application Programme (ASSENT),Assessment of the Results of the Projects, Telematics Application Programme,Deliverable 9.01, Brussels, European Commission.

Beniger, J.A. (1986), The Control Revolution. Cambridge, MA, Harvard UniversityPress.

Caminer, D., J. Aris, P. Hermon and F. Land (1996), User-Driven Innovation: TheWorld’s First Business Computer, London, McGraw-Hill.

Chandler Jr, A.D. (1962), Strategy and Structure: Chapters in the History of theIndustrial Enterprise. Cambridge, MA: MIT Press.

Cowan, R., P.A. David and D. Foray (2000), ‘The explicit economics of knowledgecodification and tacitness’, Industrial and Corporate Change, 9 (2), 211–54.

Dasgupta, P. and P.A. David (1994), ‘Toward a new economics of science’. ResearchPolicy, 97 (387): 487–521.

David, P.A. (1975), ‘Learning by doing and tariff protection: a reconsideration ofthe case of the ante-bellum US cotton textile industry’, in P.A. David, Technical


Choice, Innovation and Economic Growth, Cambridge, Cambridge UniversityPress, pp. 95–173.

David, P.A. (1991), ‘Computer and dynamo: the modern productivity paradox in anot-too-distant mirror’, in OECD (eds), Technology and Productivity: TheChallenge for Economic Policy, Paris: OECD.

David, P.A, and D. Foray (1995), ‘Accessing and expanding the science and tech-nology knowledge Base’, STI Review, 16: 13–68.

Dore, R. (1973). British Factory Japanese Factory. Berkeley, CA, University ofCalifornia Press.

Federal Reserve Bank of Dallas (1998), The Right Stuff: America’s Move to MassCustomization, The 1998 Annual Report, Dallas, TX: Federal Reserve Bank ofDallas.

Filene, E.A. (1925), The Way Out: A Forecast of Coming Changes in AmericanBusiness and Industry, New York, Doubleday.

Giedion, S. (1948). Mechanization Takes Command: A Contribution to AnonymousHistory, New York, W.W. Norton.

Granovetter, M. (1985), ‘Economic action and social structure: the problem ofembededness’, American Journal of Sociology, 91 (3), 481–510.

Gulledge Jr, T.R, and N.K. Womer (1986), The Economics of Made-to-OrderProduction Theory with Applications Related to the Airframe Industry, inM. Beckmann and W. Krelle (series eds), Lecture Notes in Economics andMathematical Systems 261, Berlin: Springer-Verlag.

Hart, P. and D. Estrin (1991), ‘Inter-organizational networks, computer integra-tion, and shifts in interdepdence: the case of the semiconductor industry’, ACMTransactions on Information Systems, 9 (4), 370–98.

Hertog, J.F.D. and E. Huizenga (2000), The Knowledge Enterprise, ImplementingIntelligent Business Strategies, London, Imperial College Press.

Hounshell, D.A. (1984), From the American System to Mass Production,1800–1932, Baltimore, MD, Johns Hopkins University Press.

Hughes, T.P. (1989). American Genesis, New York: Viking.Jeong, B.S. (1999), ‘Analysis of the Linux system, a new entrant in the operating

system market: technological innovations and business models, SPRU – Scienceand Technology Policy Research’, unpublished SPRU MSc dissertation,University of Sussex, Brighton.

Kraut, R., C. Steinfield, A. Chan, B. Butler and A. Hoag (1998), ‘Coordination andvirtualization: the role of electronic networks and personal relationships’,Journal of Computer Mediated Communications, 3 (4), http://www.ascusc.org./jcmc/ vol3/issue4/kraut.htm

Lave, J. and E. Wenger (1991), Situated Learning: Legitimate PeripheralParticipation, Cambridge, Cambridge University Press.

Mansell, R. and R. Silverstone (eds) (1996), Communication by Design: The Politicsof Information and Communication Technologies, Oxford, Oxford UniversityPress.

Mateos-Garcia, J. and W.E. Steinmueller (2003), The Open Source Way of Working:A New Paradigm for the Division of Labour in Software Development?,Falmer, East Sussex, SPRU Science and Technology Policy Research, http://siepr.stanford.edu/programs/OpenSoftware_David/NSFOSF_Publications.html

Mumford, L. (1934), Technics and Civilization, New York: Harcourt, Brace.Nevins, A. and F.E. Hill (1957), Ford: Expansion and Challenge, 1915–1933,

New York, Charles Scribner’s Sons.


Ozanne, R. (1967), A Century of Labor – Management Relations at McCormick andInternational Harverster, Madison, WI, University of Wisconsin Press.

Pavitt, K. (1984), ‘Sectoral patterns of technical change: towards a taxonomy anda theory’, Research Policy, 13, 343–73.

Pavitt, K, and W.E. Steinmueller (2000), ‘Technology, Strategy and the InformationSociety’, in A. Pettigrew, H. Thomas and R. Whittington (eds), Handbook ofStrategy and Management, London, Sage Publications.

Rosenberg, N. (1976), ‘Technological change in the machine tool industry,1840–1910’, in N. Rosenberg (ed.), Perspectives on Technology, Cambridge,Cambridge University Press, pp. 9–31.

Schmookler, J. (1966), Invention and Economic Growth, Cambridge, MA: HarvardUniversity Press.

Schonberger, R.J. (1982), Japanese Manufacturing Techniques: Nine Hidden Lessonsin Simplicity, New York, Free Press.

Smith, A. (1937), The Wealth of Nations, 5th edition (first published 1789),New York: Modern Library.

Steinmueller, W.E. (1992), ‘The economics of production and distribution of user-specific information via digital networks’, in C. Antonelli (ed.), The Economics ofInformation Networks, Amsterdam: North Holland.

Steinmueller, W.E. (1996), ‘The US Software Industry: An Analysis andInterpretative History’, in D.C. Mowery (ed.), The International SoftwareIndustry, New York, Oxford University Press.

Steinmueller, W.E. (2000), ‘Will new information and communication technologiesimprove the “codification” of knowledge?’, Industrial and Corporate Change,9 (2), 361–76.

Stigler, G. (1951), ‘The division of labor is limited by the extent of the market’,Journal of Political Economy, 59 (3): 89–90.

Young, A. (1928), ‘Increasing returns and economic progress’, Economic Journal,38 (152): 527–40.

Zuboff, S. (1988), In the Age of the Smart Machine, New York, Basic Books.


9. The economics of open technology:collective organisation andindividual claims in the ‘fabriquelyonnaise’ during the old regimeDominique Foray and Liliane Hilaire Perez

INTRODUCTION1

What we call ‘knowledge openness’ is a system in which the principles ofrapid disclosure of new knowledge are predominant, and in which anumber of procedures facilitate and reinforce the circulation not only ofknowledge such as that which is codified in instructional guides and docu-mentation, but also of tacit knowledge and research tools. It is not purechance that in this context new knowledge is codified and carefully sys-tematised in order to facilitate its transmission and discussion. But particu-lar attention is also paid to the reproduction of knowledge, that is, tolearning. It is not because knowledge flows freely – in the form of manualsand codified instructions – that it is necessarily reproduced from one placeto the next. It is also necessary to create and maintain relationships between‘masters and apprentices’, either in the context of work communities or informal processes of teaching practical knowledge. The significance ofknowledge openness is particularly important for knowledge, which is aninput for further cognitive works. In this case the principle of opennessallows external users of that knowledge to reproduce it for investigation,modification and improvement.

Systems of knowledge openness relate to public (or semi-public) spacesin which knowledge circulates. Such spaces can include areas in whichexclusive property rights cannot be granted, either constitutionally (as inthe case of open science) or within the framework of organisations spe-cially designed for the purpose (research networks where partners sharetheir knowledge) and markets whose modi operandi are conducive toefficient knowledge dissemination. In such circumstances, a fundamentaleconomic issue is the design of private incentives (to give credit to theknowledge producer) without creating exclusivity rights.

239

The economic analysis of ‘knowledge openness’ as a system has beendeveloped extensively in the field of scientific research owing to the seminalworks of Dasgupta and David (1994), David (1998; 1999) and David et al.(1999). The approach of the ‘new economics of science’ develops twoimportant arguments for theoretical analysis as well as policy implicationin the field of the economics of knowledge.

First, knowledge openness and sharing behaviours do not only expresssome kind of ethics or moral attitude (although ethical conviction certainlyplays a role). Knowledge openness is viewed, above all, as a mechanismgenerating economic efficiency that people in certain circumstances arewilling to implement and maintain in order to be players in a positive sumgame. In fact, knowledge openness that entails rapid and complete distribu-tion, facilitates co-ordination between agents, reduces risks of duplicationbetween research projects and functions as a sort of ‘quality assurance’ inso far as disclosed results can be reproduced and verified by other membersof the community. They are thus peer evaluated. Both static efficiency anddynamic efficiency are, therefore, expected to be enhanced: (1) ‘the wheel isnot re-invented’ and each ‘great’ invention will benefit from a strong col-lective focus on it; (2) propagating knowledge within a heterogeneouspopulation of researchers and entrepreneurs increases the probability oflater discoveries and inventions and decreases the risk that this knowledgewill fall into the hands of agents incapable of exploiting its potential (Davidand Foray, 1995).2

Second, knowledge openness does not mean the absence of individualincentives. There is a need for individual rewards, which are compatiblewith the complete disclosure norm. In the case of open science, a remark-able mechanism comes into play consisting of the granting of moral prop-erty rights which are not concretised in exclusivity rights.

These two features apply in the world of open science as well as in localsystems of open technology, such as the particular case of the ‘fabriquelyonnaise’ to which this chapter is devoted.

1 THE ECONOMICS OF OPEN SCIENCE

Private markets, even when equipped with a system of intellectual propertyrights, are ill-suited to the production and exploitation of certain forms ofknowledge. There is, thus, a need for some other economic institutions thatcan be relied upon to create and exploit knowledge in an efficient manner.One main institutional arrangement consists in financing knowledge pro-duction from public (or private) funds while at the same time identifyingmechanisms aimed at providing forms of self-discipline, evaluation and


competition within the beneficiary community. In return for aid received,the beneficiary is expected not so much to pursue objectives set by thefinancier, but rather to relinquish exclusive rights on knowledge produced.In concrete terms, society is responsible for covering the costs of resourcesneeded to produce knowledge. This means, however, that anything pro-duced is the property of society as a whole and cannot be privatelycontrolled. Knowledge is often disclosed through scientific publications,and since anything published can no longer be patented, it definitivelybecomes public knowledge (in the US system the grace-period mechanismallows patenting in the year following publication). Rapid communicationand sharing of knowledge are the norm, facilitating the creation ofco-operation networks.

Knowledge openness characterises, therefore, research undertaken inpublic institutions such as universities where in most cases exclusive rightscannot be granted on knowledge and where salaries and equipment arepaid from public funds. In many countries public funding of a large part ofthis system is facilitated by the close ties that exist between research andhigher education. As Arrow (1962) points out, the fact that research andteaching activities are two sides of the same profession is a ‘lucky accident’since it ensures that researchers are remunerated not on the basis of whatthey find (their income in that case would be highly irregular and only thebest would survive) but on that of regular teaching. It is because this publicsystem produces both knowledge and human capital that it easily harnessesa large proportion of public resources.3

1.1 A First Look at a Great Problem: Voluntary Spillovers and PrivateIncentives

Yet there is still a piece missing in this system. How can people be encour-aged to be efficient and effective researchers if their work is immediately dis-closed, without any possibility of private appropriation, and their salariesguaranteed? An ingenious mechanism comes into play here, consisting ofthe granting of moral property rights that are not concretised in exclusiv-ity rights (in other words, they are compatible with the complete disclosurenorm). It is the priority rule which identifies the author of the discovery assoon as he or she publishes and which thus determines the constitution of‘reputation capital’, a decisive element when it comes to obtaining grants.‘The norm of openness is incentive-compatible with a collegiate reputa-tional reward system based upon accepted claims to priority’ (David, 1998,p. 17). The priority rule creates contexts of races (or tournaments), whileensuring that results are disclosed. It is a remarkable device since it allowsfor the creation of private assets, a form of intellectual property, resulting

The economics of open technology 241

from the very act of forgoing exclusive ownership of the knowledge con-cerned. Here the need to be identified and recognised as ‘the one who dis-covered’ forces people to release new knowledge quickly and completely. Inthis sense the priority rule is a highly effective device that offers non-marketincentives to the production of public goods (Callon and Foray, 1997;Dasgupta and David, 1994).

Maximising knowledge externalities is the raison d’être of such systems(for instance, of an open science system). This is based on a set of consist-ent institutions: weak intellectual property protection; funding largely fromgovernment or private foundations; and a reward system (based on prior-ity) compatible with the fast and broad dissemination of knowledge.Moreover, management of externalities, namely, the organisation of accessto and integration of knowledge, is accomplished through norms and insti-tutions. For example, it is usual for researchers to write and share ‘surveys’aimed at making available to the rest of the community the state of theart of a particular domain. Nothing like that exists in the private propertysystem.4

Of course, the ideal world of openness described here does not excludethe possibility of bending or departing from the rules. On the contrary, thetournament contexts created by the priority rule, as well as the size ofrelated rewards, tend to encourage bad conduct. The notion of ‘openscience’ is therefore based on an ideal that is never achieved (in other words,there will always be many cases of various degrees of retention). InDasgupta and David (1994) it is argued that the norms are prescriptive, andthat beliefs that are instilled in scientists as part of the ‘culture of science’have an effect on their behaviour – making it easier to form co-operativenetworks where it is in their mutual interest (and that of society at large) toorganise research co-operatively.

1.2 Modelling Exercise

These ‘good properties’ have recently been modelled by David (1998), whoshows how the disclosure norm positively influences the cognitive perform-ance of the system under consideration. David models stochastic interac-tions in a group of rational researchers individually engaged in a continuousprocess of experimental observation, information exchange and revision ofchoices in relation to locally constituted majorities. This modelling is thenused to link micro-behaviours (being open, being closed) and macro-performances. Simulations suggest that the social norm of openness, whichinfluences micro-behaviours, favours free entry into knowledge networksand, in so doing, prevents researchers from closing in on themselves tooquickly and excluding different opinions. David shows that a system situated


beyond the critical openness threshold ensures confrontation of ideas andprovides a mechanism that guarantees the production of consensus and pre-serves the diversity of opinions. The capacity to produce scientific state-ments collectively while preserving a degree of diversity of opinions andarguments is thus an important feature in an open research network, andstandards of disclosure and openness appear to be decisive in the cognitiveperformances of the network. The advantage of such an approach is that itproduces formal results, derived from the mathematical theory of percola-tion, on the basis of which more political reflection can be envisaged:

● The size of the network is important. The smaller the network, thegreater the risk of it rapidly becoming trapped in one of those‘absorbing states’, namely, in a situation of complete agreement of allagents, from which it is difficult to collectively withdraw.

● The network can tolerate certain shortcomings and divergence fromthe openness norm. In other words, the same cognitive performanceis guaranteed as long as the network is above a certain critical thresh-old. Co-operative behaviour can emerge and be maintained withouteveryone complying perfectly with the openness standard.

2 THE ‘GRANDE FABRIQUE LYONNAISE’:KNOWLEDGE OPENNESS OUTSIDE THESCIENTIFIC FIELD

We have discussed ‘open science’ because it is probably the organisation ofscience that is closest to this standard of openness. Yet in the past there havebeen numerous cases of ‘open technology’, albeit limited in time and space.Historically, most situations of openness were linked to a specific territory:Lyons in the case of the circulation of techniques and inventions relatingto the silk industry (Hilaire Perez, 2000), Lancashire in the case of collect-ive invention in the metallurgical industry (Allen, 1983), the Clyde area inthe case of collective invention in shipbuilding (Schwerin, 2000), theCornish mining district in the case of collective invention related topumping engine technology (Nuvolari, 2004). More recent cases are thoseof mature industries (von Hippel, 1988) as well as emerging activities suchas virtual reality (Swann, 1999) or software.5

2.1 A Preliminary View of the Process of Collective Invention in Lyons6

Lyons was the second largest French town, with 143 000 inhabitants (1789)and 25 per cent were working in the silk industry (35 000). This huge sector


was fostering an important internal and foreign trade for luxury silk cloth.The dominance of the French silk industry was based upon changingfabric patterns according to taste and fashion, and upon research either toinvent new stitches or to set up (to ‘read’) more easily the drawings on thelooms in order to make the rich cloth as quickly as possible (setting up thepattern on the loom could take 25 days). Some inventions aimed to pro-gramme the patterns on the loom and to select warp threads (like Jean-Philippe Falcon’s), others were intended to change quickly parts of theweft thread to reduce their number (Philippe de Lasalle’s movable‘sample’) or to ease the pulling of the weft linked to the threads (JacquesVaucanson’s hooks). Two factors were critical to induce intensive inventiveactivities.

First, the cost of draw-girls (assistants) was a growing burden for guildfamilies. By mid-century, these girls, who came from nearby provinces, werealso very scarce.

Second, the speed and synchronisation of the work became the core ofinventions at the end of the century as the taste moved from brochés (heavysilk cloth with complicated patterns in gold and silver threads and manyshuttles) to façonnés (lighter cloth which could be manufactured withsmaller number of shuttles). This product, especially small façonnés, wasthe basis for successful research in suppressing the pulling of ropes.Jacquard loom (rewarded in 1804), which combined Falcon’s programmeand Vaucanson’s hooks, was intended for façonnés.

Nearly all Lyonnais’ inventions, which were addressed to the commerceoffice in Paris from 1700 to 1789, were related to the silk industry (181 in265, for Lyon), and more precisely to weaving (116) (generally new devicesof looms either for brochés or façonnés) and they occurred mostly after1730. Lyonnais’ artisans also represented a high proportion of inventorsapplying to the government: there were 170 inventors from Lyons, in a totalof 875 inventors (from all crafts) addressing the office of commerce.Inventor members of the Grande Fabrique were 73 and only 12 of themwere large merchants.

2.2 Institutions Promoting Knowledge Openness

This innovative context was sustained by local institutions, traditionallyinvolved in the management of innovation, since the sixteenth century, bythe means of local monopolies granted in ordonnances consulaires andfinancial rewards. In the eighteenth century, few monopolies were granted,and there was a reward fund officially established, the Caisse du droit desétoffes étrangères, created in 1711 (from a tax upon foreign silk) andintended to promote industry since 1725.


This particular mechanism was designed to reward inventors who agreedto disclose their knowledge and actively to participate in the diffusion ofthat knowledge (teaching). The setting up of a reward fund, the process ofexamining inventions and the system of financial bonuses awarded to thosewho agreed not only to disclose but also to teach their knowledge wereinstitutional mechanisms which made the system very effective. The systemof bonuses shows how well the conditions for an efficient reproduction ofknowledge, once created, were understood.

From 1752, the Intendant was at the head of the Caisse and the proced-ure involved the business community, the local council and the academy ofLyons. The management of the Caisse was based on a contradictory proofprocedure, contrived for getting more information about the invention con-sidered, so as to reduce uncertainties and secure the public investment. Thisprocedure was unique in France since it actually institutionalised the plur-ality of judgements as a method of governance: there was a double processof examination running in parallel; one involving the Intendant and amember of the academy and another involving the local council and theguild inspectors. This double procedure of examination resulted in stimu-lating exchanges between various kinds of competences, and compelledstakeholders to negotiate the rewards as they often reached contradictoryconclusion and to mobilise their own networks. The bonus system was fos-tering contacts between guild inspectors and artisans as there were manyvisits in the workshops to quantify the spread of the new looms.

Let us take the example of Michel Berthet, who, inspired by Falcon,invented a loom for easing the work of the draw-girls (an essential matterin the Lyonnaise silk industry): in 1760, the Intendant de la Michodièreagreed with the academician de Goiffon to grant him £1000: £600 immedi-ately and the rest of the sum if he taught the other artisans how to use thenew loom and if four of his looms were put in other workshops. In 1765,Berthet made a new technical improvement and the Intendant proposed£1500 in exchange for the secret and for setting up some of these improvedlooms in town. The Intendant compelled Berthet to deposit a model and adescription at the Fabrique’s office. The grants were not only rewarding thepresumed economic utility of inventions; they reflected the efforts of theinventor for sharing his knowledge within the whole community.

Thus, secrecy was actively opposed. There were few monopolies for inven-tion in Lyon: nine affairs ended with a patent, concerning seven inventors.And seven of these patents were granted before 1750. The Lyonnais elitespreferred to invest in innovation, to make inventions a common wealth, andthis was not just a fancy ideal, as the rewards were often bonuses based uponthe spreading of the inventions within the town.7 Each inventor was encour-aged to be a dynamic actor collaborating for the innovation diffusion and


the official credit (financial and symbolic) for the invention involved thechoices and decisions of the potential adopters.

3 OPEN SCIENCE AND OPEN TECHNOLOGY

The historical analysis of open technology – in the particular case of the‘fabrique lyonnaise’ – allows us to draw a parallel with the economics ofopen science.

3.1 The Basic Ingredient of Knowledge Openness

EthosIn both systems some kind of collective ethos is present, generating a sortof ‘natural’ inclination of inventors to diffuse their knowledge. In the caseof Lyons, such an ethos can be seen both at the policy/administrative leveland at the individual inventor level.

At the policy level, the municipality, following the ancien régime tradition,kept on rewarding inventions to put them into the public domain. Muchhope was placed in the free circulation of knowledge; even a slight improve-ment could bring about huge effects because all trades were viewed as natu-rally interdependent. Exclusive rights were rejected in favour of grants andbonuses for spreading knowledge and teaching. Inventors were rewardedfor the practicality of their inventions and the valuation of inventionsinvolved complex negotiations on applicability between officials and users.Liberalism, in the administrators’ eyes, meant growing exchanges betweenautonomous agents as a means of reinforcing social cohesion. Thus, col-laboration and collective inventions were strong, because very ancient pat-rimonial policy was reaffirmed by new ideals and practices. This agreementbetween artisans, merchants and elites was essential: new techniques shouldnot bring tensions nor disorders but, on the contrary, they should cementthe social cohesion through knowledge sharing and collective emulation.

At the individual level, some inventors were emblematic of this naturalinclination to reveal knowledge freely. The best example is Philippe deLasalle’s career path (1723–1804). De Lasalle was very famous in the eight-eenth century, in France and abroad, and he was largely rewarded by theGrande Fabrique and the city of Lyon (£122 000). The Lyonnais elite cher-ished him but he devoted his effort to the progress of the whole community.Enlightened administrators like Trudaine’s son and Turgot, and writers likeVoltaire, were friends of his. He belonged to the republic of arts and lettersas well as to the economic world. What he did and what he thought derivedfrom general ideals and principles he was eager to realise.


He began by learning drawing from local painters and became adraughtsman and a merchant. He was rewarded from 1758 by a pensionfor excelling in halftones for floral patterns. He also imitated tiger fur in silkcloth and he innovated by printing silk cloth like calicos. Soon after, in1760, he was asked to teach drawing in the Fabrique and his pension wasenlarged. Ten years later, his inventions for accelerating the changing ofpatterns on the looms (reversible loom and movable ‘sample’) increased hispension and he gained a bonus for helping to spread use of his looms. Aftercreating machine tools to assure better diffusion of his looms, he wasgranted £6000 in 1778. According to the administrators and to de Lasallehimself, artistic creativity, technical invention and transmitting knowledgewere closely connected. Collaborating and imitating were the main prin-ciples everywhere and the only ways to progress. Art and invention restedon a cumulative process, methods, rules, devices, lines and colours to belearnt side by side with the master, teacher, contriver or nature itself. DeLasalle had created a garden in the South of France where he sent his beststudents to train in drawing flowers. For him, there was no genius withoutcopying:

You are not unaware that art is learned through emulation and great examples.Work and my observations of the works of those who have distinguished them-selves in the career that I follow have shaped my talents. Even more ardour towarrant the protection that you grant them can afford them one day thatcelebrity which offers models to imitate and stimulates other geniuses to outdoit. Thus, amongst us, as soon as a striking piece has left the hand of a skilledartist, it is lifted up to be seen by all rivals seeking the means to acquire it, andoften provides, by its character, either the season’s fashion or the example of abeautiful subject. When in 1756 I treated a tiger skin worked with a touch of arton a golden background, one witnessed budding in each workshop tastefuldrawings representing diverse furs. The same happened on other occasions whenI introduced landscapes, birds and people.8

De Lasalle would not condemn the theft of patterns or inventions; his aimwas the circulation of knowledge and the progress of qualifications whichcould result. He was even pleased when his printed silk cloth was copiedand his workers seduced by rivals. All means were good if diffusion wereat stake: teaching, imitating, stealing and, not least of all, deeds and freeoffers. Several times, de Lasalle gave away inventions and taught about hisnew device without asking for anything in return. In 1760 he was offereda £200 bonus for each student he taught, but he refused and preferred tooffer all his knowledge freely: ‘it appears . . . that he gives up the gratifi-cation of 200 pounds for each of his six students and, moreover, that heteaches them everything learned from years of experience’ (note of theLyons’council, 1760). How such an ethos appears and becomes forceful is


a broad question, addressed for instance by Hilaire Perez (2000) in thecase of the fabrique lyonnaise.

Collective belief of being part of a positive sum gameSimilar collective belief, in both open science and open technology, of beingpart of a positive sum game plays a key role as well. A common knowledgethat open technology is a positive sum game was particularly effective and‘had force’ in the case of Lyons since the city was engaged in internationalcompetition with London and the inventors knew full well that the pros-perity of the local system to which they belonged directly influenced theirown individual prosperity.

Individual rewardsBoth collective ethics and common knowledge about the efficiency of opentechnology are not enough to sustain a system based on the free dissem-ination of knowledge. There is a need for some kind of mechanism aimingat rewarding inventors without granting exclusive rights. We have presentedabove the particular mechanisms that were designed in Lyons to rewardinventors who agreed to disclose their knowledge and actively to partici-pate in the diffusion of that knowledge.

3.2 System’s Efficiency

The efficiency of systems of open technology is similar to the efficiency ofopen science: both are a way to increase the performance of a system ofinvention by making the existing stock of knowledge more socially useful,through improved transfer, transformation and access to the existing innov-ations. In Lyons a good example is the diffusion of the Jacquard loom.

Massive diffusion of new technologiesFor the nineteenth century, A. Cottereau (1997) documented the massivediffusion of the Jacquard loom in Lyons (20 000 existed at the mid-century)and he compared this success to London where ‘sweated’ labour conditions,specialisation and private strategies obstructed its dissemination.

Cottereau (1997) explains that the London and Lyons silk manufactur-ing were based on a similar amount of machines: there were 12 000 loomsin London in 1815 and 14 500 in Lyons. Though London could competewith Lyons between 1790 and 1810, because the French revolutionary crisisdisrupted production for a while, the London silk industry began to declineas French enterprises used Jacquard’s loom to make a revival of sophisti-cated and varied silk production. In London, only 5000 looms could befound in 1853; in Lyon, there were 30 000 (and 30 000 more in rural areas


outside the town). Before the First World War, French production exceededthat of England and most of it was exported, while England was import-ing substantial amounts of silk fabrics for home consumption.

In London, Jacquard did not spread and generally speaking, there werenot many inventions in the London silk industry (Cottereau even speaks ofthe ‘backwardness of all British handlooms’). As for the Jacquard loom, itsintroduction gave rise to a frantic race between important manufacturers;one, G. Wilson, succeeded, was allowed to keep the secret of the machine,took a patent in 1821 and did not sell the invention or build new looms.Cottereau does not mention any use of licences, though they existed in thecotton industry which was the model referred to by silk manufacturers.

Cumulative knowledgeJacquard’s invention matched the needs of the Lyonnaise silk industry. Thenew loom immediately spreads and this creates a mental mobilisation anda collective focus resulting in several useful improvements. Other loom-builders made hundreds of Jacquard’s loom, compared to the first inventorwho built only 57. These networks were the basis for the pattern of innov-ation in Lyon. Inventive artisans, either weavers or not, were quicklyinformed of new devices; they watched working new looms, listened toweavers, talked with other artisans, worked on rewarded looms and con-trived improvements to them. The open system generated huge cumulativeeffect. For example, already in 1759, Berthet had presented an improvementof Falcon’s 1742 loom. In 1765, he said he had improved the new Falcon’sloom he had just acquired. And the Jacquard’s invention itself was oftencalled ‘Vaucanson–Falcon–Breton’ to show how important was knowledgerecombination in the production of this major invention. Jacquardreally stood on the ‘shoulders of giants’. There were, in Lyons, many otherexamples of cumulative progress involving successive improvements of anew loom. Moreover, inventors like Falcon and de Lasalle kept improvingtheir own devices. One invention was never definitive but always evolving,and these improvements were encouraged by the local council, which, forinstance, blamed Jacquard’s disinterest in amending his own loom.

Technical standards and intergenerational compatibilityAnother positive effect of knowledge openness was the establishment oftechnical standards. The historian Cottereau (1997) found an essay writtenin 1863 describing the networks of newly invented looms in Lyons: ‘Themost convincing proof that these successive inventions were borrowed fromone another is that a Jacquard card in use today may be applied both toVaucanson’s planchette with needles and to Falcon’s, and the match is sogood that Falcon’s initial matrix must have fixed dimensions’ (Cottereau,


1997, p. 143). According to Cottereau, the effects ‘were comparable to whatcould easily have been the case today if computer systems had been stand-ardised from the start and made cumulatively compatible as they pro-gressed’ (ibid., p. 142), even if contrived by several different firms.

Then, collaboration and open technology in Lyon was highly efficient forthe spreading of inventions, for sharing technical innovative culture and forhelping research in craftwork. This model had fostered a professional eliteof indefatigable researchers, skilled inventors and artisans, ‘artists’, whooften devoted more time to research than to their own business. Fame,excellence and performance were these inventors’ aims. But what was theboundary with self-strategies? Although Cottereau describes an equilib-rium in the Lyon industry, conflicts and private interests were very harsh.In a paradoxical way, collective innovation did usher in a disruption ofcommunity ethos; it did foster a burst of opportunism, and, most of all,claims for priority and posterity amongst inventors.

3.3 The Metaphor of the Jacquard Flight

In both cases – open science and open technology – the reward systemintroduces competition and increases the risk of disputes. Then the force ofethics as well as the effectiveness of the common knowledge about theefficiency of the system come to the fore to mitigate individual misconductsand frustrations.

Indeed, the ‘collective fabrique’ appears very fragile, and somewhat vul-nerable to individual claims, frustrations and hopes. Jacquard agreed ini-tially to give up his rights to patents and ‘left the fruit of his art to thecommunity’ (Cottereau, 1997, p. 151). The invention became the propertyof the town and quickly spread. But later on, Jacquard started to complainthat the Lyonnais administrators had not treated him well enough, con-sidering the importance (‘the social return’) of his invention. A conflictarose between the great inventor and the local council, which compelledhim to stay in Lyons, fearing that he would sell the invention to competi-tors. In 1814, Jacquard left Lyons to go to Paris where he wanted to patenthis invention. The police of Lyons were urged to take him back and tocheck if he had transmitted his invention to rivals!

The history of Jacquard’s flight is a good metaphor to capture all situ-ations where the coexistence of different incentive systems makes fragilethose not based on private property.9 There was a degree of fragility in theLyons system of knowledge openness, especially when areas close by (Paris,in this case) offered inventors the possibility of obtaining a patent. This lastpoint is particularly important. Apart from the beauty of systems of col-lective invention and the fine economic performance such systems can


produce, the individual incentive dimension remains decisive and calls forsubstantial institutional mechanisms to give credit to inventors withoutgranting them exclusivity. This is the kind of mechanism Dasgupta andDavid have explored in the case of open science and which remains uncer-tain in the case of open technology, although the case of the ‘fabrique lyon-naise’ provides some ideas about how such credit and reward incentivesmight need to be structured in order to support an ‘open technology’ envir-onment. The prize system is an efficient mechanism because it createsincentives while keeping the knowledge in the public domain. However, theamount of the reward should be equal to the social surplus afforded by theinvention. As this case shows, the ex ante prediction of the social value ofthe invention is not a trivial condition.

4 CONCLUSION

In this chapter we tried to show that there was a golden age in the regula-tion of the production and distribution of silk loom technology in Lyonwhich resembled the open science structure but made specific accommoda-tion with the practical needs of inventors to be rewarded in excess of whatthey might earn by retaining their knowledge as a secret. The result waseffective diffusion and economic growth for the whole local system ofinnovation. This system then began to break down in the nineteenthcentury as merchant pressures of various sorts increased and as verticalclass conflict (between masters and merchants) became more important.

This story also raises the issue of the type of knowledge which is relevantto the kind of reward system that has been described here.10 This wasbecause of the ability to implement the knowledge investment in superiorlooms that inventors were rewarded. Thus, la fabrique lyonnaise can bequalified as a collective enterprise in which the rewards were distributedaccording to contribution to the collective benefit. This would suggest thatthere was a distinct bias in the types of knowledge that could be relevant tothe reward system, and another possible source for decay. When it becomespossible to make inventions through knowledge rather than the practice ofthe mechanical arts, it is no longer possible for the community to retainexclusivity over the machinery. Indeed, the machinery now becomes subjectto an innovation process that lies outside the direct and co-evolving experi-ences of machine construction and use. If true, then it would follow that apeculiar condition of the ‘open technology’ is the necessity for the divisionof labour to be limited so that the externalities of knowledge production(invention) can be captured as a local externality. Once the possibility ofbroader externalities comes into play the dynamic interaction collapses and


we are reduced to the model of commodified knowledge which is what thepatent system offers.

We hope that this study opens new research avenues – historical as wellas analytical – about these classes of mechanisms which allowed credit tobe given to individual inventors while supporting strongly the disclosure aswell as the reproduction of knowledge; those mechanisms which strictlygovern the solidity and stability of open systems, as Paul David so clearlydemonstrated in the case of science and academic research.

NOTES

1. The present version of this chapter has benefited from comments received on an earlierdraft, especially those from Cristiano Antonelli, Paul A. David, Alfonso Gambardella,Bronwyn Hall, Jacques Mairesse and Peter Swann at the conference ‘In honour of PaulA. David’ (Turin, May 2000). We are grateful to all the aforementioned and we are par-ticularly grateful to Ed Steinmueller who made numerous suggestions to clarify theexposition and greatly help us to improve the final draft.

2. In economic terms, since the marginal cost of use of knowledge is nil, maximumefficiency in its use implies that there is no restriction to access and that the price of useis equal to 0. Knowledge should be a ‘free’ good; that is, the condition for optimum useof a non-rival good.

3. Note that the union between research and teaching is not always maintained, which iswhat determines the partition of the public research system between universities andnational (or regional) laboratories.

4. This argument comes from an oral comment by Iain Cockburn made at an OECD/CERImeeting (Paris, July 2001).

5. As defined in a recent research project (Cassier and Foray, 1999; Foray and Steinmueller,2002) the type of open knowledge we are dealing with in this chapter is different fromthe collusive and explicit forms of collective invention (such as high technology consor-tiums) which require explicit co-ordination mechanisms as well as the formalisation ofagreements on both the distribution of tasks and the attribution of results. Moreovercollusive forms delimit semi-private areas for the circulation and pooling of knowledge,which may in some cases be less open than informal networks we are studying here. Themain difference between these two types of collective enterprise deals with the mode ofproduction of knowledge. In the cases studied here, trading or sharing concerns know-ledge that is already available. The participants do not participate in a co-ordinatedresearch project; they trade or share existing technical data. This is an incrementalprocess based on the dissemination and reuse of knowledge available within a group offirms. In the case of collusive and explicit forms of collective invention, the actors engagein operations of knowledge production.

6. Our analysis (extensively presented in Hilaire Perez, 1994; 2000) is based upon archivalsources that were not fully exploited in the previous studies on the fabrique lyonnaise(Cottereau, 1997; Poni, 1998): the letters and reports relating to eighteenth-centuryLyonnais inventors’ claims for grants and privileges were adjudicated both in Lyons andin Paris by the Bureau du Commerce.

7. For instance, in 1760, Ringuet presented a new loom for brochés which imitated paint-ings and embroidery; he was granted a £300 (livres tournois) bonus for the 10 first loomsset up, £200 for the next 10 looms and £100 for the 100 next ones during 10 years. He wasvery successful: as early as 1760, he had set up the first 10 looms; in 1762, the next 10 andeven 17 more; in 1763, 47 others and, in 1764, 85. Thus Ringuet had even passed the


quota (169 instead of 120, and in less than 10 years). He was paid for all the looms, eventhe ones which were not planned in the grant (£19 900 instead of £15 000).

8. Translation by Liz Carey-Libbrecht. Quoted in Hilaire Perez (2000), p. 76.9. Thanks to Bronwyn Hall for her help in deciphering this case as a very powerful

metaphor for the economics of knowledge production and diffusion.10. We are grateful to Ed Steinmueller who so well captured this issue and shared it with us.

REFERENCES

Allen, R. (1983), ‘Collective invention’, Journal of Economic Behavior andOrganization, 4, 1–24.

Arrow, K.J. (1962), ‘Economic welfare and the allocation of resources for inven-tions’, in R.R. Nelson (ed.), The Rate and Direction of Inventive Activity:Economic and Social Factors, Princeton, NJ, Princeton University Press.

Callon, M. and Foray, D. (1997), ‘Nouvelle économie de la science ou socio-économie de la recherche scientifique?’, Revue d’Economie Industrielle, 79,13–35.

Cassier, M. and Foray, D. (1999), ‘The sharing of knowledge in collective, spontan-eous or collusive forms of invention’, Colline WP 02, IMRI, University ParisDauphine.

Cottereau, A. (1997), ‘The fate of collective manufactures in the industrial world:the silk industries of Lyons and London, 1800–1850’, in C.F. Sabel and J. Zeitlin(eds), World of Possibilities: Flexibility and Mass Production in WesternsIndustrialization, Cambridge, Cambridge University Press.

Dasgupta, P. and David, P.A. (1994), ‘Toward a new economics of science’,Research Policy, 23 (5), 487–521.

David, P.A. (1998), ‘Common agency contracting and the ‘emergence of “openscience” institutions’, The American Economic Review, 88 (2), 15–21.

David, P.A. (1999), ‘Patronage, reputation and common agency contracting in thescientific revolution: from keeping “nature’s secret to the institutionalization ofopen science” ’, All Souls College, Oxford, December.

David, P.A. and Foray, D. (1995), ‘Accessing and expanding the science and tech-nology knowledge base’, STI Review, 16, 13–68.

David, P.A., Foray, D, and Steinmueller, W.E. (1999), ‘The research network and thenew economics of science: from metaphors to organizational behavior’, inA. Gambardella and F. Malerba (eds), The Organization of Inventive Activity inEurope, Cambridge: Cambridge University Press.

Foray, D. and Steinmueller, W.E. (2002), ‘On the economics of R&D and techno-logical collaborations: insights from the project Colline’, Economics of Innovationand New Technology, 12 (1), 77–97.

Hilaire Perez, L. (1994), ‘Inventions et inventeurs en France et en Angleterre auXVIIIè siècle’, Doctorat de l’Université de Paris I, Atelier National deReproduction des Thèses, Lille III.

Hilaire Perez, L. (2000), L’invention technique au siècle des lumières, Paris, AlbinMichel.

Nuvolari, A. (2004), ‘Collective invention during the British industrial revolution:the case of the Cornish pumping engine’, Eindhoven Centre for InnovationStudies.


Poni, C. (1998), ‘Mode et innovation: les stratégies des marchands en soie de Lyon,XVIII%’, Revue d’Histoire Moderne et Contemporaine, 45–3, 25–58.

Schwerin, J. (2000), ‘The dynamics of sectoral change: innovation and growth inClyde shipbuilding, 1850–1900’, 8th International J.A. Schumpeter SocietyConference, Manchester, June.

Swann, P. (1999), ‘Collective invention in virtual reality’, Colline WP 05, IMRI,University Paris Dauphine.

Von Hippel, E. (1988), ‘Trading trade secrets’, Technology Review, February/March, 58–64.


10. Measurement and explanation ofthe intensity of co-publication inscientific research: an analysis atthe laboratory level*Jacques Mairesse and Laure Turner

1 INTRODUCTION

Since the scientific research system has become an essential sector in ourmodern knowledge-based economies, an important new research fieldhas opened up. The challenge is to illuminate the role of scientific institu-tions in the production, diffusion and transfer of knowledge and that ofscience in economic development and social welfare. The ‘new economicsof science’ therefore is interested in a variety of issues concerning the func-tioning of scientific institutions, the labor market, training and careersof scientists, their productiveness, the allocation of public funds to basicresearch, the design of intellectual property rights, and so on. It thus con-tributes to the understanding of the organization of science and of ways itcan be improved (Dasgupta and David, 1994; Gibbons et al., 1994;Diamond, 1996; Stephan, 1996; Callon and Foray, 1997; Shi, 2001; Foray,2004).

The analysis of co-publications between scientists presented in this con-tribution is in keeping with the main focus of the economics of science onknowledge production, and is part of a broader study of the determinantsof scientific research productivity. We believe that membership in a dynamicand productive laboratory favours collaboration between researchers andimproves their own individual productivity, and that it may be part of aprocess of cumulative advantage by which these researchers enhance theirproductivity and reputation.1 Given the substantial increase in the propor-tion of co-authored articles, it also seems that the relevant units of know-ledge production tend to be more and more specific networks of researchers,whether they belong or not to the same institutions and/or countries(Gibbons et al., 1994).

255

In the economics of science, until recently, the literature on the interac-tions that favour knowledge production and diffusion primarily concernsgeographic externalities. Authors have mainly studied such externalitieswithin industries or from universities and other public research institutionsto firms and industries, relying on the analysis of patent data.2 Our workmoves upstream to study knowledge externalities within the scientificresearch system using co-publication data. We wish to look beyond theobservation of the spatial dimensions of research activity to investigate thedeterminants of the occurrence and intensity of collaborative relationsbetween researchers. Audretsch and Stephan (1996) have done similar workbut, again, concerning the relations between public research and industry.Based on data on academic scientists collaborating with US biotechnologyfirms, they show that such collaboration between firms and researchers ismore likely when the researchers have a good academic reputation, whenthey belong to a geographically extensive network, and when they areinvolved in practice in the transfer of knowledge towards the firm (as par-ticipants in the creation of the firm or as members of the ScientificAdvisory Board). Regional and local characteristics also seem to influencethe strength of relations between scientists and firms.

In the sociology of science and in bibliometry, a number of studies havealready highlighted some of the factors facilitating collaboration withinacademic research (see Beaver and Rosen, 1978 and 1979, and Katz, 1994,for a summary presentation). They include, above all, researchers’ reputa-tion and visibility, the need to access or to share the use of specific researchinstruments and facilities, the increasing specialisation in science and geo-graphic proximity. Two types of analyses can be found, however, in this lit-erature, depending on their explicit or implicit conception of a network(Shrum and Mullins, 1988). In one line of analysis the actors in networks areidentified through their interrelations, being mainly differentiated by theirdifferent positions in the structural configuration of their networks (forexample, whether they occupy a central position or not), not by their indi-vidual characteristics such as age, gender or skills.3 By contrast, the secondline of research takes into explicit account the status, capacities and strate-gies of actors, and it is these individual characteristics that mainly determinethe position of agents in networks and the nature of interactions betweenthem.4 Yet it would be desirable to be able to include in the same analysisstructural and individual elements as determinants of network interactions,and particularly for collaboration in research. Knowledge production anddiffusion are based on the interactions of multiple agents and institutionswith diverse interests: scientists in public and private laboratories, firms,financiers, public authorities, and so on (Callon, 1999). Investigating theexistence and intensity of collaboration between researchers in relation to


their specific characteristics should afford insight into the various mecha-nisms at play.

In this chapter we present the first results of such an attempt. We proposean intensity measure of collaboration between researchers, which has anintuitive interpretation and can be simply aggregated to the laboratory levelor higher levels of aggregation. Our unit of analysis in this contribution isthe laboratory and the group of laboratories at the geographic level of a town(which to be short we will call ‘town’). Our purpose is to explain measureddifferences of intensity of collaboration as revealed by co-publications byvarious potential determinants: precisely the geographic distance betweenlaboratories, their thematic specialisation, their size, their productivity interms of average number of publications per researcher, their quality interms of average citation impact factor per publication, and their inter-national openness.5 In particular, to what extent does the geographic dis-tance between researchers and their laboratories strongly impede, or not,their scientific collaboration?

Our approach is basically descriptive. We measure the intensity of co-publications among the researchers of the French Centre National de laRecherche Scientifique (CNRS) in the field of condensed matter physics,during the six-year period 1992–97.6 We first estimate the intensity of co-publication among these researchers, within their laboratories and betweenthem, and also within and between the towns in which these laboratoriesare located. Next we consider by means of simple correlations the possibleinfluence of geographical distance and other determinants on the occur-rence and intensity of co-publication. We then try to better assess the spe-cific impacts of these different factors by estimating their relative weight ina regression analysis.

The chapter is organised as follows. In section 2, we give necessary infor-mation on the scope of our study, the construction of our sample, and somedescriptive characteristics of co-publication. In section 3, we define ourmeasure of intensity of collaboration, giving a detailed example of its com-putation. In section 4, we present our correlation and regression results,and comment on what they tell us of the respective importance of thevarious determinants of co-publication we have been able to consider. Webriefly conclude in section 5.

The intensity of co-publication in scientific research 257

2 SCOPE OF THE STUDY AND GENERALCHARACTERISTICS OF CO-PUBLICATION

2.1 Scope of the Study: The Collaboration between CNRS Researchers inCondensed Matter Physics

In this chapter we study the determinants of collaboration among a groupof 493 physicists belonging to the condensed matter section at the CNRS,over the six-year period 1992–97. This sample consists of practically all theCNRS physicists in this field who were born between 1936 and 1960 andwere still working at the CNRS in 1997.7 Condensed matter physics inves-tigates, at various scales (atom, molecules, colloids, particles or cells), allstates of matter from liquids to solids in which molecules are relatively closeto each other. It is based on a heritage of traditions, both experimental(crystallography, diffusion of neutrons and electrons, magnetic resonanceimagery, microscopy, and so on) and theoretical (solid state physics). It hasrecently developed a closer relation with industry, contributing to the devel-opment of materials used in electronics, plastics, food or cosmetic gels, andso forth. We chose condensed matter physics for three main reasons. First,the characteristics of this field are particularly well suited to our study: it isa domain of basic research, which is clearly defined and where the journalswith a sound reputation are easily identifiable. Second, condensed matteris a fast-growing field, honoured by the Nobel Prize for Physics awarded toPierre-Gilles de Gennes in 1991, and currently accounting for close to halfof all French research in physics. Third, there is relatively little mobilityamong CNRS researchers outside of the field to other fields of research inCNRS, or out of CNRS towards academia or industry.

The sample of 493 physicists studied here represents a majority of allCNRS researchers in the field. The CNRS and higher education institu-tions are the only public research institutions in this field in France. In 1996,there were a total of 654 condensed matter physicists in CNRS, as against1475 in universities and ‘Grandes Ecoles’ (Barré et al., 1999).

The fact that our study is limited to researchers belonging to the sameinstitution, the CNRS, comes as an advantage. It implies a strong organisa-tional proximity between the researchers, characterised by the sharing ofcommon knowledge and implicit or explicit rules of organisation that favourinteraction and co-ordination (Rallet and Torre, 2000; Foray, 2004). Becausethey all belong to the same scientific community within the same institution,they work in a context directly conducive to co-operation that does notinvolve prior agreement on rules of behaviour. The existence of such strongorganisational proximity thus makes it possible to isolate more clearly theeffects on collaboration of geographic distance proper and other factors.


The indicator of collaboration that we use in this study is co-publication.It seems to be a reliable indicator of collaboration without being an exhaus-tive measurement, in so far as collaboration can have results other thanpublications. Our database has been compiled on the basis of all the pub-lications drawn from the Science Citation Index (SCI), for 518 CNRS con-densed matter physicists over the period 1992–97, of whom 493 publishedat least one co-authored article during this six years.8 Of the remaining 25physicists, 21 published no articles in this period, and the other four pub-lished only a total of five non-co-authored articles. Collaboration appearsto be the main mode of publication for the 493 researchers. Only 132 ofthem also wrote articles without co-authors over the period (for a total of252 articles) and, from the total corpus of 7784 articles they wrote over theperiod, 7532 (97 per cent!) are co-authored.

In order to improve the measurement of the intensity of collaboration inour analysis, we thought it appropriate to weight co-authored articles inproportion to the number of pairs of co-authors they involve. In otherwords, we simply chose to study the network of collaboration ‘link by link’,that is, by pairs of co-authors. In practice, this means that an article appearsin the database we constructed as many times as the number of differentpairs of its CNRS co-authors.9

We also chose to centre our study at the level of the laboratory, and evenat the more aggregate level of groups of laboratories in the same towns orlocalities (‘towns’). We thus consider networks of collaboration betweenlaboratories and towns rather than directly between individual researchers.When two researchers belonging to different laboratories (towns) collabor-ate, we consider that these laboratories (towns) collaborate, and on thisbasis we can measure the intensity of collaboration between laboratories(towns). When two researchers belonging to the same laboratory (town)collaborate, we also simply consider it as a case of collaboration ‘within’this laboratory (town), and likewise we compute the intensity of collabora-tion within laboratories (within towns). We can also similarly computeintensity of collaboration between laboratories-within towns. Carrying outour study at the aggregate level of laboratories and towns simplifies theanalysis and makes the use of our measure of collaboration intensityperhaps more convincing, since networks of collaboration are, of course,much denser at these levels than at the individual researcher level. But, aswe shall see, it also has the advantage that it allows for a direct characteri-sation of the influence on collaboration of working in the same laboratoryor town, and thus of the importance of spatial proximity and easy face-to-face relations.


2.2 Two Configurations of Co-publication

The co-authors of the articles of our group of 493 CNRS researchers,whom we will simply call ‘CNRS researchers’ from now on, can be (these)CNRS researchers themselves, or other researchers, mainly belonging touniversities or other institutions, either French or foreign, whom we willcall ‘external researchers’. In our analysis, we are led to distinguish betweentwo configurations of co-publications, depending on whether a publicationinvolves at least two CNRS researchers and possibly other researchers(CNRS or external), or whether it concerns at most one CNRS researcherand one or more external researchers. An important reason for this dis-tinction is a practical one. We not only preferred a priori to focus our analy-sis on the collaboration among CNRS researchers in the same field, butalso we could not extend it in practice to the external researchers in thisfield. The CNRS researchers were the only ones for whom we could haveaccess to the name, location and some characteristics of their laboratories,in addition to their individual characteristics (age, gender, seniority, and soon).10 This was not possible for the external researchers since we could noteven retrieve the name and location of their laboratories with sufficient reli-ability from the SCI.11 We were thus left with much more limited informa-tion for them than for the CNRS researchers and their laboratories.

Our group of 493 CNRS researchers generally co-publish both with theother CNRS researchers in the group and with external researchers. Asindicated in Figure 10.1, 38 of them are collaborating only with CNRSresearchers (never with external researchers), and 69 are collaborating onlywith external researchers (not with the other CNRS researchers), and thus386 ( 493�38�69) are collaborating in both ways. The first configura-tion of co-publication (involving at least two CNRS researchers and possi-bly other researchers) corresponds to ‘Group 1’ with a total of 1823 articles( 1741�82), while the second (involving only one CNRS researcher withexternal researchers) corresponds to ‘Group 2’ with a total of 5709 articles( 5012�697). Group 1 thus concerns 424 of our CNRS researchers (493�69), while Group 2 concerns 455 of them ( 493�38).

Table 10.1 shows the two-way distribution of the number of articles inGroup 1 and Group 2 with respect to the number of their CNRS authorsand that of their external authors (see also the two related distributionsshown in Figure 10.2). We observe immediately that, for both Group 1 andGroup 2 articles, collaboration generally involves several ‘external’researchers (82 articles are written by CNRS researchers only!). We can alsonote that most articles of Group 1, which have at least two CNRS co-authors, do not involve a third (or more) CNRS co-author (1498 out of1823). Thus, the average number of authors per article for Group 1 is 5.9,


261

493

CN

RS

res

earc

hers

38

rese

arch

ers

69

rese

arch

ers

386

rese

arch

ers

• 7

532

artic

les

c

o-au

thor

ed•

252

art

icle

s al

one

Col

labo

ratio

n on

lyw

ith ‘o

ther

s’ a

ndne

ver

with

CN

RS

rese

arch

ers

***

697

artic

les

Col

labo

ratio

n in

bot

hm

odes

:A

t lea

st 2

CN

RS

and

‘oth

ers’

(G

roup

1)

orat

mos

t 1 C

NR

S a

nd‘o

ther

s’ (

Gro

up 2

)**

*67

53 a

rtic

les

(174

1 fo

r G

roup

1 a

nd 5

012

for

Gro

up 2

)

Col

labo

ratio

n on

lyw

ith C

NR

Sre

sear

cher

s an

dne

ver

with

‘oth

ers’

***

82 a

rtic

les

Fig

ure

10.1

Cho

osin

g th

e sa

mpl

e

262

Tab

le 1

0.1

Num

ber

ofar

ticl

es in

Gro

up 1

and

Gro

up 2

by

num

ber

ofC

NR

S a

nd e

xter

nal c

o-au

thor

s

Num

ber

of‘e

xter

nal’

01

23

45

67

89

Tota

lco

-aut

hors

:

Gro

up 1

(at

leas

t 2

8223

032

437

526

020

912

781

5616

118

23C

NR

S co

-aut

hors

)(4

24*)

Of

whi

ch:

2C

NR

S co

-aut

hors

6419

626

830

021

817

210

661

4766

1498

3C

NR

S co

-aut

hors

1531

4560

3133

1515

67

257

4C

NR

S co

-aut

hors

32

912

84

64

34

555

or m

ore

CN

RS

01

23

30

01

03

13co

-aut

hors

Gro

up 2

(on

ly 1

072

610

8711

1497

670

844

124

112

828

857

09C

NR

S co

-aut

hor)

(455

**)

Tota

l (G

roup

182

956

1411

1489

1236

917

568

322

184

367

7532

and

2)(4

93)

Not

es:

( )

The

thr

ee n

umbe

rs in

par

enth

eses

bel

ow t

he n

umbe

rs o

far

ticl

es a

re t

he n

umbe

rs o

fC

NR

S re

sear

cher

s co

-aut

hori

ng t

hese

art

icle

s.*

Incl

udin

g 38

CN

RS

rese

arch

ers

who

nev

er p

ublis

hed

wit

h ex

tern

al r

esea

rche

rs a

nd w

ho a

ccou

nt fo

r 82

pub

licat

ions

of

Gro

up 1

art

icle

s.**

Inc

ludi

ng 6

9 re

sear

cher

s w

ho n

ever

pub

lishe

d w

ith

othe

r C

NR

S re

sear

cher

s an

d w

ho a

ccou

nt fo

r 69

7 pu

blic

atio

ns o

fG

roup

2 a

rtic

les.

132

CN

RS

rese

arch

ers

who

als

o pu

blis

hed

co-a

utho

red

pape

rs h

ave

wri

tten

252

art

icle

s al

one

(not

incl

uded

in t

he fi

rst

grou

p or

sec

ond

grou

p of

arti

cles

).

of which 2.2 are CNRS researchers and 3.7 external researchers, andfor Group 2 it is 4.9 (that is, 1 CNRS researcher and 3.9 externalresearchers).

2.3 The Selected Sample of Co-publications and Some Characteristics

Four main reasons determine our choice of limiting our analysis to the firstconfiguration of co-publications and to the Group 1 sample of articles. Thefirst reason, which we have already stressed, is analytical. By studying co-publication between couples of CNRS researchers, we control for institu-tional and organisational proximity resulting in ‘common knowledge’ ofrules and practices and strongly favouring collaboration. Organisationalproximity and geographical proximity usually being confounded, this hasthe great advantage of allowing us to unravel clearly the impact of the latteron collaboration. The second reason, which we also mentioned, is simplythat we cannot identify precisely enough the laboratories of the ‘external’researchers, and thus cannot locate them or characterise them, as we cando for the laboratories of the CNRS researchers.

But there is a third important reason of an empirical nature for focusingour investigation on the collaboration between CNRS researchers. Theoccurrence of co-publication between a CNRS and an external researcheris extremely low, while it is much higher, as we would expect, betweencouples of CNRS researchers. The 1823 articles in Group 1, written by 424CNRS authors and about 3500 external co-authors, actually involve only880 different couples of CNRS researchers out of the 89 676, ( 423'424/2) number of potential couples (that is one out of 100). The 5709 art-icles in Group 2, written by 455 CNRS authors with close to 10 000 exter-nal co-authors, involve by contrast as much as 17 500 couples of a CNRSresearcher with an external researcher, out of the 4 550 000 potentialcouples (that is only four out of 1000). Thus, the average number of articlesper effective couple of co-authors is 2.1 in Group 1 and only about 0.3 inGroup 2. Likewise the probability (frequency) of a CNRS author havinganother CNRS co-author (in Group 1) is much higher than having an exter-nal co-author (in Group 2): 0.021 as against 0.001.

A last consideration arises from the fact that some characteristics of co-publication in Group 1 and Group 2 are nonetheless close enough. Thissuggests that, hopefully, a number of the results we find in the analysis ofco-publication between CNRS researchers might not be too different fromthose we would have obtained if we had been able to extend the analysisto the co-publication with non-CNRS researchers. This is clear for thethree following characteristics that we can compute for the sample of 6753articles published by the 386 CNRS researchers involved in both types of


publication (see Figure 10.1). The first of these is the frequency distribu-tion of the number of articles per number of external co-authors. Asshown in Figure 10.2 the probability (frequency) that an article is co-authored by a given number of external researchers is nearly the same inthe two groups of articles. The second very close characteristic concernsthe degree of concentration of the number of articles published in the twogroups of articles by the CNRS researchers. As shown in Figure 10.3 theconcentration curves practically coincide in both cases, with nearly 40 percent of the articles being co-authored by 10 per cent of the most produc-tive CNRS researchers, and about 80 per cent by the more productive halfof them.

Yet, as can be seen in Figure 10.4, the distribution of the number of art-icles written per CNRS researcher (our third characteristic) differs some-what for the two groups of articles. During the six-year period 1992–97, thecumulative probability that a CNRS researcher publishes less than six art-icles in Group 1 is 50 per cent, while it is about 35 per cent in Group 2.Likewise, during this period, a CNRS researcher published an average of9.9 articles in Group 1, as against 13 in Group 2.12

2.4 Other Restrictions on the Selected Sample

In practice, in order to avoid having laboratories and towns with too fewCNRS researchers we thought it better to put two further restrictions on our


Figure 10.2 Frequency of the number of articles written by CNRSresearchers with external co-authors in Groups 1 and 2 ofarticles

Fre

quen

cy o

f art

icle

s in

%

02468

101214161820

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Number of external co-authors

Group 1 Group 2

sample. We imposed that laboratories in our sample had at least five CNRSresearchers, and towns had at least nine CNRS researchers. Our final samplethus consists of 470 CNRS researchers in condensed matter physics (out ofthe initial group of 493), located in 34 laboratories and 17 towns. Likewise,


Figure 10.3 Concentration curves of the number of articles written byCNRS researchers in Groups 1 and 2 of articles

0

10

20

30

40

50

60

70

80

90

100

0 10 20 30 40 50 60 70 80 90 100 Cumulative frequency of the number of CNRS

researchers in %

Cum

ulat

ive

freq

uenc

y of

the

num

ber

ofar

ticle

s in

%

Group 1 Group 2

Figure 10.4 Distribution of the number of articles written by CNRSresearchers in Groups 1 and 2 of articles

0

10

20

30

40

50

60

70

80

90

100

0 5 10 15 20 25 30 35 40 45 50 55 60

Number of articles of CNRS researchers

Cum

ulat

ive

freq

uenc

y of

artic

les

in %

Group 1 Group 2

in our analysis, we thought it better to avoid characterising collaborationbetween two laboratories, or collaboration between two towns, on the basisof too few co-publications between their CNRS researchers. We thusdefined collaboration between a couple of laboratories as involving morethan four co-publications over the six-year study period, and between acouple of towns as involving more than six co-publications. These two typesof restrictions had the consequence of limiting also the number of Group 1articles (with at least two CNRS co-authors), on which our analysis con-centrates, to 1634 articles (out of 1741). To summarise, our investigation isthus mainly based on a sample of 470 CNRS condensed matter physicists(located in 17 towns and 34 laboratories) and a sample of 1634 articles theyhave co-published over the period 1992–97.

3 MEASUREMENT OF INTENSITY OFCOLLABORATION

The behaviours of agents in networks is determined by ‘intrinsic’ individ-ual characteristics such as age, gender, skills, motivations and objectives,and by more ‘structural’ variables such as the density of their networks,their more or less central or peripheral situation, geographic distance, andso on. As a result, the form and functioning of networks differ. If the actorswere not differentiated and if they collaborated with all the others withequal probability, we would expect to observe a uniform structure of rela-tions between all the individuals. We take this extreme case of ‘homogene-ity’ as a reference. At the aggregate level of entities such as the laboratoriesand groups of laboratories (towns) on which we centre our analysis, thecase of homogeneity corresponds to a configuration in which the frequencyof collaboration of agents, the CNRS researchers, is the same, irrespectiveof the entities to which they belong, their geographic localisation and othercharacteristics. Our simple measure of (relative) intensity of collaborationbetween two entities is simply based on the comparison between the realnetwork as portrayed by the data and the network that would be observedin the hypothetical case of homogeneity. We generally define this measurein sub-section 3.1, and comment on its aggregation properties and on theweighting issues in sub-sections 3.2 and 3.3. In sub-section 3.4 we thenprovide a detailed example of its calculation.

3.1 Definition

In this sub-section we assume for simplicity that collaboration alwaysinvolves at the most two (CNRS) researchers (this assumption is discussed


in the next sub-section). The overall or ‘complete’ network of collaborationstudied has a finite number of entities (laboratories or towns) consisting intotal of N researchers who can form C collaboration pairs, or couples,where by definition CN(N�1)/2, the total number of possible pairs. Letn be the total number of articles produced in collaboration between the Nresearchers, then p, the frequency of the number of co-publications per pairin the complete network, is the ratio between the total number of articlesn, and the number of possible pairs C, that is, pn/C.

Using similar notations at the level of the network’s entities, considernow two entities X and Y, where NX and NY are the numbers of researchersworking in them respectively. The numbers of possible pairs of researcherswithin X and within Y are respectively CXNX(NX�1)/2 and CYNY(NY �1)/2, and the number of possible pairs that can be formed betweenresearchers from X and Y is CXYNXNy. If the total numbers of articleswritten jointly within X and Y are respectively nX and nY, and the totalnumber of articles written in common by researchers in X and Y is nXY,the frequencies of collaboration within the entity X and Y, noted as pX andpY, are the corresponding ratios between the total number of articles nX andnY written together by researchers from X or Y, and the number of possi-ble pairs CX and CY of researchers in entity X and Y, that is pXnX/CX andpYnY/CY. Similarly the frequency of collaboration pXY between the twoentities X and Y is the ratio between the total number of articles nXY writtenin common by researchers in X with researchers in Y, and the number ofpossible pairs CXY of researchers from the two entities, that is pXYnXY /CXY.

The intensity of collaboration relates the frequencies obtained at theentities’ level to the frequency p obtained for the complete network. We thusrespectively define the two intra- or within-intensity and the inter- orbetween-intensity as:

Note that in what follows we will be using indifferently the expression intra-or within-intensity, and inter- or between-intensity.

In the reference case of homogeneity of the network we have pXpYpXY for all X and Y, and consequently we can see that pXpYpXYp, orin terms of the intensity measure: iX iY iXY1. In the case of homo-geneity, the frequencies of collaboration intra- and inter-entities are allequal to the overall frequency p for the network, and the intra- and inter-intensities of collaboration are all equal to unity. Otherwise, in the case ofa real network, as the one we are considering, various factors influence

iX pX

p

nX �CX

n �C� iY

pY

p

nY �CY

n �C� iXY

pXY

p

nXY �CXY

n�C


intensities of collaboration; we can expect them, of course, to be verydifferent from unity, which can be viewed as an average benchmark value.

Note that another way of looking at our measure of intensity of collab-oration of an entity is to interpret it as its contribution of co-authored arti-cles nX to the total number n of co-authored articles in the network,normalised by its size relative to that of the complete network measured interms of possible pairs of co-authors, that is, iXpX /p(nX/n)/(CX/C).Note also that the structure of intensity of a network of E entities can berepresented by means of a symmetrical matrix E by E with positive or zerocoefficients where the diagonal terms are equal to the intra-entity intensi-ties and the off-diagonal terms are equal to the inter-entity intensities.13

Appendix 1 gives this matrix for the 17 towns in our sample.

3.2 Aggregation Properties

The (relative) intensity of collaboration as defined above has the advantageof being easy to aggregate at different levels of analysis. In order to see this,suppose that V is a town with two laboratories, X and Y. The total numberof co-authored articles written in V is the sum of co-authored articles byresearchers from X and Y separately, and from X and Y jointly. Likewise,the number of possible pairs of researchers in V is the sum of the possiblepairs of researchers in X and in Y separately, and between X and Y. We thuscan write:

or in terms of frequencies and intensities of collaboration:

pVwX pX�wYpY�wXYpXY or iVwXiX�wYiY�wXYixy

where

with

wX�wY�wXY1

This formula can easily be extended to groups of more than two laborato-ries. Aggregating over the entire network, we have

�I

wIiI � �I,J�I

wIJiIJ 1

wX CX

CV wY

CY

CV wXY

CXY

CV

nV

CV

nX � nY � nXY

CX � CY � CXY

nX

CX'

CX

CV �

nY

CY'

CY

CV �

nXY

CXY'

CXY

CV


with

3.3 Remark on the Weighting

Until now, we have considered for simplicity that the articles were co-authored by two (CNRS) researchers. In reality, they can also be written bythreesomes or foursomes of (CNRS) researchers, and so on. But, as alreadyindicated (in sub-section 2.1), we thought it appropriate to study thenetwork of collaboration ‘link by link’, that is, by couples or pairs of co-authors. In practice, this means that an article is repeated in our database(and thus counted) as many times as there are pairs of different (CNRS) co-authors. For example, for an article published by three CNRS researchers,one belonging to a laboratory X and the two others to a laboratory Y, wecount three co-publications – two between X and Y and one within Y.14

Note that, if we follow this procedure, the aggregation formula (as wesimply write it in the previous sub-section) applies more generally in the casewhere there are more than two (CNRS) co-authors for an article. Note alsothat in practice in our case, since only 20 per cent of the articles in Group 1are co-authored by more than two CNRS researchers, the choice of theweighting assumption should not make an important difference.

3.4 Practical Calculation: An Example

Let us take the concrete example of the town of Marseille to describe indetail the calculation of our measure of the intensity of collaboration,using the information displayed in Table 10.2, which also gives the resultsof this calculation for the other towns. Marseille (as indicated in column 1)is a town with 18 CNRS researchers (among the 470). These researchers areinvolved in 34 co-publications among themselves (column 3) and in 18 co-publications with the CNRS researchers from two other towns (column 4),ten of them with Grenoble and eight with Strasbourg.15 The number ofpossible couples of researchers working in Marseille is 18'17/2, or 153.The frequency of collaborations per couple of researchers in Marseille istherefore 34/153 or 0.22. Given that the numbers of researchers in Grenobleand Strasbourg are 105 and 14, the number of possible couples ofresearchers linking Marseille and Grenoble and Marseille and Strasbourgare respectively 1890 ( 105'18) and 252 ( 14'18). The correspondingfrequencies of collaborations per couple are therefore 0.0053 ( 10/1890)and 0.0317 ( 8/252).

�I

wI � �I,J�I

wIJiIJ 1


270

Tab

le 1

0.2

Des

crip

tive

sta

tist

ics

and

wit

hin-

and

bet

wee

n-to

wn

inte

nsit

y of

co-p

ublic

atio

n at

the

tow

n le

vel#

Tow

nsN

umbe

r of

Num

ber

ofN

umbe

r of

Num

ber

ofN

umbe

r of

Inte

nsit

yIn

tens

ity

Inte

nsit

yC

NR

Sla

bora

tori

espa

rtne

rar

ticl

esar

ticl

esw

ithi

n-to

wn

betw

een-

tow

nbe

twee

n-to

wn

rese

arch

ers

(*)

per

tow

n (*

)to

wns

(**

)‘w

ithi

n’‘b

etw

een’

(ave

rage

(ave

rage

com

pute

d on

com

pute

dal

l oth

eron

par

tner

16 t

owns

)to

wns

onl

y)

Bag

neux

91

651

171

63.0

1.3

3.5

Gif

sur

Yve

tte

161

311

404.

10.

20.

9G

reno

ble

105

612

666

449

5.4

0.7

0.9

Mar

seill

e18

12

3418

9.9

0.1

0.8

Meu

don

91

227

1933

.30.

10.

5M

ontp

ellie

r20

37

4783

11.0

0.3

0.8

Orl

éans

101

07

06.

90.

00.

0O

rsay

663

917

419

23.

60.

20.

4P

alai

seau

182

415

454.

40.

20.

9P

aris

866

724

914

83.

00.

30.

6Po

itie

rs11

10

310

25.1

0.0

0.0

Sain

t M

arti

n31

25

161

193

15.4

0.3

0.9

d’H

ères

Stra

sbou

rg14

12

7220

35.2

0.1

0.9

Tal

ence

91

08

09.

90.

00.

0To

ulou

se29

24

8863

9.6

0.4

1.6

271

Vill

eneu

ve d

’Asc

q10

13

3931

38.5

0.6

3.0

Vill

eurb

anne

91

235

5843

.20.

21.

4To

tal

470

3468

1715

a76

5b—

——

Mea

n27

.62.

04.

0—

—18

.90.

31.

0

Not

es:

# T

he o

vera

ll fr

eque

ncy

ofco

-pub

licat

ions

for

the

sam

ple

of47

0 C

NR

S re

sear

cher

s is

p

0.02

25.

* To

wns

wit

h le

ss t

han

nine

CN

RS

rese

arch

ers

and

labo

rato

ries

tha

n le

ss t

han

five

CN

RS

rese

arch

ers

are

not

cons

ider

ed.

** P

artn

er t

owns

are

defi

ned

as h

avin

g m

ore

than

six

art

icle

s co

-pub

lishe

d by

the

ir C

NR

S re

sear

cher

s ov

er t

he s

ix-y

ear

peri

od,1

992–

97 (

that

is,a

tle

ast

an a

vera

ge o

fon

e co

-pub

licat

ion

per

year

).a

Eac

h ar

ticl

e is

wei

ghte

d by

the

num

ber

ofpa

irs

ofau

thor

s th

at c

ontr

ibut

e to

its

publ

icat

ion,

othe

rwis

e th

e nu

mbe

r of

arti

cles

wou

ld b

e 12

22.

b E

ach

arti

cle

is w

eigh

ted

by t

he n

umbe

r of

pair

s of

auth

ors

that

con

trib

ute

to it

s pu

blic

atio

n,ot

herw

ise

the

num

ber

ofar

ticl

es w

ould

be

412.

In order to compute the intensities of collaboration, we have also to cal-culate p, the overall frequency of collaboration per couple of researchersfor the complete set of the 17 towns. It is the ratio between the totalweighted number of articles, 2480 ( 1715�765), and the number of pos-sible couples that can be formed by the 470 CNRS researchers, that is,110 215 pairs ( 470'469/2). We thus have p0.0225. This overall fre-quency p is also the intra- (or within-) and inter- (or between-) frequencyof collaboration that would have been obtained for Marseille and all theother towns in the hypothetical case of homogeneity. In fact, the intra-frequency for Marseille is much higher (0.22) than this reference value,the inter-frequency of collaboration with Grenoble much lower (0.0053),and that with Strasbourg relatively closer (0.0317). Finally, the intra-townintensity for Marseille is of 0.22/p or 9.88 (column 5). Likewise, theMarseille–Grenoble and Marseille–Strasbourg inter-intensities are 0.24and 1.41 respectively, yielding a mean inter-intensity of collaboration ofMarseille with all the other 16 towns of (0.24�1.41)/16 or 0.1 (column 6),and a mean inter-intensity of Marseille with only its two effective partnersof (0.24�1.41)/2 or 0.82 (column 7).

4 RESULTS: THE IMPORTANCE OFGEOGRAPHICAL PROXIMITY AND QUALITYOF SCIENTIFIC ENVIRONMENT

We look first at the estimated intensities of co-publication between theCNRS researchers at the town level (Table 10.2 and Appendices 1 and 2).Next, we consider in detail the statistical evidence on the potential deter-minants of co-publication we have been able to measure, which is mainlyprovided by simple correlations computed both at the town and labora-tory levels (Tables 10.3 to 10.5 and Appendix 3). Finally, we assessthe robustness of these results by examining the multivariate regressionsof the occurrence and intensity of co-publication on these various deter-minants (Table 10.6).

4.1 Intensity of Co-publication at the Town Level

The estimated inter-town intensities of co-publication among all thedifferent couples of towns, as we can see from the matrix of co-publicationintensity in Appendix 1 and from the graph of the co-publication networkin Appendix 2 (and also from their averages by towns computed in Table10.2), are extremely dispersed. Of the 136 possible couples of towns, only34 are effectively collaborating.16 Grenoble, Orsay and Paris are the main


273

Tab

le 1

0.3

Des

crip

tive

sta

tist

ics

for

the

mai

n de

term

inan

ts o

fco

-pub

licat

ion

at t

he t

own

leve

l

Tow

nN

umbe

r of

Num

ber

ofN

umbe

r of

Stoc

k of

Mea

nM

ean

dist

ance

Mea

nM

ean

Mea

nsc

ient

ists

poss

ible

poss

ible

publ

icat

ions

geog

raph

icof

prod

ucti

vity

qual

ity

ofpr

opor

tion

coup

les

coup

les

betw

een

1992

dist

ance

to

spec

ialis

atio

npu

blic

atio

nsof

arti

cles

‘bet

wee

n’‘w

ithi

n’an

d 19

97pa

rtne

rsco

-aut

hore

dw

ith

fore

igne

rs

Bag

neux

94

149

3632

834

43.

0236

.44

3.68

0.12

Gif

sur

Yve

tte

167

264

120

246

171

3.93

15.3

83.

070.

14G

reno

ble

105

3832

55

460

187

042

110

.19

17.8

13.

390.

50M

arse

ille

188

136

153

235

361

11.1

713

.06

2.84

0.44

Meu

don

94

149

3699

208

5.97

11.0

02.

630.

20M

ontp

ellie

r20

900

019

036

554

85.

6318

.25

3.47

0.21

Orl

éans

104

600

4563

08.

246.

303.

540.

32O

rsay

6626

664

214

592

233

44.

9713

.97

3.69

0.25

Pal

aise

au18

813

615

327

429

74.

3215

.22

4.77

0.33

Par

is86

3302

43

655

985

291

6.72

11.4

53.

750.

48Po

itie

rs11

504

955

880

5.17

8.00

2.34

0.26

Sain

t M

arti

n31

1360

946

543

841

84.

4414

.13

3.78

0.27

d’H

ères

Stra

sbou

rg14

638

491

248

449

6.45

17.7

13.

690.

16T

alen

ce9

414

936

193

011

.07

21.4

43.

940.

43To

ulou

se29

1278

940

637

941

05.

3013

.07

2.71

0.24

Vill

eneu

ve d

’Asc

q10

460

045

184

305

3.12

18.4

04.

130.

28V

illeu

rban

ne9

414

936

139

188

4.55

15.4

43.

020.

23To

tal

470

194

176

1312

77

056

——

——

—M

ean

27.6

——

—27

96.

1315

.71

3.44

0.29


Table 10.4 Correlations at the town level with the occurrence and intensityof co-publication

Intensity Occurrence Intensitywithin-town between-town between-town

(N17) (N136) (N34)

Geographic distance — �0.09 �0.16

Distance in specialisationOverall profile — �0.02 �0.21Physics-chemistry �0.18** �0.14General physics �0.20*** �0.20Solid-state physics �0.14 �0.24Applied physics 0.06 �0.08Materials science �0.02 �0.06Crystallography 0.16* �0.17Other �0.00 �0.16

SizeNumber of researchers

NI �0.48* — —Maximum (NI, NJ) — 0.52*** �0.41**Minimum (NI, NJ) 0.42*** �0.31*Average (NI�NJ )/2 0.56*** �0.45***

Publications 1992–97SI �0.29 — —Maximum (SI, SJ) — 0.52*** �0.29*Minimum (SI, SJ) 0.54*** �0.21Average (SI�SJ)/2 0.60*** �0.33*

Number of couples of �0.39 0.49*** �0.32*researchers CIJNI * NJ

ProductivityPI 0.62*** — —Maximum (PI, PJ) — 0.11 0.67***Minimum (PI, PJ) 0.26*** 0.47**Average (PI�PJ)/2 0.19*** 0.69***

Quality of publicationsQI �0.11 — —Maximum (QI, QJ) — 0.03 0.09Minimum (QI, QJ) 0.23*** 0.03Average (QI�QJ)/2 0.16* 0.08

nodes in the network of collaboration, being respectively linked to twelve,nine and seven other towns, whereas Poitiers, Orléans and Talence appearto be isolated.17 Among all effectively linked couples of towns (or partnertowns), the intensity of co-publication ranges from a lowest value of 0.24for Grenoble–Marseille and Grenoble–Meudon to a highest value of 7.40for Bagneux–Villeuneuve d’Ascq. As could be expected (and checked bylooking at the number of CNRS researchers in our sample per town, givenin column 2 of Table 10.2), the towns with the largest number of CNRSresearchers are those which tend to have more links with other towns butalso lower inter-town intensity estimates.

The estimated intra-town intensities of co-publication (given in column6 of Table 10.2) are much higher than the inter-town intensities, with veryfew exceptions. These are always greater than one, and on average equal to18.9, as compared with an average inter-intensity of 0.3 when computedover all couples of towns and of 1.0 when computed only over the partnertowns. This strongly points to a major influence of geographical proximityon the intensity of collaboration. Note also that intra-town intensity tendsto be high in towns with few partners, like Meudon, Poitiers, Strasbourg,Villeneuve-d’Ascq or Villeurbanne, compared with towns with many part-ners like Grenoble, Orsay and Paris, which have among the lowest intra-town intensities (5.4, 3.6 and 3.0 respectively). This result could mainly beexplained by the larger size of the CNRS research community in Grenoble,Orsay and Paris, and the fact that these towns host several laboratories,both characteristics entailing numerous potential links among which rela-tively many do not occur. As a matter of fact, the co-publication intensitiesestimated at the laboratory level for Grenoble, Orsay and Paris are muchhigher, being on average equal to 19.4, 33.4 and 34.5 respectively, and quitecomparable to that of the laboratories of the other towns (see column 3 inTable 10.5).


International opennesspeI �0.37 — —Maximum (peI, peJ) — 0.14* �0.26Minimum (peI, peJ) — 0.34*** �0.12Average (peI�peJ)/2 — 0.26*** �0.22

Note: The stars ***, ** and * indicate that the correlations are statistically significant at aconfidence level of 1 per cent, 5 per cent and 10 per cent, respectively.

4.2 Determinants of the Occurrence and Intensity of Co-publication:Correlation Evidence

We consider six a priori influential determinants of collaboration at the lab-oratory or town levels, which we have been able to approximately measureor proxy: geographic distance, specialisation, size, productivity, quality ofpublications and international openness. Apart from geographic distanceand distance in specialisation which are directly defined for a couple of lab-oratories or towns, it is somewhat problematic to adopt a priori a singlemeasure for our four other variables, such as their average over the twolaboratories or towns concerned, say, for example, (SI�SJ)/2 where S is ameasure of size of the laboratories or towns I and J. Thus, in addition to


Table 10.5 Town averages of within- and between-town intensity ofco-publication at the laboratory level

Town Number of Intensity of Intensity of Intensity oflaboratories collaboration collaboration collaboration

per town within- between- between-laboratory laboratory- laboratory-

within-town* between-town*

Bagneux 1 58.4 — 2.0Gif sur Yvette 1 9.2 — 0.4Grenoble 6 19.4 2.7 0.5Marseille 1 11.6 — 0.1Meudon 1 30.9 — 0.1Montpellier 3 28.1 0.0 0.3Orléans 1 6.4 — 0.0Orsay 3 33.4 1.4 0.2Palaiseau 2 11.9 0.0 0.2Paris 6 34.5 0.2 0.2Poitiers 1 23.2 — 0.0Saint Martin d’Hères 2 37.1 0.0 0.4Strasbourg 1 38.0 — 0.1Talence 1 15.7 — 0.0Toulouse 2 25.7 1.5 0.2Villeneuve d’Ascq 1 98.9 — 0.5Villeurbanne 1 40.0 — 0.4Mean* 2 30.7 0.8 0.3Mean** 2 30.7 3.2 4.0

Notes:* Mean computed on all the laboratories.** Mean computed on partner laboratories only.

the average, we used the maximum and minimum values, say SI and SJ forthe couple of laboratories or towns I and J. Note that for size we also havethree different possibilities: the number of CNRS researchers, say NI for lab-oratory or town I; the number of possible pairs of CNRS co-authors, say(NI 'NJ) for the couple (I, J) of laboratories or towns – or (NI'NI�1)/2within the laboratory or town I – and the number of publications over thesix-year period, say SI for laboratory or town I.

We will examine all these variables in turn. Their means at the town levelare given in Table 10.3; and their correlations with the binary indicator ofoccurrence of co-publication and with our measures of both intra- andinter-intensity, also at the town level, are displayed in Table 10.4. These cor-relations are also recorded in Appendix 3 at the laboratory level, bothoverall (for the 34 laboratories) and within towns (for the seven towns outof 17 which have more than one laboratory). The statistical evidence isquite consistent at both the town and laboratory levels, with the notableexception of the correlations of the occurrence of co-publication with spe-cialisation at the laboratory level within towns. It is also, in general, quali-tatively comparable for the occurrence and intensity of co-publication, theone major exception being the size variable positively correlated withoccurrence and negatively with intensity.

4.2.1 Geographic distanceThe average distance of a town from its partners can vary widely (see Table10.3). At the two extremes, Montpellier collaborates with seven other towns,situated at an average distance of 550 km (kilometres), while Gif sur Yvetteis related to three towns much closer, at an average distance of 170 km, twoof these being also located in the Parisian region. Four of the five towns situ-ated less than 300 km on average from their partners are in the Parisianregion (besides Paris intramuros, Bagneux, Gif sur Yvette, Meudon, Orsayand Palaiseau). The geographic distance apparently plays a negligible role,or only a slightly negative one, in the occurrence of collaboration, as well ason its intensity. This is shown by the correlations computed at the town levelbut also at the laboratory level. The relevant two correlations at the townlevel (�0.09 and �0.16) are negative but both statistically non-significant,and the two at the laboratory level (�0.09 and �0.06) are also negative, withonly the first moderately significant (at a 5 per cent confidence level).

In fact, as we have already noted in comparing the values of the intra-and inter-town intensities of co-publication presented in Table 10.2, prox-imity has a major influence on collaboration. This is confirmed by the com-parison of the intra- and inter-laboratory intensities shown in Table 10.5.But more interestingly, this can also be qualified, since by comparing theinter-laboratory–intra-town intensity to the inter-laboratory–inter-town


intensity we distinguish between what we shall call immediate proximityand local proximity. Immediate proximity, which is that of researchersworking in the same laboratory, usually located in a common building,favours frequent face-to-face interactions and can be expected to induceand facilitate collaboration. Local proximity, which is that of researchersworking in different laboratories but still relatively close, as when in thesame town, can be expected to be less conducive to collaboration thanimmediate proximity. Nonetheless, local proximity should be morefavourable to collaboration than when researchers are working in reallydistant laboratories, as when located in different towns. This is, indeed, veryclearly what we find. For example, for Grenoble and its six laboratories, theaverage intra-laboratory intensity is 19.4, which is about seven times higherthan the inter-laboratory–intra-town intensity (2.7), itself about 5 timeshigher than the average inter-laboratory–inter-town intensity (0.5). Onaverage, for all 17 towns the pattern is the same, the three average intensi-ties being 30.7, 0.8 and 0.3 respectively.

One can thus distinguish three scales of geographic distance, which influ-ence collaboration very differently. Immediate proximity has a considerableimpact on collaboration and local proximity is also favourable but muchless so. By contrast, beyond proximity, geographic distance stronglyhinders collaboration, but per se and only slightly, if at all, in proportionto real distance (say in kilometres). Such findings are well corroborated byprior studies on knowledge flows between public laboratories and indus-trial firms, which show that proximity allows face-to-face interactions andeasy exchanges of tacit knowledge between actors, inducing them to builda common understanding rather than referring only to a ‘common text’(see, for example, Zucker et al., 1998a and 1998b, and Leamer and Storper,2000). New communication technologies have certainly contributed to the‘death of distance’, by helping researchers to collaborate in research muchmore easily and faster: however, they have not done away with the crucialimportance of proximity.

4.2.2 SpecialisationProximity in specialisation, not only geographic proximity, should alsostrongly influence collaboration in a field as diverse and large as condensedmatter physics. The network of co-publication presented in Appendix 2 isby itself suggestive of such an influence. Orsay and Grenoble, which appearto be two central nodes of the network, are indeed the location of the twoFrench storage rings, which are very large facilities used by physicists ofcondensed matter.18

We have tried to take into account the specialisation of laboratories (ortowns), although there is no easy and good way to do so. We have defined a


profile of specialisation of a laboratory (or town), based on the classificationby the main theoretical and/or experimental ‘sub-domains’ of the journals inwhich the CNRS researchers of this laboratory (or town) are publishing.Such classification is difficult but seems carried out relatively well by theScience Citation Index (SCI). We found that the most frequently listed sub-domains of the journals in which the articles of the CNRS condensed matterphysicists are published were physics-chemistry, general physics, solid-statephysics, applied physics, materials science and crystallography. We then char-acterised an ‘overall specialisation profile’ of each laboratory (or town) bythe [7, 1] column vector defined by the proportions of publications of itsCNRS researchers in these six main sub-domains and the group of othersub-domains. We also considered the seven ‘specific specialisation profiles’corresponding to the seven [2, 1] column vectors defined by the proportionsof publications in each of the seven sub-domains and all the six others.

Next, to measure the distance in specialisation between all couples of lab-oratories (or towns) we adopt the chi-squared distance between their spe-cialisation profiles. To facilitate interpretation we also normalised thismeasure in such a way that the average distance between any one laboratory(or town) and all others will at most be equal to one if the laboratories (ortowns) had specialisation profiles which were not statistically different at the1 per cent confidence level.19 The average specialisation distances (in termsof the overall profiles) between towns, given in Table 10.3, show that we arein fact far from this hypothetical situation. Specialisation profiles are quitediverse, and the specialisation distances as we measure them vary widely(the lowest town average distance being about 3 for Bagneux and Villeneuved’Ascq and the highest one about 11 for Marseille and Talence).

With the major exception of the puzzling positive and significantbetween-laboratory-within town correlations with the occurrence of co-publication (and some minor ones concerning particularly the specialisationindicator in crystallography and again the occurrence of co-publication), allother correlations of our overall and specific specialisation distance meas-ures with the occurrence and intensity of co-publication are consistentlynegative (see Table 10.4 and Appendix 3). The correlations with the existenceof co-publication, however, are mostly small and not significant, while thecorrelations with the intensity tend to be sizeable and statistically significant.The two between-laboratory-between-town and between-laboratory-within-town correlations of intensity of co-publication and overall special-isation distance are, for example, as large as �0.3 and �0.4 respectively.Note also that the puzzling exception of the between-laboratory-within-town correlations with the occurrence of co-publication may largelyreflect the correlations of our measure of specialisation with the other deter-minants of co-publication, since the corresponding coefficient in our


regression analysis is not statistically significant (see sub-section 4.3 andTable 10.6).

On the whole, our evidence thus tends to show that proximity in special-isation favours strongly the intensity of collaboration, but much moreweakly so, or not, its existence. This latter result is not what we expected. Itis clear, however, that our attempt here at measuring specialisation is crude,and that much remains to be done to better characterise what it is and toassess its impact on collaboration.

4.2.3 SizeWe considered three different measures of size. The first relies on thenumber of CNRS researchers in the laboratories (or towns) concerned, andthe second on the total number (or stock) of their publications over the sixyears 1992–97.20 The third, which is particularly well suited to our defini-tion of intensity of co-publication, is the number of possible couples ofCNRS researchers for each couple of laboratories (or towns). For the firsttwo measures, as already explained, we experimented with the average, themaximum and minimum of the number of researchers and of the stock oftheir publications for each couple of laboratories (or towns). Not surpris-ingly, these three types of size measures are overall quite consistent, asshown by the descriptive statistics in Table 10.3. Grenoble comes first of alltowns, with 22 per cent of the total number of CNRS researchers involved,20 per cent of the total number of possible couples of co-authors amongthem, and 27 per cent of their total publications. Paris comes second andOrsay third (with about 18 per cent and 14 per cent respectively of thenumber of researchers and of all possible couples of co-authors, and foreach of them roughly 14 per cent of all their publications).

A priori one would expect that the size of laboratory (or town) wouldimpact positively on the chance of collaboration, but not its intensity, sinceby construction our measure of intensity already takes into account sucha size effect. One would even think it likely that the larger the laboratories(or towns) involved in collaboration, the smaller its intensity. This is indeedwhat we see clearly when looking at the correlations in Table 10.4 and inAppendix 3 for the different indicators of size we used. The correlations ofnearly all of them, both at the town level and the laboratory level (within-and between-town), are thus very significantly positive and substantial(ranging from 0.2 to 0.6) with the occurrence of co-publications, while verysignificantly negative in a comparable range (from �0.2 to �0.6) with itsintensity. Note that it is also the case that the correlations of the intensityof co-publication of the researchers within their own laboratories (ortowns) with the size of their laboratories (or towns) tend to be significantlynegative.


4.2.4 ProductivityThe productivity of the laboratories (or towns) is simply measured as thestock of publications of their CNRS researchers in the period 1992–97 perresearcher (that is, as the ratio of our two first measures of size just definedin sub-section 4.2.3).21 As can be seen in Table 10.3, productivity varieswidely from one town to another, from a minimum of 6.3 articles perresearcher over six years for Orleans to a maximum of 36.4 for Bagneux,the overall mean being 15.7 articles per researcher (that is, 2.6 articles peryear). In contrast with size, it seemed a priori likely that both the correla-tions of productivity with the occurrence of co-publication and its inten-sity should be positive. This is definitely what we find. Nearly all thesecorrelations, including the two of within-town and within-laboratory inten-sities with productivity, are very significantly positive and of a sizeableorder of magnitude, from 0.2 up to 0.7 (see Table 10.4 and Appendix 3).

4.2.5 Quality of publicationsOur measure of the quality of publications of laboratories (or towns) isconsistent with that of their productivity. It is the average impact factor (orimpact score) per publication of their stock of publications over the period1992–97. Precisely, it is the weighted mean of the impact factors of the jour-nals in which these publications have appeared (the weights being thenumbers of publications in the different journals). The impact factors ofthe journals are provided by the SCI; they are defined and computed as theaverage number of citations per article received by the articles published inthe journals over a period of two and five years. We used here the two-yearimpact factors, but using the five-year impact factors did not make adifference in our results. Our measure of the quality of publications of alaboratory (or town) is thus an estimate of the expected number of citationsthat the publications of its researchers will on average receive over twoyears. In Table 10.3, we see that this average number overall is 3.4 citationsper article over two years, and that it can differ by a factor of 2 at the townlevel, being lowest for Poitiers, with a citation rate of 2.3, and highest forPalaiseau, with a citation rate of 4.8.

Although we expected that the quality of publications, like productivity,would be positively correlated with both the occurrence and intensity ofcollaboration, the evidence (recorded in Table 10.4 and Appendix 3) ismixed. There are no statistically significant correlations with the intensityof co-publication at the town or laboratory levels. We find a significantlypositive correlation with the occurrence of co-publication only when we useas our quality indicator the minimum value for the couples of laboratoriesor towns involved (0.23 at the town level and 0.06 at the laboratory level).This suggests, interestingly but tentatively, that what matters in establishing


a collaboration is a minimum quality requirement on the two partnersinvolved. To confirm such a proposition would, of course, need a moredetailed analysis and should be performed at the researcher level, not onlyat the aggregate level of the laboratory. Note, however, that the between-laboratory-within-town correlations with the occurrence of co-publicationare all very significantly negative, raising a similar puzzle as the one we havewith our specialisation indicators.

4.2.6 International opennessAs we indicated in sub-section 2.2, we cannot precisely locate the laborator-ies of the very many ‘external’ researchers who are co-publishing with oursample of 470 CNRS physicists. However, it is possible to identify theforeign (non-French) addresses among all those listed in the SCI electronicrecords for all their articles (both in Group 1 and Group 2). In spite of theimprecision of such information, we can thus build an indicator of the inter-national openness of the laboratory as the proportion of articles of theirCNRS researchers (over the six years, 1992–97) involving a least one foreignco-author. In Table 10.3, we see that this proportion is overall about 30 percent, and that it is the highest, about 50 per cent, for Grenoble and Paris, andthe lowest, about 15 per cent, for Bagneux, Gif sur Yvette and Strasbourg.

Our a priori thought was that international openness would also gotogether with greater occurrence and intensity of collaboration between theCNRS researchers themselves and their laboratories (or towns). This iswhat we observe, although the evidence is not strong. Many of the correl-ations given in Table 10.4 and in Appendix 3 are not significantly differentfrom zero, but those that are significant are all positive.

4.3 Regression Confirmatory Evidence

To assess the robustness of the evidence provided by our analysis of simplecorrelations, we did a number of regressions of both the occurrence and theintensity of co-publication on the six a priori influential variables we havebeen able to consider. All of these mainly told the same story, confirmingour observations, which were already strongly supported by the correlationevidence.22 We present in Tables 10.6 and 10.7 the regressions we did at thelaboratory level, which include all six variables measured in the simplest way(that is, for specialisation in terms of overall profile, for size as the averageof the numbers of researchers in all couples of laboratories, and similarlyfor productivity, quality of publications and international openness as theaverages of the corresponding indicators in all couples of laboratories).

The geographic distance between laboratories does not influence theintensity of collaboration and has only a small, significantly negative


impact on its occurrence – an increase of 100 km in the distance betweentwo laboratories corresponding to a decrease in the frequency of co-publi-cation of less than 1 per cent (0.8 per cent). As expected, the distance in spe-cialisation has a negative impact on the intensity of co-publication, whichseems sizeable although statistically not very significant. An increase of onestandard deviation in the distance of specialisation, as we characterise it,will thus imply a fall of nearly 30 per cent in the intensity of co-publicationbetween-laboratory-between-town.

Laboratory size has a very significant and large impact on collaboration:positive in its occurrence, while negative in its intensity. A 10 per centincrease of the average size of each laboratory will thus entail an increaseof the frequency of collaboration within-town and between-town of about15 per cent and 25 per cent respectively, while it will correspond to a declineof the intensity of co-publication within-laboratory of nearly 10 per cent,and between-laboratory-between-town of about 10 per cent. Laboratoryproductivity has positive and mostly significant effects on both the


Table 10.6 Regression results at the laboratory level on the occurrence ofco-publication

Variables Occurrence of co-publication between laboratories

Within-town Between-town(N39) (N522)

Geographic distance — �0.008**(0.004)

Distance in specialisation 0.02 �0.009*(0.04) (0.005)

Size 0.04*** 0.015***(0.01) (0.001)

Productivity 0.03* 0.013***(0.02) (0.003)

Quality of publications �0.18*** �0.017(0.02) (0.029)

International openness �0.08 0.17(1.5) (0.22)

Adjusted R2 0.374 0.107

Notes:The standard errors of the estimated coefficients are given in parentheses.The stars ***, ** and * indicate that they are statistically significant at a confidence level of1 per cent, 5 per cent and 10 per cent, respectively.

occurrence and intensity of collaboration, which are of the same or evenlarger orders of magnitude than the size effects. A 10 per cent increase ofproductivity will thus involve an increase of 15 per cent and 30 per cent inthe frequency of collaboration within-town and between-town respectively,and will result in a rise of about 20 per cent in the intensity of co-publica-tion within-laboratory.

The quality of laboratory publications does not seem to have a signifi-cant impact on collaboration, except one which is negative, contrary toour a priori expectation, on the frequency of co-publication between-laboratory-within town (confirming the puzzling simple correlations wealready noted). Clearly the international openness of laboratories, at leastin the way we can proxy for it, has also apparently no significant influenceon collaboration.


Table 10.7 Regression results at the laboratory level on the intensity ofco-publication

Variables Intensity Intensity of co-publicationwithin-laboratory between-laboratory

(N34)

Within-town Between-town(N15) (N41)

Geographic distance — — �0.02(0.19)

Distance in specialisation — �0.78* �0.41*(0.43) (0.23)

Size �1.51*** �0.31 �0.24**(0.50) (0.34) (0.09)

Productivity 3.31*** 0.32 0.19(1.09) (0.19) (0.12)

Quality of publications 6.88 �0.83 0.48(5.41) (2.42) (1.82)

International openness 40.1 33.4 �0.93(58.5) (29.8) (12.1)

Adjusted R2 0.312 0.162 0.369

Notes:The standard errors of the estimated coefficients are given in parentheses.The stars ***, ** and * indicate that they are statistically significant at a confidence level of1 per cent, 5 per cent and 10 per cent, respectively.

5 CONCLUSION

In order to study networks of collaboration between researchers, we pro-posed a simple measure of the intensity of collaboration, which can beintuitively interpreted in terms of relative probability and easily aggregatedat the laboratory level. We first used this measure to characterise the rela-tions of collaboration between the scientists of the French Centre Nationalde la Recherche Scientifique (CNRS) in the field of condensed matterphysics, as defined in terms of co-publication during the six-year period,1992–97. We then used it to investigate the importance of various factorsof collaboration: mainly the geographical distance between laboratories,but also their specialisation and size, their productivity, the quality of theirpublications and their international openness.

We find that the average intensity of co-publication of researchers withinlaboratories is about 40 times higher than the average intensity between lab-oratories if they are located in the same town, and about 100 times higherthan the intensity between laboratories if they are not located in the sametown. Yet, geographical distance does not have a significant impact, or hasa very weak one, on the existence and intensity of co-publication betweenlaboratories located in different towns. There are basically three scales ofgeographic distance. Immediate proximity, which allows easy face-to-faceinteractions, has a considerable impact on collaboration, while local prox-imity is also relatively favourable but much less so. Geographic distanceper se, that is, beyond proximity, remains by contrast a strong obstacle tocollaboration, but only slightly, if at all, in proportion to real distance.

Although our measure of specialisation between laboratories remainscrude, we find that proximity in specialisation has also a large positive influ-ence on the intensity collaboration. The size of laboratories and their pro-ductivity in terms of number publications per researcher appear to beinfluential determinants of collaboration, having both a positive impact onthe occurrence of co-publication, but a negative impact for size and a posi-tive one for productivity on the intensity of co-publication. Contrary to ourexpectations, we do not really observe significant effects on collaborationof the average quality of publications and of international openness of lab-oratories. However, this may be due, at least in part, to the fact that thesetwo indicators, as we have been able to construct them, are at best imper-fect proxies.

In future work, it will thus be necessary to improve the measurement ofthe potential determinants of collaboration that we have been able to con-sider, as well as to extend the list of these determinants. Clearly it will alsobe important to broaden the scope of our study, which remains mainly illus-trative. In particular, although we think it is appropriate and interesting to


analyse collaboration at the level of the laboratory, as we did here, it will beboth enlightening and challenging to carry out this analysis together withan investigation at the individual researcher level. This will undoubtedlylead to an assessment of the role of ‘star scientists’ in the scientific perfor-mance of their own laboratories and in the formation and development ofnetworks of collaboration. By focusing on the co-publication betweenresearchers working in the same institutional setting, that of the FrenchCNRS, we have been able to control for organisational proximity.Comparing similar studies in different research environments could beinstructive by itself. But, of course, trying more generally to integrate insti-tutional and organisational characteristics in the analysis and to understandhow they can enhance or hinder collaboration should be a central objectivein the research agenda – one that will keep up with the high standards ofPaul David.


287

App

endi

x 1

Mat

rix

ofin

tens

itie

s of

co-p

ublic

atio

n be

twee

nn C

NR

S c

onde

nsed

mat

ter

phys

icis

ts a

t th

e to

wn

leve

l

Bag

neux

Gif

sur

Gre

nobl

eM

arse

ille

Meu

don

Mon

t-O

rléa

nsO

rsay

Pal

aise

auP

aris

Poit

iers

Sain

t St

ras-

Tal

ence

Toul

ouse

Vill

e-V

illeu

r-Y

vett

epe

llier

Mar

tin

bour

gne

uve

bann

ed’

Hèr

esd’

Asc

q

Bag

neux

62.9

63.

531.

730.

372.

575.

627.

4

Gif

sur

4.07

0.53

0.38

1.70

Yve

tte

Gre

nobl

e5.

420.

240.

240.

490.

440.

280.

131.

840.

360.

152.

50

Mar

seill

e9.

881.

41

Meu

don

33.3

30.

80

Mon

tpel

lier

11.0

00.

810.

990.

130.

790.

38

Orl

éans

6.91

Ors

ay3.

610.

520.

290.

330.

350.

37

Pal

aise

au4.

36

Par

is3.

030.

380.

36

Poit

iers

25.0

5

Sain

t 15

.39

1.29

Mar

tin

d’H

ères

Stra

sbou

rg35

.16

Tal

ence

9.88

Toul

ouse

9.63

Vill

eneu

ve38

.52

d’A

scq

Vill

eur-

43.2

1ba

nne

AP

PE

ND

ICE

S

288

Poi

tiers

Mar

seill

e

Meu

don

Mon

tpel

lier

Ors

ay

Par

is

Str

asb

ourg

Vill

eneu

ve d

’Asc

q

Tale

nce

Gif

sur

Yve

tte

Orlé

ans

Vill

eurb

anne

Bag

neux

Pal

aise

au

Sai

nt M

artin

d’H

ères

Toul

ouse

Gre

nob

le

LEG

EN

D: T

he in

tens

ity o

f co-

pub

licat

ion

bet

wee

n tw

o to

wns

is in

dic

ated

by

a lin

e:•

thic

k gr

ey if

mor

e th

an 1

•

thic

k b

lack

if b

etw

een

0.4

and

1•

thin

bla

ck if

less

tha

n 0.

4.

App

endi

x 2

Net

wor

k of

co-p

ublic

atio

ns a

t th

e to

wn

leve

l


Appendix 3a Correlations at the laboratory level with the occurrence ofco-publication

Occurrence of co-publicationsbetween-laboratory-

between-town within-town(N522) (N39)

Geographic distance �0.09** —

Distance in specialisationOverall profile �0.02 0.40***Physics-chemistry �0.05 0.30***General physics �0.09*** 0.35***Solid-state physics �0.04 0.24**Applied physics �0.01 0.05Materials science 0.05* 0.41***Crystallography 0.06** 0.10Other �0.03 �0.11


NI — —Maximum (NI, NJ) 0.21*** 0.45***Minimum (NI, NJ) 0.22*** 0.57***Average (NI�NJ )/2 0.24*** 0.57***

Stock of publications between 1992 and 1997SI — —Maximum (SI, SJ) 0.23*** 0.58***Minimum (SI, SJ) 0.35*** 0.63***Average (SI�SJ)/2 0.31*** 0.63***

Number of couples of researchers CIJNI * NJ 0.27*** 0.61***

ProductivityPI — —Maximum (PI, PJ) 0.19*** 0.35***Minimum (PI, PJ) 0.30*** 0.40***Average (PI�PJ)/2 0.26*** 0.39***

Quality of publicationsQI — —Maximum (QI, QJ) �0.03 �0.46***Minimum (QI, QJ) 0.06** �0.30***Average (QI�QJ)/2 0.02 �0.44***


Appendix 3b Correlations at the laboratory level with the intensity ofco-publication

Intensity Intensity ofwithin-laboratory co-publication

(N34) between-laboratory

between towns within towns(N41) (N15)

Geographic distance — �0.06 —

Distance in specialisationOverall profile �0.32*** �0.42**Physics-chemistry �0.18 �0.10General physics �0.29*** �0.30Solid-state physics �0.25** �0.05Applied physics �0.05 �0.26Materials science �0.21* �0.35*Crystallography �0.26** �0.16Other �0.11 �0.26


NI �0.42** — —Maximum (NI, NJ) �0.52*** �0.34*Minimum (NI, NJ) �0.51*** �0.02Average (NI�NJ )/2 �0.60*** �0.27

Appendix 3a (continued)

Occurrence of co-publicationsbetween-laboratory-

between-town within-town(N522) (N39)

International opennesspeI — —Maximum (peI, peJ) 0.02 0.07Minimum (peI, peJ) 0.03 �0.06Average (peI�peJ)/2 0.03 0.01



Stock of publications between 1992 and 1997SI �0.06 — —Maximum (SI, SJ) — �0.08 0.18Minimum (SI, SJ) �0.14 0.13Average (SI�SJ)/2 �0.12 0.17

Number of couples of researchersCIJNI * NJ �0.34** �0.49*** �0.23

ProductivityPI 0.36** — —Maximum (PI, PJ) — 0.51*** 0.33*Minimum (PI, PJ) 0.42*** 0.36**Average (PI�PJ)/2 0.54*** 0.36**

Quality of publicationsQI 0.20 — —Maximum (QI, QJ) — �0.03 �0.25Minimum (QI, QJ) 0.12 0.11Average (QI�QJ)/2 0.05 �0.01

International opennesspeI 0.14 — —Maximum (peI, peJ) — 0.34*** �0.005Minimum (peI, peJ) — 0.17 0.35*Average (peI�peJ)/2 — 0.29*** 0.22


NOTES

* We wish to thank Michèle Crance, Serge Bauin and the members of the Unité desIndicateurs de la Politique Scientifique (UNIPS, CNRS), for their help in compiling thedatabase and their advice. We have also benefited from remarks by Ajay Agrawal, Marie-Laure Allain, Céline Allard-Prigent, Anne Crubelier, Dominique Foray and ClaudineHermann, and are particularly grateful, for detailed comments, to Bronwyn Hall andEdward Steinmueller.

1. For a simulation analysis of this process in the institutional context of the USA, seeDavid (1994), and for a first attempt of an econometric analysis on the same data as theone used in the present work, see Turner and Mairesse (2002).

2. Three of these studies can be mentioned here. Jaffe (1989) shows that there is in the USAa close relationship at the state level between the number of patents and the importanceof university research, which he interprets as evidence of geographic externalities. Jaffeet al. (1993) investigate the localisation of knowledge externalities using patent citationdata. The authors show that citing and cited patents belong to the same geographicregion with a very high probability. Jaffe and Trajtenberg (1998), also on the basis ofpatent citation data, study the localisation of flows of knowledge at an internationalscale. They find that patents cite much more frequently patents whose inventors live inthe same country than patents whose inventors live in different countries.

3. For example, by adopting a definition of a network as a ‘clique’ in the sense of the theoryof graphs (that is, a set of points which are connected or such that the intensity of theirinterconnections exceeds a certain threshold), Blau (1973) makes the following observa-tions for a group of 411 physicists. Members of large networks are often young, work innew and innovative specialties, have a teaching post and are relatively well known; bycontrast, members of small networks are older, work in established specialties, in presti-gious university departments and are involved in administration. These findings seem toreflect the existence of a cycle in research careers, leading the most productive scientiststo be also part of the administrative elite.

4. The analysis by Cole and Cole (1973) on stratification in science is typical of thisapproach. The authors classify physicists in terms of different criteria such as age, pres-tige within university departments, productivity and scientific awards. They thenmeasure the impact of these characteristics on the researchers’ ranking in terms of scien-tific reputation and visibility. They finally use their results in an attempt to assess the exis-tence and the extent of discrimination possibly arising from differences of race, genderand religion.

5. In future work, a possibility will be to carry on this research at the level of individualresearchers as well. In addition to what we can already observe at the laboratory level,this would allow the analysis of the role of productive and well-known ‘star’ scientists inshaping research networks. See for example the work by Crane (1969 and 1972),Crawford (1971) and Zucker et al. (1998a and 1998b).

6. The Centre National de la Recherche Scientifique (CNRS) is the main French publicorganisation for basic research. With 25 000 employees (11 000 researchers and 14 000engineers, technicians and administrative staff) and over 1200 research and service units(laboratories) throughout the country, the CNRS covers all fields of science. Directlyadministered by the ministry responsible for research, which is also usually responsiblefor higher education, the CNRS has very close links with academic research, researchersfrom the CNRS and from universities often working in the same laboratories.

7. These criteria are mainly based on two practical considerations: researchers had to be‘not too young’ so that we had a history of their publications (the youngest researchersborn in 1960 had already been publishing for a few years in 1992, when they were 32 yearsold), and 1997 was the year for which we could know precisely the laboratories in whichthe researchers were working, when we first started compiling our database. Note thatwe have been able to follow up female researchers who happened to change names duringthe study period because they married.


8. The Science Citation Index (SCI) is produced by the Institute for Scientific Information(ISI). It encompasses all the (‘hard’) scientific disciplines and is constructed on the baseof a compilation of over 3200 of the most cited international periodicals. The quality ofthe data is remarkable and, in particular, the coverage of scientific publications by CNRSunits is very satisfactory (UNIPS, 1999). Ninety-five per cent of the scientific articleswritten by the CNRS researchers are in English and these are fully covered by the SCI.

9. Another solution would be to count each article only once, by simply weighting them bythe inverse of the number of pairs of authors concerned. This point is discussed in sub-section 3.3. It seems that the main results of our analysis would have been qualitativelyunchanged.

10. This information on the individual researchers and their laboratories was provided to usby the Unité des Indicateurs de la Politique Scientifique (UNIPS) of CNRS.

11. There is no strict rule for correspondence between authors and addresses in the SCI,since the number of authors recorded for a scientific article often differs from the numberof addresses listed for them. It is possible that several co-authors have the same address,in which case their address may be listed only once. Or when the collaboration involvesdifferent laboratories, the correspondence between authors and addresses is not alwaysclear. Another possibility is that of multiple affiliations, with one co-author mentioninghis or her affiliation to two or more laboratories, which may result again in a problem ofattribution.

12. Note that for Group 1 articles this is a weighted average in which each article is countedas many times as the number of pairs of CNRS co-authors. The simple average forGroup 1 is 4.5 ( 1741/386). See sub-sections 2.1 and 3.3 and note 9.

13. Note that this matrix is similar to the adjacency matrix used in the graph theory. Thecoefficients of the adjacency matrix are equal to 1 when there is a link between the enti-ties corresponding to rows and those corresponding to columns; otherwise it is 0. Theadjacency matrix thus characterises only the occurrence of collaboration between enti-ties but not their intensities.

14. It is possible, of course, to proceed otherwise; that is, in this example we could havecounted the article for one article giving rise to ‘two-thirds’ of a co-publication betweenX and Y and ‘one-third’ within Y. But, as we said, since we are interested in the analysisof collaboration relations, we deemed it better to consider that the more co-authors, thegreater the weight of an article.

15. Marseille in fact has also relations with Poitiers, Gif sur Yvette, Orsay, Toulouse andVilleurbanne, but these are not taken into account because they all involve less than sixco-publications (see sub-section 2.4).

16. As explained in sub-section 2.4, we only estimated the intensity of co-publicationbetween any two towns (or two laboratories) when the actual number of co-publica-tions between the CNRS researchers in these two towns (or two laboratories) was nottoo small, that is, less than six (or four) over the six-year period, and set it to zerootherwise.

17. The average number of links per town with other towns is only four. Although we do notknow of such a result in another study with which we could compare this estimate, it mayseem somewhat on the low side for towns with at least nine CNRS researchers inour sample (and given our adoption of a rather small threshold of at least six co-publications over six years for the definition of an actual link between two towns).

18. Storage rings have become of great importance throughout the world. They are used tocurve or oscillate the trajectory of light-charged particles (electrons or positrons) thatemit ‘synchrotron radiations’. They thus constitute an extraordinary source of radi-ations of varying wavelengths, especially X-rays. The European ring of the ESRF(European Synchrotron Radiation Facility) is situated at Grenoble and employs as manyas 500 people on a permanent basis. France has two other rings situated at Orsay at theLURE. About 30 outside laboratories collaborate on a permanent basis with the LURE,as do 20 industrial partners, in the field of physics but also chemistry, biology and envir-onmental science, micro-production, lithography and astrophysics. The LURE rings


should soon be replaced by the ‘SOLEIL’ ring, which will constitute a source of ‘super’synchrotron radiation (several thousand times brighter).

19. The chi-squared distance between the specialisation profile column vectors 1 and 2 oftwo laboratories 1 and 2 is thus defined as:

,

where 1i and 1i denote the coefficients of vectors 1 and 2 with i varying from 1 to7 (or 1 to 2) for the overall (specific) profiles, and where n1 and n2 are the numbers ofpublications of the two laboratories, and ^ is the specialisation profile column vector ofthe two laboratories taken together (or weighted average profile). This chi-squareddistance (as shown in Table 10.3) is normalised by dividing it by the 99 percentile valueof the chi-squared statistic of 6 degrees of liberty for the test of equality of the overallspecialisation profile vectors (of dimension 7) for any two given laboratories or towns.

20. Note that the stock of publications not only includes the co-publications of Group 1 ofour final sample of 470 CNRS researchers, but also their co-publications of Group 2 andtheir (few) publications alone. Note also that each publication is counted only once, irres-pective of the number of co-authors.

21. Our measure of productivity thus corresponds to all the publications of our 470 CNRSresearchers in the period 1992–97 (see previous note).

22. Note that estimating a generalised Tobit regression model of both the occurrence andintensity of co-publication also provides practically the same picture as the two corres-ponding separate linear regressions (the estimated correlation between the probit occur-rence equation and the linear intensity equation being not statistically different fromzero).

BIBLIOGRAPHY

Audretsch, D. and Stephan, P. (1996), ‘Company-scientist locational links: the caseof biotechnology’, American Economic Review, 86 (3), 641–52.

Barré, R., Crance M. and Sigogneau, A. (1999), ‘La recherche scientifiquefrançaise: situation démographique’, Etudes et dossiers de l’OST, 1.

Blau, J. (1973), ‘Patterns of communication among theoretical high energy physi-cists’, Sociometry, 37, 391–406.

Beaver, D. and Rosen, R. (1978), ‘Studies in scientific collaboration. Part 1. Theprofessional origins of scientific co-authorship’, Scientometrics, 1, 65–84.

Beaver, D. and Rosen, R. (1979), ‘Studies in scientific collaboration. Part 2.Scientific co-authorship, research productivity and visibility in the French scien-tific elite’, Scientometrics, 1, 133–49.

Callon, M. (1999), ‘Le réseau comme forme émergente et comme modalité decoordination: le cas des interactions stratégiques entre firmes industrielles et lab-oratoires académiques’, in M. Callon et al., Réseau et coordination, Paris:Economica.

Callon, M. and Foray, D. (1997), ‘Introduction: nouvelle économie de la scienceou socio-économie de la recherche scientifique?’, Revue d’Economie Industrielle,79, 13–37.

Cole, J. and Cole, S. (1973), Social Stratification in Science, Chicago IL: Universityof Chicago Press.

Crane, D. (1969), ‘Social structure in a group of scientists: a test of the invisiblecollege hypothesis’, American Sociological Review, 34, 335–52.

�2 �i�n1

(1i � ^

i)2

^

i

� n2

(2i � ^

i)2

^

i


Crane, D. (1972), Invisible Colleges: Diffusion of Knowledge in ScientificCommunities, Chicago IL: University of Chicago Press.

Crawford, S. (1971), ‘Informal communication among scientists in sleep research’,Journal of the American Society for Information Science, 22, 301–10.

Dasgupta, P. and David, P.A. (1994), ‘Toward a new economics of science’,Research Policy, 23 (5), 487–521.

David, P. (1994), ‘A science simulation for studying the US and similar institutionalsetups (SCISIMUS)’, mimeo.

Diamond, A. (1996), ‘The economics of science’, Knowledge and Policy, 9 (2–3),6–49.

Foray, D. (2004), The Economics of Knowledge, Cambridge, MA: MIT Press.Gibbons, M., Limoges, C., Novotny, H., Schwartzman, S., Scott, P. and Trow, M.

(1994), The New Production of Knowledge, London: Sage.Jaffe, A. (1989), ‘Real effects of academic research’, American Economic Review,

79 (5), 957–70.Jaffe, A. and Trajtenberg, M. (1998), ‘International knowledge flows: evidence from

patent citations’, NBER Working Paper 6507.Jaffe, A., Trajtenberg, M. and Henderson, R. (1993), ‘Geographic localization of

knowledge spillovers as evidenced from patent citations’, Quarterly Journal ofEconomics, 108 (3), 557–98.

Katz, J.S. (1994), ‘Geographical proximity and scientific collaboration’,Scientometrics, 31 (1), 31–43.

Leamer, E. and Storper, M. (2000), ‘The economics of geography at the Internetage’, mimeo, University of California at Los Angeles and Management andPublic Policy, November.

Rallet, A. and Torre, A. (2000), ‘Is geographical proximity necessary in the innov-ation networks in the era of global economy?’, mimeo, Université ParisDauphine et Institut National de la Recherche Agronomique.

Shi, Y. (2001), The Economics of Scientific Knowledge, Cheltenham, UK: EdwardElgar.

Shrum, W. and Mullins, N. (1988), ‘Network analysis in the study of science andtechnology’, in Handbook of Quantitative Studies of Science and Technology, ed.A.F.J. Van Raan, Amsterdam: North Holland.

Stephan, P. (1996), ‘The economics of science’, Journal of Economic Literature,34, 1199–235.

Turner, L. and Mairesse, J. (2002), ‘Explaining individual productivity differencesin public research: how important are non-individual determinants? An econo-metric analysis of French physicists (1986–1997)’, Working Paper – Cahiers de laMSE 2002-66, Université Paris I – Panthéon-Sorbonne.

Unité des Indicateurs de la Politique Scientifique (UNIPS) (1999), ‘Les publicationsdes laboratoires du CNRS et leur impact’, March.

Zucker, L., Darby, M. and Armstrong, J. (1998a), ‘Intellectual human capitaland the birth of U.S. biotechnology enterprises’, American Economic Review,88, 290–306.

Zucker, L., Darby, M. and Brewer, M. (1998b), ‘Geographically localized know-ledge: spillovers or markets?’, Economic Inquiry, 36, 65–86.


11. Epistemic communities andcommunities of practice in theknowledge-based firm1

Patrick Cohendet and Ash Amin

INTRODUCTION

As Fransman (1994) has argued, the traditional theoretical approaches tothe firm – transaction costs theory in particular – considers the firm as a‘processor of information’. For these traditional approaches, the behaviourof the firm can be understood as an optimal reaction to external environ-mental signals detected by the firm. The focus is thus on the process of allo-cation of resources needed to cope with this adaptation. In this perspective,the establishment of incentive schemes to avoid informational asymmetriesis the main driving force for firm governance mechanisms.

The competence-based approach, in contrast, relies on a different pointof view: the firm is conceived as ‘a processor of knowledge’, that is, as alocus of construction, selection, usage and development of knowledge.This vision strongly differs from the above ‘information-based’ theories ofthe firm. Considering the firm as a processor of knowledge leads to therecognition that cognitive and related processes are essential, and that rou-tines play a major role in keeping the internal coherence of the organisa-tion. In this perspective, the governance of the firm is no more focused onthe resolution of informational asymmetries but on the co-ordination ofdistributed pieces of knowledge and distributed learning processes. Thefocus of theory now falls on the process of creation of knowledgeresources, as evidenced through the work of, among others, Cyert andMarch (1963), Cohen et al. (1972), Cohen (1991), Loasby (1976; 1983),Eliasson (1990), Dosi and Marengo (1994), and Marengo (1994; 1996).

The existence of two alternative visions of the firm (processor of infor-mation versus processor of knowledge) raises some fundamental theor-etical questions. As Langlois and Foss (1996) have pointed out, we areconfronted with the choice between, on the one hand, a contractualapproach based on transaction costs, where ‘firms and other institutions

296

are alternative bundles of contracts understood as mechanisms for creat-ing and realigning incentives’, and, on the other hand, ‘a qualitativeco-ordination, that is helping co-operating parties to align not their incen-tives but their knowledge and expectations’ (pp. 10–11). But, does the com-petence approach represent a complementary or competing approach totraditional theories of the firm, in particular the dominant transactionbased approach? Which of the two approaches is the best equipped toexplain what Casson (2000) considers to be the main issues addressed bymodern theories of the firm, namely: (1) the boundary of the firm; (2) theinternal organisation of the firm; (3) the formation growth and diversifica-tion of the firm, and (4) the role of the entrepreneur?

The aim of this contribution is to shed light on some aspects of this the-oretical debate, by addressing two issues. First, we argue that the twoapproaches to the firm are more complementary than substitutes. Weassume that firms simultaneously manage competences and transactions,but they do so according to a specific order of priorities. They rank theiractivities according to their intensity of knowledge. Within the domain ofcore competences, the governance mechanisms are specifically devoted toknowledge co-ordination. Then, as we move away from the core, firms tendto allocate resources and adapt to the environment in accordance with gov-ernance mechanisms that are well analysed by the transaction costapproach. Second, we argue that the growing need to co-ordinate knowledgeand achieve coherence within the firm (to avoid an excessive tension betweendifferent governance mechanisms) requires a reconsideration of the actualprocess of production and circulation of knowledge. We emphasise the roleplayed by two specific communities, epistemic communities and commu-nities of practice, in the formation and evolution of the routines of the firm.

1 COMPETENCES AND TRANSACTIONSRECONSIDERED

1.1 The Firm as a Processor of Information

Traditional theories of the firm, especially the transaction-based approachand the principal/agent theory, share some common essential features. Forthem, the behaviour of the firm can be understood as an optimal reactionto information from the external environment. The firm as a rationalprocessor of information implies that the same signals will through timegive rise to the same pattern of action, provided that the technical condi-tions (as expressed by the production function) remain unchanged. Theneoclassical theory of the firm, in particular principal/agent theory, have

Epistemic communities and communities of practice 297

basically reduced co-ordination principles to a bundle of bilateral contractswhich are meant to achieve co-ordination through appropriate incentiveschemes to align self-interested action with common organisational goals.An important ommission of this approach is any serious treatment of theproduction process as a collective activity. The transaction cost approach,even if conceptually different and with a specific focus on the boundariesof the firm, comes to a similar fundamental conclusion: the firm could beseen as a nexus of contracts. Its very reason for existence is to correctmarket failures, when the operation of market mechanisms in terms ofinformation processing is too costly. Transaction cost theory agrees withthe principal/agent vision that information is imperfect and that the exist-ence of potential asymmetries of information authorises unproductiverent-seeking behaviour. The firm is thus conceived as an institutional mech-anism offering a governance structure to solve the problem of misalignedincentives related to imperfect information. It should be emphasised that,in such a vision, the unit of analysis is the transaction, with the productiveactivities which a firm undertakes out of focus. The transactional approachis by definition a defensive one in so far that it reacts to an environmentcharacterised by imperfect information and agents who are opportunistic.Thus, it is in a sense a ‘theory of frictions’ to quote Favereau (1989).

The focus on adaptation to imperfect information signals from the envir-onment does not imply that the contractual approaches are unable toincorporate some aspects of cognition and learning among economicagents. The transaction cost approach assumes action based on boundedrationality, which is to admit the existence of cognitive constraints uponindividuals, and to analyse key learning processes as learning by doing.However, the scope of analysis is extremely narrow, since the cognitive capa-bilities of agents are taken as given. Agents do not change their representa-tion of the world through time, they do not differ in their perception of theenvironment, and they do not pay attention to the definition and evolutionof common sets of rules, codes and languages within the organisation. Tosome extent, it can also be said that the traditional approach tackles theproblem of knowledge. But here, too, the analysis is restricted to a verylimited conception of knowledge: that knowledge is a mere stock resultingfrom the accumulation or loss of information considered as a flux. This is anarrow conception of information that does not acknowledge the cognitiveand non-cognitive mechanisms involved in the production of knowledge.

1.2 The Firm as a Processor of Knowledge

A substantial volume of research, from different disciplines (economichistory, industrial organisation, sociology of organisation, evolutionary


theory, management science, and so on) has seriously questioned the con-tractual vision of the firm (see Penrose, 1959; Richardson, 1960; 1972;Chandler, 1962; 1992; Nelson and Winter, 1982; March and Simon, 1993).The criticisms converge towards a common view that the concept of com-petence is a leading variable for explaining the organisation of firms, as wellas their diversity and persistence (Dosi and Marengo, 1994). The conceptof competence, which relies on that of routines and rules, centres on a viewof the firm as a social institution, the main characteristic of which is toknow (well) how to do certain things. Competences are coherent sets ofcapabilities used in an efficient way. Some of the competences are strategic(‘core-competences’ according to Teece, 1988) and constitute the mainsources of the competitiveness of a firm (‘what a firm does well and betterthan the others’). They are the product of a selection process both internaland external to the firm. How these competences are constructed, com-bined and managed is critical for understanding the boundaries of the firmas well as the co-ordination and incentive structure of the firm.

According to this alternative approach, the firm is conceived as an insti-tution where competences are actively and consciously built, shaped, main-tained and protected. This is a cumulative and strategic process that reliesintensively on the management of knowledge. This has important conse-quences in particular for co-ordination by the firm. First because know-ledge of complex production processes is necessarily distributed (cf. Hayek,1937) and cannot be fully grasped and controlled by a single individual, aprimary role of organisation becomes that of co-ordinating this dispersedknowledge. Second, co-ordination in this case generally involves the cre-ation of commonly shared bodies of knowledge: sets of facts, notions,‘models of the world’, rules, procedures which are – at least partly – knownto all the members of the organisation involved in a given interaction. In asense this kind of co-ordination is a precondition for the co-ordination ofactions that are examined by the literature which implicitly assumes that allthese mechanisms for the co-ordination of dispersed knowledge are alreadyin place. It is most unlikely that incentive mechanisms alone could besufficient to promote this kind of co-ordination. But, and perhaps evenmore important, the focus on knowledge issues highlights the question ofhow such knowledge is generated, maintained, replicated and modified(and possibly also lost) – that is, the question of learning and its nature. Asrepeatedly argued (see, for example, Nelson and Winter, 1982; Dosi andEgidi, 1991), innovative activities involve a kind of learning that is quitedifferent from Bayesian probability updating and regression estimation. Itrequires agents to build new representations of the environment, whichremains largely unknown, and to develop new skills which enable themboth to explore and exploit this world of ever-expanding opportunities.


Such representations are embedded in the routines which characterise theorganisation.

1.3 Competences and Transactions Viewed as ComplementaryMechanisms

In our view the core statement of the competence theory is that the firmmust be seen in primis as a processor of knowledge, not just as a mere infor-mation processing device. But in this perspective, the relationships betweencompetences and transactions may be viewed as complementary. To bemore specific:

1. The firm will first focus its limited attention on its core competences.Within this set of core competences, the firm functions as a knowledgeprocessor that gives full priority to the creation of resources. Such afocus signifies that the activities that belong to the ‘core’ of the firm arenot considered as tradable in the market: they are ‘disconnected’ fromany ‘make or buy’ trade-off as suggested by transaction cost theory.However, the scope of the set of core competences is limited; focusingon core competences is by definition very costly. This requires specificsunk costs, forging and managing co-operation with institutions thathave complementary forms of knowledge, accessing and absorbing themost recent scientific results related to the core competences, and so on.For a given firm, in terms of the exchange of knowledge, this zone ischaracterised by ‘partners’ or ‘quasi-integrated’ suppliers, whoproduce high-value components or systems that are highly strategic.These could be wholly owned suppliers or partly owned suppliers inwhich the firm holds an equity stake and typically transfers personnelto work on a part-time or full-time basis. These suppliers participate inlong-term strategic plans, capital investments and capacity planning,and personnel transfers. The formal duration of the typical contract islong-term, and most contracts are renewed automatically. The sup-pliers also tend to participate in building the knowledge base of thefirm, and benefit from the absorptive capacities accumulated by thefirm. But it is also important for the firm to enhance the absorbingcapacities of the suppliers themselves. The firm provides assistance tosuppliers not only in the areas of quality, cost reduction, factory layoutand inventory management, but also in terms of increasing technolog-ical competences and research facilities. What is essentially transferredin this zone are creative ideas through multiple functional interfaces(manufacturing to manufacturing, engineering to engineering, and soon). This requires permanent capabilities benchmarking within the


group of partners, a substantial investment in inter-firm knowledgeshared routines, and regular activities of socialisation. In contrast therelationship with competitors in this zone is highly unstable and con-flicting, and leads generally to acquisitions or mergers.

2. In decreasing order of attention, next comes the domain of non-corecompetences. This is the domain of activities, which the firm ‘knows’well to do, but does not necessarily invest much in the systematic pro-duction of new knowledge to place it at the leading edge of competition.To make the knowledge effective, the firm mostly has to function withinnetworks and diverse types of alliances where it can access the comple-mentary forms of knowledge required to makes its own knowledge valu-able. In terms of transfer of technology what is at stake in this zone isthe mutual exchange of complementary forms of knowledge. Networksoffer precisely such an opportunity. What differentiates a given eco-nomic agent from another is its specific body of tacit knowledge.Through networks, agents can organise an efficient circulation of codi-fied knowledge through a structure that renders compatible differentsegments of agent-specific tacit knowledge. Agents agree to specialise ina given area of tacit knowledge, because they are confident that the otheragents will increase their specialisation in complementary forms. Thisreduces the risks of overspecialisation, but relies centrally on buildingmutual trust and reciprocity in the production of knowledge. Takinginto account the degree of trust raises an important issue, which has todo with the choice between specialisation and co-operation in the pro-duction of knowledge. As argued by Zuscovitch (1998: 256):

Trust is a tacit agreement in which rather than systematically seeking out thebest opportunity at every instant, each agent takes a longer perspective to thetransactions, as long as his traditional partner does not go beyond somemutually accepted norm. Sharing the risks of specialization is an aspect ofco-operation that manifests an important trust mechanism in network func-tioning. Specialization is a risky business. One may sacrifice the ‘horizontal’ability to satisfy various demands in order to gain ‘vertical’ efficiency in aneffort to increase profitability. Any specializing firm accepts this risk, networkor not. A risk-sharing mechanism is essential because, while aggregate profitsfor participating firms may indeed be superior to the situation where firmsare less specialized, the distribution of profits may be very hazardous. Tomake specialization worthwhile, the dichotomous (win-lose) individualoutcome must be smoothed somehow by a cooperative principle of risksharing.

Trust is relevant for the reliability of other specialised producers ofcomplementary knowledge.2 In such a perspective, one should notworry too much about excessive uncontrolled spillovers and risks of


excessive imitation, precisely because of significant transaction costs.Imitating is very costly, and loose co-operation in informal networksallowing a certain control of the diffusion of spillovers between agents,can be an efficient form of collaboration. One of the key determinantsof innovative networks is the constant trade-off by agents between, onthe one hand, delimiting property rights and, on the other hand, ter-mination of rights of access to complementary forms of knowledge.

3. Finally, away from the focus of attention on competences, one finds theperipheral activities. Once the set of activities that belong to the coreand non-core competences has been chosen, the other activities that donot belong to the domain of competences are managed under tradi-tional methods which may rely on the transaction cost approach. Theseactivities are necessary to support the domain of competences and theygenerally correspond to the larger number of activities and employ-ment positions in the firm. These activities do not require by definitiona strong commitment in terms of knowledge. The firm just needs to ‘beinformed’ of the best practices of external firms and organisations thatcan offer equivalent support services and, if it appears that these activ-ities are too costly to be run within the firm compared to market mech-anisms (according to transaction costs criteria), they will beoutsourced. This is a zone of ‘quasi-market’ relations, where the degreeof supplier–buyer interdependence is generally low. Products are stan-dardised, and require few interactions with other inputs. Contracts areat arm’s length, the duration of which depends on the classical trans-actional parameters. For a given firm, in terms of supplier managementpractices, this zone requires minimal assistance to suppliers, accompa-nied by single functional interfaces (sales to purchasing, for instance),and the practice of price benchmarking. In terms of technology trans-fer what is at stake in this zone is the exchange of an artefact, ratherthan innovative ideas or new tacit knowledge.

Such a ranking of activities is suggested by Langlois and Foss (1996), whenthey argue that beyond core competences, firms will rank their activitiesaccording to an index of growing distance. As we move away from the core,we enter a domain which is more and more regulated by the classical needto process information. A rendering along the lines of the modern eco-nomics of organisation may be: as firms move increasingly away from theircore businesses, they confront increasing adverse selection and moralhazard, since management becomes increasingly unable efficiently tomonitor employees or to evaluate their human capital. Agency costs risecorrespondingly, producing the net profitability disadvantage associatedwith further integration (for a similar story, see Aghion and Tirole, 1995).



Figure 11.1 illustrates the above ranking of knowledge activities withinthe firm. The first zone (Zone 1) is the core itself. The second zone (Zone 2)is one where the firm holds significant pieces of knowledge, but needs toaccess complementary forms of knowledge held by other firms to be ableto develop and use the knowledge efficiently. This zone is characterised by‘networks’. The third zone (Zone 3) is the peripheral zone, where the firmdoes not hold any specific advantage in terms of knowledge.

1.4 Economic Consequences of the Ranking of Activities from the Core tothe Periphery

The consequences of this ‘lexicographic’ choice first focusing on compe-tences, then managing the periphery, are significant. Let us emphasise twoof them:

1. In terms of dynamic evolution: the above representation is essentiallystatic. It corresponds to the actual ranking of activities within the firmat a given moment of time. However, the dynamic functioning of activ-ities could be interpreted along the lines of the evolutionary theory ofthe firm (Cohendet et al., 2000). Through the combined mechanismsof selection and variation in the body of existing routines, lies always

Intensity offocus of

knowledge

PartneringZone 1

NetworkZone 2

MarketZone 3

Distance fromthe core

Source: Amesse and Cohendet (2000).

Figure 11.1 Ranking of activities of the firm from core competences

the possibility of transforming a set of secondary routines situatedin the periphery into a new competence. Naturally, the reverse mecha-nism is also possible. For instance, routines that belong at a givenmoment in time to the core domain could be ‘declassified’ to the com-petence or the peripheral domain over time if they happened not to besuccessful. It must be emphasised that in the selection process thatoperates on routines, in addition to the classical external competitiveenvironment, the attention of the firm operates as an internal elementof selection.

2. In terms of governance structures: the ranking of activities emphasisesthe need for the firm to define at least two distinct governance prior-ities; first, a structure to manage competences in order to align dis-persed knowledge and expectations, and second, a structure conceivedalong the transaction costs criteria to manage the periphery. In a recentarticle, Nonaka and Konno (1998) underline the importance of dom-inant learning practices for managing the core of the organisation(concept of ‘ba’). Within this ‘core’ structure, some contractualschemes may naturally be implemented (for example, stock options, orspecific rewards for inventors within the organisation), but these arenot essential when compared with the priority given to the stimulationof collective learning processes. In the second structure of governance,classical contractual schemes are dominant to ensure the informationprocessing that is central to the functioning of the periphery.3

1.5 Summary

We have argued thus far that a competence perspective on the firm – asopposed to a contract or transaction-based perspective – opens up consid-erably the scope for exploring how firms learn and adapt in complex andchanging business environments. The acquisition and renewal of know-ledge – tacit and formal – is crucial for survival, it cannot be taken as pre-given, and it occurs at a variety of levels through a variety of means. For suchreasons, the firm has to be seen as much more than an allocating mechanism.It is primarily a generator of resources, defined as distinctive knowledge andorganisational routines, locked in core and non-core competences. We havealso argued that the cognitive set-up of firms – more specifically their organ-isational rationality – is crucial for framing expectations and outcomes.Thus, for example, substantive rationality, based upon the principle of rule-following behaviour, is efficient for decision-making in stable and simple-response situations (for example, mass production for planned markets), butinappropriate for a continually changing environment. In contrast, aprocedural rationality favours learning through continual adjustment as


agents modify their behaviour to external circumstances, but is ill-equippedfor strategic action in the context of radical change. In practice, individualfirms tend to draw on these rationalities selectively depending on the natureof the task in hand, but routines often settle around one dominant cognitiveset-up, which is precisely why firms vary in their learning potential.

We have also claimed that a key governance challenge for firms is to rec-oncile organisational arrangements for transactional efficiency with thosefor sustaining learning. This is much the same as the exploration versusexploitation trade off signalled by James March (1991) to mark thedilemma faced by firms in balancing the short-term exploitation of exist-ing competence and the long-term exploration of new competence. Thus inthe ‘learning domain’ organisational and management practices mustfacilitate the creation and circulation of knowledge as well as strengthendecision-making. The challenge is to build trust, long-term commitmentsand knowledge externalities, to encourage experimentalism, variety andcreative friction, and to facilitate the conversion of knowledge (betweentacit and explicit, between individual to collective, between local andglobal). All this tends to privilege de-centred management, distributedcapability and, to a degree, organisational ‘excess’ (Nohria and Ghoshal,1997). In contrast, the transactional domain of routine activities (forexample, securing supply, achieving scale economies, make–buy trade-offs)demands the efficient allocation of resources, largely through substantiveor procedural responses to the environment. Here, as ever, the governancechoice is between hierarchy and market, based on the cost-efficiency oftransactions largely of a contractual nature.

2 THE COHERENCE OF THE FIRM AS APROCESSOR OF KNOWLEDGE:ORGANISATIONAL LEARNING BETWEENEPISTEMIC COMMUNITIES AND COMMUNITIESOF PRACTICE

The perspective that we have set out, based on matching two distinct oper-ational domains and governance imperatives, might imply that the firm isa neatly divided entity. The aim of this section is to avoid this risk byframing the firm as a set of overlapping learning practices, involving recur-sive interchange between the domains and solutions emerging from thepractices themselves. Ongoing learning in ‘epistemic communities’ and‘communities of practice’ within firms bridges different types of rational-ity in separate domains. Consequently, in practice, the resolution of thegovernance problem of reconciling creation and allocation of resources is


less a matter of formally integrating two separate models of governance,than a question of how meaningful links are generated in daily practiceacross the distributed mechanisms of governance.

2.1 Bridging Epistemic Communities and Communities of Practice

In the traditional vision of the firm, the distinction between epistemic com-munities and communities of practice is based on a linear representation ofthe process of transformation of knowledge. This process is viewed asevolving from separate departments in charge of producing new (deliber-ate) knowledge or handling and distributing information to the otherdepartments that assimilate and use this new knowledge to improve theircurrent activities. These latter departments can to some extent producesome new knowledge from their routine activities, but this would be a non-deliberate form of production of knowledge that emerges as a by-productof learning by using or learning by doing. We will argue that thedifferentiation between deliberate and non-deliberate forms of knowledgeproduction is becoming strongly blurred. In a knowledge-based context theessence of the coherence of the firm precisely relies on the ways these twotypes of communities deliberately interact and organise simultaneously theproduction and circulation of knowledge. But, first the main characteris-tics of each community need to be detailed:

1. Epistemic communities are involved in the deliberate production ofknowledge. They comprise ‘agents who work on a mutually recognizedsubset of knowledge issues, and who at the very least accept some com-monly procedural authority as essential to the success of their collect-ive building activities’ (Cowan et al., 2000: 234). The existence ofprocedural authority aids in the resolution of potential disputes andprovides a reference point for achieving ‘closure’ in various stages ofthe codification process. In these communities the knowledge basefrom which the agents work is generally highly codified, but ‘paradox-ically, its existence and contents are matters left tacit among the groupunless some dispute or memory problem arises’ (Cowan et al., 2000:234). What characterises the knowledge activities within these com-munities is that they are deliberately focused on the production of newknowledge, with a priori little reference to the different contexts inwhich the new knowledge produced will be used.

2. Communities of practice involve learning in doing, or the non-deliberate production of knowledge. Wenger (1998) defines a commu-nity of practice as one marked by three dimensions, which take shapethrough repeated interaction, rather than rule or design. The first is


mutual engagement among participants, involving negotiating diver-sity, doing things together, mutual relationships and community main-tenance. The second is joint enterprise, involving the negotiation ofdiversity among members, the formation of a local code of practiceand a regime of mutual accountability. The third dimension is ashared repertoire that draws on stories, artefacts, discourses, concepts,historical events, discourses, reflecting a history of mutual engage-ment, ambiguity of meaning through new metaphors, and dynamicco-ordination through the latter ‘tools’ of engagement. Thus a com-munity of practice – drawing on the subconscious, interaction, par-ticipation and reified knowledge to act, interpret, innovate andcommunicate – acts as ‘a locally negotiated regime of competence’(ibid.: 137), as ‘shared histories of learning’ (ibid.: 86).4 What charac-terises these communities is that the production of knowledge is aby-product of the practice.

In practice, increasingly the frontier between the two communities isbecoming blurred. As Lundvall (2000) has pointed out, the emergence ofnew forms of learning such as ‘experimental learning’ makes thedifferentiation between ‘on-line’ and ‘off-line’ learning activities less andless relevant. The importance of experimental learning has been firstemphasised by Paul David in a contribution with Warren Sanderson(David and Sanderson, 1996). For Lundvall (OECD, 2000: 25) experimen-tal learning may start ‘on-line’, that is to say, during the process of pro-ducing a good, but consists in deliberately experimenting during theproduction process:

By doing so, one creates new options and variety. This form of learning is basedon a strategy whereby experimentation allows for collecting data, on the basis ofwhich the best strategy for future activities is chosen. For example, a professorcan undertake pedagogical experiments; the craftsman can seek new solutionsto a problem even during the fabrication process. The possibility of moving thistype of learning in many activities represents an important transition in the his-torical emergence of the knowledge-based economy. In effect, as long as anactivity remains fundamentally based on learning processes that are routineadaptation procedures and leave no room for programming experiments duringeconomic activity, there remains a strong dichotomy between those who delib-erately produce knowledge and those who use and exploit it. When an activitymoves to higher forms of learning, and where the individual can programmeexperiments and obtain results, the production of knowledge becomes muchmore collectively distributed . . . With the emergence of experimental learning,the feedback and reciprocal links that tie ‘on-line’ learning processes and inhouse R&D together – and whereby a potential creative activity effectively con-tributes to the production of knowledge – become crucial.


This implies a complete reconsideration of the separation between epis-temic communities and communities of practice.

The management of the collectively distributed knowledge within theorganisation that brings together epistemic communities and communitiesof practice is one of the cornerstones of the coherence of the firm in aknowledge-based context. The development of different modes of inter-action between the two types of communities (that rely on particularprocesses of codification of knowledge) becomes critical. Devices areneeded to overcome the problem of knowledge fit or integration acrossboundaries (for example, longer-term ‘languaging’ devices, informal cross-overs and deployment of ‘boundary spanning’ individuals such as brokersand intermediaries). Nooteboom (1999a) has emphasised the role of thirdparty ‘go-betweens’ as vital brokers of innovation who help to sedimenttrust, maintain unique secrets, resolve conflicts, reveal mutual advantageand introduce innovation without destabilising established competenceswithin each firm.

Learning, then, is a fine-grained and grounded process (Gibbons et al.,1994) that straddles across the deliberate and the routine, across the epis-temic community and the community of practice. It involves trial and manyerrors, chance discoveries, and mistakes, in a context of firms operating as‘experimental learning machines’ (Eliasson, 1994) in uncertain circum-stances. Their daily hazard is to ‘act prematurely on a very incompleteinformation base’ (Eliasson, 1994: 184). This is not to say, of course, thatall is left to chance as we shall see below, but it does mean that distributedor varied learning, is neither guaranteed nor that easily ‘arranged’(Metcalfe, 1998).

In many regards, the daily practice of learning across the boundaries ofthe firm helps adaptation and continuity. And here, the ‘anthropological’literature on firms, stressing the generation of knowledge through practiceand social interaction, is particularly insightful. For example, in theirseminal article on communities of practice, John Seely Brown and PaulDuguid (1996) argue that learning and innovation only too often are situ-ated practices in the everyday humdrum of interaction with one’s peers andwith the environment. They explain:

Alternative worldviews, then, do not lie in the laboratory or strategic planningoffice alone, condemning everyone else in the organisation to a unitary culture.Alternatives are inevitably distributed throughout all the different communitiesthat make up the organisation. For it is the organisation’s communities, at alllevels, who are in contact with the environment and involved in interpretativesense making, congruence finding, and adapting. It is from any site of such inter-actions that new insights can be co-produced. If an organisational core over-looks or curtails the enacting in its midst by ignoring or disrupting its


communities-of-practice, it threatens its survival in two ways. It will not onlythreaten to destroy the very working and learning practices by which it, know-ingly or unknowingly, survives. It will also cut itself off from a major source ofpotential innovation that inevitably arises in the course of the working andlearning. (Ibid.: 76)

It is clear that, for Brown and Duguid, organisational innovation is relatedto tolerance for diversity (as most evolutionary economists would agree),but the important point here is the emphasis placed on the development ofalternatives (that is, exploration) within autonomous communities of prac-tice within the organisation. Once again, they explain:

Within an organisation perceived as a collective of communities, not simply ofindividuals, in which enacting experiments are legitimate, separate communityperspectives can be amplified by inter-changes among communities. Out of thisfriction of competing ideas can come the sort of improvisational sparks neces-sary for igniting organisational innovation. Thus large organisations, reflectivelystructured, are perhaps well positioned to be highly innovative and to deal withdiscontinuities. If their internal communities have a reasonable degree of auton-omy and independence from the dominant worldview, large organisations mightactually accelerate innovation. (Ibid.: 77–8)

In this account of innovation, the impression we have is that groups withinorganisations act, at one and the same time, recursively (or reflectively) andprocedurally. The variety and redundancy necessary for organisationallearning is situated within each part of an organisation or firm. Everyorganisation is made up of many communities of practice in which learn-ing is not a matter of conscious design or recognisable rationalities and cog-nitive frames, but a matter of new meanings and emergent structuresarising out of common enterprise, experience and sociability – learning indoing.5

Consistent with the argument thus far that learning is distributed, com-posite and realised in communities of practice, might also be the sugges-tion, increasingly peddled in the organisational literature, that the firm isan autopoetic cognitive system, a site for developing knowledge in a self-referential manner, usually through internal communication and ‘languag-ing’ procedures at different scales of organisation (von Krogh and Roos,1995; Magalhães, 1998). Indeed, Salvatore Vicari and Gabriele Toniolo(1998) argued that the firm should be seen as a cognitive system whichenacts and makes sense of the environment only from its own individualpoint of view. As such, it makes no sense to theorise learning as a linear orprocedural reaction to an environment ‘out there’, since, at least accordingto Vicari and Toniolo, the latter is known only through the firm’s cognitiveschemata (scripts, maps, data and stories), which define the structure of its


knowledge. The firm has knowledge only of itself but, equally, it needs toproduce knowledge in order to survive. For Vicari and Toniolo, a firm doesthis via enactment, that is, by constructing maps and images of realitybased on raw data that are monitored and interpreted, expectations aboutcustomer and competitor behaviour, time and motion studies, and so on.The performance of these enactments, rather than ‘in here’ anticipation of‘out there’ signals, is the basis upon which firms know and act.

Such an interpretation of knowledge formation in the firm considerablyopens up the scope for seeing learning and innovation as a fragile, experi-mental and uncertain process. So much so, that Vicari and Toniolo (1998)claim that innovation and learning is made possible via the ‘production oferrors’ (ibid.: 211) – errors which could imperil the firm if the gap betweenthe enacted knowledge scripts and say customer preferences and expect-ations remains wide. They distinguish between chance errors, generated byalterations in the external environment (for example, new competitors ortechnical changes) or transformations in the cognitive structure (forexample, the discovery of unusual behaviours or product shifts) and inten-tional errors, produced through intentional search for disturbing events(for example, customer surveys or firing management) or experimentation(for example, via acquisitions or the launch of new product lines).

2.2 Governance Implications

The management implication of stressing the role of error could betwofold: either that there is only so much that can be done in order to over-come the ever-present gap between autopoietic knowledge and marketpreferences; or, as argued by Vicari and Toniolo, that firms should fosterlearning through error production, for example, by developing proceduresto detect and amplify market signals (for example, benchmarking to detectthe size of error, or acknowledgement of error throughout the firm).Perhaps both implications are valid, especially if we take seriously the claimthat learning occurs in communities of practice which combine proceduraland recursive knowledge, exploration and exploitation.

The broader implication of the emphasis on learning in communities ofpractice is that the ‘management by design’ of learning is not feasible.Instead, as already indicated earlier, the imperative is to maximise thepotential for de-centred and distributed learning. This includes the removalof organisational barriers designed for time- and resource-efficiency in thefield of innovation, or, put differently, root-and-branch evaluation of thevaried tasks of the organisation followed by the application of transactionalrules of governance only to a restricted set of functions and corporate pri-orities. It requires, above all, the identification and full acknowledgement of


the communities of practice in play within and across the boundaries of thefirm. It also requires analysis of the learning processes within these com-munities, and actions designed to facilitate new learning within them as anongoing process – not dispositions from the centre.

Part of this recognition of de-centred learning involves mobilising the‘soft’ infrastructure for learning – both procedural and recursive – andrecognising its integral relationship with the formal infrastructure forlearning (from courses and texts to artefacts and technologies). We havealready had a glimpse of some of the elements. Slack, memory and forget-ting are clearly of crucial importance: slack, facilitated by the retention ofskills and capabilities in excess of the minimum necessary for immediateuse, helps ‘innovative projects to be pursued because it buffers organisa-tions from the uncertain success of these projects, fostering a culture ofexperimentation’ (Nohria and Ghoshal, 1997: 52); memory, facilitated byremembering, inter-generation mixture, and stories among long-servingemployees, mobilises the fruits of experience, including knowledge of pasttrials and errors; forgetting, stimulated by employee rotation and mobility,new training and new routines, helps to weed out practices that are notsuited for changing circumstances.

Another aspect of the soft infrastructure, so clearly evident from thework of Wenger and Brown and Duguid is the practice itself of commu-nity. This does not necessarily mean consensus or trust and loyalty (if thelatter two are meant to imply lack of conflict and dissonance). Instead, itrefers to the daily practices of a group bound together by common purposeand expertise (for example, insurance claims processors, middle mangers,R&D workers). Learning and innovation is the product of shared expertise,social discourse, sociability, argument, disagreement, negotiation skills,and so on. It is the realisation of potential through the practices of jointinteraction that is important. Of course, not all communities of practice arelearning communities, leaving much to be done by management to stimu-late ‘learning in doing’: decisional autonomy, allocation of complex tasks,away days and group reflections, opportunities for socialising, group work,opportunities for dispute resolution, and encouragement of talk, exchangeof ideas and ‘manageable’ disagreement.

A third aspect of soft learning is dissonance and experimentation.Creative communities are those which are able to mobilise difference,variety and counter-argument. Thus, to accompany exploitation of currentopportunities through procedural efficiency, they need to regularise explor-ation, as a means of generating new routines, goals and possibilities.Exploration, or learning to learn, is routinely sacrificed in most organisa-tional cultures largely for the threat it poses to current repertoire and com-petitive advantage. It therefore requires active pursuit, perhaps through the


encouragement of ideas within groups, competition for ideas betweengroups, use of brokers and intermediaries to manage external exposure andadoption of new routines, encouragement of scenario-building and experi-mental games, and so on. Such a pursuit cannot be restricted to individualcommunities of practice, but has to involve the entire organisation, for asRichard Nelson (1994: 238) notes: ‘devising and learning to use effectivelya significantly new organisational form involves much the same kind ofuncertainty, experimental groping, and learning by making mistakes . . .that makes technological invention and innovation’.

What remains for the central management in a de-centred organisationalenvironment, other than actions such as those above to support theongoing practices of distributed learning? Perhaps the central imperative isto hold the network of distributed competences in place. As Nohria andGhoshal (1997: 87) explain in the context of the multinational corporationwhich relies on its subsidiaries for the creation, adoption and distributionof innovations: ‘the real leverage lies in creating a shared context andcommon purpose and in enhancing the communication densities withinand across the organisation’s internal and external boundaries’.

It is interesting to note that at the level of the organisation as a whole,here too, it is the soft infrastructure for learning that is increasinglystressed. What seems to be important is to find ways of binding thedifferent subsidiaries, nodes in a network of alliances, or communities ofpractice, into a common enterprise, and perhaps in ways which go beyondthe traditional mechanisms of propriety rights and the employmentcontract. Nohria and Ghoshal seem to want to emphasise the role ofsocialisation (for example, via corporate encounters, conferences andrecreational clubs) and mechanisms that lead to normative integration (forexample, membership incentives such as access to health care or travel con-cessions, company rituals, slogans, and mission statements, inculcation ofcorporate or brand quality standards, and so on). No doubt there are manymore integration mechanisms to ensure that the parts do somehow belongto the same whole.

Nohria and Ghoshal also stress the architecture of communication asa central management concern in a system of de-centred learning. Suchcommunication, however, is no longer a simple matter of information flowwithin and beyond a firm, as it is in the classical contract-based modeldesigned to minimise transaction costs and other frictions impeding infor-mation processing. Now it is a matter of ensuring that there is effectivecommunication between self-governing but interdependent units. It is cru-cially a matter of relational or cognitive proximity (Nooteboom, 1999b),implying linguistic and semantic equivalence, shared tacit knowledge,rapid flow and processing of information, trust or other conventions of


negotiation. Indeed, perhaps the better word is interaction, with its activepromotion and management a key management priority.

The use of state-of-the-art communications technologies to facilitatereal-time and two-way interchange across the boundaries of the firm is, ofcourse, one aspect of the architecture of interaction, but it is not the onlyaspect. Von Krogh et al. (2000) for instance have argued the centrality of‘knowledge enablers’ as boundary managers in learning-based networks.For example, they stress the role of interaction through dialogue with cus-tomers based on shared tacit knowledge as a way of constructing markets,they argue the importance of entering into the domain of suppliers or usersof technology through personnel exchanges, they emphasise the need todevelop organisational conversations through block conferences, personnelrotation, and exchange of business plans among decentralised units, andthey recommend knowledge managers as a new professional cadre dedi-cated, among other things, to managing relationships across internal andexternal boundaries.

Governance of the architecture for dissipated learning, thus, is thecentral management challenge in the domain of resource mobilisation,alongside, of course, as argued earlier, managing the domain of resourceallocation. On this point, to return to the original paradox of reconcilingexploration and exploitation, the major conflict may well lie in the cultureclash between procedural and recursive thinking within whatever is desig-nated as ‘management’ in an organisation. At the level of communities ofpractice, we suggested that everyday ongoings tend to generate a hybridculture that is capable of both experimentation and routine response. Atthe level of designated corporate or net-wide management, groups tend tobe specialised in either the preservation of routine or the search for novelty.How to de-centre and reintegrate the centre?

One of the main issues to solve the tension between centralisation anddecentralisation in the organisational learning process is the building of acommon knowledge, specific to the firm, that integrates the decentralisedbodies of knowledge held by members of the firm. As has been said, firmsrequire both centralisation and decentralisation to operate successfully in achanging environment. Decentralisation in the acquisition of knowledge isa source of diversity, experimentation and ‘ultimately’ of learning. But,eventually, knowledge has to be made available for exploitation by the entireorganisation. When agents differ with regard to their representations of theenvironment and their cognitive capabilities, a body of common (or collect-ive) knowledge must exist, within the organisation, which guarantees thecoherence of the various learning processes (Crémer, 1990). This is a pre-requisite for an efficient management of the competences. In order to copewith changing environments, the process of generation and modification of


such a body of common knowledge, although fed by decentralised learn-ing processes, has to undergo some forms of centralisation, even if it isbasically maintained by decentralised learning processes (Cohendet andLlerena, 1991).

CONCLUSION

This contribution has stressed that the production and circulation of col-lective organisational knowledge is a key determinant of the capability ofthe organisation to innovate. It considers the ‘cognitive architecture’ ofknowledge within the firm (the way knowledge is produced, stored,exchanged, transmitted and retrieved) to strongly influence the process oforganisational learning and, in turn, the creation process. Two main fea-tures of the cognitive architecture of the firm have been emphasised: firstthe ‘ranking of activities’ by the firm in terms of the level of attentionrequired in a knowledge context; second, the nature of interactions betweenepistemic communities and communities of practice within the firm.

Of course, many other types of community participate in the productionand circulation of knowledge within the firm. For instance, project teamshave always been considered as key organisational devices to bring togetherdifferent forms of heterogeneous functional knowledge held by agents,in order to develop new knowledge. Our purpose has not been to coverall knowledge actor-networks. However, we are assuming that in theknowledge-based economy, of growing importance will be the mutualinteractions between communities of practice and epistemic communities,as a tool for both extensive learning and governance cohesion. This per-spective on knowledge communities opens a new avenue of research in thecompetence-based theory of the firm. In this contribution we have concen-trated on governance mechanisms. It certainly would be of major interestto investigate, within such a perspective, individual incentive mechanisms,the relationships between managers and stakeholders, the role of the entre-preneur and many other key features of the learning and governance envi-ronment of the firm.

NOTES

1. This contribution has been particularly inspired by ideas brought to the TIPIK project byPaul David on epistemic communities and the collective building of knowledge. It drawson our joint work, in particular, Amin and Cohendet (2000).

2. The institutionalisation of incentives for validation (peer refereeing , for instance) in epis-temic communities may vary widely. The choice for an agent to specialise in one domain


of knowledge (and to bear the sunk costs) in co-operation with others agents that acceptspecialising in complementary types of knowledge is an important line of research tounderstand the management of knowledge by organisations.

3. The possibility of a dual structure of governance within the same firm raises the problemof internal coherence which could also impact on the internal process of technologicaltransfer. The conflicting tensions within a firm that is at the same time trying to organiseand manage its information flow and its knowledge base have been analysed by Marengo(1994), through a discussion of the limits of the well-known multi-divisional form(‘M-form’) and functional form (‘U-form’). He notes that these traditional forms are con-ceived to solve information problems (information overload by managers, in particular),but that they are not appropriate for creating and transferring knowledge. He argues, forexample, that the U-form centralises competences in inter-functional co-ordination anddecentralises instead to functional departments competences in many strategic issues con-cerning products and diversification. With the growing multiplicity of products the func-tional structure does not seem that of information overload, but that of mismatchbetween competences and tasks. Chief executives are unable to do their job effectively, notbecause they are burdened by an excess of information, but rather because the organiza-tional structure does not enable them to develop the necessary competences. Chief exec-utives should respond to environmental changes, but when such changes push towardsproduct diversification, many of the competences that are necessary to promote andmanage diversity remain, in the U-form, at the level of functional departments. In thesame vein, he also argues that the M-form, by preventing cross-fertilisation of innovativeideas between separate departments, does not favour the internal transfer of knowledge.

4. Learning, for Wenger, occurs through ongoing practice and draws on social energy andpower generated through interaction in joint enterprises with some history. He identifiesthree infrastructures of learning – corresponding to the three dimensions of a communityof practice – which potentially have enough novelty, perturbation and emergence in themto sustain incremental and discontinuous learning, as well as procedural adaptation andgoal monitoring. These are infrastructures which draw upon a staggeringly broad rangeof facilities, tools, practices and conventions (lest we are inclined to think that learning inaction is a simple process). One infrastructure is engagement, composed of mutuality(supported by such routines as joint tasks and interactive spaces), competence (supportedby training, encouragement of initiative and judgement) and continuity (supported byreified memory locked in data, documents, files as well as participatory memory unlockedby storytelling and inter-generation encounters). Another is alignment, composed of con-vergence (facilitated by common focus, shared values and leadership), co-ordination(helped by such devices such as standards, information transmission, feedback, divisionof labour, and deadlines) and arbitration (facilitated by rules, policies and conflict reso-lution techniques). The third infrastructure is imagination, composed of orientation(helped by visualisation tools, examples, explanations, codes and organisational charts),reflection (supported by retreats, time-off, conversations and pattern analysis) and explor-ation (facilitated by scenario-building, prototypes, play, simulations, experimentation).These are embedded infrastructures of learning built into the routines and daily practicesof members, and the facilities put into place through experience or management decision.They are features of all the communities of practice that are to be found within and acrossorganisations. Wenger’s example of insurance claim processors should not be taken tomean that learning of the sort he describes does not apply to top management, strategistsand scientists. Indeed, one of the remarkable early insights of applications of actornetwork theory in the literature on sociology of science, was to show that incremental andradical learning in the R&D laboratory is no different from the processes described byWenger, locked as it is in routines, conversations, artefacts, things, memory, stories and soon. It has been argued recently that the resource-based perspective, with its emphasis onthe centrality of knowledge creation for competitive advantage, ‘does not provide anexplanation about how resources such as organisational knowledge develop over time’(Probst et al., 1998: 243). Thus, the sophistication with which competitive advantage is


explained at any point in time seems not matched in diachronic analysis of learning tra-jectories. Probst et al. provide a partial corrective by mobilising metaphors from evolu-tionary economics (for example, imitation, selection, variety, replication, and so on). Forexample, they relate changes in organisational knowledge to the cumulative effects ofindividual and group learning: when such learning becomes routinised and synthesisedthrough agent interaction and shared beliefs over time, the organisational knowledge setcan be said to be transformed. Thus, replication at one level becomes a precondition forinnovation at another level.

5. The above perspective properly contextualises also the processes that feed into radicalinnovations forced by dramatic events. A revealing example is provided by EdwinHutchins (1996) in his study of how a ship’s navigation team arrived at a new stable pro-cedure when, upon entering a harbour, a large ship suffered an engineering breakdownthat disabled a vital piece of navigational equipment. Following a chaotic and unsuccess-ful search for a solution from reflection involving thought experiments to computationaland textual alternatives, the team developed an answer through acting. As local task werefound for individuals distributed across the ship, the ensuing sequence of actions and con-versations, drawing on experience and experimentation, led to the construction of a solu-tion based on trial and testing. On this occasion, a solution was found on time, and thereis every possibility that the ship could have got into very serious trouble. The point,however, is that other alternatives such as learning by design, once a snap solution failedto emerge, were not viable.

BIBLIOGRAPHY

Aghion, P. and Tirole, J. (1995), ‘Some implications of growth for organizationalform and ownership structure’, European Economic Journal, 39 (3), 440–55.

Alchian, A. (1951), ‘Uncertainty, evolution and economic theory’, Journal ofPolitical Economy, 57, 211–21.

Alchian, A. and Demsetz, H. (1972), ‘Production, information costs and economicorganization’, American Economic Review, 62, 777–95.

Amesse, F. and Cohendet P (2000), ‘Technology transfer revisited from the per-spective of the knowledge-based economy’, paper presented at the conference onEconomics and Management of Knowledge, HEC Montreal, 5–6 May.

Amin, A. and Cohendet, P. (1999), ‘Learning and adaptation in decentralized busi-ness networks’, Society and Space, 17, 87–104.

Amin, A. and Cohendet, P. (2000), ‘Organisational learning and governancethrough embedded practices’, Journal of Management and Governance, 4, 93–116.

Andersen, E.S. (1998), ‘The evolution of the organisation of industry’, AalborgUniversity, DRUID Working Paper 98–13.

Aoki, M. (1986), ‘Horizontal vs. vertical information structure of the firm’,American Economic Review, 76, 971–83.

Aoki, M. (1988), Information, Incentives and Bargaining in the Japanese Economy,Cambridge, Cambridge University Press.

Aoki, M. (1990a), ‘Toward an economic model of the Japanese firm’, Journal ofEconomic Literature, 28, 1–27.

Aoki, M. (1990b), ‘The participatory generation of information rents and thetheory of the firm’, in Aoki, M., Gustafsson, B. and Williamson, O.E. (eds), TheFirm as a Nexus of Treaties, London, Sage, pp. 26–52.

Aoki, M., Gustafsson, B. and Williamson, O.E. (eds) (1990), The Firm as a Nexusof Treaties, London, Sage.


Arrow, K.J. (1974a), The Limits of Organization, New York and London, W.W.Norton.

Arrow, K.J. (1985), ‘Informational structure of the firm’, American EconomicReview, 75, 303–7.

Avadikyan, A., Cohendet, P. and Llerena, P. (1995), ‘Coherence, diversity of assets:towards an evolutionary approach’, Revue International de Systémique, 7 (5),505–31.

Brown, J.S. and Duguid, P. (1996), ‘Organizational learning and communities-of-practice’, in Cohen, M. and Sproull, L. (eds) Organizational Learning, London,Sage.

Casson, M. (2000), ‘An entrepreneurial theory of the firm’, in Foss, N. and Mahnke,V. (eds), Competence, Governance, and Entrepreneurship, Oxford, OxfordUniversity Press, pp. 116–46.

Chandler, A.D. Jr (1962), Strategy and Structure: Chapters in the History ofIndustrial Enterprise, Cambridge, MA, MIT Press.

Chandler, A.D. Jr (1992), ‘Corporate strategy, structure and control methods in theUnited States during the 20th century’, Industrial and Corporate Change, 1 (2),263–84.

Coase, R.M. (1937), ‘The nature of the firm’, Economica, 4, 386–405.Coase, R.M. (1988), ‘Lectures on “The nature of the firm” 50 years after’, Journal

of Law, Economics and Organization, 4, 3–47.Cohen, M.D. (1984), ‘Conflict and complexity: goal diversity and organizational

search effectiveness’, American Political Science Review, 78, 435–51.Cohen, M.D. (1987), ‘Adaptation of organizational routines’, mimeo, Ann Arbor,

University of Michigan, Institute of Public Policy Studies.Cohen, M.D. (1991), ‘Individual learning and organizational routine: emerging

connections’, Organization Science, 2, 135–9.Cohen, M.D. and Bacdayan, P. (1991), ‘Organizational routines are stored in pro-

cedural memory: evidence from a laboratory study’, mimeo, Ann Arbor,University of Michigan, Institute of Public Policy Studies.

Cohen, M.D., Burkhart, R., Dosi, G., Egidi, M., Marengo, L, Warglien, M. andWinter, S. (1996), ‘Routines and other recurring action patterns of organiza-tions: contemporary research issues’, Industrial and Corporate Change, 5 (3),653–98.

Cohen, M.D., March, J.G. and Olsen, J.P. (1972), ‘A garbage can model of organ-izational choice’, Administrative Sciences Quarterly, 17, 1–25.

Cohendet, P. (1996), ‘Information, connaissances et théorie de la firme évolution-niste’, séminaire Economie de l’Information, Commissariat au Plan, Paris,16 April.

Cohendet, P. and Llerena, P. (1991), ‘Integration and learning processes’, inBourgine, P. and Walliser, B. (eds), Economics and Artificial Intelligence, Oxford,Pergamon Press.

Cohendet, P., Llerena, P. and Marengo, M. (2000), ‘Is there a pilot in the evolu-tionary firm?’, in Foss, N. and Mahnke, V. (eds), Competence, Governance, andEntrepreneurship, Oxford, Oxford University Press, pp. 95–116.

Coriat, B. and Dosi, G. (1995), ‘Learning how to govern and learning how to solveproblems: on the coevolution of competences, conflicts and organizational rou-tines’, IIASA Working Paper No. 95–06 Laxenburg (Austria).

Coriat, B. and Weinstein, O. (1995), Les nouvelles theories de l’entreprise, Livre dePoche, Collection Référence, Paris: La Découverte.


Coriat, B. and Weinstein, O. (1996), ‘Sur la théorie évolutionniste de la firme’,Colloque ‘L’évolutionnisme, fondements, perspectives et réalisations’, Sorbonne,Paris, 19–20 September.

Cowan, R., David, P. and Foray, D. (2000), ‘The explicit economics of knowledgecodification and tacitness’, Industrial and Corporate Change, 9 (2), 212–53.

Crémer, J. (1990), ‘Common knowledge and the co-ordination of economic activ-ities’, in Aoki, M., Gustafsson, B. and Williamson, O.E. (eds), The Firm as aNexus of Treaties, London, Sage, pp. 53–75.

Crémer, J. (1993), ‘Corporate culture and shared knowledge’, Industrial andCorporate Change, 2 (3), 351–86.

Crémer, J. and Riordan, M.H. (1987), ‘On governing multilateral relationships withbilateral contracts’, RAND Journal of Economics, 18, 436–51.

Cyert, R.M. and March, J.G. (1963), A Behavioural Theory of the Firm, EnglewoodCliffs, NJ, Prentice-Hall.

David, P. (1994), ‘Why are institutions the “carriers of history”? Path dependenceand the evolution of conventions, organizations and institutions’, EconomicDynamic and Structural Change, 5 (2), 205–20.

David, P. and Foray, D. (1996), ‘Accessing and expanding the science and techno-logy knowledge base’, STI Review, 16, 87–116.

David, P. and Sanderson, W. (1996), ‘Making use of treacherous advice: cognitivelearning, Bayesian adaptation; and the tenacity of unreliable knowledge’,MERIT working paper, 2/96–004, University of Limburg.

Dosi, G. (1988), ‘Sources, procedures and microeconomic effects of innovation’,Journal of Economic Literature, 26, 1120–71.

Dosi, G. and Egidi, M. (1991), ‘Substantive and procedural uncertainty. An explor-ation of economic behaviours in complex and changing environments’, Journalof Evolutionary Economics, 1, 145–68.

Dosi, G. and Marengo, L. (1994), ‘Toward a theory of organizational competences’,in England, R.W. (ed.), Evolutionary Concepts in Contemporary Economics, AnnArbor, MI, University of Michigan Press.

Dosi, G., Freeman, C., Nelson, R., Silverberg, G. and Soete, L. (eds) (1988),Technical Change and Economic Theory, London, Pinter.

Dosi, G., Teece, D.J. and Winter, S.G. (1991), ‘Toward of theory of corporate coher-ence’, in Dosi, G., Giannetti, R. and Toninelli, P.A. (eds), Technology and theEnterprise in a Historical Perspective, Oxford, Oxford University Press.

Egidi, M. (1994), ‘Routines, hierarchies of problems, procedural behaviour: someevidence from experiments’, IIASA Working Paper No. 94–58, Laxenburg(Austria).

Egidi, M. and Marengo, L. (1995), ‘Division of labor and social coordinationmodes: a simple simulation model’, in Gilbert, G.N. and Conte, R. (eds),Artificial Societies: The Computer Simulation of Social Life, London, UCL Press.

Egidi, M. and Ricotilli, M. (1997), ‘Co-ordination and specialisation’, WorkingPaper 1997–03, University of Trento.

Eliasson, G. (1990), ‘The firm as a competent team’, Journal of Economic Behaviourand Organization, 13, 275–98.

Eliasson, G. (1994), ‘The theory of the firm and the theory of economic growth’, inMagnusson, L. (ed.), Evolutionary and Meo-Schumpeterian Approaches toEconomics, Dordrecht, Kluwer.

Farma, E. (1980), ‘Agency problems and the theory of the firm’, Journal of PoliticalEconomy, 88, 288–307.


Favereau, O. (1989), ‘Organisations et marchés’, Revue Française d’Économie, 4,273–328.

Favereau, O. (1993), ‘Suggestions pour reconstruire la théorie du salaire sur unethéorie des règles’, WP Laetix, University of Paris X, April.

Favereau, O. (1995), ‘Apprentissage collectif et coordination par les règles: applica-tion à la théorie des salaires’, in Lazaric, N. and Monnier, J.M. (eds), Coordinationéconomique et apprentissage des firmes, Paris, Economica, pp. 23–38.

Fontana, W. and Buss, L.W. (1994), ‘What would be conserved “if the tape wereplayed twice”?’, Proceedings of the National Academy of Science USA, 91,757–61.

Foss, N.J. (1993), ‘Theories of the firm: contractual and competence perspectives’,Journal of Evolutionary Economics, 3, 127–44.

Foss, N.J. (1996), ‘The “alternative” theories of Knight and Coase, and the moderntheory of the firm’, Journal of the History of Economic Thought, 18 (Spring), 1–20.

Fransman, M. (1994), ‘Information, knowledge, vision and theories of the firm’,Industrial and Corporate Change, 3 (3), 1–45.

Geanakoplos, J. and Milgrom, P. (1991), ‘A theory of hierarchies based on limitedmanagerial attention’, Journal of the Japanese and International Economies, 5,205–25.

Gibbons, M., Limoges, C., Novotny, H., Schwartzman, S., Scott, P. and Trow, M.(1994), The New Production of Knowledge, Beverly Hills, CA, Sage.

Hammond, T.H. (1990), ‘The agenda of the firm: structure and incentive in insti-tutional design’, conference on Fundamental Issues in Strategy: A ResearchAgenda for the 1990s, Napa, CA, December.

Hart, O. (1988), ‘Incomplete contracts and the theory of the firm’, Journal of Law,Economics and Organization, 4, 119–39.

Hayek, F.A. von (1937), ‘Economics and knowledge’, Economica, 4, 33–54.Hutchins, E. (1996), ‘Organizing work by adaptation’, in Cohen, M. and L. Sproull

(eds), Organizational Learning, London, Sage.Itoh, H. (1987), ‘Information processing capacities of the firm’, Journal of the

Japanese and International Economies, 1, 299–326.Jensen, M.C. and Meckling, W.H. (1976), ‘Theory of the firm: managerial behav-

iour, agency costs and ownership structure’, Journal of Financial Economics, 3,305–60.

Kauffman, S.A. (1993), The Origins of Order. Self-Organization and Selection inEvolution, New York and Oxford, Oxford University Press.

Klepper, S. (1997), ‘Industry life cycles’, Industrial and Corporate Change, 6, 145–81.Knight, F.H. (1921), Risk, Uncertainty and Profit, Boston, MA, Houghton Mifflin.Kreps, D. (1992), ‘Corporate culture and economic theory’, in Alt, J. and Shepsle,

K. (eds), Positive Perspectives on Political Economy, Cambridge, CambridgeUniversity Press.

Langlois, R. and Foss, N. (1996), ‘Capabilities and governance: the rebirth of pro-duction in the theory of economic organisation’, working paper, University ofConnecticut/Copenhagen Business School.

Levinthal, D. (1997), ‘Adaptation on rugged landscapes’, Management Science, 43,934–50.

Levinthal, D.A. and March, J.G. (1981), ‘A model of adaptive organizationalsearch’, Journal of Economic Behaviour and Organization, 2, 307–33.

Levitt, B. and March, J.G. (1988), ‘Organizational learning’, Annual Review ofSociology, 14, 319–40.


Loasby, B.J. (1976), Choice, Complexity and Ignorance, Cambridge, CambridgeUniversity Press.

Loasby, B.J. (1983), ‘Knowledge, learning and the enterprise’, in Wiseman, J. (ed.),Beyond Positive Economics? London, Macmillan, pp. 104–21.

Loasby, B.J. (1991), Equilibrium and Evolution: An Exploration of ConnectingPrinciples in Economics, Manchester, Manchester University Press.

Lounamaa, P.H. and March J.G. (1987), ‘Adaptive coordination of a learningteam’, Management Science, 33, 107–23.

Lundvall, B.A (2000), ‘Unlocking knowledge in society’, in OECD reportKnowledge in the Learning Economy, OECD.

Machlup, F. (1983), Knowledge, its Creation, Distribution, and EconomicSignificance, Princeton, NJ, Princeton University Press.

Magalhães, R. (1998), ‘Organizational knowledge and learning’, in von Krogh, G.,Roos, J. and Kleine, D. (eds), Knowing in Firms, London, Sage.

March, J.G. (1991), ‘Exploration and exploitation in organizational learning’,Organization Science, 10 (1), 71–87.

March, J.G. and Simon, H. (1993), ‘Organizations revisited’, Industrial andCorporate Change, 2 (3), 299–316.

Marengo, L. (1994), ‘Knowledge distribution and coordination in organizations:on some social aspects of the exploration vs. exploitation trade-off’, RevueInternationale de Systémique, 7, 553–71.

Marengo, L. (1996), ‘Structure, competence and learning in an adaptive model ofthe firm’, in Dosi, G. and Malerba, F. (eds), Organisation and Strategy in theEvolution of the Enterprise, London, Macmillan, pp. 124–54.

Marengo, L. (1998), ‘Interdependencies and division of labour in problem-solvingtechnologies’, presented at the conference, Competences, Governance andEntrepreneurship, DRUID, Bornholm, June.

Marschak, J. and Radner, R. (1972), Economic Theory of Teams, New Haven, CT,Yale University Press.

McGuire, C.B. and Radner, R. (eds) (1986), Decision and Organization, 2nd edition,Minneapolis, MN, University of Minnesota Press.

Metcalfe, S. (1998), ‘Innovation as a policy problem: new perspectives and old onthe division of labour in the innovation process’, mimeo, Centre for Research onInnovation, University of Manchester.

Milgrom, P. and Roberts, J. (1988), ‘Economic theories of the firm: past, present,future’, Canadian Journal of Economics, 21, 444–58.

Nelson, R. (1994), ‘The role of firm difference in an evolutionary theory of tech-nical advance’, in Magnusson, L. (ed.), Evolutionary and Neo-SchumpeterianApproaches to Economics, Dordrecht, Kluwer.

Nelson, R.R. and Winter, S.G. (1982), An Evolutionary Theory of Economic Change,Cambridge, MA, Harvard University Press.

Nohria, N. and Ghoshal, S. (1997), The Differential Network: OrganizingMultinational Corporations for Value Creation, San Francisco, CA, Jossey-Bass.

Nonaka, I. (1994), ‘A dynamic theory of organizational knowledge creation’,Organization Science, 5 (1), 14–37.

Nonaka, I. and Konno, N. (1998), ‘The concept of Ba: building for knowledge cre-ation’, California Management Review, 40 (3), 40–54.

Nonaka, I. and Takeuchi, H. (1995), The Knowledge Creating Company, New York,Oxford University Press.

Nooteboom, B. (1999a), Inter-Firm Alliances, London, Routledge.


Nooteboom, B. (1999b), ‘The triangle: roles of the go-between’, mimeo, Faculty ofManagement and Organization, Groningen University.

OECD (2000), Knowledge Management in the Learning Economy: Education andSkills, Paris: Organisation for Economic Cooperation and Development.

Pavitt, K. (1990a), ‘What we know about the strategic management of technology?’,California Management Review, 32, 17–26.

Pavitt, K. (1990b), ‘Some foundations for a theory of the large innovating firm’,mimeo, Brighton, SPRU, University of Sussex.

Pavitt, K. (1997), ‘Technologies, products and organization in the innovative firm:what Adam Smith tells us and Joseph Schumpeter doesn’t’, mimeo, SPRU,September, presented at the conference, Competences, Governance andEntrepreneurship, DRUID, Bornhom, 9–11 June 1998.

Penrose, E. (1959), The Theory of the Growth of the Firm, Oxford, OxfordUniversity Press.

Postrel, S. and Rumelt, R.P. (1992), ‘Incentives, routines and self-command’,Industrial and Corporate Change, 1 (3), 397–425.

Prahalad, C.K. and Hamel, G. (1990), ‘The core competence of the corporation’,Harvard Business Review, 68 (May–June), 79–91.

Probst, G., Büchel, B. and Raub, S. (1998), ‘Knowledge as a strategic resource’, invon Krogh, G., Roos, J. and Kleine, D. (eds), Knowing in Firms, London, Sage.

Radner, R. (1986), ‘The internal economy of large firms’, Economic Journal, 96(supplement), 1–22.

Reynaud, B. (1996), ‘Types of rules, interpretation and collective dynamics: reflec-tions on the introduction of a salary rule in a maintenance workshop’, Industrialand Corporate Change, 5 (3), 699–723.

Richardson, G.B. (1960), Information and Investment, Oxford: Oxford UniversityPress.

Richardson, G.B. (1972), ‘The organization of industry’, Economic Journal, Sept.,883–96.

Schelling, T. (1978), Micromotives and Macrobehavior, New York: W.W. Norton.Simon, H.A. (1957), Models of Man, New York, Wiley.Simon, H.A. (1976), Administrative Behaviour, 3rd edition, New York, Free Press.Simon, H.A. (1979), ‘Rational decision making in business organizations’,

American Economic Review, 69, 493–513.Teece, D.J. (1988), ‘Technological change and the nature of the firm’, in Dosi, G.,

Freeman, C., Nelson, R., Silverberg, G. and Soete, L. (eds), Technical Change andEconomic Theory, London, Pinter, pp. 256–81.

Teece, D.J. and Pisano, G. (1994), ‘The dynamic capabilities of firms: an introduc-tion’, Industrial and Corporate Change, 3, 537–56.

Teece, D.J., Rumelt, R.P., Dosi, G. and Winter, S.G. (1994), ‘Understanding corpor-ate coherence: theory and evidence’, Journal of Economic Behaviour andOrganization, 23, 1–30.

Tell, F. (1997), ‘Knowledge and justification – exploring the knowledge based firm’,thesis, Department of Management and Economics, University of Linköping.

Vicari, S. and Toniolo, G. (1998), ‘Errors and learning in organizations’, in vonKrogh, G., Roos, J. and Kleine, D. (eds), Knowing in Firms, London, Sage.

Von Krogh, G., Ichijo, K. and Nonaka, I. (2000), Enabling Knowledge Creation,Oxford, Oxford University Press.

Von Krogh, G. and Roos, J. (1995), Organizational Epistemology, Basingstoke,Macmillan.


Wenger, E. (1998), Communities of Practice: Learning, Meaning, and Identity,Cambridge, Cambridge University Press.

Williamson, O.E. (1967), ‘Hierarchical control and optimum firm size’, Journal ofPolitical Economy, 75, 123–38.

Williamson, O.E. (1975), Markets and Hierarchies: Analysis and AntitrustImplications, New York, Free Press.

Williamson, O.E. (1985), The Economic Institutions of Capitalism: Firms, Markets,Relational Contracting, London, Macmillan.

Williamson, O.E. (1993), ‘Transaction costs economics and organisation theory’,Industrial and Corporate Change, 2 (2), 107–56.

Winter, S.G. (1982), ‘An essay on the theory of production’, in Hymans, H. (ed.),Economics and the World around It, Ann Arbor, MI, University of MichiganPress, pp. 55–93.

Zuscovitch, E. (1998), ‘Networks specialization and trust’, in Cohendet, P., Llerena,P., Stahn, H. and Umbhauer, G. (eds), The Economics of Networks, Berlin,Springer Verlag, pp. 243–64.


12. Markets for technology: ‘panda’sthumbs’, ‘calypso policies’ andother institutional considerationsAshish Arora, Andrea Fosfuri andAlfonso Gambardella

1 INTRODUCTION

We have argued elsewhere that markets for technology have becomeimportant (see Arora and Gambardella, 1994; 1998; Arora et al., 2001b;2001c; Arora and Fosfuri, 2003). Our research in this field is related to ourintellectual legacy from Paul David. It was Paul who originally suggestedto us that the ‘technology of technical change’ (an expression that Ashishand Alfonso heard from him when they were graduate students, and thatAndrea heard from them) was an important phenomenon to understand.Moreover, Paul’s work on the nature of knowledge directed some of ourwork, especially when he emphasised that differences in the nature ofknowledge can give rise to different opportunities of exchanging knowl-edge or technologies among parties that do not belong to the same organ-isation (for example, David, 1993a).

But while one may think that our intellectual debt to Paul David is largelyassociated with the study of the nature and the role of technology, hisemphasis on institutions and norms, and their role in affecting economicgrowth and technology, probably influenced our way of thinking even moredeeply. In this chapter, we therefore touch upon several topics about normsand institutions at large that Paul addressed in several ways in his work.Specifically, we discuss how standards, intellectual property rights, institu-tions and the related social, cultural or political norms shape the rise anddevelopment of markets for technology.

We draw inspiration from Paul’s work on the scientific community as aninstitution (for example, Dasgupta and David, 1987, 1994; David, 1993a;1998; 2003) and his concerns about the threats to the norms of open science.As we shall discuss in this chapter, while markets for technology may onthe one hand encourage greater diffusion of knowledge and technology,

323

the implied ‘privatisation’ of knowledge has serious side effects for openscience.

To set this discussion within the specific context of the markets for tech-nology, a natural starting point is to note that markets do not arise or func-tion in a vacuum. They need a supporting infrastructure. This is evident inthe current context of on-line markets and the fortunes of firms thatprovide the technical infrastructure for such markets. But no less importantis the institutional and policy infrastructure. This includes not just theformal laws and policies that govern such markets, but also norms and‘rules of the game’ that determine the transaction costs of participating inthem.

But this also suggests that rather than an ‘exhortative’ approach topolicy analysis, whose objective is a list of recommendations for policychanges, one can more usefully rely on a more ‘institutionalist’ approach,which as we shall see, appears to be particularly suited for the contextthat we set forth in this chapter. As Douglass North (for example, 1990),among others, put it, such an approach focuses on understanding howadequate institutions for supporting economic growth are created, howthey function, and the role of governments in creating, supporting andshaping these institutions. Several studies in the new institutionalistapproach, especially those developed in a historical context, deal specifi-cally with the genesis and development of new markets, and their effectson economic growth (for example, Rosenberg and Birdzell, 1986; Alstonet al., 1996).

In this chapter, we argue that the support that policy can provide to thegrowth of markets for technology is relatively more important when theyhave to be created, compared with supporting their functioning when theyare in place.1 As those creating the new electronic virtual market places wellunderstand, market creation may require explicit transfers between marketparticipants that enter at different times. Put differently, creating marketsoften implies large externalities, and public policy can play an importantrole in subsidising those creating the positive externalities. For instance, thedifficulties in establishing standards hinder the development of newmarkets. These standards need not be merely technical standards, but theycan also pertain to contracts or the forms in which commercial informationis gathered and recorded, and these are typically the domain of legal andpolicy infrastructures.

Further, these standards need not be set by governments or other publicbodies. As Rosenberg and Birdzell (1986) show, critical institutions for theexpansion of markets and commerce in the fifteenth and sixteenth centuriesin Europe, such as a bill of exchange, insurance or double entry bookkeep-ing, were produced independently by the economic agents. Even so, public


policy can play an important role in encouraging compliance. For instance,as Bordo et al. (1999) describe, the flow of foreign capital to the USA andthe development of an equity market were greatly facilitated by the adop-tion, by US firms, of English accounting practices and the regular report-ing of financial results. State legislatures, particularly those in influentialstates such as New York, played an important role in the widespread adop-tion and diffusion of these accounting standards. In this chapter, we discussthe role of standards in section 2.2.

Specialised technology suppliers can play an important role in thecontext of markets for technology. Some societies tend to providegreater opportunities for new firm formation and for the extent to whichfirms, particularly new firms, can enter new market niches. Specialisedtechnology suppliers are a special case, albeit a very important specialcase, of the latter category. Needless to say, as we discuss in section 2.3,public policy can facilitate or discourage new technology-based startupsand risk-taking activities more generally. To say this is not necessarily toadvocate such policies. Whether such policies are appropriate and neces-sary is itself something that needs to be examined more closely on itsown merits.

The most obvious relevant institution for technology markets is intellec-tual property rights. As Paul David noted in several papers (for example,David, 1993a; 1993b), intellectual property rights are a social institution.Although much of the discussion has focused on their role in providingincentives for innovation, a market for technology perspective focusesattention on their role in facilitating transactions in technology. Simply put,intellectual property rights are almost a precondition for a market for tech-nology to exist. That said, there are some subtleties about how these rightsare defined and interpreted that have important implications for the func-tioning of these markets.

For example, as noted by Heller and Eisenberg (1998), and as we discussin section 3, fragmentation of intellectual property rights may hinder theefficient working of markets for technology. Excessive fragmentation(or excessive overlap) of intellectual property rights may prevent the devel-opment of innovations when complementary property rights are owned byindependent agents and no one is capable of collecting or co-ordinating allthe rights to develop the innovation. A related consequence of the increas-ing strength of intellectual property rights that we observe, for instance, inthe USA today is an increase in litigation costs. Thus, mechanisms likestronger intellectual property rights, which, as we argued elsewhere (forexample, Arora, 1995; Arora et al., 2001c), may help the formation ofmarkets for technology by reducing some of their transaction costs, maycreate other transaction costs.2

Markets for technology 325

An institutional perspective also points to an important interactionbetween the growth of markets for technology and other institutions insociety, most notably universities. Academic research has contributed sig-nificantly to the growth of new scientific and technological knowledge.Further, universities have played a crucial role in codifying and standardis-ing the language and terminology used to communicate scientific findingsand, more generally, in creating and sustaining scientific communities.A natural result has been the development of conceptual categories that aregeneral and universal in scope, an important requirement for an effectivemarket for technology.

These activities of the research university system form an important‘public good’ upon which all participants in a market for technology drawand for which they make no direct recompense. These public good creatingactivities of universities are greatly enhanced by norms of disclosure andcollegiality that appear to have arisen in response to specific historicalfactors unrelated to their present function but which, as Dasgupta andDavid (1994) have noted, appear to be critical to the role research univer-sities play in modern economies. But, as we discuss in section 4, the greater‘privatisation’ of knowledge and the ability to directly value knowledge thatmarkets for technology make possible often clash with these academicnorms. In times of weakening public support, this may seriously attenuatethese norms. In turn, this presents a serious challenge for public policybroadly construed.

Finally, in a globalising world, markets for technology are also likely tobe global. But the globalisation of the markets for technology may limit thescope and effect of national policies. It also means that a country maybenefit from the rise of markets for technology elsewhere, even though littleeffort was made domestically to create them.

The direct implications of this argument for policy are twofold. First, inthe European context, policies for encouraging markets for technology arebest considered at the level of the European Union rather than the individ-ual nation-state. Second, particularly in developing countries, science andtechnology policies should be mindful of whether markets for technologyexist and, if they do, their efficiency and, working. In turn, this argues forsector- or industry-specific policies, rather than a ‘one size fits all’ policy.Where markets for technology do exist, the policy questions may wellnarrow down to how best to take advantage of the ongoing growth in theworldwide technology trade. We discuss these issues in section 5 of thischapter. Section 6 concludes the chapter.


2 TRANSACTION COST-REDUCING MECHANISMSIN THE MARKETS FOR TECHNOLOGY

2.1 Institutions and the Rise of New Markets

Whenever new markets arise, and as long as they develop and grow, therewill be some private incentive to create the institutions that support thesemarkets. Thus, as Rosenberg and Birdzell (1986) note, the expansion ofcommerce, particularly international commerce, in the fifteenth and six-teenth centuries was accompanied by several institutional innovations.Some of these were made by the economic agents themselves (for example,insurance and bills of exchange), while others required supra-individualpolicy interventions (for example, taxation). In addition, there is a ques-tion of timing. Institutions are less ‘plastic’ than technology or industrystructure. As a result, new technologies diffuse rapidly compared to thedevelopment of complementary organisational innovations (for example,David, 1990). Since the formation of complementary institutions takestime, these institutions are more likely to follow rather than anticipate thenew markets.

With these broad remarks in mind, the institutional settings that arerequired for the markets for technology to arise are in many respects similarto those that are required for any new market to arise. There are strikingsimilarities between the institutional innovations that gave rise to thegrowth of new commercial markets in sixteenth-century commerce andthose needed for the development of technology markets today. Just as thegrowth of sixteenth-century commerce required well-defined propertyrights, so also markets for technology require better defined intellectualproperty rights (further discussed in the next section).

Similarly, markets require standards. In the special case of the marketsfor technology, technological standards reduce the market risks of newtechnological developments by ensuring the compatibility of specific tech-nological components with existing technological architectures and com-plementary technologies. In other words, provided that their devices areconsistent with a commonly known architecture, they only face the techno-logical risk that their device may not function, that it might not be cost-effective or that it might be inferior to competitors.

Finally, new markets, and their expansion, are typically associated withgreater economic experimentation and risk. Thus, we also enquire into theinstitutions that support economic or technological experimentation.


2.2 Standards

The absence of standards can significantly increase transaction costs. Theimportance of standards for technology markets or the exchange of know-ledge should not be confined to ‘hard’ technological standards. ‘Soft’ stand-ards can also be crucial for enabling independent parties, often located faraway from one another, to exchange knowledge or information. A strikingexample is the scientific community. The scientific community has devel-oped for many centuries a common language, and a common set of normsand customs. We discuss the importance of such norms and customs for thediffusion of knowledge in section 5. Here we want to highlight a relatedeffect. As David (1991; 1993a) and Dasgupta and David (1987; 1994) pointout, the current norms and customs of the scientific community requirethat scientific discoveries should be reproducible by peers, and the resultsbe cast in a ‘language’ that is commonly understood by the scientificcommunity.

The common language of science, in its various fields and disciplines, hasacted as a natural standard for research. The growing use of mathematicaland computer modelling has furthered standardisation. In many ways, thescientific system is a prototypical case of widespread effective exchange ofknowledge among individuals or groups who have never had any directcontact with one another, but who exchange ‘codified’ knowledge throughjournal articles and build upon the work of others.3 Accordingly, onewould expect that a greater diffusion of standardised R&D techniques,such as standardised software and simulation tools (for example, for testingproducts or for computerised product design) can promote methodologicalstandards, and therefore create opportunities for exchange among inde-pendent parties (Arora and Gambardella, 1994).

Of course, ‘hard’ technological standards are also critical for the growthof markets for technology. The argument is not new (for example, see Davidand Greenstein, 1990, for a survey), and Langlois and Robertson (1992),among others, have shown how the development of an ‘open architecture’,based on well-defined architectural interfaces and standards, was criticalfor giving rise to widespread innovation activities and experimentation withpersonal computer (PC) components by many independent suppliers.Established architectural interfaces meant that independent innovators didnot face the risk of creating incompatible innovations, but only the risk offailing to develop and commercialise the innovation.

Technological standards have become common in many markets fortechnology. One example is the development of a component-managementsoftware technology, called CORBA, produced by a US-based non-profitconsortium – Component Management Group (CMG) – with 20 000


members (individuals and companies) in the USA, Canada, Italy, Japanand India. CORBA provides standard interfaces for designing softwarecomponents that can be plugged and played into the systems without inter-fering with the basic structure of the architecture, along with other stand-ard setting operations and procedures. Moreover, the rise of suchstandard-setting institutions is complementary with the creation of openelectronics markets for software components. Web-based markets existtoday to help match buyers and sellers, thereby reducing search and relatedtransaction costs. Standards for components and their reuse further reducethe transaction costs, and therefore enhance the value of such markets.

Linden and Somaya (2003) have shown that institutional developmentsin the semiconductor industry have helped create standards that facilitatelicensing and cross-licensing of design modules (see also Hall and Ziedonis,2001). Linden and Somaya note that a critical event in the 1980s was theestablishment of the silicon-based CMOS technology as the dominantdesign in semiconductor process technology. In the 1990s, when fablessdesign firms arose, they could take advantage of this standard processingtechnology by focusing on designing integrated circuits for users and relyupon merchant foundries for manufacturing.

The existence of a standard process like CMOS reduced transactioncosts and was critical for an independent fabless design industry to arise.Put differently, the existence of a standard manufacturing process impliedthat it was easier to de-link product design from process requirements and,therefore, to separate the skills, knowledge and activities that were neededto design the chips and those that were required to fabricate them. The sep-aration of the knowledge and related domains for designing the productcould be achieved to a greater extent by ensuring the compatibility of theinterfaces between the design and the process.

During the 1990s, the semiconductor industry gave rise to two majorstandard setting alliances. The first, called the Virtual Socket InterfaceAlliance (VSIA), was established in 1996 by 35 founding members, whichincluded Electronic Design Automation (EDA) software firms, fablesssemiconductor design companies and electronics companies. The goal ofVSIA was to define and establish open compatibility standards (‘virtualsockets’) in semiconductor design. Virtual Socket Interface Alliance bothreleases the specifications, and it actively tries to encourage their use by theparticipating firms. Reusable Application-Specific Intellectual PropertyDevelopers (RAPID) aims at improving access to information about designmodules. Thus, for instance, RAPID developed a standard catalogue forfeaturing commercially available design modules on the Internet. Similarly,the Virtual Component Exchange (VCX), was created in 1998 by theScottish economic development agency, ‘Scottish Enterprise’, and a few


major players from VSIA. VCX is addressing business and legal issuesrelated to trade in design modules by developing standard contracts, mon-itoring systems, a matchmaking service and customised arbitration services(see Linden and Somaya, 2003).

There are several noteworthy aspects of the foregoing. First, in thecase of semiconductors, it is interesting that coalitions for establishingstandards have not been confined to the setting of purely technologicalstandards. Virtual Component Exchange, in particular, is addressing stand-ardisation in areas such as contracts for design modules. Thus, standardsrelated to the efficient functioning of various aspects of the market for tech-nology, including norms about legal settings, have been central to the for-mation of such institutions.

Second, some of these coalitions are private initiatives, with no evidentgovernment intervention. This suggests that direct policy intervention isnot strictly necessary for addressing or solving co-ordination problemsinvolved in standard setting. However, public agencies like the ‘ScottishEnterprise’ (for VCX) can help catalyse such initiatives. Thus, indirectpolicy interventions, possibly through such agencies, may be important toencourage the industry to coalesce in order to promote standards. At thesame time, as noted earlier, VCX deals with the creation of standard con-tracts. This suggests that such decentralised institutions may even embraceareas that once might have been thought the preserve of government alone.

Clearly, government intervention may be critical when private initiativesare in conflict with each other, or absent. But rather than directly creatingsuch standards, it is probably more effective if governments encouraged‘private’ coalitions for standard setting, which would be closer to the actualneeds and information of the specific industry producers and stakeholders,and therefore more likely to develop valid and competent solutions.Governments may still have important roles to play, and in this respectthere may be different remarks to be made for governments with differentpolicy traditions, like the USA and Europe. For example, in the USA,where antitrust is more stringent, antitrust considerations could be tradedoff against the need to promote effective standards and prevent fragmenta-tion of intellectual property rights, as for instance in the case of patentpools described later in this chapter. This also requires that such alliancesare monitored closely to verify, for example, whether industry pricesincrease after the alliance.

By contrast, in Europe, where inter-firm alliances have attracted lessantitrust scrutiny, greater attention should be paid to these coalitions tomake sure that they actually act as institutions for the creation of standardsrather than as means for price-fixing. At the same time, the substantial bar-riers to information exchange, mobility and interaction among companies


and other institutions located in different European countries suggest thatthe European Commission should take a more active role in favouring theformation of such coalitions.

2.3 Financial and Other Institutions

Specialised technology suppliers play a key role in the market for tech-nology. Although firms operating in standard product markets can, and do,supply technology, there are some obvious ways in which specialised tech-nology suppliers are important. For one, unlike technology producers thatare also suppliers in the final markets, technology suppliers that specialiseonly in the production of the technology do not face the inherent conflictof having to compete with their customers. On the one hand, this encour-ages them to supply their ‘best’ technology, because of the natural marketincentives to provide the highest possible value to the market, as opposedto second-rate technologies to avoid that competitors catch up with them.On the other hand, the buyers themselves feel more comfortable whenbuying technologies from firms that have no vested interests in the finalmarkets. Both factors reduce problems of asymmetric information andother transaction costs or sources of opportunism. A related reason is thatthe independent technology suppliers are not locked into an installed baseof products and need not fear cannibalising their existing market by devel-oping new technology. Finally, their corporate culture is likely to be moreflexible and open to communication, and their management less likely to bedistracted by the needs of manufacturing and distribution.

For such specialised firms, many of which are startups and small, thereare many other types of barriers and transaction costs to contend with.Obtaining finance is a particularly important one. Financial institutionscan play a critical role in fostering or hindering the markets for technology.A general discussion of the role of financial institutions and policy isbeyond the scope of the present discussion. Instead, we will focus uponpolicy initiatives for promoting risk-taking, to encourage specialised tech-nology suppliers and startups.

Venture capital, initial public offerings and ‘new’ financial and equitymarkets have grown in parallel with the rise of new business opportunitiesand innovation in high-tech industries. Institutional innovations likeventure capital have proved to be extremely flexible. They have adapted invarious ways to the actual needs and conditions for supporting new busi-ness activities. For example, many analysts have noted that not only doventure capitalists provide finance and managerial support, but, mostimportant of all, they provide their startups with connections to theirbroader networks of people and resources (see, for example, Gomperts,


1999). This networking capability has been critical for maximising theexploitation of the external economies that exist in areas like California orelsewhere (see also Bresnahan and Gambardella, 2004). To a large extent,these institutions are themselves the result of private responses to profit-making opportunities. In successful cases, policy has mainly played the roleof creating the general ‘ambience’, rather than direct intervention.

Even with venture capital, technology-based startups face a high degreeof risk. Thus, in many countries institutions for supporting technologicalrisk have taken the form of direct policy interventions as well. For example,many European governments, often with the financial and political supportof the European Union, have invested sizeable public resources in the cre-ation of science and technology parks, especially in less developed regions.The stated objective is to provide the physical and business infrastructureneeded to nurture startup firms, particularly, R&D-intensive firms.

These initiatives have had mixed success. Rather than enter into a dis-cussion of the pros and cons of such initiatives, a more interestingapproach is to note that other, more indirect ways, of reducing risks fortechnology-based new business exist. Compared with direct measures, indi-rect measures have the advantage that the benefits from the measure accrueprimarily to those who have attained some independent achievement,rather than indiscriminately to everybody.

A report by the European Commission (ETAN, 1999) suggests threeareas for institutional developments that would encourage risk-takingbehaviour – changes in fiscal policy, the creation of security interests inintellectual property rights and changes in insolvency laws. As far as thefiscal incentives are concerned, R&D and innovation tax credit can beuseful institutional innovations to favour startup companies and themarkets for technology. These objectives, however, have to be taken expli-citly into account in designing the schemes. For example, the US Researchand Experimentation Tax Credit issued in the early 1980s was amended in1993 to properly extend the underlying incentives to smaller firms. Smallfirms, new startups or, more generally, firms that could not claim a basis ofR&D expenditures in the previous three years on which to computethe incremental R&D tax credit, were assigned a fixed percentage increaseof 3 per cent for the first five taxable years beginning after 1993.

Another problem with startups is that they may not be able to enjoy thebenefits of the R&D credit because of having no taxable income in the year.In order to make an unused R&D credit a valuable asset, the 1993Amendment established that firms could carry back the credit three yearsand carry it forward up to 15 years. In so doing, the credit becomes a hiddenasset that can be unlocked in the future when the company becomes pro-fitable or is sold. Venture capitalists and lenders understand the importance


of these hidden assets and may grant more favourable terms if they knowa credit exists and can be deployed in the near future.

There is a more general perspective to this point. A well-known problemwith technological activities is how to account for the value of these act-ivities – and of the technology-based companies more generally – given thatsuch a value is often made up of intangible rather than tangible assets. Thevaluation of intangible assets, and specifically the valuation of technology,is particularly relevant in cases where the firm lacks downstream assets tocommercialise the technology. This is a complex issue and well beyond thescope of the present discussion. The point is that accounting practices andnorms, decided through a complicated interaction between private andpublic bodies, can affect the fortunes of technology-based firms, particu-larly startups, in important ways. Current accounting practices and norms,derived as they are from times when measuring tangible and material assetswas their crucial task, will have to be modified in order for technologymarkets to flourish.

This point is not new, and other writers have emphasised this as well (Lev,2000; Hand and Lev, 2003). There has also been some interesting researchusing firm-level R&D and patent and patent citation data to developimplicit measures of the technological assets of firms (Deng et al., 2003).What is less well understood is the role that technology markets themselvescan play in improving the accounting for intangible technological assets. Amarket for technology improves the accuracy of any valuation attempt.It does so in the most obvious way, by providing an objective measure ofthe value, if the asset has been traded in the past or if similar assets havebeen traded. Needless to say, technology is highly differentiated, and its‘price’ is likely to reflect factors idiosyncratic to the buyer and the seller.Thus, any monetary measure is likely to be imperfect. That said, such prob-lems are not unique to the measurement of the value of technology.Further, when investing in R&D, firms are implicitly making such measure-ments, as do investors when they value the firms in capital markets. Marketsfor technology allow for the possibility of valuing the contribution oftechnology separately from the value of other valuable assets the firmmay possess. In turn, such valuation may enable firms to specialise in thedeveloping technology without necessarily having to acquire downstreamcapabilities.

This problem is also closely related to attempts in the USA and Europeto remove legal obstacles to the creation of security interests in intellectualproperty rights, as discussed in the ETAN report (1999: 46). The point isthat once lenders, investors or the entrepreneurs themselves can meaning-fully assess the value of these assets, the assets can be used in a variety ofways, including as collateral to obtain financing. When such assets can be


‘securitised’ and traded in a market, this is likely to further encourage thegrowth of firms specialising in developing technology.4

The third area identified by the ETAN report (1999) is that of the insol-vency laws. In fact, the introduction of limited liability in the sixteenthcentury is a major institution for limiting the risk of the entrepreneur whosets up a new business. However, as the report notes, the limited liability canbe in many practical cases more apparent than real. For example, on manyoccasions, the life savings and dwelling house of the entrepreneurs have beenused as collateral for company debts. The issue is at least as serious in thecase of technological business where companies may take even higher risks.This could limit, for instance, their ability to invest in expensive equipmentto carry out research experiments, computerised product designs and thelike. The ETAN report argues that one ought to encourage institutionalchanges that would further reduce liability for failure in technologicallyrisky industries. The report also argues that in the USA the FederalBankruptcy Code Chapter 11 is more favourable to the setting up of newcompanies for marketing technological innovations than existing laws inmost European Union member countries. This is clearly a quite complicatedarea for intervention, as one has to properly balance the need for encourag-ing risky business against the need to discourage excessive or imprudentrisks. But this only increases the importance of carefully crafted policies.

As with any other market, intermediating institutions for reducinginformation search costs will improve the working of technology markets.However, the reduction of information search costs in the market fortechnology may often be carried out by agents who would either specialisein acquiring and diffusing the information, or who would do so jointlywith other supporting activities for these markets. For example, patentattorneys and patent agents did play such a role in the development of themarket for technology in the USA during the nineteenth century and earlytwentieth century (see Lamoreaux and Sokoloff, 1996; 1999). In additionto providing patent counselling and related services, they also helpedmatch demand and supply. A similar role is played today by several inde-pendent firms and technology traders on the Internet (see for instanceBTG, 1998). Public institutions can step in where private initiative islacking. Thus, for instance, the European Union has created its ownInternet information providing service, CORDIS, which collects informa-tion about potential technologies for licensing, as well as requests for tech-nological partnerships and the like. Similarly, the German government,along with the Länder and private investors, has created the SteinbeisFoundation – a large network of German firms, research institutions andacademic professors, whose task is to co-ordinate and match technologydemands with the supply of technologies and related competencies.


3 INTELLECTUAL PROPERTY RIGHTS AND THEMARKET FOR TECHNOLOGY IN ANON-COASIAN WORLD

3.1 Fragmentation of Intellectual Property Rights and Related Issues

It is difficult to think that a market, or at least what economists mean byemploying this word, could ever function properly without property rightson the object of the transaction. Markets for technology are no exception.Indeed, intellectual property rights are to the markets for technology whatproperty rights are to the markets for cars or for PCs. They recogniseenforceable claims to some benefits or use to the technology. Intellectualproperty rights consist of patent rights, copyrights, design rights, trade-mark rights, trade secret rights and a few other special property rights incontemporary law. Our focus here is on patents.

Patents are the most important type of property right for technology,although as we have repeatedly pointed out, technology is much more thanwhat is covered by patents. However, selling a technology without anenforceable patent would be difficult, although not impossible. Themoment one has to disclose a piece of information in order to sell it, one isrunning the risk of being cheated. Potential buyers and other third partiescan therefore appropriate these ideas and knowledge without having to payfor them. Anticipating this, potential sellers may be reluctant to show theobject of the transaction and let it be evaluated by the prospective buyers.The latter will not pay money for something whose value they cannotappraise up front. The net result is that such transactions may not takeplace at all.5 Therefore, patents, or, more generally, well-defined enforceableintellectual property rights, are not only critical to protect the incentives forinnovation but are also the supporting institutions for the existence andfunctioning of markets for technology.

Arora and Merges (2004) use the incomplete contracting approach(Grossman and Hart, 1986; Hart and Moore, 1990) to argue that well-defined enforceable patents reduce transaction costs, and thereby helpincrease transactions in technology. Efficient contracting for technologyenhances the opportunity to profit from innovation through licensing.Patents can be used to structure technology transfer contracts, therebyplaying an important role in determining the efficiency of knowledgeflows. Arora and Merges (2004) also find support for the argument putforward in Arora and Gambardella (1994) that patents are likely to havea greater value for small firms and independent technology suppliers ascompared to large established corporations. Whereas the latter haveseveral means to protect their innovations – for instance, through their


extensive downstream manufacturing and commercialisation assets – theformer can only appropriate the rents to their innovation by leveraging theprotection that patents provide.6

The role of patents in facilitating transactions in technology has largelybeen ignored in formal economic analysis. The focus there has been on thetrade-off between the ex ante incentives to innovate and the ex post advan-tages of innovation diffusion, with the limitation to ex post diffusion beingthe price that society has to pay in order to encourage market-based inno-vation ex ante. The major policy question related to the optimum length(and, later, length and breadth) of the temporary monopoly to be granted(see, for instance, Gilbert and Shapiro, 1990; Klemperer, 1990). Onefeature common to both views of patents – as inducement for innovationand as the basis for markets for technology – is that they treat patentsas covering inventions with well-defined and clearly de-limited scope ofapplication.

In fields such as chemicals, biology, materials and electronics, the growthin our understanding of the underlying physical phenomena makes it pos-sible to represent the invention succinctly and effectively, through the useof abstract generalisation that a scientific approach makes possible.However, the very same growth in scientific understanding, and thegrowing power and use of the abstraction that this understanding makespossible, also makes it possible to relate knowledge created in a specificcontext to a much broader array of applications.

This growth in generality, which has spurred the growth of many science-based startups, especially in biotechnology, has also created several chal-lenges for the patent system and for the role of intellectual property rights.For the first of the challenges, consider the example raised by patenting ofparts of the human genome. This is a controversial and emotive issue butour focus here is somewhat different. Understanding the structure of a geneprovides information about the proteins it codes for. If one also under-stands the role of the protein in the context of some disease, then under-standing the structure of the gene provides an opportunity to try to preventor cure the disease. A patent on the gene would therefore allow the patent-holder to share from the economic rents created by this therapy. However,these rents would have to be shared with the firm that uncovers the role ofthe protein coded for by the gene as well as with the firm that uses thatknowledge to develop a cure, to test the cure in clinical trials and to manu-facture, market and distribute it.

This raises the question of how the different contributors should berewarded. One might expect that the relative bargaining power of theparties involved would determine the rewards. In principle, the situation isnot very different from a landowner bargaining with a real estate developer


who will put up a shopping mall on the land. In simple terms, one wouldexpect higher transaction costs as the various parties to the negotiation tryto get the best deal for themselves. The major difference is that the knowl-edge of the structure of the gene (and the working of the protein, for thatmatter) is a non-rival good in that it may be applied in other contextswithout reducing the economic value derived from its application in thefirst context. In other words, there is a strong ‘public goods’ character.7

Indeed, applying the knowledge about the structure of the gene to cure onedisease in no way reduces the value of applying the same knowledge to cureother diseases. In this sense, knowledge is non-rival. As is evident uponfurther thought, the key here is that the knowledge has multiple potentialapplications, so that users do not compete.

When knowledge is non-rival, protecting that knowledge throughpatents creates potential inefficiencies. For instance, in the case of ex antecontracting, a number of different potential users may have to get togetherto invest in creating the knowledge. Such contracts are problematic becauseusers will differ in the value they place upon the enterprise and, con-sequently, are likely to under-report their value. Similar problems are likelywith ex post contracting, with different users being charged different prices.Moreover, the closer a patent comes to covering knowledge that amountsto a basic understanding of the physical phenomena involved, the broaderthe likely sweep of the patent and the further in time its applications.

In other words, the implicit one-to-one relationship between a patent andan innovation that many economic models assume, although analyticallyconvenient, has obscured the point that in cumulative or systemic tech-nologies, a commercialisable innovation may require many different piecesof knowledge, some of which may be patented and owned by people withconflicting interests.8 In turn, an agent holding a patent on an importantcomponent may cause severe ‘hold-up’ problems, retarding the develop-ment of the technology (see also Scotchmer, 1991, and Green andScotchmer, 1995, for further discussion.) In a similar vein, Merges andNelson (1990; 1994) argue that broad patents increase the likelihood thatan innovator would try to control future innovations based upon its owninnovation, thereby slowing down the pace of technological progress.

However, the essential problem is not caused by patents, but by factors(such as negotiation costs) that prevent agents from entering into contractsfor the use of patents. In a Coasian world with no transaction costs, agentswill bargain for a Pareto superior solution given any initial distribution ofproperty rights over the fragments. More realistically, the required collec-tion of property rights, although socially efficient, might not occur becauseof transaction costs and ‘hold-up’ problems. An agent holding a patent onan important fragment (‘blocking patent’) may use his or her patent as a


‘hold-up’ right in an attempt to extract as much of the value of his or herinnovation as possible.

Thus the issue at stake is the impact that strengthening and expansion ofpatent rights – which is what is happening today, particularly in the USA –would have on transaction costs. An especially problematic case is when theproperty rights are defined around very narrow fragments of knowledgeand owned by separate entities. In this case, each patent-holder has the rightto exclude the others from the use of his or her piece of knowledge. In otherwords, when several pieces of intellectual property rights have to be com-bined, the transaction costs implied could be so high as to prevent someotherwise productive combinations. This problem has been studied in abroader context as the ‘anti-commons’ problem (Heller, 1998; Heller andEisenberg, 1998).

To fix ideas, suppose that the development of a new technology involvesthe use of N fragments invented and patented by separate firms. In addi-tion, the technology innovator has to pay ex ante a fixed cost, I, whichmight be thought of as expenditures in R&D. In order to assemble thenew technology either the innovator has to buy licences on the fragmentsor alternatively he or she has to invent around them. The cost of invent-ing around depends, among other things, on the strength on intellectualproperty rights defined around the single fragments.9 By definition, a‘blocking patent’ implies that such cost is extremely large. Let us assumethat in the process of collection of the rights, the parties agree to sign alicensing contract which stipulates an up-front fee negotiated throughbilateral bargaining.

There are two straightforward results that one can derive. First, thehigher the cost of ‘inventing-around’ the fragments, the weaker the bar-gaining power of the innovator in the licensing negotiations to collect therights for the use of the different fragments. This is simply because the inno-vator’s outside option – ‘inventing around’ the patents on the fragments –becomes less attractive. Second, the larger the number of fragments, thehigher the number of contracts that should be signed to guarantee the useof the innovation. If transaction costs are increasing with the number oftransactions, a larger N is likely to increase the total transaction costs toassemble the fragments.

A more interesting and less straightforward result emerges when one con-siders opportunistic behaviour of firms holding ‘blocking patents’ on thefragments. Indeed, the further the innovator goes in the collection of therights for use of the fragments, the more resources are subject to an irre-versible commitment and, therefore, the weaker is his or her bargainingpower in future licensing negotiations for the collection of the remainingfragments. This implies that in subsequent negotiations he or she will have


little chance to recoup all costs sunk up to that moment, that is, the fixedinvestment, I, and the fees paid for the rights on the fragments alreadybought. Furthermore, firms might also try to delay selling their blockingclaims in order to hold out for more ‘quasi-rents’ that become available. Putdifferently, since the last firm that is going to negotiate its ‘blocking patent’has the strongest bargaining power vis-à-vis the innovator and can capturethe largest amount of rents, all firms have incentives to be the last. This islikely to introduce some further inefficiencies and delays in assembling thetechnology. Obviously, all these ‘hold-up’ problems can only be exacerbatedwhen the number of ‘blocking patents’ on separate fragments increases.10

So far we have analysed a scenario with basically no uncertainty.Technologies, by definition, are characterised by a high level of uncertainty.This can only worsen the picture. First, it can sometimes be very difficult toknow a priori N. In other words it is a hard task to determine whichdomains of technology have a legitimate bearing on the commercialproduct and who all the relevant intellectual property rights holders are.Ensuring access to all potentially blocking rights can therefore becomeextremely cumbersome.

Second, when the market value of the innovation is uncertain firmsmight agree to sign royalty-based payment for the use of the fragments. Inprinciple, these offer the advantage to the innovator of delaying the pay-ments till the moment profits from the innovation start to materialise andto the patent-holders they confer the chance at larger pay-offs from sales ofdownstream products rather than certain, but smaller, up-front fees.However, the presence of such royalty-stacking negotiated on individualbasis might imply that the total amount of royalties per unit of output isinefficiently high both from a private and social point of view.11

Third, as Langlois (2002) pointed out, in environments marked byKnightian uncertainty, transaction costs include not only the problems ofhold up, bargaining and imperfect contracts. Rather, they include theproblem that Langlois calls dynamic transaction costs. For instance,Langlois argues that Henry Ford’s consolidation of all production steps ina vertically integrated company was critical to the successful introductionof the Model T in the 1920. Had the various stages of production remainedunder separate ownership, Ford would have had difficulty experimentingwith new techniques, machines and parts, all of which had to fit with eachother. In other words, till the overall architecture of the product, in thiscase, the Model T, was settled, the costs of co-ordinating the actions ofindependent parts suppliers and machine-makers would have been veryhigh. Similarly, when an innovation is based on combining the intellectualproperty of several independent agents, the costs of persuading them to‘rent’ or part with their property for a particular application will be high,


because each person’s pay-off will be contingent on all relevant partiescoming on board. The net results might be that the parties involved areunable to reach an agreement.

One example of what might occur when several companies hold patentson different components is provided by the early development of the radio(Merges and Nelson, 1990). The Marconi Wireless and TelegraphCompany, AT&T, General Electric and Westinghouse all held importantpatent positions in the early stages of the development of the industry. Theensuing fragmentation of property rights is said to have caused seriousdelays in the pace of technological innovation. For instance, the basicpatent on the diode was granted to Marconi, while the patent on the triodevacuum tube was assigned to AT&T. Marconi’s patent was needed for usingthe triode technology, yet neither party would license the other and, as aconsequence, no one used the revolutionary triode for some time.

Software, semiconductors and computers are other good examples ofindustries where the nature of innovation is systemic and cumulative, andwhere the intellectual property is very fragmented. In these industries theopportunities for hold up are enormous. Indeed, as reported in Grindley andTeece (1997) and Hall and Ziedonis (2001), this has led industry actors tosign cross-licensing agreements covering whole portfolios of patents relatedto an entire technical field (including both existing and future patents).

These concerns have been echoed by industry participants as well. CecilQuillen, former Senior Vice President and General Counsel of the EastmanKodak Company, claims that since the early 1980s, the legal costs ofintellectual property protection has risen dramatically to the point ofsubstantially raising the cost of innovation itself. Michael Rostoker, formerhead of LSI Logic, a semiconductor manufacturer, has also suggested that,due to stronger patent protection, firms holding old technology have beenin a position to command licensing fees from a current generation of inno-vators even while the original patent-holders have long ceased advancingthe state of the art, leading to a stacking of licensing fees that impede thedevelopment of new generations of chips (Hadley, 1998).

Similar situations arose in the early stages of development of theautomobile and aircraft industry, and in the chemical process technologyindustry. Biomedical research may provide another possible example. Aparticular concern raised by Heller and Eisenberg (1998) and the NationalResearch Council (1997, ch. 5) was the prospect that, by potentially increas-ing the number of patent rights corresponding to a single gene, patents onexpressed sequence tags would proliferate the number of claimants toprospective drugs and increase the likelihood of bargaining breakdowns.

Although plausible, the available evidence, limited as it is, suggests thatanticipated problems have not yet materialised. For instance, based on


about 70 interviews, Walsh et al. (2003b) report that although there hasbeen an increase in the 1990s in patents on the inputs to drug discovery,drug discovery has not been substantially impeded by these changes. Theyreport some evidence of delays associated with negotiating access topatented research tools. Further, there are instances where patents overtargets limit access, and where access to foundational discoveries can berestricted. There are also cases where research is redirected to areas withmore intellectual property freedom. Still, the vast majority of respondentssay that there are no cases where valuable research projects were stoppeddue to intellectual property (IP) problems.

One does not observe as much breakdown or even restricted access toresearch tools as one might expect because firms have been able to develop‘working solutions’ including licensing, inventing around and legal chal-lenges. Importantly however, institutional responses, particularly new PTOguidelines, active intervention by NIH and some shift in courts’ viewstowards research tool patents, appear to have further reduced the threat ofbreakdown and access restrictions.12

Interestingly enough, the institutional response had important privatecomponents as well (Walsh et al., 2003b). For example, public databases(for example, GenBank, or the Blueprint Worldwide Inc. venture to createa public ‘proteomics’ database ) and quasi-public databases (such as theMerck Gene Index and the SNPs Consortium) have been created, with sub-stantial public, private and foundation support. Merck has sponsored an$8 million programme to create 150 patent-free transgenic mice to be madeavailable to the research community at cost, without patent or use restric-tions.13 These initiatives represent a partial return to the time beforethe genomics revolution, when publicly funded university researchersproduced a body of publicly available knowledge that was then used bypharmaceutical firms to help guide their search for drug candidates.14

Scientific journals have also pushed for access to research materials, andbiology journals require that authors deposit sequences in public databasessuch as GenBank or Protein Data Bank (Walsh and Bayma, 1996).Similarly, when Celera published their human genome map findings,Science’s editors were able to gain for academics largely unrestricted accessto Celera’s proprietary database.

3.2 Policy Responses

This leads to some straightforward implications for patent offices as well.The main purpose of the Patent Office is to issue patents that the legisla-tion permits or deems desirable. The Patent Office evaluates whetherthe claims are enabled under the terms of the statute or the patent has the


statutory novelty in light of the prior art. The role of the Patent Office is toassure that only well-specified applications receive a patent. The better thedefinition of the claims, the less uncertainty about their scope and validity.This translates into lower transaction costs for technology trade andhence better functioning of markets for technology. Moreover, owing tothe growing importance of patents, the social cost of ‘bad’ patents hasincreased, along with the number of patents themselves, as Kortum andLerner (1999) have documented. This argues for more resources to be madeavailable to patent offices for examining patents and, in particular, forsearching for prior art.15

One may conjecture that larger patent offices and stronger incentives topatent might increase patent disputes and litigation, creating deadweightloss. The spurt in litigation activity that we have witnessed in recent yearsis not a consequence of a greater number of granted patents alone. Indeed,it also points to the likelihood that patent offices, particularly in the USA,have issued poorly defined patents, with overly broad scope or of dubious‘non-obviousness’ and novelty over prior art.

Often, broad and imprecise patents are issued because patent offices areunder-funded, and the patent examiners inadequately trained and lackingthe necessary capabilities to search for the prior art. In software, forinstance, the US Patent Office has issued what are widely seen as overlybroad patents, in large measure because the examiners rely very heavilyupon previous patent applications to discover prior art. Since softwarepatents are relatively new (copyrights having been the typical way of pro-tecting software until recently), the result is bad and socially harmfulpatents, which nonetheless carry with them the presumption of validity.

By the same token, patent offices should pay more attention to patent-ing requirements. Specifically, in the USA, the patentee is required to‘reduce to practice’ the invention, show that it ‘possesses’ the invention,demonstrate the best known way the invention is to be used or ‘enabled’and show the usefulness or ‘utility’. In the late 1990s, there was a percep-tion that these requirements were not been enforced very seriously, at leastin certain well-known cases. For instance, some early patents on gene frag-ments (ESTs) were issued without any clear knowledge of what proteinsthe gene fragment was coded for and what functions the proteins per-formed. In principle, these fragments could prove to be useful in a broadspectrum of applications, as yet unknown. If granted, the patent-holdermight be able to demand a large share of the rents from any such appli-cations or even block such applications, without having contributed totheir discovery. Thus, the fear was that such patents could perverselyblock genomic research and, particularly, the commercialisation of suchresearch.


As indicated, these fears have, for the most part, not been realised.It appears that the institutional response to the concerns about the possi-bility of intellectual property rights impeding research and its commercial-isation has been very important. In January 2001 the US Patent Officeissued guidelines clarifying that a clear and specific utility would have tobe indicated in applications for ESTs. Similarly, the US Patent Officeannounced that it would undertake a more careful examination of prior artfor business method patents involving the Internet (www.uspto.gov, postedApril 2000). In addition, in a series of judgments, the Court of Appeals ofthe Federal Circuit (CAFC) in the US (the so called ‘patent court’) nar-rowed the scope of many patents and limited the ability of patent-holderson upstream discoveries to block downstream development.16

In addition, there is room for policy-oriented interventions to facilitatethe functioning of markets for technology even under very fragmentedintellectual property. One possibility to modify the traditional stance ofantitrust authorities on patent-pooling agreements. A patent-poolingagreement involves typically two or more companies with similar or over-lapping patents. Rather than pursuing interference proceedings, or engag-ing in long and costly litigation to determine issues such as patent validityor infringement, they put their collective efforts to more productive use. Forexample, they may form a separate entity to which they assign or licensetheir patents. The entity collects money for the service or product and paysout a royalty to each of the patent owners, according to the terms of theagreement. A similar argument applies to cross-licensing agreements wherefirms agree to license each other the use of their respective fragments.17

Traditionally, antitrust authorities in the USA have aggressively scrutini-sed patent pools and cross-licensing agreements, because such kind ofagreements were sometimes used for restricting entry, controlling pricesand market shares. However, recently the antitrust stance appears to havechanged somewhat, favouring the emergence of market-based responses tothe problem of excessive fragmentation of intellectual property rights. Forinstance, Grindley and Teece (1997) attribute the extensive use of cross-licensing agreements in electronics and semiconductors, where innovationsare typically based on hundreds of different existing patents, to the largetransaction costs required to bundle together patent portfolios. Recently,the Department of Justice has given the green light to a group of nine com-panies and one university to create a pool of patents that are essential tothe MPEG-2 video standard. Another patent pool involving 11 patent-holders has been agreed for the IEEE 1394 bus, a popular solution fortransferring audio and video data.18

In the chemical process industry, technology-sharing agreements have along history and were established to alleviate the transaction costs involved


in market relationships. The case of the chemical process industry is inter-esting for another reason as well. The specialised engineering firms (SEFs)mentioned earlier, which supplied chemical process engineering and design,have sometimes acted as technology integrators which have helped ingetting around the ‘hold-up’ problem of fragmented property rights. Thus,another potential benefit of specialised technology suppliers is that theycan act as technology integrators limiting the hold-up problem created bythe fragmentation of intellectual property rights.

A final set of, more controversial, policies whose merits remain underdebate is the extension of ‘eminent domain’ (that is, the legal doctrine thatallows the government to take over private property for public purpose) tointellectual property. In principle, the threat that the government may stepin and buy out a patent-holder at a ‘fair’ price can be a powerful deterrentto the sort of opportunism that underlies the fragmentation problem. Butgovernments may not be the best agencies to take over a technology wherepublic good considerations might be quite indirect. Determining the pricefor the patent is an important challenge. Kremer (1998) suggests using anauction as mechanism to determine the private value of patents. The gov-ernment would use this price to buy out the patents and place them in thepublic domain. Alternatively, the law may simply allow for ‘efficientbreach’ – that is, let people ‘infringe’ the patent and leave the courts decideabout a ‘fair’ royalty. The latter is very similar in spirit to the compulsorylicensing provisions and provisions that require the patent to be ‘worked’.Both these provisions have been present in many countries, especially inthe past, and require courts to intervene more aggressively than is proba-bly desirable.

4 THE PRIVATISATION OF KNOWLEDGE:CRUMBLING ACADEMIC NORMS?

The growth of markets for technology and the concomitant strengtheningof intellectual property rights raises another fundamental challenge, thistime at an institutional level. As basic scientific knowledge, such as thestructure of genes, becomes eligible for patenting, universities are bothpulled and pushed into entering the market for technology. In the USA,the decline of the cold war reduced government funding for researchand, hence, an important source of revenues for research universities.David (1997) describes the trend in policy discourse towards a more instru-mentalist view of university research, moving from national defence tonational competitiveness, to wealth creation in the most recent stage. Thesetrends have contributed to an increasing pressure upon universities and


university researchers to demonstrate the economic value of their work.Concomitantly, universities have increasingly resorted to patenting andcopyrighting their research. As Mowery et al. (2001) suggest, this is partly(but not only) a consequence of the passage of the Bayh–Dole Act in theUSA in 1980, which allowed universities and government laboratories toclaim patents on federally funded research.

Other factors have encouraged the trend towards greater privatisation ofuniversity-based research. A general expansion of patentable subjectmatter and a more patent-friendly legal environment (with the establish-ment of the so called ‘patent court’, the Court of Appeals of the FederalCircuit) in 1981 meant that the patents that universities could obtain wouldbe more valuable.19 The result was a great expansion in university patent-ing, as well as the establishment of university licensing offices and encour-agement of ‘spin offs’ to commercialise university-based research. In short,universities are beginning to resemble, albeit only partially, both a sourceof technology as well as incubators for developing independent suppliersof technology.

The number of patents issued to the US universities increased signifi-cantly between 1980 and the late 1990s. The share of patents assigned toUS universities grew from less than 1 per cent of all the patents assigned toUS inventors in 1975 to almost 4 per cent in 1997 (Mowery et al., 2001).Since US patents have grown rapidly during this period as well, it impliesthat university patents have grown faster still. Moreover, Mowery et al.report that while the patents per $1 billion R&D spending (in constantterms) declined from 780 in 1975 to 429 in 1990, university patenting hasshown an increase in this ratio from 57 to 96 over the same period. Theseincreases reflect more systematic attempts by universities to assert rightsover inventions, including attempts by university licensing offices to elicitthe disclosure of such inventions. Alongside this, US universities steppedup their efforts to license their patents. Mowery et al. report that licensingrevenues of US universities (those that are members of the Association ofUniversity Technology Managers) have increased, in real terms, from$222 million in 1991 to $698 million in 1997. This is a notable increase,which is confirmed by their case studies of three leading US universities –Columbia, the UC system and Stanford.

These trends raise three concerns. First, to some there seems to be some-thing wrong with the notion of publicly subsidising research, whose resultswill later be monopolised. Put differently, even if one accepts that the tem-porary patent-based monopoly is a necessary evil to provide incentives forinvestments in research, then surely research that is publicly funded doesnot require the lure of patents. Patents are both unfair and inefficient. Thisline of reasoning, however, ignores the other role that patents can play,


namely, in encouraging the efficient transfer of knowledge between theinventors and the commercialisers. It has been argued that without suchpatenting, much of the research now being commercialised would lie fallowand unused. In so far as this is important, one benefit of university patent-ing is that university researchers can effectively benefit from the inventionby licensing the technology and the know-how, instead of attempting tocommercialise the innovation themselves. If the latter were to happen, theuniversity would lose the services of, potentially, very valuable researchersand teachers.20 Patents affect the commercialisation of university tech-nology through another route as well. They provide incentives, not forinvention, but for development. Indeed, at the heart of the Bayh–Dole Actwas the belief that without some measure of exclusivity, companies wouldnot invest in developing university research.

University licensing also seem to rely heavily on exclusive licensingcontracts. For instance, Mowery et al. (2001) report that for the period1986–90, the fraction of licensed disclosures that were through exclusivecontracts was 58.8 per cent, 59.1 per cent and 90.3 per cent for Stanford,Columbia and University of California, respectively. They seem to worryabout the rather heavy use of exclusive contracting, arguing that non-exclusive licensing would balance the need for exclusivity with the publicinterest of broad dissemination of knowledge. The heavy use of exclusivelicences should raise preoccupations, particularly if they cover technologywith broad applicability.

The second concern relates to the nature of university research. Forinstance, might the push to earn revenues from research increase emphasison more applied research at the cost of basic, or fundamental, research?21

Paradoxically enough, the analysis of patent citations provides one sourceof empirical evidence. Henderson et al. (1998) found that the rise of uni-versity patenting after 1980 was associated with a decline in the ‘impor-tance’ and ‘generality’ of university patents. Since patents relating to morefundamental and broad-ranging discoveries were likely to score higher onthe proxies used to measure importance and generality, this evidence seemsto support the concern about a move towards more short-term and appliedresearch at universities.

However, a more recent study (Mowery and Ziedonis, 2002), which con-trols more carefully for the changing composition of universities thatpatent, reaches a different conclusion. They analyse citations to patentsheld by Stanford and the University of California and find that, relativeto a control group of similar patents by industry, there has been no declinein the importance and generality of patents from these universities. Usingdata on all university patents, they find that the decline in importance andgenerality is largely as a result of the increased share of university patents


due to universities that are new to patenting. Thus Mowery and Ziedonisconclude that the evidence does not support a major shift in the contentor culture of university research. As before, we lack systematic evidenceto come to a definite conclusion on this issue. However, a shift towardsmore applied research and a neglect of more basic and fundamentalresearch is certainly a possibility, and universities and governments haveto be watchful.

The third, and perhaps most important, concern raised by the growinginvolvement of universities and university faculty and researchers inpatenting and commercialising their discoveries is the impact on academicnorms, and the consequences for the growth of new knowledge. Dasguptaand David (1994) and David (1993a) have highlighted both the featuresthat distinguish the production and dissemination of university researchfrom that carried out in firms and also provided an economic frameworkfor understanding these distinguishing features. They argue that thedifference is not so much in the methods of enquiry or the nature of theknowledge obtained. Rather, it is the nature of the goals accepted aslegitimate within the two communities of researchers, the norms regard-ing disclosure of knowledge and the reward systems that are held tobe the distinguishing features. Roughly speaking, university research isundertaken with the intent of disclosure and the rewards include theapproval and respect of a broad invisible college of peers. Inevitably, thesedifferences are associated with differences in the types of questionstackled and the methods for representing and communicating the resultsof the research. As Walter Vincenti (1990) put it, scientists are interestedin understanding how things are, whereas engineers focus on how theyought to be.22

These differences have evolved along with the corresponding institutionsin response to some specific features of research as an economic activity.For instance, David (1991; 1993b) argues that an inability on the part ofEuropean princes and noblemen in the Middle Ages to monitor the qualityand effort of the scientists they patronised provided an impetus for opendisclosure of research findings. With open disclosure and peer review, themerits of the research findings and, hence, the quality of the researcher,would be easy to establish. As noted earlier, this also required consensus onthe methodology and terminology. The open and rapid dissemination ofresearch findings and the associated academic norms of scientific conduct,especially peer review, academic freedom and an apprenticeship-type rela-tionship between research students and professors, have become part andparcel of what we think of as university research. These norms are sus-tained by the public subsidy for university research, allocated principallythrough a peer-review based mechanism. In essence, the community of


researchers decides upon how funds should be allocated, subject to somegeneral guidelines and constraints.

Perhaps even more important are the norms of cooperation and colle-giality. As Dasgupta and David (1994) point out, the high importanceattached to priority creates a tension between complying with the norm offull disclosure and the individual urge to be the first to publish.23 Since thesolution of scientific problems typically requires research into several sub-problems, full disclosure yields a better outcome for the community as awhole in the long run. However, individual researchers have an incentive tofree-ride by learning from others but not co-operating in turn. Althoughco-operative behaviour can be sustained in repeated games by threat ofexclusion, Dasgupta and David (1994) argue that scientific norms greatlyincrease the likelihood that networks of co-operative information sharingwill arise because of an increased trust. More generally, these norms arecritical for the formation and sustenance of these communities. In turn,these scientific communities act as an agent of society at large, punishingthose that violate co-operation (by withholding findings), reviewing thevalidity of results as well as training new researchers, and providing somedegree of verification of the quality of the researchers themselves. Thatfunding for scientific research is allocated by the community itself is com-plementary to these other functions and reinforces the ability of the scien-tific community to enforce scientific norms.

Privatisation of knowledge weakens these norms by reducing the abilityof the scientific community to sanction violators and by increasing therewards to violators. To hark back to Dasgupta and David’s model of thesituation as a repeated prisoner’s dilemma game, privatisation of knowl-edge increases the pay-off to withholding co-operation when secrecy fol-lowed by patenting of the results will yield large monetary pay-offs. Thereare indirect effects as well. Focusing universities on earning revenuesthrough research leads to a dilution of the role of the larger scientific com-munity in the allocation of funds for scientific research, weakening thepower of the community and, hence, also weakening the hold of the normsof disclosure and collegiality.

The foregoing is certainly closer to the Platonic ideal of universityresearch rather than an accurate description of existing reality. The pointis that these norms are valuable and useful, and need to be reinforcedrather than weakened. Moreover, although resilient, norms are easier todestroy than build – once a sufficiently large fraction of the research com-munity moves away from them, it will be hard to sustain them anywhere.Indeed, the growing commercial applicability of scientific research inbiotechnology has been accompanied by growing anecdotal evidence ofthe violation of these norms by scientists, including the withholding of


important information, delays in disclosure, refusal to co-operate withother researchers, as well as tales of abuse of graduate students and post-doctoral students (for example, Kenney, 1986).

As noted earlier, the available evidence, limited as it is, does not supportany significant shift in norms associated with patenting. Mowery et al.(2001) conclude that the bulk of the increased patenting appears tohave been associated with the increased importance in software and bio-medical research, which, though more amenable to patenting, is no lessbasic or fundamental than other research at universities. Based on theircase study of patenting at Columbia University, they also concludethat the impacts on norms are in any case likely to be limited to a fewdepartments most heavily involved in patenting: electrical engineering,computers science and the medical school. Even so, it is likely that in sofar as the privatisation of knowledge adversely affects universities, it willbe by diluting and degrading the norms of openness and collegiality.Moreover, the ‘few’ departments that are most involved in patenting arelikely to be quite important ones for the production of ‘open’ academicresearch and public domain knowledge that is critical for socio-economicgrowth. In this respect, the fact that two quite important fields such as soft-ware and biomedical research are more involved in patenting and mayloose their reliance on open academic settings, can be by itself a source ofconcern.

A broader question is whether good research can flourish anywherewithout such norms. This is a question of institutional design that we raisebut to which we do not know the answer. Although the academic model ofresearch has been very successful in the last 100–150 years, to our knowl-edge, alternatives to that model have not been tried. Indeed, there is evi-dence that successful research-oriented firms have tended to adopt theacademic mode of organising at least some of their research.24 In otherwords, the prevailing Western university model has implicitly been assumedto be the only sensible way of organising research, especially basic or fun-damental research. With the growing privatisation of knowledge, thisassumption may well be tested in the future.

5 ‘GLOBAL’ MARKETS FOR TECHNOLOGIES ANDNATIONAL POLICIES

Along with other markets, markets for technology are becoming global. Insome ways, this is only to be expected, given the smaller ‘transport’ costsand the greater appreciation, by even otherwise protectionist governments,of the benefits of technology. Rapid advances in communications, with


the Internet being only the most recent, has only hastened the process ofglobalisation.

Markets for technology are far more likely to arise in large and tech-nologically and economically advanced regions, than in developing coun-tries. The latter, therefore, need not focus on developing such markets.Instead, they can focus on developing institutions that will enable theirfirms to participate more effectively in these markets. The example of theWestern European chemical industry in the years after the Second WorldWar is a case in point. Prior to the war, the European chemical industry wastechnologically far ahead of the USA industry. The disruption due to thewar and the rise of the petrochemical industry, and the associated processtechnologies, in the USA, ought to have provided the US chemical indus-try with a decisive advantage over its European rivals whose expertise layin coal-based processes. Yet, in a period of a few years, the German, Britishand French chemical industries had largely switched over to using petro-leum and natural gas as basic inputs. The availability of US-developedrefining and chemical engineering expertise made this switch possible.Further, the specialised engineering producers, the SEFs, played an impor-tant role in integrating and supplying technology to European customers.In the 1960s, the SEFs played a similar role in Japan. Japanese industrialpolicy, which tended to restrict access to the Japanese market for foreignfirm, was far more receptive to foreign technology imports. Indeed, thepolicy focus in this context was in creating the ability to absorb and adaptforeign technology (Arora and Gambardella, 1998.)

The point is simple and well known: global markets tend to circum-scribe the role of policy in being able to improve market outcomes.For smaller countries like the individual European countries or the lessdeveloped countries, the impact of their own policies, if they are notco-ordinated with those of other countries, is likely to be small. Forexample, policies by smaller countries to develop standards or othertypes of supporting institutions are unlikely to induce the development oftechnology markets on substantial scale. Similarly, strengthening or weak-ening intellectual property rights will probably have little effect on theglobal market for technology, although this may affect the extent to whichtechnology flows into their country or technology trades takes placewithin it.

Policies for encouraging, co-ordinating or controlling the markets fortechnology will be most effective when they are developed by large coun-tries (for example, the USA), or by sets of countries (for example, theEuropean Union). Such policies require co-ordination among countries,and this requires super-national interventions in international policy set-tings. But it is precisely at this super-national level that policy decisions are


harder to take because of the many conflicting interests involved, and thelack of strong enforcement mechanisms. And this is why policies developedby a large homogeneous country like the USA (for example, in intellectualproperty rights, in the development of standards) can have a strong impacton the world development of markets for technology, as we are observingtoday with the effects and the debate that has been raised worldwide by theUS attitude towards stronger patents. Likewise, the European Union canplay a significant role, especially if it can harmonise the policies of the indi-vidual member states, and avoid that individual member states adoptdifferent rules and standards.

For most other countries, the key policy question may be how to takeadvantage of the worldwide growth in technology trade. This will requireencouraging the effective use of existing technologies, rather than the cre-ation of new ones. Also, policies aimed at monitoring international tech-nological developments increase in importance, as do institutions forenhancing the efficiency of contracts and reducing search costs. In thisview, countries may increase the emphasis on the ability to identify andselect technology, and develop complementary capabilities.

In sectors where markets for technology develop, and technology can betraded more effectively, countries or regions should specialise according tocomparative advantages. This does not imply that countries should ceaseto invest in research and development. Rather, it implies that they shouldbe more selective in terms of which sectors they focus on, and more select-ive in terms of the types of activities they focus on, at least in the short tomedium term.

It is well known that R&D and technology production is quite concen-trated worldwide. The rich countries, and in particular the USA andWestern Europe, have a head start in terms of basic research, developing‘generic’ technologies like semiconductors and genetics. Their advantagelies not only in first-mover advantage, but also the broader industrial baseover which to apply these findings. These advantages are less salient whentechnologies and products need to be adapted for local uses and needs. Ifone accepts that companies or industries located ‘near’ users have anadvantage in communicating with their markets and in acquiring the rele-vant information for adapting the technologies, this suggests that firms inother parts of the world could seize this niche. Thus, even if the productionof more basic technologies is concentrated regionally, if markets for tech-nology work well, other regions can get access to these basic technologiesand exploit their proximity to users, or their comparative advantage indeveloping complementary technologies (for example, the growing devel-opment of software components for leading software producers worldwideby Indian firms today – see Arora et al., 2001a).


These recommendations are not new and in some quarters, are viewed asa prescription for perpetual technological ‘backwardness’, and some coun-tries may resist such an international division of labour in technology pro-duction and adaptation. The reasons may range from national pride to thewillingness to control strategic technologies.

Thus, some form of the ‘not-invented-here’ syndrome, at the level ofcountries, is likely to operate. Whether justified or not, it is important toknow that markets for technology, where they exist, increase the opportu-nity cost of such an attitude. Simply put, if others have already paid thefixed cost of developing technology, and competition among sellers impliesthat the price of the technology is related to the marginal cost of tech-nology transfer, a strategy of developing the technology in-house andincurring the fixed cost all over again must provide some additional bene-fits over mere ownership of the technology. There is little point in nationalpolicies aimed at ‘reinventing the wheel’, except where such reinvention is apart of the process of building ‘absorptive capacity’ or as a part of a long-run strategy to develop international technological leadership.

Second, in a dynamic setting, the international division of labour, withimplied specialisation in technology production and adaptation, meansthat countries that specialise in the latter need not give up the possibility ofbecoming technology producers, at least in some well-defined areas. Forexample, by starting with a policy of developing technologies that are com-plementary to those developed by some leading areas or regions, the localfirms and industries may gradually learn about the basic technology as well,and they could possibly escalate up to becoming the producers of some keytechnologies. The Indian software industry, for instance, started as a low-end supplier of software components to the major software companiesespecially in the USA, and this has proven to be an overly successful strat-egy by many standards (employment growth, exports, and so on, see Aroraet al., 2001a). A similar argument can be made for the Irish software com-panies, which seem to have improved their ability to become producers ofnew software products in some niches of this market (Arora, et al., 2004).In short, in a dynamic setting, the pattern of specialisation is notimmutable. With proper technology policies, the advantages of specialisa-tion in lower-end technological activities (adaptation) could even becomethe springboard for a move up the value chain. Learning through system-atic interactions with the markets or the technology producers of moreadvanced countries may be critical for this process to take place.

Indeed, some countries, like Russia and Israel, and to a lesser extent,India, have a relatively well-developed scientific and engineering infra-structure. However, they lack the market size and the complementary tech-nological and economic infrastructure that could best exploit their


scientific and engineering infrastructure. In this respect, they are similar tospecialised technology suppliers. A well-developed and globalised marketfor technology will enable firms from these countries to derive more valuefrom their investments in science and engineering, by supplying technologyto others that can develop and commercialise it more effectively. Here, too,one may encounter opposition from those who would see this as ‘givingaway the store’. Once again, our objective is not to act as advocates, for theappropriate policy will depend on the specifics of the situation, but to high-light the option that markets for technology would create.

6 CONCLUSIONS

Markets do not arise simply because the benefits of having them outweighthe costs. They also require institutions that support them. Further,markets develop over time, along with these complementary institutions.As Paul David has long taught us, this development has to be understoodas a historical process, with the pace and form of the development condi-tioned by starting conditions and chance. Further, the rise of a new marketaffects other markets and other existing social and economic institutions.Their development raises new challenges for policy-makers but creates newpolicy options as well. So also with markets for technology. Policy can playa relatively more important role for encouraging these markets when theyarise rather than after they start functioning. Further, policies that encour-age ‘decentralized’ institutional innovations, along the lines of the new‘market-enhancing view’ of policy are likely to be more successful thandirect policy attempts to create such institutions (Aoki et al., 1996).

In this chapter we have highlighted some of the major policy challengesposed by the development of markets for technology. Intellectual propertyrights are a sine qua non for the development of markets for technology.But given the nature of knowledge, property rights in knowledge, such aspatents, can create problems. In some cases, they can retard the develop-ment and commercialisation of innovations, as for instance when such a userequires combining the intellectual property rights controlled by a numberof independent agents. How serious this problem is in practice is uncertainand further research in this area would be very valuable.

The privatisation of knowledge can also undermine an important insti-tution of modern capitalism, namely, the research university, by weakeningacademic norms of open disclosure and collegiality. Weakening publicsupport for academic research exacerbates the problem by forcing univer-sities to look to generate additional resources by patenting and licensingtheir research findings. As we note, empirical research on this topic is just


beginning and the available evidence suggests that the situation is not irre-vocable. However, by their nature, norms are easier to destroy than to createand it seems sensible to try to modify only very slowly a system that appearsto have worked well as a way of organising basic research.

With markets becoming global, the exercise of national policy has to bemore circumscribed. Especially for ‘smaller’ countries like the individualEuropean countries, or the less developed countries, markets for tech-nology imply a focus on how best to benefit from the growth of thesemarkets. We suggest that this would mean becoming more open to outsidetechnology and re-examining arguments for investments based on nationalpride. It would mean participating in an international division of labour,by increasing the emphasis on using technology and building complemen-tary capabilities, possibly at the cost of investing in basic research. In adynamic setting, the learning potential that is embedded in a division oflabour with more advanced technology producers can create the opportu-nities for later specialisation in some of the more basic technology areas.

NOTES

1. Aoki et al. (1996: 6) make a similar point. They argue that technological and other com-plementarities (for example, of institutions, or expectations) are more likely to occur atthe early stages of development of new markets or technologies. Hence, governmentintervention and the need for co-ordination are more useful earlier than later when suchmarkets function, or the technologies are more mature, and there is greater competition(and substitutability) among them.

2. Heller and Eisenberg’s concerns were specifically directed to the growth in genomicpatenting, and more specifically, on patenting of Expressed Sequence Tags (ESTs). Theavailable evidence suggests that a variety of factors, including institutional responses bythe Patent and Trademark Office (PTO), the courts and by the National Institute Health(NIH), as well as private responses by firms, have largely avoided the problem. Furtherdetails are contained in Walsh et al. (2003a; 2003b).

3. Clearly, conferences, seminars, visiting positions in other schools or departments showthat the exchange of papers is not enough, and many scholars have pointed to theimportance of tacit communication. Even so, one striking example of what a commonlanguage can do is open-source software, whereby software programs are developed bydevelopers located far away from each other and communicating principally over theInternet and by exchanging the software code itself.

4. In this respect, another area for further study is the effective prohibition on lawyersundertaking patent infringement cases on a ‘contingency’ basis.

5. This is the same argument put forward in Arora (1995) to explain the problems in trans-ferring technology based on tacit, unpatented knowledge. Anton and Yao (1994) developa clever model which shows that in principle, a technology holder can sell the technologywithout any intellectual property protection, in effect, by threatening to put it in thepublic domain (and destroying its value) if the buyer reneges on it. Anton and Yao (2002)show that a seller can partially reveal the technology to signal its value, mitigating theasymmetric information problem.

6. In some cases, policies designed in the naive hope of encouraging small inventors haveencouraged the abuse of the patent system. In the USA, for instance, there have been


well-known cases where patents filed in the 1950s ultimately were issued more than20 years later. In the mean time, the patentee could legally amend the application so thatit covered inventions made well after the filing date. Since patents in the USA are pub-lished only upon issue, many established firms have been surprised by such patents(sometimes referred to as ‘submarine’ patents because they are not visible for longperiods after they are filed). The move towards patent harmonisation, which will requirepublication of patent applications after a certain period, will be helpful in this respect.

7. The point is not that information can be reproduced at low cost or that information isnon-rival in the sense that one person’s knowing something does not preclude anotherfrom possessing the same information. A familiar counter-example is as follows. If onlyone person knows what is going to happen to the price of a stock, he or she is likely tobenefit greatly. But if all (or sufficiently many) were to have the same information, noneis going to benefit. Thus information can be rival in use, although in the physical sense,it is non-rival.

8. As Paul David (1993a) noted, knowledge is different from the prototypical public goodssuch as lighthouses and airport beacons. One important point of differentiation is thatthe acquisition of knowledge is cumulative and interactive: knowledge itself is animportant input into the production of knowledge.

9. Other factors such as standards may also raise the cost of inventing around.10. A related consequence is that non-manufacturing firms that hold patents on key com-

ponents are likely to bargain more aggressively for licensing fees. The strategies of firmsthat have significant market shares in the downstream markets (in which the technologyis applicable) are more complex. However, they are likely to co-operate, particularly ifthere is a stable group of such firms. Interestingly enough, the ownership of mutuallyblocking patents can actually support licensing in this context, since each party will havethe ability to block commercial development by the other.

11. Think for instance of two patent holders fixing separately their royalty rates for sellingtheir patents to a unique licensee vis-à-vis the case in which the two patents are pooledand a single royalty is set up. This is a distortion similar to the one generated by thedouble marginalisation in a chain of monopolies.

12. For instance, as Walsh et al. (2003b) note, NIH negotiated with DuPont to provide morefavourable terms for transgenic mice for NIH and NIH-sponsored researchers, to relaxrestrictions on publication and sharing of animals and eliminate reach-through provi-sions. The NIH has also begun a ‘mouse initiative’ to sequence the mouse genome andcreate transgenic mice. One of the conditions of funding is that grantees forgo patent-ing on this research. Then NIH also pushed for broader access to stem cells, as well asfor a simplified, one-page Material Transfer Agreement (MTA) without reach-throughclaims or publication restrictions.

13. The firms in the SNPs Consortium include Bayer, Bristol-Meyers Squibb, GlaxoWellcome, Hoechst Marion Roussel, Monstanto, Novartis, Pfizer, Roche, SmithKlineBeecham and Zaneca. Each firm contributed $3 million, and Wellcome Trust addedanother $14 million to the effort.

14. One could speculate that the shift of Human Genome Science (HGS) away from thedatabase business and toward the drug development business may be a response to boththe higher returns available to drug companies and the lower returns available togenomics companies that are competing with increasingly developed public databases.

15. Merges (1999) shows that the US Patent Office has about $3000 to spend on each patentapplication. Further research is needed to assess whether this is the optimal amount tospend. Any such assessment should take into account the impact of intellectual propertyrights on the functioning and development of markets for technology.

16. Cockburn et al. (2002) find that the CAFC went from upholding the plaintiff in about60 per cent of the cases to finding for the plaintiff in only 40 per cent of the cases in recentyears. Similarly, in University of California v. Eli Lilly and Co, the court ruled against theUniversity of California’s argument that its patent on insulin, based on work on rats,covered Lilly’s human-based bio-engineered insulin production process. Similarly, it


ruled against the University of Rochester in its attempt to enforce its COX-2 patentagainst Searle.

17. More nuanced institutional arrangements are also possible. For copyrights, organisa-tions such as ESCAP that hold the copyrights of individual song writers and singers andcollect fixed royalty payments for their use on behalf of the artists, have worked well.

18. However, in other circumstances firms failed to reach a satisfactory agreement forpooling together the patents. This is the case, for instance, for DVDs, where agreementsto cross-license the rights or to pool them together have not been reached for quite a fewyears.

19. Needless to say, ours is a hindsight view, not a reconstruction of how university leaderssaw the situation at the time.

20. We should also note that there may be many benefits for university researchers who leavethe university to start ‘spin-off ’ firms. These benefits may take many forms, includingproviding the researchers with better information on financially (and economically)promising areas of research, and in providing teachers with better information on thetypes of skills and competencies students need. However, for the most part, universityspin-offs are celebrated as evidence of the university’s contribution to the national andregional economy, ignoring the potentially much greater contributions in terms of train-ing and other types of technology transfer, such as faculty consulting with industry.

21. In this context, one must note that American universities have historically been veryresponsive to industry needs. Collaborative research relationships between universityand industry in a broad range of fields have been a distinctive hallmark of the Americanuniversity system (Rosenberg and Nelson, 1994). Rosenberg (1992) in particular, hasconvincingly argued the critical role that American universities have played in support-ing innovation, often by helping in the solution of very practical, and sometimes, scien-tifically mundane problems.

22. Dasgupta and David (1987;1994) distinguish between what they call the realm ofScience and the realm of Technology, associating the first with open, university-typeresearch and the latter with research in firms. It is tempting to interpret this as implyingthat researchers in firms never participate in open research, or that university research isnever applied nor with immediate practical utility. This interpretation is incorrect.Rosenberg and Nelson (1994) have argued, for instance, that, in the USA at least, uni-versity researchers have also performed a variety of important applied activities, such assimple chemical assays, or development of instruments like instruments for ascertainingthe fat content of milk at the University of Wisconsin.

23. One reason is that the results of one project feed into the next. Full disclosure of theresults of one project would put all researchers on the same footing in terms of being thefirst to complete the next. By contrast, by only imperfectly disclosing the researchfinding, a researcher completing a stage ahead of others would get a head start on com-pleting the next stage as well.

24. See Gambardella (1995) for a discussion of this point in the context of the pharmaceu-tical industry.

REFERENCES

Alston, L., Eggertsson, T. and North, D. (1996) Empirical Studies in InstitutionalChange, Cambridge University Press, Cambridge.

Anton, James J. and Dennis A. Yao (1994) ‘Expropriation and inventions: appro-priable rents in the absence of property rights’, American Economic Review, 84,190–209.

Anton, James J. and Dennis A. Yao (2002) ‘The sale of ideas: disclosure, propertyrights, and incomplete contracts’, Review of Economic Studies, 69, 513–31.


Aoki, M., Murdock, K. and Okuno-Fujiwara, M. (1996) ‘Beyond the East Asianmiracle: introducing the market-enhancing view’, in Aoki, M., Kim, H.-K. andOkuno-Fujiwara, M. (eds), The Role of Government in East Asian EconomicDevelopment, Oxford University Press, Oxford.

Arora, A. (1995) ‘Licensing tacit knowledge: intellectual property right and themarket for know-how’, Economics of Innovation and New Technologyi, 4, 41–59.

Arora, A. and Fosfuri, A. (2003) ‘Licensing the market for technology’, Journal ofEconomic Behavior and Organisation, 52, 277–95.

Arora, A. and Gambardella, A. (1994) ‘The changing technology of technologicalchange: general and abstract knowledge and the division of innovative labour’,Research Policy, 23, 523–32.

Arora, A. and Gambardella, A. (1998) ‘Evolution of industry structure in thechemical industry’, in Arora, A., Landau, R. and Rosenberg, N. (eds), Chemicalsand Long-Term Economic Growth, John Wiley, New York.

Arora, A. and Merges, R. (2004) ‘Intellectual property rights, firm boundaries andR&D inputs’, Industrial and Corporate Change, 13 (3), 451–75.

Arora, A., Arunachalam, V., Asundi, J. and Fernandes, R. (2001a) ‘The Indian soft-ware service industry’, Research Policy, 30, 1267–87.

Arora, A., Fosfuri, A. and Gambardella, A. (2001b) ‘Specialised technology sup-pliers, international spillovers and investment: evidence from the chemical indus-try’, Journal of Development Economics, 65, 31–54.

Arora, A., Fosfuri, A. and Gambardella, A. (2001c) Markets for Technology: TheEconomics of Innovation and Corporate Strategy, MIT Press, Cambridge, MA.

Arora, A., Gambardella, A. and Torrisi, S. (2004) ‘In the footsteps of the SiliconValley? Indian and Irish software in the international division of labor’, inBresnahan, T. and Gambardella, A. (eds), Building High Tech Clusters: SiliconValley and Beyond, Cambridge University Press, Cambridge.

Bordo, M., Eichengreen, B. and Irwin, D. (1999) ‘Is globalization today reallydifferent from globalization a hundred years ago?’, NBER Working Paper 7195,NBER, Cambridge, MA.

Bresnahan, T. and Gambardella, A. (eds) (2004) Building High Tech Clusters:Silicon Valley and Beyond, Cambridge University Press, Cambridge.

British Technology Group (BTG) (1998) ‘IPR market benchmark study’,www.btgplc.com

Cockburn, I., Kortum, S. and Stern, S. (2002) ‘Are all patent examiners equal? Theimpact of examiner characteristics’ NBER Working Paper 8980, www.nber.org

Dasgupta, P. and David, P. (1987) ‘Information disclosure and the economics ofscience and technology’, in Feiwel, G. (ed.), Arrow and the Ascent of EconomicTheory, New York University Press, New York.

Dasgupta, P. and David, P. (1994) ‘Towards a new economics of science’, ResearchPolicy, 23, 487–521.

David, P. (1990) ‘The dynamo and the computer’, American Economic ReviewPapers and Proceedings, 80 (2), 355–61.

David, P. (1991) ‘Reputation and agency in the historical emergence of the institu-tions of “open science,” ’ CEPR Working Paper N.261, CEPR, StanfordUniversity, Stanford, CA.

David, P. (1993a) ‘Knowledge, property, and the system dynamics of technologicalchange’, Proceedings of the World Bank Annual Conference on DevelopmentEconomics 1992, World Bank, Washington, DC.


David, P. (1993b) ‘Intellectual property institutions and the panda’s thumb:patents, copyrights, and trade secrets in economic theory and history’, inWallerstein, M., Mogee, M. and Schoen, R. (eds), Global Dimensions ofIntellectual Property Rights in Science and Technology, National Academy Press,Washington, DC.

David, P. (1997) ‘From market magic to calypso policy: a review of Terence Kealey’sThe Economic Laws of Scientific Research’, Research Policy, 26, 229–55.

David, P. (1998) ‘Common agency contracting and the emergence of “open science”institutions’, American Economic Review, 88 (2), 15–21.

David, P. (2003) ‘The economic logic of “open science,” ’ SIEPR Working Paper02–30, SIEPR, Stanford University, Stanford, CA; forthcoming in The Role ofthe Public Domain in Scientific and Technical Data and Information, NationalAcademy Press, Washington, DC.

David, P. and Greenstein, S. (1990) ‘The economics of compatibility standards: anintroduction to recent research’, Economics of Innovation and New Technology, 1,3–41.

Deng, Z., Lev, B. and Narin, F. (2003) ‘Science and technology as predictors ofstock performance’, in Hand, J. and Lev, B. (eds), Intangible Assets: Values,Measures and Risks, Oxford University Press, Oxford.

European Technology Assessment Network (ETAN) (1999) ‘Strategic dimensionsof Intellectual property rights in the context of science and technology policy’,ETAN Expert Working Paper, European Commission, Directorate GeneralScience and Technology, Brussels, June.

Gambardella, A. (1995) Science and Innovation: The US Pharmaceutical Industry inthe 1980s, Cambridge University Press, Cambridge, MA.

Gilbert, R. and Shapiro, C. (1990) ‘Optimal patent length and breadth’, RandJournal of Economics, 21, 106–12.

Gompers, P. (1999) ‘Resource allocation, incentives and control: the importance ofventure capital in financing entrepreneurial firms’, in Acs, Z., Carlsson, B. andKarlsson, C. (eds), Entrepreneurship, Small and Medium Sized Enterprises and theMacroeconomy, Cambridge University Press, Cambridge.

Green, J. and Scotchmer, S. (1995) ‘On the Division of Profits in SequentialInnovations’, Rand Journal of Economics, 26 (1), 20–33.

Grindley, P.C. and Teece, D.J. (1997) ‘Licensing and cross-licensing in semiconduc-tors and electronics’, California Management Review, 39 (2), 8–41.

Grossman, S. and Hart, O. (1986) ‘The costs and benefits of ownership: a theory ofvertical and lateral integration’, Journal of Political Economy, 94, 691–719.

Hadley, W. (1998) ‘Rapporteur’s report’, conference on intellectual property rightsand competition policy, Stanford Law School, Stanford, CA, April.

Hall, B.H. and Ziedonis, R. (2001) ‘The patent paradox revisited: determinants ofpatenting in the US semiconductor industry, 1980–1994’, Rand Journal ofEconomics, 32 (1), 101–28.

Hand, J. and Lev, B. (eds) (2003) Intangible Assets: Values, Measures and Risks,Oxford University Press, Oxford.

Hart, O. and Moore, J. (1990) ‘Property rights and the nature of the firm’, Journalof Political Economy, 98, 1119–58.

Heller, M. (1998) ‘The tragedy of the anticommons: property in the transition frommarx to markets’, Harvard Law Review, 111, 621–88.

Heller, M. and Eisenberg, R. (1998) ‘Can patents deter innovation? The anticom-mons in biomedical research’, Science, 280, 698–701.


Henderson, R., Jaffe, A. and Trajtenberg, M. (1998) ‘Universities as a source ofcommercial technology: a detailed analysis of university patenting, 1965–1988’,Review of Economics and Statistics, 80, 119–27.

Kenney, M. (1986) Biotechnology: The University–Industrial Complex, YaleUniversity Press, New Haven, Ct.

Klemperer, P. (1990) ‘How broad should the scope of patent protection be?’, RandJournal of Economics, 21, 113–30.

Kremer, M. (1998) ‘Patent buyouts: a mechanism for encouraging innovation’,Quarterly Journal of Economics’, 113, 1137–67.

Kortum, S. and Lerner, J. (1999) ‘What is behind the recent surge in patenting’,Research Policy, 28, 1–22.

Lamoreaux, N. and Sokoloff, K. (1996) ‘Long term change in the organisation ofinventive activity’, Proceedings of the National Academy of Science USA, 93,12686–92.

Lamoreaux, N. and Sokoloff, K. (1999) ‘Inventors, firms, and the market for tech-nology: US manufacturing in the late nineteenth and early twentieth centuries’,in Lamoreaux, N., Raff, D. and Temin, P. (eds), Learning by Firms, Organisations,and Nations, University of Chicago Press, Chicago, IL.

Langlois, R. (2002) ‘Modularity in technology, organisation, and society’, Journalof Economic Behavior and Organisation, 49, 19–37.

Langlois, R. and Robertson, P. (1992) ‘Networks and innovation in a modularsystem: lessons from the microcomputer and stereo component industries’,Research Policy, 21, 297–313.

Lerner, J. (1995) ‘Patenting in the shadow of competitors’, Journal of Law andEconomics, 38, 563–95.

Lev, B. (2000) Intangibles: Management, Measurement and Reporting, BrookingsInstitutions, Washington, DC.

Linden, G. and Somaya, D. (2003) ‘System-on-a-chip integration in the semicon-ductor industry: industry structure and firm strategies’, Industrial and CorporateChange, 12, 545–76.

Merges, R. (1999) ‘As many as six impossible patents before breakfast: propertyrights for business concepts and patent system reform’, Berkeley Technology LawJournal, 14, 577–615.

Merges, R. and Nelson, R. (1990) ‘On the complex economics of patent scope’,Columbia Law Review, 90 (4), 839–916.

Merges, R. and Nelson, R. (1994) ‘On limiting or encouraging rivalry in technicalprogress: the effect of patent scope decisions’, Journal of Economic Behavior andOrganisation, 25, 1–24.

Mowery, D., Nelson, R., Sampat, B. and Ziedonis, A. (2001) ‘The growth of patent-ing and licensing by US universities: an assessment of the effects of the Bayh-Dole Act of 1980’, Research Policy, 30, 99–119.

Mowery, D. and Ziedonis, A. (2002) ‘Academic patent quality and quantitybefore and after the Bayh–Dole Act in the United States’, Research Policy, 31,399–418.

National Research Council (1997) Intellectual Property Rights and Research Toolsin Molecular Biology, National Academy of Sciences, Washington, DC.

North, D. (1990) Institutions, Institutional Change and Economic Performance,Cambridge University Press, Cambridge.

Rosenberg, N. (1992) ‘Scientific instrumentation and university research’, ResearchPolicy, 21, 381–90.


Rosenberg, N. and Birdzell, L. (1986) How the West Grew Rich, Basic Books,New York.

Rosenberg, N. and Nelson, R. (1994) ‘American university and technical advance inindustry’, Research Policy, 23, 323–48.

Scotchmer, S. (1991) ‘Standing on the shoulders of giants: cumulative research andthe patent law’, Journal of Economic Perspectives, 5 (1), 29–41.

Vincenti, W.G. (1990) What Engineers Know and How They Know It, Johns HopkinsUniversity Press, Baltimore, MD.

Walsh, J., Arora, A. and Cohen, W. (2003a) ‘Working through the patent problem’,Science, 299 (5609), 1021.

Walsh, J., Arora, A. and Cohen, W. (2003b) ‘Research tool licensing and patentingand biomedical innovation’, in Cohen, W. and Merril, S. (eds), Patents in theKnowledge Based Economy, NAS Press, Washington, DC.

Walsh, J.P. and Bayma, T. (1996) ‘Computer networks and scientific work’, SocialStudies of Science, 26, 661–703.


13. The key characteristics of sectoralknowledge bases: an internationalcomparison*

Stefano Brusoni and Aldo Geuna1

1 INTRODUCTION

This chapter builds upon and extends existing studies of scientific and tech-nological specialisation by proposing a unifying theoretical framework inwhich to compare sectoral knowledge bases across countries. In conductingthis comparison, we elaborate upon the large body of literature that analy-ses national systems of innovation (NSI) (Lundvall, 1992; Nelson, 1993).An NSI is defined as ‘being comprised of those elements of social organi-sation and behaviour, and the relationships among them, that are eitherlocated within or rooted inside the borders of a national state, and thatinteract in the production, diffusion and use of new, and economicallyuseful knowledge’ (David and Foray, 1995, p. 14). The concept of NSIgained wide popularity that goes beyond the boundaries of the academiccommunity as it became (often unwillingly) entangled with ‘techno-nation-alistic’ positions that have animated the industrial policy debate throughoutthe 1980s and 1990s. As stressed by David and Foray (1995), such positionsare based upon two related (and nowadays widely held) assumptions. First,technical capabilities lie at the core of a country’s international competi-tiveness. Second, the development of such capabilities is influenced by issuesof national localisation and can be managed via proper government action.

Recent research has challenged the relevance of the national dimension.In particular, it stresses that firms and researchers are entwined in thick net-works of international relationships that cut across national boundaries.National systems of innovation come under increasing strain, as theresearch and development (R&D) activities of large firms are progressivelyinternationalised. Such internationalisation is caused by emerging imbal-ances between what a country’s science base has to offer and the knowledgerequirements of innovative processes. However, despite their undeniable

361

increase, R&D linkages have not developed on a global scale, but ratherthey involve mainly US, European Union (EU) and, to a lesser extent,Japanese firms (Patel and Pavitt, 2000).

In this situation of internationalised rather than globalised R&D activ-ities, it is very important to understand why specific countries lie at the coreof such international networks. Standard explanations refer to a number offactors considered to be key determinants of ‘national competitiveness’(Porter, 1990). Following a well-established tradition (Fagerberg et al.,1999), this chapter acknowledges that a country’s specialisation pattern inspecific scientific and technological fields plays a key role: firms establishR&D facilities for which they perceive they have the relevant capabilities.However, most studies that empirically explore specialisation patterns atcountry level focus on a rather narrowly defined concept of specialisation.The emphasis falls squarely on the fields in which countries and/or firmspatent. Classic specialisation studies focus on the cumulative evolution ofcountries’ technological capabilities and, in most cases, scientific special-isation is not analysed. The stability of specialisation patterns over time(what we will term ‘knowledge persistence’) is well established; however,persistence and cumulativeness are not the only dimensions relevant to astudy of knowledge bases.

It is well known that design and development activities capture a relevantshare of the R&D funded by companies (Rosenberg, 1994). A country’sknowledge base may have a strong science base but lack the engineeringcapabilities to embody scientific results in profitable products. Or it canhave strong development capabilities that are not supported by robust basicscientific knowledge. Different typologies of knowledge are complemen-tary and interrelated. A strong presence in each typology of researchinduces an easier multi-directional flow of knowledge that can facilitatethe production of successful innovation. Micro-level innovation studiesstrongly support this view, for example, Pisano (1997). Therefore, whattype of research is carried out in each field (for example, basic versusengineering-oriented research) becomes a key issue.

The chief aim of this chapter is to develop a framework in which toanalyse knowledge specialisation both over time and across researchtypology. We put forward this framework as a way to approach questionsrelated to industry decisions to source knowledge internationally. In par-ticular, we want to link these decisions to specific characteristics of thesectoral knowledge base that is drawn on. The chapter identifies and oper-ationalises, at sectoral level, the relevant dimensions that make the com-parison of the knowledge bases of different countries a meaningfulexercise. Particular attention is devoted not only to examining whethereach country’s specialisation is stable over time (knowledge persistence),


but also to whether specialisation by field is similar across differenttypologies of research (knowledge integration).

The operationalisation of these two dimensions is based upon the designof a comprehensive data set of peer-reviewed papers that was obtained bycombining the standard Institute for Scientific Information (ISI) classifica-tion by science field with the Computer Horizons Inc. (CHI) classificationby type of research (that is, Applied Technology and Engineering, AppliedResearch and Basic Research). The result is an original data set encompass-ing some 630 000 papers in 11 different sub-fields of chemistry and pharma-cology published between 1989 and 1996. The limitations of peer-reviewedpublications as an indicator of the knowledge bases is discussed. This data-set will allow for a quantitative analysis of the characteristics and evolutionof the specialisation profile of the four largest European countries (the UK,Germany, France and Italy), the EU as a whole, the USA and Japan.

This data-set is analysed in combination with the Policies, Appropriationand Competitiveness in Europe (PACE) survey (Arundel et al., 1995). Theresults of the PACE questionnaire pinpoint the pharmaceutical industry asbeing a highly internationalised industry. The PACE survey data show thatnot only do EU R&D managers in the pharmaceutical sector value theresults of public research, but also that they rely upon internationalresearch much more than those in the chemical sector and in other manu-facturing industries. Also, PACE stresses that the pharmaceutical industryrelies more on North American research than on EU research. The ques-tions that demand explanation are why do EU pharmaceutical firms rely tosuch a great extent on North American research? What makes it attractiveto EU firms? In attempting to answer these questions, we discuss some evi-dence related to the existence of a ‘European paradox’ in the case of trad-itional pharmaceuticals. To do this, we compare sectoral knowledge basesacross countries by developing a grid designed along the two dimensionsidentified above: integration and persistence.

The chapter is organised as follows. Section 2 discusses the concept ofknowledge persistence and integration. Section 3 presents an empiricalexploration of the concepts developed in section 2 in the case of the phar-maceuticals and chemicals knowledge bases. Finally, section 4 offers con-cluding observations and raises a few policy issues.

2 TOWARDS A THEORETICAL FRAMEWORK OFKNOWLEDGE SPECIALISATION

Although the recent literature has devoted increasing attention to analysisof the economics of science and its implication for the innovation process

The key characteristics of sectoral knowledge bases 363

(Dasgupta and David, 1994; Mansfield, 1991; Narin et al., 1997), the analy-sis of national science and technology specialisation profiles has remained,so far, largely independent. Despite token acknowledgement of the com-plexity and intricacy of the relationships between the science and the tech-nology domains, specialisation studies tend to focus either on science or ontechnology. The former traditionally rely on bibliometric indicators; thelatter on patent studies. The former are dominated by sociologists ofscience; the latter by economists who study technical change. The rhetoricof the linear model still determines the intellectual division of labour in thisarea of research. This chapter represents a first step towards the redefin-ition of such a division of labour.

This is achieved by complementing the analysis of knowledge special-isation over time with an analysis of knowledge specialisation across typeof research. First, we briefly review a few classic specialisation studiesdeveloped in the historical, sociological and economic literature to stressthe cumulative and path-dependent process of knowledge production andaccumulation. The concept of knowledge persistence (that is, of special-isation over time) is based upon these notions. Second, micro-level analysisof technical change will inspire the introduction of the concept of know-ledge integration (that is, specialisation across type of research). In thisrespect, and somewhat paradoxically, this chapter is a first attempt todevelop the ‘macro-foundations of innovation management studies’. Wewill argue that the combined study of specialisation, in terms of both spe-cialisation over time and specialisation across types of research, enablesanalysts to start addressing issues such as why different industries utilisenational versus international sources of knowledge to different extents.

2.1 Specialisation Patterns over Time: Knowledge Persistence

Research in the history of science has stressed the cumulative and socialaspects of scientific endeavour. Historians have provided a number of accu-rate case histories that reveal how the accumulation of results over timeinfluences the rate and direction of the discovery process. For instance,Conant and Nash (1964) describe the process of accumulation of quantita-tive results in physics that led to Lavoisier’s revolution in modern chemistry.Such a process did not entail the substitution of inaccurate explanationswith more accurate ones; rather, it involved the re-conceptualisation ofexisting findings to deliver a new, more general, explanation. In addition, itis particularly interesting that scientific advancement is often focused on acommon frontier. The evidence for this is the incidence of multiple discov-eries that Merton characterised as endemic rather than isolated features ofscience (Merton, 1965). The cumulative development of science has also


been studied following the seminal work of Price (1963). Price sketches amacro ‘growth of knowledge’ approach that highlights the acceleration ofscientific publication that accompanied the growth of the scientific com-munity. This approach is probably more congenial to economists who canadvance a number of established theoretical propositions to explain Price’sempirical results. First, the increasing size of the scientific communitywould enable increasing division of labour and generate network externali-ties so that ‘increasing returns’ in scientific endeavour would be activated.Second, the growth of the scientific community stimulates the race for pri-ority in discovery. This would create a powerful incentive to publish moreprolifically in order to share some of the credit for ‘discovery.’ Scientificadvance would then occur in smaller steps with greater overlap and dupli-cation. Third, as the scientific community grows it becomes more difficult toassess individual contributions which in turn provides an incentive toproduce more publications in order to make claims about ‘productivity’. Asthe three mechanisms are not mutually exclusive, cumulativeness is the mostlikely outcome.2

On the basis of the above-mentioned literature, studies in the fields ofbibliometrics and the sociology of science have analysed the scientific baseof individual countries in terms of publications share (Braun et al., 1995).However, the analysis of absolute shares does not allow for meaningfulcross-country comparisons. Only recently has the methodology used toanalyse technological specialisation (based upon relative specialisationindicators) been applied to the publication output of countries in anattempt to develop a comparative analysis of scientific specialisation pat-terns (European Commission, 1997; Geuna, 2001; Godin, 1994; OST, 1998;Pianta and Archibugi, 1991).

The works of Soete (1981), Pavitt (1989) and Cantwell (1989) provide thebuilding blocks for the analysis of stability of technological specialisationpatterns at the country level. Following these studies a large body of liter-ature has been devoted to the study of technology and trade specialisation.The analysis of country-level technological specialisation patterns is nowa-days a methodology commonly used to study the relationship betweeninnovation and performance in terms of international trade and/or growth.In a nutshell, as technical change is a cumulative process that generates clus-tersof innovations, it isnot indifferent towhichtechnologicalareascountriesare specialised in (Meliciani, 2001). Different technical fields are charac-terised by different degrees of innovative opportunities and appropriabilityconditions (Carlsson, 1997; Malerba and Orsenigo, 1997). Furthermore, thelearning processes that underpin technical change tend to be localised andcumulative (Pavitt, 1992): it is easier to learn in the proximity of what onealreadyknows, sotospeak.Therefore, if one is specialised inthe ‘wrong’(that


is, low opportunity) technical or scientific fields, one should not expect to beable to refocus one’s own specialisation pattern in the short term. Trade andgrowth indicators will reflect such ‘bad’ specialisation. Scholars of technicalchange have therefore devoted much effort to matching technological spe-cialisation indicators and countries’ growth indicators (Fagerberg et al.,1999). Although there is some consensus about the importance of the knowl-edge base (or science base) of a country in the process of economic growth,the empirical and theoretical analyses have focused almost entirely on tech-nology (especially patents) and, generally, do not attempt to provide mea-surement of the scientific base of the country. The work of Archibugi andPianta in the early 1990s (see, for example, Archibugi and Pianta, 1992) is arare example of the combination of patent studies and bibliometric analysisto examine national specialisation in EU countries. Expanding upon thebodies of literature discussed above, we define knowledge persistence as thestability of the knowledge specialisation pattern over time.

2.2 Specialisation across Research Typologies: Knowledge Integration

Persistence and cumulativeness are not the only dimensions relevant to astudy of the knowledge bases of firms or countries. Micro-level studies oftechnical change have highlighted how the integration of different typesof research plays a crucial role in the process of innovation. Integrationissues have been studied at length in the innovation management liter-ature. Pavitt (1998) stresses that the key role played by modern firms is tomap an increasing range of relevant disciplines into products. Integrationefforts at firm level have been thoroughly discussed by a number ofauthors. Granstrand et al. (1997) studied the distributed capabilities thatenable firms to monitor and integrate technologies. Iansiti (1998) analysedintegration issues in the mainframe industry. Prencipe (1997) studiedsimilar problems in the aero-engine industry. Engineering disciplines arecommonly stressed as being powerful, although often overlooked,enablers of such integration. They provide the problem-solving techniquesto handle complex problems by decomposing them into simpler sub-tasks,which can be solved and then integrated back into a consistent whole. Forinstance, Patel and Pavitt (1994) studied the pervasiveness of mechanicalengineering skills across a variety of sectors. Landau and Rosenberg(1992) analysed chemical engineering as the key engine of growth in themodern chemical industry. Vincenti (1990) stressed the key role played byengineers and engineering sciences in solving the problems and finding theexplanations that led to the birth of the aircraft industry. Pisano (1996;1997) studied in detail a sample of pharmaceutical development projectsin order to conclude that success is related to the capability to carry out,


in a co-ordinated and timely manner, a number of activities that go wellbeyond the traditional boundaries of the R&D laboratory. The develop-ment of economically viable routes to produce drugs on an industrialscale is fraught with complex engineering issues, particularly where newuntested routes are being explored.

Nevertheless, more aggregated studies continue to have a focus on indi-cators that do not make it possible to analyse whether the country possessesa strong knowledge base that spans basic, applied and developmentresearch activities in any specific sector. A country’s sectoral knowledgebase may have a strong science base, but lack the engineering capabilities toembody scientific results within profitable products, or strong developmentcapabilities, but a not sufficiently robust base of scientific knowledge. Ineither case, firms may need to access those capabilities that are lacking,from where they exist, for example, from another country. This view is notbased on a simple linear model that sees basic research as the source of thewhole knowledge that is then transformed into technology. On the con-trary, what we want to stress here is that the various typologies of know-ledge are complementary and interrelated. A strong base in each typologyof research induces an easier multi-directional flow of knowledge that canfacilitate the production of successful innovation.

Such intuition is consistent with the theoretical framework developed byDavid and Foray (1995: 40), when they argue that ‘an efficient system ofdistribution and access to knowledge is a sine qua non condition forincreasing the amount of innovative opportunities’. Consistent with theresults of micro-level studies of technical change, we argue that the suc-cessful exploitation of such combinations requires the existence of cap-abilities spanning a range of disciplines that go beyond the traditionalboundaries of scientific endeavour. Knowledge bases that are too narrowlyfocused around core scientific disciplines (with no competencies in therelated, but different, engineering sciences) may fail to close the feedbackloop between the science and the technology domains. Such failure wouldseriously hamper the ‘ “distribution power” of the system’ (David andForay, 1995: 46). In other words, in order to close the feedback loopbetween the science and the technology domains, countries (as well asfirms) need to maintain distributed (rather than narrowly focused) compe-tencies at sectoral level. As a national bias seems to exist in terms of theeffectiveness of the linkages between business practitioners and academicresearch (Arundel and Geuna, 2004; Malo and Geuna, 2000; Narin et al.,1997), it is likely that such a bias exists also with respect to the linkagesbetween the scientific and engineering communities. Thus, particular atten-tion should be devoted not only to examining whether each country’s sec-toral specialisation is stable over time (knowledge persistence), but also to


whether each country’s sectoral specialisation cuts across different types ofresearch (knowledge integration). A sectoral knowledge base with highknowledge integration would have similar specialisation by field acrossdifferent typologies of research.

To conclude, what we propose is a simple analytical framework capableof combining the analysis of specialisation profiles over time with the spe-cialisation across type of research. This framework, built upon the notionsof knowledge persistence and knowledge integration, should shed light onthe ‘morphological’ characteristics of different countries’ knowledge basesin certain sectors, and thus help to explain firms’ international knowledgesourcing decisions. Figure 13.1 summarises the above discussion. In whatfollows, we will argue that this typology can be usefully deployed to studya number of issues related to the characteristics and evolution of countries’sectoral specialisation profiles, as well as firms’ decisions about where tosource useful knowledge and capabilities.

With respect to any specific sector, a country can be positioned in one ofthe four quadrants of the matrix of knowledge specialisation (Figure 13.1).Country A, in the top right-hand quadrant is characterised by a persistentpattern of scientific specialisation and high level of knowledge integration.In the fields where it is positively specialised, Country A has developedcapabilities in basic, applied and engineering research. Country B (top left-hand quadrant) would be persistently specialised in one or more fields, butits capabilities would be focused on, say, basic research only. Country C(bottom right-hand quadrant) would be characterised by integrated,although somewhat erratic, scientific and technological skills. Finally,Country D would be both erratic and unfocused in terms of research types:the fields of positive specialisation would change frequently and would bedifferent in different types of research.


Knowledge integration

Low High

Knowledge

persistence

High Country B Country A

Low Country D Country C

Figure 13.1 Matrix of knowledge specialisation

3 AN EMPIRICAL EXPLORATION OFKNOWLEDGE PERSISTENCE AND KNOWLEDGEINTEGRATION

In this section, we argue that by introducing the notion of knowledge inte-gration alongside the more traditional one of persistence we can actuallyquantitatively explore problems that are only partially answered by moretraditional specialisation studies. Specifically, in this chapter we focus ontwo main issues. First, we are interested in understanding the differencesbetween patterns of internationalisation of knowledge sourcing activitiespursued by different industries. Or, in other words, why firms in differentindustries appear to rely more on foreign innovation systems. Traditionalexplanations of this type of behaviour stress that firms go abroad wheneverthey (think they) can access ‘better’ capabilities relevant to their innovativeand manufacturing efforts (Cantwell, 1995). Implicitly, these explanationsassume that firms go abroad when their home knowledge base is not spe-cialised in the ‘right’ fields. Second, the notion that a European paradoxexists has gained wide support in the public policy arena (EuropeanCommission, 1996). According to this position, in some sectors EU firmswould be very good at developing new ideas, but would tend to fail toexploit them commercially. Something would be ‘missing’ from the EUsystem of innovation (or its national components) that would leave EUfirms at a disadvantage to their US counterparts. While the anecdotal evi-dence is abundant, rigorous empirical studies to prove (or disprove) theexistence of such a problem are scant. Tijssen and van Wijk (1999) provideone of the few systematic efforts to solve this difficulty using robust empir-ical data in the specific case of the ICT sector.

To contribute to this debate, we operationalise our framework in the caseof the international pharmaceutical industry, using the chemical industryas a yardstick. The pharmaceutical industry is an interesting case for ourpurposes for a number of reasons. First, it relies heavily on basic, highlycodified research at the forefront of human knowledge; thus, the scientificand technological knowledge base contributes to the development of thisindustry in a crucial way. Second, the pharmaceutical industry appears tobe one of the most internationalised manufacturing sectors, not only interms of product markets, but also, specifically, in terms of the knowledgesourcing strategies pursued by the major players (Patel and Pavitt, 2000).Third, and consistent with the previous point, the results of the PACEsurvey (Arundel et al., 1995)3 show that ‘general and specialised know-ledge’ produced by public research institutes is particularly valuable topharmaceutical firms (much more than to other manufacturing sectors),and that these firms consider scientific publications to be the key channel


to internationally access this knowledge. Thus, publications can be used asa proxy for the measurement of the characteristics of persistence and inte-gration of the knowledge base.4

Chemicals are used as a benchmark for establishing the divergence ofpharmaceuticals. This benchmark is appropriate because the chemicalindustry, overall, behaves similarly to the other industrial sectors (see also,Geuna, 2001). If all sectors are similar, why choose the chemical industryas a benchmark, and not some other? The principal reason is that we wishto make inter-country and inter-sectoral comparisons simultaneously.

Countries differ in the nature and extent of their development of specificindustries. Since these differences are very difficult to capture, it is useful tochoose industries that share a common knowledge base as the point of ref-erence, but that rely on knowledge generated outside their home countriesto different extents. Therefore, differences in sectoral behaviour may berelated to the country-specific characteristics of the foreign NSI. Needlessto say, the knowledge bases of the chemical and pharmaceutical industriesdiffer greatly. The key difference is the increasing reliance of pharmaceuti-cals on biology and biotechnology, rather than chemistry (Gambardella,1995; Orsenigo, 1989; Pisano, 1997). The chemical sector seems not to haveseriously explored the potential of biotechnologies, although recent devel-opments in combinatorial chemistry and biology provide evidence of thepossibility for convergence (Malo and Geuna, 2000). By leaving aside thebiotechnological knowledge base, the portion of the pharmaceuticalsknowledge base that relies on the more traditional chemical processes canbe analysed. This knowledge base is fairly similar to the knowledge baserelied on by the chemical industry.

3.1 International Knowledge Sourcing Activities and the EuropeanParadox

The results of the PACE survey reveal the sources of the public researchactivities most useful to EU R&D managers and the specific channels usedto find out about such research activities. Studying the frequency withwhich EU firms source knowledge from different regions (with respect toeach method for learning about public research) reveals the geographicorigin of the most useful research activities (Table 13.1).

In the chemical industry, respondents obtained the results of researchconducted by public research institutes or universities from the whole rangeof sources at levels similar to those for all the other industrial sectors.However, a few particularities are worth mentioning. Chemical firmsobtain information from conferences in their own country and those inother European countries with the same frequency (about 88 per cent of


371

Tab

le 1

3.1

Sou

rces

of

usef

ul p

ublic

res

earc

h ac

tivi

ties

by

geog

raph

ic a

rea

and

sect

or:

perc

enta

ge o

fre

spon

dent

s th

atre

port

ed o

btai

ning

res

ults

fro

m t

he li

sted

sou

rces

Sect

orA

ll in

dust

rial

Che

mic

als

Pha

rmac

euti

cal

Met

hod

ofkn

owle

dge

Kno

wle

dge

Hom

eE

UN

AJa

pH

ome

EU

NA

Jap

Hom

eE

UN

AJa

pac

quis

itio

nso

urce

Pub

licat

ions

and

tec

hnic

al r

epor

ts91

.70

82.7

077

.50

54.8

090

.60

82.3

084

.40

61.5

095

.10

92.7

095

.10

68.3

0C

onfe

renc

es a

nd m

eeti

ngs

87.0

081

.00

66.2

036

.70

87.5

087

.50

67.7

040

.60

87.8

092

.70

90.2

053

.70

Info

rmal

per

sona

l con

tact

s89

.10

71.9

059

.40

31.9

089

.60

68.8

066

.70

36.5

092

.70

87.8

090

.20

58.5

0H

irin

g tr

aine

d sc

ient

ists

and

71

.00

32.7

019

.80

4.40

75.0

034

.40

33.3

010

.40

85.4

061

.00

46.3

019

.50

engi

neer

sT

empo

rary

exc

hang

es o

fpe

rson

nel

34.0

016

.50

11.1

04.

8028

.10

15.6

014

.60

4.20

48.8

031

.70

24.4

04.

90C

ontr

act

rese

arch

don

e by

the

76

.00

33.2

020

.30

4.00

79.2

039

.60

30.2

08.

3082

.90

58.5

056

.10

17.1

0in

stit

ute

or u

nive

rsit

yJo

int

rese

arch

pro

ject

s73

.20

43.8

019

.70

5.10

72.9

042

.70

28.1

04.

2080

.50

48.8

048

.80

14.6

0A

ffilia

ted

firm

s45

.70

36.5

033

.80

11.1

039

.60

45.8

050

.00

24.0

048

.80

51.2

056

.10

26.8

0Jo

int

vent

ures

29.7

024

.90

16.3

012

.20

28.1

028

.10

20.8

024

.00

24.4

017

.10

24.4

022

.00

Sou

rce:

Aut

hors

’ela

bora

tion

of

PAC

E d

ata.

respondents). In the case of publications, informal contacts and hiring,respondents from the chemical industry attribute equal weights toother European countries and North America (exhibiting a lower EU-localisation effect than all industrial sectors combined).

In the pharmaceutical industry, the home country localisation effecttends to vanish. About 90 per cent of respondents obtained informationthrough publications, informal contacts and conferences within their owncountry, other European countries and North America with only a smallpercentage receiving information from Japan. In seeking to source publicresearch results EU R&D managers approach the North American sciencesystem, the EU and domestic sources with similar frequency. In particular,North American papers are used with the same frequency as home-countrypublications (95 per cent) and more frequently than papers from other EUcountries (92 per cent).

The behaviour of the pharmaceutical industry is of particular import-ance, if only because it is widely considered as one of the main areas ofstrength for the EU (Sharp et al., 1997). However, despite past successfulperformance, the EU industry seems to be rather pessimistic about thefuture (Sharp et al., 1997). Rising levels of R&D, decreasing profitmargins and the struggle to refocus research efforts toward biotechnol-ogies have been undermining the competitive position of the industry.United States pharmaceutical companies are often considered to be wayahead of their EU competitors, particularly with respect to the adoptionof biotechnologies.

It is often argued that the comparative success of the US industry isrelated to its capability to effectively transform the results of basic researchinto blockbuster drugs, rather than its ability to generate such results per se.In this respect, a European paradox is commonly evoked to emphasise thegap between the seemingly good performance of EU basic research and therelatively bad commercial performance of EU firms (when compared withtheir US competitors): European firms are not particularly good at trans-forming brilliant ideas into successful products. While most observers nowagree on this last point, explanations of the reasons for this are in shortsupply. Generally speaking, the firms themselves are often blamed for notapplying sensible management practices that would enable them to fullyexploit the wealth of insights provided by the EU system of innovation andits national components. British firms are too short-sighted; German firmsare too slow in making decisions and pursuing new research routes; Frenchfirms are sheltered from the pressures of the global economy by a compla-cent state-managed health system, and so on.

While appealing, such propositions fail to consider the complementarypossibility that there is something systematically different between the EU


and the US knowledge bases that enables US firms to be more competitiveand induces EU firms to look to the US scientific knowledge base to sourcenew knowledge. In what follows, we operationalise our framework to inter-pret the results of the PACE survey and assess the anecdotal evidence ofthe European paradox.

3.2 Mapping and Measuring Countries’ Sectoral Knowledge Bases

We examined the publication profiles of different countries in the fields ofchemistry and pharmacology. Following Geuna (2001) we used the ScienceCitation Index (SCI) database of the Institute for Scientific Information toanalyse the publication output of the four largest European countries, theEU, Japan and the USA in the period 1989–96 (see Geuna, 2001, for adescription of the data). Eleven scientific fields relevant to the chemical andpharmaceutical industries are identified.5 Each publication in these fields isclassified in a typology of research using the CHI journal classification:Applied technology, Engineering and technological sciences, Appliedresearch, and Basic research. Godin (1994), who studied a sample of largeinnovating firms in order to analyse the complementarities between scienceand technology, proposed a similar approach. He developed a database ofpublications that were then divided into four groups in a spectrum thatvaried from very applied to basic (‘untargeted’) research. Unlike thischapter, his work focused on firm-level activities, rather than sectors.

To develop a comparative analysis of the knowledge base in chemistryand pharmacology, the relative specialisation of a country was studied.The symmetric Relative Specialisation Index (RSI) (see Appendix I formethodological issues concerning the RSI) is calculated on the basis ofdata from the SCI database for six countries and the EU, 11 scientificfields and three research areas between 1989 and 1996 (Balassa, 1965;Soete, 1981). The statistical results are used to operationalise the theoret-ical framework of knowledge specialisation for the pharmaceutical andchemical industries.

3.2.1 Knowledge persistence (stability of specialisation patternsover time)

In the eight-year period under examination the specialisation of the EUand the six countries considered has changed, in some cases quite substan-tially. To verify the stability (or lack thereof) of overall specialisation pat-terns we examined how all 11 specialisation indices had changed over time.Following the work of Pavitt (1989), we calculated the Pearson correlationcoefficient for each country at the start and at the end of the period con-sidered. Positive and significant coefficients would hint at the cumulative


and path-dependent nature of knowledge accumulation processes. We dis-covered that the knowledge specialisation in France, Germany, the UK,Japan and the USA is positively correlated in the two periods, while no sig-nificant correlation was found for Italy and the EU.

Furthermore, in order to analyse the path of specialisation or de-specialisation of a country, we regressed the symmetric specialisation indexin 1996 on the 1989 value, country by country. Such a methodology wasoriginally proposed by Cantwell (1989) and consists of a simple country-by-country regression at two different points in time. The dynamic path,therefore, cannot be studied. Also, nothing can be said about the determin-ants of the initial pattern of specialisation. Despite these limitations, thismethodology has been widely used in specialisation studies. Its mainadvantage is its simplicity.

If the � coefficient is equal to 1, then the country specialisation patternhas remained unchanged over the period. If ��1 then the country isincreasing its positive specialisation in fields where it was already spe-cialised. If 0��1, the country has decreased its non-specialisation inthose fields where it was negatively specialised at the beginning of theperiod (or decreased its positive specialisation where it was positively spe-cialised). In all cases, variations in specialisation occur in a cumulative way,as ��0. In the case that � is not significantly different from zero, thehypothesis that changes in specialisation are either not cumulative or arerandom cannot be excluded. If � is negative we are witnessing a process ofreversion in the specialisation. The case where � �1 is often referred to as�-specialisation (Dalum et al., 1998).

Cantwell (1989: 31–2) argues that ��1 is not a necessary condition forincreasing specialisation. Therefore, we have also analysed the so-calleds-specialisation (Dalum et al., 1998). The dispersion of a given distri-bution does not change if �R; if ��R the specialisation increases(s-specialisation) and if ��R the specialisation decreases (s-specialisation).

For each country we ran regressions on all fields, basic research only,applied research only, and applied technology and engineering researchonly. For the most general regressions, we found that Germany was thecountry with the most stable specialisation pattern (� .96, R .724).Italy and EU do not have significant coefficients. All the other countriesde-specialised cumulatively (that is, 0��1) in terms of both � ands-specialisation, with the UK (� .843, R .897) being the most cumula-tive, followed by the USA (� .779, R852), Japan (� .659, R926)and France (� .42, R799). All coefficients are significant at the 1 percent level (2 per cent for Germany). In terms of basic research, only theUSA, Japan and the UK have � coefficients with a significance levelhigher than 5 per cent – respectively 1.04 (1 per cent), 0.95 (2 per cent) and


0.44 (4 per cent). The USA, with both ��1 and �/R�1, increased spe-cialisation in sectors where it was already specialised, and became lessspecialised where initially specialisation was low. Japan, with both �1and � /R1, showed a high stability in its specialisation patterns. In par-ticular, the USA deepened its specialisation in fields related to the phar-maceutical industry: medical chemistry (C6) and pharmacology (C10).The four largest European countries saw an increase in the dispersion oftheir basic research specialisation. European Union countries, especiallyGermany and France, show a tendency to remain more focused on tradi-tional chemistry fields.

3.2.2 Knowledge integration (specialisation across research typology)Table 13.2 presents a summary of the relative specialisation of the EU andthe six countries under consideration over the entire eight-year period. Itlists, by type of research, the chemical fields in which each country exhib-ited positive specialisation (top two outside square brackets). The firstobservation that can be made is that there is some degree of overlapbetween the positive specialisation in applied research and that in basicresearch, while the area of applied technology and engineering tends todiffer from the other two areas of research. If knowledge integration isdefined as the presence of positive specialisation in the same scientific fieldsin the three typologies of research, it can be stated that the USA has a much


Table 13.2 Fields of positive specialisation by type of research

Applied technology Applied research Basic researchand engineering

EU15 C10 C3 [C8] C4 C7 [C10 C8 C2 C1] C2 C5[C4 C8 C6 C7]France C8 C4 [C6 C10] C7 C4[C8 C6 C9] C7 C4 [C8]Germany C3 C4 [C11] C1 C5[C4 C9 C8] C5 C4 [C8]Italy C10 C6 [C8] C7 C6 [C10] C2 C6 [ C7 C5 C4]UK C3 C10 C4 C7[C10 C3 C6] C6 C4[C5 C2]USA C1 C2 [C9 C6 C10] C6 C10 [C3 C1] C6 C10 [C2 C8 C3 C7]Japan C4 C11 C10 C9 C1 C10 [C7]

Notes:Top two positive specialisation fields out of brackets.C1 General chemistry, C2 Analytical chemistry, C3 Applied chemistry, C4 Crystallography,C5 Inorganic and nuclear chemistry, C6 Medical chemistry, C7 Organic chemistry,C8 Physical chemistry, C9 Polymer science, C10 Pharmacology and pharmacy,C11 Chemical engineering.

Source: Authors’ elaboration of ISI and CHI data.

higher degree of knowledge integration than the EU. Indeed, the USA hasa positive specialisation in medical chemistry (C6) and pharmacology andpharmacy (C10) in all research typologies. Among the four largest EUcountries, France has positive specialisation in crystallography (C4),organic chemistry (C7) and physical chemistry (C8) in all three researchtypologies. Similarly, Germany is positively specialised in all typologies incrystallography (C4) and inorganic chemistry (C5). Finally, Italy is consist-ently positively specialised in medical chemistry (C6) and organic chem-istry (C7).6

A simple indicator of integration is calculated by dividing the fields inwhich a country is positively specialised in all research typologies by thetotal number of fields in which a country is positively specialised. So, foreach country we have:

This indicator varies from 0 to 1. It is 0 when the country considered doesnot exhibit any overlap between the three types of research. It is 1 whenthe country considered is fully integrated across all types of research in allthe fields in which it exhibits positive specialisation. The USA is positivelyspecialised in medical chemistry and pharmacology and pharmacy in alltypologies of research, and positively specialised in a total of eight fields.Its indicator of integration is 2/80.25. France, Italy and Germany are lessdispersed – that is, more integrated – than the USA. France scores 3/60.5(where the fields of full integration are crystallography, physical chemistryand inorganic chemistry out of a total of six fields of positive specialisa-tion). Italy and Germany are integrated in two fields (medical chemistryand organic chemistry versus crystallography and inorganic chemistry) outof a total of seven fields of positive specialisation, giving an integrationcoefficient of 2/70.29. Japan and the UK are not integrated at all and donot exhibit any overlap across the different typologies of research.

3.2.3 A taxonomy of knowledge specialisationBy combining the results on persistence and integration, it is possible to mapthe science and engineering bases of different countries in a two-dimen-sional space that summarises the results sketched above. We have mappedthe indicator of integration on the horizontal axis and the indicator of per-sistence on the vertical axis. For persistence, we have used the results of theregressions for all chemistry and pharmacology fields, first and last year,country by country. We have also used the coefficients for Italy and the EUalthough, as stated above, they are not significant. As all coefficients except

INT � no. of fields of positive specialisation in all types of research

� no. of fields of positive specialisation


those for Germany (whose � equals 1) are 0��1 (�-de-specialisation), weset 0.5 as the threshold. For integration, we have used the simple indicatorsketched above. The threshold between high and low integration is given bythe arithmetic average of the indicator (0.22). Figure 13.2 reports the resultof such a combination.

It is fairly apparent that the USA and Germany combine high levels ofboth integration and persistence. France, despite a high level of integra-tion, exhibits low persistence over time. Neither Japan nor the UK showsany integration, but the pattern of specialisation in the UK is more stable.Italy and the EU are somewhere in between. The EU as a whole is charac-terised by both average integration and low persistence (this lattercoefficient was not significant in the regression). Italy appears to be rela-tively integrated, but exhibits low persistence (Italy’s coefficient for persist-ence is not significant).

It is worth combining the results of this taxonomy with the analysis ofthe specific fields of specialisation listed in Table 13.2. Despite the highpersistence and integration exhibited by both the USA and Germany, theirspecialisation profiles appear to be very different. In particular, Germany’sspecialisation revolves around traditional chemistry fields, such ascrystallography (C4) and inorganic chemistry (C5). The USA is specialisedin those fields more directly related to pharmaceuticals: medical chemistry(C6) and pharmacology and pharmacy (C10). The other EU countriesstudied also are more specialised in ‘chemistry for chemicals’, rather thanpharmaceuticals. Furthermore, it is evident from the regressions we ran by


Knowledge integration

Knowledge

persistence

* France (0.5)

* Italy (0.29)

* EU15 (0.22)

*USA (0.25)

*Germany (0.29)

Below .5

Above .5

* Japan (0.00)

* UK (0.00)

Below average (< .22) Above average (> .22)

Figure 13.2 Integration and persistence: a rough-and-ready map

type of research that the EU countries’ specialisation in medical chemistryand pharmacology decreases as we move away from development typeresearch towards applied and then basic research.

Such results are consistent with other studies of specialisation that rely ontraditional methodologies. So, for Germany, specialisation in ‘traditional’chemistry (that is, inorganic and organic) is confirmed by Sternberg (2000,p. 98) who also highlights the German disadvantage in medical sciences.The Office of Science and Technology (OST, 1998) confirms both the inte-gration of the German pattern of specialisation and its focus on chemistry.Furthermore, the UK seems to be more specialised in medical research thanFrance and Germany. The OST (1998) also confirms the strong EU special-isation in chemistry and its relative disadvantage (in terms of publications)in biology (basic research).

These different specialisation profiles hint at a possible explanation forthe results of the PACE questionnaire. The PACE survey revealed thatpublic research carried out in North America was valued and used exten-sively (even more than public research carried out in other European coun-tries) by the largest EU R&D firms in the pharmaceutical sector. The PACEquestionnaire does not allow speculation about why this happens, though.We argue that the reliance of EU firms on the North American knowledgebase is consistent with the fact that the USA exhibits a persistent as well asan integrated specialisation pattern in medical chemistry and pharmacyand pharmacology. The results for the chemical industry confirm this.European Union chemical firms do not use US-generated research to thesame extent as pharmaceutical firms. Their home-country knowledge baseis relatively more specialised in a persistent and integrated manner in thosefields that are particularly relevant to the innovative efforts of the chemicalindustry. Thus, they rely heavily on the public research of their own countryor other European countries.

Particular attention should be devoted to specialisation by type ofresearch in EU countries. It was noted above that they are positively spe-cialised in either medical chemistry or pharmacology at the level of appliedand development research. However, those two fields do not show up asareas of positive specialisation in basic research (Table 13.2). Also, theresults of the regression by type of research clearly show that only theUSA and Japan are increasing their specialisation in basic research. Noclear pattern is discernible for EU countries except for the UK, which is�-despecialising. Therefore, these data do not allow us to talk about a‘European paradox’, according to which EU firms would not be capable ofexploiting an efficient basic research system because of lack of ‘develop-ment’ capabilities. Our data seem to point to the fact that these types ofcapabilities do exist. What is missing is the basic research bit, with the result


that EU pharmaceutical firms have to source research results from theUSA. The pattern of sourcing is consistently different when chemical firmsare considered, as their home-country knowledge bases seem more capableof providing basic research capabilities.

Despite the limitations of the data and the simplicity of this analysis, thelocation of different countries along the grid defined by the measures ofpersistence and integration matches with a few things we know about theinstitutional structure of each country, and also raises some interestingquestions. For instance, the results concerning the 15 countries of the EUas a whole are hardly surprising. An EU-wide system of innovation is stillin the process of formation. National industry and science and technology(S&T) policies still heavily influence country-level specialisation patterns,preventing them from converging toward a homogeneous whole.

4 CONCLUSIONS

The evolution of country-level sectoral specialisation has been conceptu-alised by the discussion on knowledge persistence and knowledge integra-tion. Persistence is related to the evolution of specialisation over time. Ithints at the cumulative, path-dependent nature of the learning processes.Integration is related to the evolution of specialisation across differenttypologies of research. It suggests the complex, non-linear interdepend-encies that link the scientific and technological domains. The interaction ofthe concepts of knowledge persistence and knowledge integration providea head start for the development of a robust conceptual framework inwhich to compare countries’ sectoral knowledge bases. It is quite signifi-cant that the conceptualisation proposed in terms of persistence is consis-tent with the results of micro-studies of technical change that pinpointlearning processes as cumulative and path dependent (David, 1985). Also,innovation studies hint at the key role played by distributed (rather thannarrowly focused) capabilities in enabling technical change (Granstrandet al., 1997). This is captured by the concept of knowledge integration.

This chapter represents a first attempt to operationalise this frameworkon the basis of a statistical analysis of a huge, original and custom-builtdata-set that describes the scientific and engineering knowledge base inchemistry and pharmacology in the four largest European countries, theEU as a whole, the USA and Japan during the period 1989 to 1996.Analysis of the relationships between core positive and negative special-isation, and of the typology of research (Applied technology and engi-neering, Applied research and Basic research) has shown that the countriesconsidered have different degrees of knowledge integration and knowledge


persistence. Specifically, the USA and Germany exhibit the highestcoefficients of persistence and integration. However, the USA is moreheavily specialised in fields related to pharmaceuticals (that is, medicalchemistry and pharmacology and pharmacy) than Germany and the otherEU countries, which appear to be more specialised around traditionalchemistry. These results are consistent with the views expressed by the EUR&D managers that responded to the PACE questionnaire. They stressedthat public research developed in North America was particularly useful totheir innovative efforts in pharmaceuticals. In contrast, domestic and EUlocalisation effects prevail in the case of the chemical industry.

As for the policy implications, the empirical results presented (althoughpreliminary) allow us to make two main observations. First, our data-setdoes not identify any ‘European paradox’ in pharmacology. EuropeanUnion countries exhibit capabilities in terms of applied and engineeringresearch, but not in basic research. Instead, the USA only increases its�-specialisation in basic research in pharmacology and medical chemistry.No clear pattern is discernible for EU countries, with the exception of theUK, which is �-despecialising. Such lack of basic research capabilities maywell explain the frequency with which EU R&D managers in the pharma-ceutical industry approach the US knowledge base. For chemicals, thepattern of sourcing is different. As their home-country knowledge basesseem more capable of providing a more integrated pattern of research capa-bilities, EU chemical firms rely chiefly on their home-country knowledgebase and then approach that of the EU. At least for pharmacology andmedical chemistry we found no evidence of paradoxes.

Second, our approach hints at the possibility that government can actu-ally influence the rate of technical change by fostering the development ofan ‘integrated’ specialisation profile. Empirically, one can identify the NSIthat firms consider to be more helpful to their innovative activities (forexample, the USA for pharmaceuticals), analyse it in terms of integrationand then target the type of research that is lacking in the home country. Wemay call this the ‘policy for integration’ option. In fact, despite the enor-mous resources devoted by policy-makers to the exploration of emergingtechnologies, ‘picking a winner’ remains a rather hazardous activity. Thegreatest successes of recent years are the unintended consequences of poli-cies aimed at fostering other paths of research – for example, biotechnologybeing the unintended offspring of US cancer research programmes and thebeneficiary of military research for the bio-war (Martin et al., 1990). Whichspecific scientific field will be responsible for the next revolution continuesto be difficult to predict. We argue that our approach would not allow gov-ernments to pick the winners, but would allow them to support the devel-opment of an integrated knowledge base once a new path has emerged.


The limitations of this work open up a number of challenging questionsfor future research. First, publications are a very good way to trace the sci-entific knowledge base of a country, but are less successful as far as engi-neering research is concerned. Merging traditional data-sets of patentactivities with our data-set of publications would provide a better picture ofthe interactionbetweenscientificandengineeringspecialisation.Second, theconcepts of knowledge integration and persistence can also be applied to thestudy of firms’ knowledge bases to further confirm the consistency betweenmicro- and macro-level dynamics. It is important to expand such analysis totest the existence of a correlation or causation between knowledge integra-tion and knowledge persistence in certain fields on the one side, and techno-logical and economic performances of firms and countries, on the other.7

The qualitative indications provided by the PACE questionnaire are but afirst step. Third, this analysis should be extended to a sample of ‘small coun-tries’. These may be much less integrated and persistent than large countriesas they may find it more convenient (or just more feasible) to exploit theadvantages of flexibility by specialising narrowly in terms of fields and/ortypes of research and then switching when new research trajectories emerge.Finally, on a more theoretical note, PACE reveals that firms can sourceknowledge not available in their home country by looking abroad. However,there are costs attached to such a choice. Traditionally, costs are related tothe geographic distance between source and user. This chapter hints at thepossibility that there might be costs attached also to the relative position inthe ‘knowledge spectrum’, so that the farther from a typology of research themore expensive it will be to develop knowledge exchange.

APPENDIX 1: METHODOLOGY

The symmetric Relative Specialisation Index (RSI) is given by the ratiobetween the share of the given scientific field in the publication of the givencountry and the share of the given scientific field in the world total of pub-lications (activity index – AI) minus one, divided by AI plus one. It maytake values in the range [�1, 1]. It indicates whether a country has a higher-than-average activity in a scientific field (RSI�1) or a lower-than-averageactivity (RSI�1).

(13.1)AI �pij ⁄�i

pij��j

pij ⁄�ijpij�


where pnumber of publications, i1 . . . nnumber of scientific fields11 and j1 . . . mnumber of countries7.

(13.2)

As the denominator of AI is the share of the given scientific field in theworld total of publications, the number and choice of the countries in thecomparative analysis does not influence the robustness of this indicator.

NOTES

* A version of this chapter has appeared in Research Policy, 32 (2003).1. Comments and suggestions made by Anthony Arundel, Dominique Foray, Ben Martin,

Richard Nelson, Ammon Salter, Ed Steinmueller and Robert Tijssen are gratefullyacknowledged. The participants in the New Frontiers in the Economics of Innovationand New Technology: A Conference in Honour of Paul A. David and the SixthInternational Conference on Science and Technology Indicators conferences providedinsightful comments. We are grateful to Anthony Arundel for the PACE data, and toDiana Hicks for the CHI classification. The financial support of the Commission of theEuropean Communities, TSER project ‘From Science to Products’ and IST project‘NewKind – New Indicators for the Knowledge-based Economy, no. IST 1999–20782’ isacknowledged.

2. We are indebted to Ed Steinmueller for the development of the discussion on the cumu-lativeness of science.

3. The PACE questionnaire surveyed the largest R&D performing industrial firms in 1993in 12 of the EU countries. The responses are from 414 large manufacturing firms acrossnine EU countries (Belgium, Denmark, Germany, Ireland, Italy, Luxembourg, theNetherlands, Spain and the UK).

4. Before proceeding, an important caveat related to the use of publications as a descrip-tor of a knowledge base needs to be discussed. We fully acknowledge that by adoptingpeer-reviewed publications as a descriptor of a country’s knowledge base we limit ouranalysis to the most codified (and codifiable) bits of this base. This limitation is deter-mined by obvious data constraints (tacit knowledge is rather difficult to capture ‘alive’),and also by the responses to the PACE questionnaire, which pinpoint scientific papers asa key mechanism to locate relevant sources of knowledge. Hicks (1995) thoroughly dis-cusses the role of scientific papers as signals of information about the presence of valu-able ‘hidden’ tacit skills. This chapter considers publications as elements of a signallingsystem whose morphological characteristics reveal something of the deeper structure ofa country’s sectoral knowledge base. Publications would be a sort of observable‘sufficient statistics’ of the underlying unobservables. Finally, as publications represent apreliminary and incomplete proxy of the knowledge base, more inclusive indicators orcombinations of indicators should be developed in the future to operationalise the inter-pretative framework.

5. C1: general chemistry, C2: analytical chemistry, C3: applied chemistry, C4: crystallogra-phy, C5: inorganic and nuclear chemistry, C6: medical chemistry, C7: organic chemistry,C8: physical chemistry, C9: polymer science, C10: pharmacology and pharmacy andC11: chemical engineering.

6. A problem emerged with respect to inorganic chemistry (C5) and organic chemistry(C7). For these fields, no publications are recorded in applied technology and engineer-ing. Thus, for these two fields, we considered as integrated those countries that exhibitspositive specialisation in applied research and basic research only.

RSI AI � 1AI � 1


7. A first attempt in this direction has been made within the project NewKind. All results,deliverables and data sources are available at http://www.researchineurope.org/newkind/index.htm (accessed on 21 March 2005).

REFERENCES

Archibugi, D. and M. Pianta (1992), The Technological Specialisation of AdvancedCountries, Dordrecht, Kluwer Academic.

Arundel, A. and A. Geuna (2001), ‘Proximity and the use of public science by inno-vative European firms’, Economics of Innovation and New Technology, 13 (6),559–80.

Arundel, A., G. van de Paal and L. Soete (1995), Innovation Strategies of Europe’sLargest Industrial Firms: Results of the PACE Survey: 23, Brussels, EuropeanInnovation Monitoring System, European Commission.

Balassa, B. (1965), ‘Trade liberalization and “revealed” comparative advantage’,The Manchester School of Economic and Social Studies, 32, 99–123.

Braun, T., W. Glänzel and H. Gruup (1995), ‘The scientometric weight of 50nations in 27 science areas. 1989–1993. Part I All fields combined, mathematics,engineering, chemistry and physics’, Scientometrics, 33, 263–93.

Cantwell, J.A. (1989), Technological Innovation and Multinational Corporations,Oxford, Basil Blackwell.

Cantwell, J.A. (1995), ‘Multinational corporations and innovatory activities:towards a new evolutionary approach’, in J. Molero (ed.), TechnologicalInnovation, Multinational Corporations and New International Competitiveness,Chur, Harwood, pp. 21–57.

Carlsson, B. (ed.) (1997), Technological Systems and Industrial Dynamics, Boston,MA, Kluwer Academic.

Conant, J.B. and L.K. Nash (1964), Harvard Case Histories in ExperimentalScience, vol. 1, Cambridge, MA, Harvard University Press.

Dalum, B., K. Laursen and G. Villumsen (1998), ‘Structural change in OECDexport specialisation patterns: de-specialisation and “stickiness” ’, InternationalReview of Applied Economics, 12, 447–67.

Dasgupta, P. and P.A. David (1994), ‘Towards a new economics of science’,Research Policy, 23, 487–521.

David, P.A. (1985), ‘Clio and the economics of QUERTY’, American EconomicReview, 75, 332–37.

David, P.A. and D. Foray (1995), ‘Accessing and expanding the science and tech-nology knowledge base’, STI Review, 16, 13–68.

European Commission (1996), Green Paper on Innovation, Brussels: EuropeanCommission.

European Commission (1997), Second European Report on S & T Indicators 1997,Brussels, European Commission.

Fagerberg, J., P. Guerrieri and B. Verspagen (eds) (1999), The Economic Challengefor Europe, Cheltenham, Edward Elgar.

Gambardella, A. (1995), Science and Innovation: The US Pharmaceutical IndustryDuring the 1980s, Cambridge, Cambridge University Press.

Geuna, A. (2001) ‘The evolution of specialisation: public research in the chemicaland pharmaceutical industries’, Research Evaluation, 10, 67–79.


Godin, B. (1994), ‘The relationship between science and technology’, unpublishedDPhil thesis, University of Sussex, Brighton.

Granstrand, O., P. Patel and K. Pavitt (1997), ‘Multi-technology corporations: whythey have “distributed” rather than “distinctive core” competencies’, CaliforniaManagement Review, 39, 8–25.

Hicks, D. (1995), ‘Published papers, tacit competencies and corporate managementof public/private character of knowledge’, Industrial and Corporate Change, 4,401–24.

Iansiti, M. (1998), Technology Integration: Making Critical Choices in a DynamicWorld, Boston, MA, Harvard Business School Press.

Landau, R. and N. Rosenberg (1992), ‘Successful commercialization in the chem-ical process industries’, in N. Rosenberg, R. Landau, and D.C. Mowery (eds),Technology and the Wealth of Nations, Stanford, CA, Stanford University Press.

Lundvall, B.A. (ed.) (1992), National Systems of Innovation. London: Pinter.Malerba, F. and L. Orsenigo (1997), ‘Technological regimes and sectoral pattern of

innovative activities’, Industrial and Corporate Change, 6, 83–118.Malo, S. and A. Geuna (2000), ‘Science-technology linkages in an emerging

research platform: the case of combinatorial chemistry and biology’,Scientometrics, 47, 303–21.

Mansfield, E. (1991), ‘Academic research and industrial innovation’, ResearchPolicy, 20, 1–12.

Martin, B.R., J. Irvine and P.A. Isard (1990), ‘Trends in UK government expend-iture on academic and related research: a comparison with the Federal Republicof Germany, France, Japan, the Netherlands and United States’, Science andPublic Policy, 17, 3–13.

Meliciani, V. (2001), Technology, Trade and Growth in OECD Countries. DoesSpecialisation Matter? London, Routledge.

Merton, R.K. (1965), On the Shoulders of Giants – a Shandean Postscript,New York, Free Press.

Narin, F., K.S. Hamilton and D. Olivastro (1997), ‘The increasing linkage betweenU.S. technology and public science’, Research Policy, 26, 317–30.

Nelson, R.R. (ed.) (1993), National Systems of Innovation: A Comparative Study,Oxford, Oxford University Press.

Office of Science and Technology (OST) (1998), Science & Technologie Indicateurs.Paris: Economica.

Orsenigo, L. (1989), The Emergence of Biotechnology: Institutions and Markets inIndustrial Innovation, London, Pinter.

Patel P. and K. Pavitt (1994), ‘Uneven (and divergent) technological developmentamongst advanced countries: evidence and a framework of explanation’,Industrial and Corporate Change, 13, 759–87.

Patel, P. and K. Pavitt (2000), ‘National systems of innovation under strain: theinternationalisation of corporate R&D’, in R. Barrell, G. Mason andM. O’Mahoney (eds), Productivity, Innovation and Economic Performance,Cambridge, Cambridge University Press.

Pavitt, K. (1989), ‘International patterns of technological accumulation’, inN. Hood and J.E. Vahlne (eds), Strategies in Global Competition, London, CroomHelm.

Pavitt, K. (1992), ‘The key characteristics of the large innovating firm’, in G. Dosi,M. Giannetti and L. Toninelli (eds), Technology and Enterprise in HistoricalPerspective, Oxford, Clarendon Press.


Pavitt, K. (1998), ‘Technologies, products and organisation in the innovating firm:what Adam Smith tells us that Schumpeter doesn’t’, Industrial and CorporateChange, 7 (3), 433–52.

Pianta, M. and D. Archibugi (1991), ‘Specialisation and size of scientific activities:a bibliometric analysis of advanced countries’, Scientometrics, 22, 341–58.

Pisano, G.P. (1996), ‘Learning-before-doing in the development of new processtechnology’, Research Policy, 25, pp. 1079–1119.

Pisano, G.P. (1997), The Development Factory: Unlocking the Potential of ProcessInnovation, Boston, MA, Harvard Business School Press.

Porter, M.E. (1990), The Comparative Advantage of Nations, New York, Free Press.Prencipe, A. (1997), ‘Technological competencies and a product’s evolutionary

dynamics: a case study from the aero-engine industry’, Research Policy, 25,1261–76.

Price, D. de S. (1963), Little Science, Big Science, New York, Columbia UniversityPress.

Rosenberg, N. (1994), Exploring the Black Box: Technology, Economics and History,Cambridge, Cambridge University Press.

Sharp, M., P. Patel and K. Pavitt (1997), Europe’s Pharmaceutical Industry:An Innovation Profile, Brussels, report prepared for DG XIII–D-4.

Soete, L.L.G. (1981), ‘A general test of the technological gap trade theory’,Weltwirtschaftliches Archiv, 117, 638–66.

Sternberg, R. (2000), ‘University–industry relationships in Germany and theirregional consequences’, in Z.J. Acs (eds), Regional Innovation, Knowledge andGlobal Change, London, Pinter.

Tijssen, R.J.W. and E. van Wijk (1999), ‘In search of the European paradox: aninternational comparison of Europe’s scientific performance and knowledgeflows in information and communication technologies research’, Research Policy,28, 519–43.

Vincenti, W.G. (1990), What Engineers Know and How They Know: AnalyticalStudies from Aeronautical History, Baltimore, MD, and London, Johns HopkinsUniversity Press.


PART IV

The Diffusion of New Technologies

14. Uncovering general purposetechnologies with patent data1

Bronwyn H. Hall and Manuel Trajtenberg

1 INTRODUCTION

In ‘The computer and the dynamo,’ Paul David (1991) makes a persuasivecase for considering the process by which the electric dynamo spreadthroughout the economy during the turn of the twentieth century and theprocess by which the use of information technology (specifically, comput-ing technology) is currently being spread throughout different industries assimilar manifestations of the diffusion of ‘general purpose technologies,’a term introduced into the economics literature by Bresnahan andTrajtenberg (1995). All these authors, as well as Helpman and Trajtenberg(1998a; 1998b), emphasise the singular contribution to economic growthmade by this type of technology, because of its ability to transform themeans and methods of production in a wide variety of industries.

At the same time and using historical data, David (1990; 1991),Rosenberg (1976), and others have argued that the diffusion of these tech-nologies throughout the economy may take decades rather than yearsbecause of co-ordination problems and the need for complementary invest-ments (both tangible and intangible) in using industries. For this reason itmay take some time for the benefits of the technologies to be manifest ineconomic growth. On the theoretical side, Bresnahan and Trajtenberg(1995) have studied the non-optimality of innovation and diffusion when adecentralised market system is called upon to try to solve the co-ordinationproblem between technology-innovating and technology-using industries.However, there has been relatively little empirical and econometric workthat incorporates the insights of these various authors to analyse specifictechnologies.

Our modest goal in this chapter is to see what might be learned about theexistence and technological development of general purpose technologies(GPTs) through the examination of patent data, including the citationsmade to other patents. Such measures would be useful both to help iden-tify GPTs in their early stages of development and also as proxies for the

389

various rates of technical change called for in a fully developed growthmodel such as that in Helpman and Trajtenberg (1998b). In doing thisexploration we are also motivated by the observation that not all technolo-gies or, indeed, R&D dollars are equal, but that economists too oftenignore that fact, primarily because of data limitations. As has been pointedout by others before us patenting measures have the potential to allowmore detailed analysis of the ‘direction’ as well as the ‘rate’ of technicalchange.2

Although such an exploration might be made using data from a varietyof countries, our focus here is on the use of US patent data, where thecitations have a well-defined meaning and also where they have been com-puterised since 1977, enabling us to work with them relatively easily. Giventhe importance of the USA as a locus of technical change in the late twen-tieth century, we do not feel that this limitation to US patenting activity isa serious drawback for a preliminary investigation of this kind.

We begin with the definition (description) of GPTs offered by Helpmanand Trajtenberg (1998a):

1. They are extremely pervasive and used in many sectors of the economy.Historical examples are the steam engine and the electric dynamo (theengine of electrification). Contemporary examples are the semi-conductor and perhaps the Internet.

2. Because they are pervasive and therefore important, they are subject tocontinuous technical advance after they are first introduced, with sus-tained performance improvements.

3. Effective use of these technologies requires complementary investmentin the using sectors; at the same time, the GPT enhances the product-ivity of R&D in the downstream sector. It is these points that areemphasised by David.

Using this definition, the contribution of the effort described here is todefine measures using patents and citations that quantify the insights ofDavid and Trajtenberg and their co-authors.

Our study is subject to a variety of limitations, however. First, it is basedon patent data, which provides imperfect coverage of innovative activity, asnot all innovations are patented or patentable.3 Second, it relies heavily onthe US Patent and Trademark Office (USPTO) classification system fortechnology, treating each three-digit patent class as roughly comparable tothe ‘size’ of a technology. An examination of the classes suggests that thisis unlikely to be strictly true (for example, the chemistry of inorganic com-pounds is a single class, whereas there are multiple optics classes). Makinguse of the sub-classes to refine the class measures would be a formidable

390 The diffusion of new technologies

task, because sub-classes are spawned within the three-digit class ad libitumand may descend either from the main class or from another sub-class.Thus some sub-classes are more ‘important’ than others, but this fact hasto be uncovered by a tedious search of the text on the USPTO website.Rather than attempting to construct our own classification system in thisway, we chose to look at measures based on the primary InternationalPatent Classification (IPC) class of the patent.4 Finally, because patentcitation data is only available in computerised form in 1975, and becausesevere truncation due to the application-grant lag and the citation lag setsin by around 1995 our period of study is necessarily fairly short and empha-sises the 1980s and 1990s. However, truncation in the later years means thatwe are unable (yet) to explore fully the implications of changes in informa-tion processing technology during the very recent past.

Because of time and resource constraints, we have focused the large partof our analysis on an extremely small subset of nearly 3 million patentsavailable to us, the 780 most highly cited patents that were granted between1967 and 1999. There is considerable evidence that the value or importancedistribution of patents is highly skewed, with most patents being unimport-ant and a few being highly valuable.5 We expect that one reason for thisfinding is that true GPT patents are concentrated among the highly citedpatents, so the current endeavour is centred on those patents which repre-sent the extreme tail of a very skew distribution.

In order to understand how patent data might help us identify GPTs andexplore their development and diffusion, it is necessary first to understandsomething more about patent citations. This is the subject of the nextsection. We then discuss the GPT-related measures we have constructedfrom the patent data, and show how our sample of highly cited patentsdiffers in various dimensions from the population as a whole.

2 PATENT CITATIONS6

A key data item in the patent document is ‘References Cited – US PatentDocuments’ (hereafter we refer to these just as ‘citations’). Patent citationsserve an important legal function, since they delimit the scope of the prop-erty rights awarded by the patent. Thus, if patent B cites patent A, itimplies that patent A represents a piece of previously existing knowledgeupon which patent B builds, and over which B cannot have a claim. Theapplicant has a legal duty to disclose any knowledge of the ‘prior art’, butthe decision regarding which patents to cite ultimately rests with thepatent examiner, who is supposed to be an expert in the area and hence tobe able to identify relevant prior art that the applicant misses or conceals.

Uncovering general purpose technologies with patent data 391

The presumption is thus that citations are informative of links betweenpatented innovations. First, citations made may constitute a ‘paper trail’for spillovers, that is, the fact that patent B cites patent A may be indica-tive of knowledge flowing from A to B; second, citations received may betelling of the ‘importance’ of the cited patent.7 The following quote pro-vides support for the latter presumption:

the examiner searches the . . . patent file. His purpose is to identify any prior dis-closures of technology . . . which might be similar to the claimed invention andlimit the scope of patent protection . . . or which, generally, reveal the state ofthe technology to which the invention is directed. If such documents arefound . . . they are ‘cited’ . . . if a single document is cited in numerous patents,the technology revealed in that document is apparently involved in many devel-opmental efforts. Thus, the number of times a patent document is cited may bea measure of its technological significance. (OTAF, 1976: 167)

The aspect of citations that is important for the present effort is that theyprovide a record of the link between the present invention and previousinventions. Thus they can tell us both the extent to which a particular lineof technology is being developed (if they are made to patents in the sametechnology area) and whether a particular invention is used in wide varietyof applications (if they are made to patents in different technology areas).In principle, given that we know which firm owns the relevant patents, it ispossible to ask these question both using the technology field, which is aclassification made by the USPTO,8 and using the industry in which thepatent falls, as indicated by the firm to which it is assigned.

3 MEASURES OF GPTS

The definition of GPTs paraphrased in the introduction suggests that thefollowing characteristics (observations) apply to the patents associatedwith GPT innovations: (1) they will have many citations from outside theirparticular technology area or perhaps from industries outside the one inwhich the patented invention was made; (2) they will have many citationswithin their technology area, and the citations will indicate a pattern ofcumulative innovation, or trace out a technology trajectory; (3) morespeculatively, citing technologies will be subject to a burst of innovativeactivity as complementary goods are developed; and (4) given the length oftime it takes for a GPT to pervade the economy, citation lags for patents inthis area may be longer than average. In this section we report on the con-struction of a number of proxies for these characteristics. We use theseproxies to identify patents that are in the extreme tail of the distribution of


patent characteristics, in an effort to identify some candidate GPTs. Notsurprisingly, we find that looking at a single characteristics may be mis-leading, so in the later sections of the chapter we use a more multivariateapproach to refine the analysis.

It is well known that the distribution of patent values and patentcitations is very skewed with almost half of all patents receiving zero or onecite and less than 0.1 per cent receiving more than 100 cites (see Hall et al.2005, for evidence on both points). Observations (1) and (2) above alsosuggest that GPT patents are likely to be highly cited. Therefore, we beganour investigation by focusing on highly cited patents. We identified thesepatents by requiring that the number of citations the patent received begreater than three times the number received by the patent in the 99th per-centile of the distribution. The results of this selection process are shownin Table 14.1. It yielded 780 patents granted between 1967 and 1999 thatwere ultimately granted, together with the name and type of their assignee(owner), the three-digit patent classification, and similar information on allthe patents issued between 1975 and 2002 that cited this patent, for a totalof 159 822 citations. Table 14.1 also makes it clear how skewed the citationdistribution is: our sample of 780 patents is about one out of 3700 patents,whereas the 160 000 citations are one out of 100 citations (there are16 million citations in all). Thus our patents are 37 times more likely to becited than predicted by the average probability.

3.1 Generality

Observation (1) above suggests the use of a measure that is similar to theTrajtenberg, Jaffe and Henderson ‘generality’ measure, which is defined inthe following way:

where sij denotes the percentage of citations received by patent i that belongto patent class j, out of ni patent classes (note that the sum is the Herfindahlconcentration index). Thus, if a patent is cited by subsequent patents thatbelong to a wide range of fields the measure will be high, whereas if mostcitations are concentrated in a few fields it will be low (close to zero).9

Observation (2) suggests that even if generality is relatively high, theabsolute number of citations should also be high, implying that there maystill be a large number of citations in the patent’s own technology class. Italso suggests that ‘second-generation’ citations be examined. We imple-ment this using two variables, the average number of citations to the citingpatents and the average generality of the citing patent.

Generality � Gi 1 � �ni

js2

ij



Table 14.1 Selecting the sample of highly cited patents

Grant Cutoff Highly cited patents All patents Highlyyear no.

Number Median cites Number Median citescited

citations share (%)

1967 78 20 103.0 65 652 2 0.0301968 84 15 131.0 59 104 2 0.0251969 87 28 115.0 67 559 3 0.0411970 93 19 105.0 64 429 3 0.0291971 99 22 165.0 78 317 3 0.0281972 108 30 136.0 74 810 4 0.0401973 111 25 132.0 74 143 4 0.0341974 117 24 136.5 76 278 5 0.0311975 126 19 152.0 73 690 5 0.0261976 132 22 156.5 72 015 5 0.0311977 135 17 182.0 66 883 5 0.0251978 138 16 179.0 67 862 5 0.0241979 141 7 190.0 50 177 5 0.0141980 144 26 204.5 63 371 5 0.0411981 147 17 177.0 67 373 5 0.0251982 150 19 232.0 59 462 5 0.0321983 156 20 224.5 58 435 5 0.0341984 159 22 188.5 69 338 5 0.0321985 159 25 199.0 73 824 5 0.0341986 168 21 204.0 72 977 6 0.0291987 180 29 213.0 85 522 6 0.0341988 177 29 252.0 80 345 6 0.0361989 177 27 258.0 98 567 5 0.0271990 177 27 215.0 93 290 5 0.0291991 171 38 204.5 99 789 5 0.0381992 174 41 218.0 100 760 5 0.0411993 174 33 231.0 100 980 4 0.0331994 171 26 214.5 104 317 4 0.0251995 156 21 178.0 104 091 4 0.021996 141 13 171.0 112 832 3 0.0121997 114 28 136.0 115 337 3 0.0241998 90 33 103.0 151 745 2 0.0221999 63 21 69.0 153 486 1 0.014

All years 780 183.0 2 756 760 3 0.028

Note: *Patents with zero cites 1975–2002 are excluded.

In actual measurement, the preceding two predictions interact in wayswhich make our task a bit more complex. Like patents, citation counts area discrete random variable bounded below by zero. This means that fewercitations in total imply that fewer classes will be observed to have citationsthan should be observed were the total number of citations larger. That is,ni is biased downward by the fact that fractional citations are notobserved, and generality will tend to be lower when there are fewercitations. This is quite visible in the graph of average generality over timeshown in Figure 14.1, where we show two different versions of generality,one based on US patent classes, and another based on the InternationalPatent Classification, as assigned to these patents by the USPTO. Notethat the average of either generality measures begins to decline fairlysteeply in 1993–95, at the same time as our measure of average citationsper patent turns down sharply due to the effects of lag truncation (seeFigure 14.2). In this case, this is a spurious rather than real decline in gen-erality, owing to the fact that our patent grant data ends in 2002, andtherefore our application-dated data ends around 1999, so that patents inthe years after about 1994 have had less chance to receive citations.10

Using a simple binomial model of the probability of observing a citationin a given cell, Hall (2002) shows that an unbiased estimate of the general-ity of the ith patent can be computed using the following correction:

where Ni is the number of citations observed. Note that this measure is notdefined when Ni � 2 and will be fairly noisy when Ni is small. We have usedthis bias correction for the first three generality measures described below.

The US patent classification system has grown over time in ways thatmake it not ideal for the purpose we have in mind here. Generality measuresessentially assume that all categories are equidistant from each other if theyare to be compared, but this is not the case for the US patent class system.Therefore, we explore the use of generality measures based on five differentclassification systems:

1 US patent class (approximately 400 cells).2 Hall–Jaffe–Trajtenberg technology subcategories (36 cells).3 Main International Patent Class (approximately 1200 cells).4 Industry classification based on Silverman’s IPC-SIC concordance

(Silverman, 2002) for industry of manufacture, aggregated to the Hall–Vopel (Hall and Vopel, 1997) level (37 cells).

5 Industry classification based on Silverman’s IPC-SIC concordance forindustry of use, aggregated to Hall–Vopel level (37 cells).

Gi Ni

Ni � 1 Gi


396

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

1977

1979

1981

1983

1985

1987

1989

1991

1993

1995

1997

1999

Pat

ent a

pplic

atio

n ye

ar

Index

0.00

1.00

2.00

3.00

4.00

5.00

6.00

7.00

8.00

9.00

10.0

0

Cites/patent

Gen

eral

ity(I

PC

cla

ss)

Cite

s pe

r pa

tent

(as

of 1

996)

Cite

s pe

r pa

tent

(as

of 2

002)

Gen

eral

ity(U

S c

lass

)

Cite

s/pa

t (20

02)

Gen

eral

ity

Cite

s/pa

t (19

96)

Fig

ure

14.1

Ave

rage

mea

sure

s of

gene

ralit

y an

d or

igin

alit

y

397

Fig

ure

14.2

Cit

atio

n la

g di

stri

buti

on (

1976

–94)

.Tra

jten

berg

,Jaff

e an

d H

ende

rson

Met

hodo

logy

0.0

0.1

0.2

0.3

0.4

0.5

0.6

12

34

56

78

910

1112

13 1

415

1617

1819

2021

2223

2425

2627

2829

3031

Cita

tion

lag

Relative citation probability

The rationales for these various choices are the following. First, in additionto using the US patent classification system (measure 1), we also con-structed generality based on the more equal groupings of technologies con-structed by Hall et al. (2002) from the patent classes (measure 2) and fromthe International Patent Classification system main four-digit classes(measure 3), which is more detailed that the US patent classification system.

Second, it could be argued that a GPT is not likely to manifest itself asa series of citations by patents in different technology classes, but rather ascitations by firms in different industries. To consider this possibility, wewould like to base a measure on the shares of citations that come from firmsin different industries at the roughly two and one-half digit level. That is,our fourth measure is a Herfindahl for patent citation dispersion acrossindustries rather than across technologies. Based on the discussion of GPTdiffusion to using industries in the introduction, an industry-based measurewould seem to be intrinsically preferred for this exercise. There are basicallytwo ways to construct such a measure: the first uses the industry of owner-ship of the patents based on identification of the patent assignees and thesecond determines the industry for each patent class/sub-class from sometype of industry-patent class concordance. The first approach is difficult toimplement in practice, given the number of patents that are unassigned andthe number of assignees that are not identifiable, either because they aresmall firms or because they are foreign firms for which we do not yet havea match to other data sources.11

Therefore we used the SIC-technology class concordances of Silverman(2002) to assign these patents and their citations to industries of manufac-ture and of use. Then we collapsed the distribution of citations by SICcodes into a 37-element vector of industries using the SIC-industry corres-pondence given in Appendix Table A1, and used this vector to construct thegenerality measures 4 and 5. The computation of these measures was ableto make use of all the patents rather than just those held by US industry.12

One major drawback of using the Silverman concordance ought to be men-tioned, especially in the light of our subsequent findings: it is based onassignments to industry of manufacture and use made by Canadian patentexaminers between 1990 and 1993. This means that it will do a poor job onpatents in technologies related to the growth of the Internet and software,because there were unlikely to be many of these in the Canadian patentsystem prior to 1994.

Figures 14.3 and 14.4 show the distribution of two of the computed gen-erality indices for the highly cited patents, Figure 14.3 the US patent classindex and Figure 14.4 the index based on Silverman’s industry of manu-facture map. As the figures show, these indices range from zero to one andthe measure based on the industry of manufacture has a somewhat



Fre

quen

cy

Generality index0 0.5 1

0

20

40

60

Figure 14.3 Distribution of US patent class-based generality

Figure 14.4 Distribution of industry of manufacture-based generality

Fre

quen

cy

Generality based on sic of mfg

0 0.5 1

0

20

40

60

different distribution from that based on the US classification system. InAppendix Table A3, we show the correlation matrix for all five generalitymeasures for our highly cited sample. Although they are generally fairlyhighly correlated, the industry of use measure (5) is not very correlated withthe US class-based measures (1 and 3), and the industry of manufacture (4)not very correlated with the US class measure (1).

Table 14.2 shows the 20 highly cited patents which also have the highestgenerality, where generality is measured by each of the five measures. Ingeneral, the most general patents are those in chemicals, especially when weconsider the industry of use. Looking at the industry of manufacture, thosein other technologies seem to be the most general. However, looking by USclass, we can see the drawback of this generality measure: there are a numberof chemical classes that are all essentially the same large class (the series532–570), whereas in the case of some of the physics-based classes, there isonly a single class. This fact will tend to bias the index towards generality inthe chemicals case; however, the fact that the IPC classification which doesnot have this structure produces a similar result is somewhat reassuring.

3.2 Patenting Growth

Observation (3) suggested that we look at patent classes with rapid growthin patenting. Using the entire patent database aggregated to patent class,we constructed three sub-periods (1975–83, 1984–92 and 1993–99) andcomputed the average growth within class for each of the periods.13 Theresults are shown in Table 14.3. As might be expected, in all three periods,the patent classes with rapid growth are dominated by the information anddata processing classes (395 and the 700 series), with the addition of thenew multicellular biotechnology class 800 in the latter two periods. Highlycited patents are slightly more common in rapidly growing classes,although only a few of these classes have significant numbers of highlycited patents and the difference may not be very significant. It does appearthat the patent classes that are growing rapidly include technologies thathave more of the character of what we think of as GPTs, but that althoughhighly cited patents are two to three times more likely to be found in rapidlygrowing classes (as we might expect if citations tend to come from the sameclass), they do not seem to be disproportionate in these classes.

Another way of looking at the growth in patenting following the intro-duction of a GPT is to look at the growth of the patent classes that cite sucha technology. The hypothesis is that innovations which build on a GPT-likeinnovation will themselves spawn many new innovations. Table 14.4 showsthe patent classes for the top 20 patents in terms of the growth of their citingpatent classes, both for the highly cited patents and for all patents, excluding


401

Tab

le 1

4.2

Num

ber

ofto

p 20

hig

hly

cite

d pa

tent

s

HJT

sub

-U

S pa

tent

cla

ssC

lass

des

crip

tion

Gen

eral

ity

mea

sure

cate

gory

US

US

Indu

stry

of

Indu

stry

of

clas

sIP

Csu

bcat

egor

ym

anuf

actu

reus

e

1144

2T

exti

les:

Web

or

Shee

t C

onta

inin

g0

14

13

Stru

ctur

ally

De

Coa

ting

s0

01

10

1210

6C

ompo

siti

ons:

Coa

ting

or

Pla

stic

00

01

011

8C

oati

ng A

ppar

atus

00

10

014

540

556

568

Org

anic

Com

poun

ds –

Par

t of

the

33

01

0C

lass

532

–570

Ser

ies

1552

152

352

452

8Sy

nthe

tic

Res

ins

or N

atur

al R

ubbe

rs –

3

51

11

Par

t of

the

Cla

ss 5

20 S

erie

s19

Mis

cella

neou

s ch

emic

als

00

02

615

6A

dhes

ive

Bon

ding

and

Mis

cella

neou

s 0

00

02

Che

mic

als

366

Agi

tati

ng0

00

01

430

Rad

iati

on I

mag

ery

Che

mis

try:

Pro

cess

,0

00

12

Com

posi

tion

,or

Pro

duct

The

reof

510

Cle

anin

g C

ompo

siti

ons

for

Solid

Sur

face

00

01

1T

OT

AL

CH

EM

ICA

LS

69

66

10

2121

Com

mun

icat

ions

02

10

034

0C

omm

unic

atio

ns:E

lect

rica

l0

01

00

342

Com

mun

icat

ions

:Dir

ecti

ve R

adio

Wav

e 0

10

00

Syst

ems

and

Dev

ices

(e.

g.,R

adar

,R

adio

Na)

455

Tel

ecom

mun

icat

ions

01

00

0

402

Tab

le 1

4.2

(con

tinu

ed)

HJT

sub

-U

S pa

tent

cla

ssC

lass

des

crip

tion

Gen

eral

ity

mea

sure

cate

gory

US

US

Indu

stry

of

Indu

stry

of

clas

sIP

Csu

bcat

egor

ym

anuf

actu

reus

e

2334

5 Se

lect

ive

Vis

ual D

ispl

ay S

yste

ms

10

00

0T

OT

AL

CO

MP

UT

ING

12

10

0

3212

8Su

rger

y1

22

21

604

Surg

ery

20

20

2T

OT

AL

DR

UG

S A

ND

ME

DIC

AL

32

42

3IN

STR

UM

EN

TS

4117

4E

lect

rici

ty:C

ondu

ctor

s an

d In

sula

tors

11

11

146

257

Act

ive

Solid

-Sta

te D

evic

es (

e.g.

,Tra

nsis

tors

)1

11

00

49M

isce

llane

ous

elec

tric

al4

00

00

348

Tel

evis

ion

30

00

038

6T

elev

isio

n Si

gnal

Pro

cess

ing

for

Dyn

amic

1

00

00

Rec

ordi

ng o

r R

epro

duci

ngT

OT

AL

EL

EC

TR

ICA

L6

22

11

5126

4P

last

ic a

nd N

onm

etal

lic A

rtic

le S

hapi

ng3

34

21

5435

9O

ptic

s:Sy

stem

s (i

nclu

ding

Com

mun

icat

ions

)0

00

02

5949

Mov

able

or

Rem

ovab

le C

losu

res

01

00

0T

OT

AL

ME

CH

AN

ICA

L3

44

23

6713

8P

ipes

and

Tub

ular

Con

duit

s0

01

10

6853

Pac

kage

Mak

ing

00

01

069

248

Supp

orts

11

26

3T

OT

AL

OT

HE

R1

13

83

403

Tab

le 1

4.3

US

pat

ent

clas

ses

wit

h ra

pid

grow

th

Pat

ents

inH

ighl

y ci

ted

pate

nts

Ann

ual g

row

thC

lass

des

crip

tion

Cla

ss19

9919

93–9

919

92–9

9

Num

ber

Num

ber

Shar

e (%

)(%

)

335

00.

0022

.2D

ata

Pro

cess

ing:

Des

ign

and

Ana

lysi

s of

716

Cir

cuit

or

Sem

icon

duct

or M

ask

900

0.00

22.2

Inte

ract

ive

Vid

eo D

istr

ibut

ion

Syst

ems

725

311

00.

0022

.1D

ata

Pro

cess

ing:

Soft

war

e D

evel

opm

ent,

717

Inst

alla

tion

,or

Man

agem

ent

152

127.

8921

.6D

ata

Pro

cess

ing:

Stru

ctur

al D

esig

n,70

3M

odel

ing,

Sim

ulat

ion,

and

Em

ulat

ion

523

71.

3415

.6M

ulti

cellu

lar

Liv

ing

Org

anis

ms

and

800

Unm

odifi

ed P

arts

The

reof

and

Rel

ated

Pro

cess

es1

061

00.

0013

.1D

ata

Pro

cess

ing:

Dat

abas

e an

d F

ile

707

Man

agem

ent,

Dat

a St

ruct

ures

,or

Doc

umen

t P

roce

ssin

g74

20

0.00

13.0

Sem

icon

duct

or D

evic

e M

anuf

actu

ring

:43

8P

roce

ss25

50

0.00

12.1

Am

usem

ent

Dev

ices

:Gam

es46

337

00.

0012

.0F

ound

atio

n G

arm

ents

450

420

0.00

11.7

Che

mis

try:

Fis

cher

–Tor

psch

Pro

cess

es;

518

or P

urifi

cati

on o

r R

ecov

ery

ofP

rodu

cts

The

reof

754

819

0.25

Tota

l for

sel

ecte

d cl

asse

s14

604

514

80.

10A

ll cl

asse

s

404

Tab

le 1

4.3

(con

tinu

ed)

Pat

ents

inH

ighl

y ci

ted

pate

nts

Ann

ual g

row

thC

lass

des

crip

tion

Cla

ss19

9219

84–9

219

83–9

2 (%

)

Num

ber

Num

ber

Shar

e (%

)(%

)

240

00.

0022

.2Su

perc

ondu

ctor

Tec

hnol

ogy:

505

App

arat

us,M

ater

ial,

Pro

cess

204

00.

0021

.8D

ata

Pro

cess

ing:

Art

ifici

al I

ntel

ligen

ce

706

910

0.00

18.2

Mul

tice

llula

r L

ivin

g O

rgan

ism

s an

d80

0U

nmod

ified

Par

ts T

here

ofan

d R

elat

ed P

roce

sses

300

41.

3317

.6E

lect

rica

l Com

pute

rs a

nd D

igit

al

709

Pro

cess

ing

Syst

ems:

Mul

tipl

e C

ompu

ter

or P

roce

ss27

34

1.47

17.2

Dat

a P

roce

ssin

g:D

atab

ase

and

File

70

7M

anag

emen

t,D

ata

Stru

ctur

es,o

rD

ocum

ent

Pro

cess

ing

264

41.

5216

.1In

form

atio

n P

roce

ssin

g Sy

stem

39

5O

rgan

izat

ion

180

0.00

15.9

Tex

tile

s:C

loth

Fin

ishi

ng26

221

41.

8115

.7E

lect

rica

l Com

pute

rs a

nd D

igit

al

713

Pro

cess

ing

Syst

ems:

Supp

ort

320

0.00

14.2

Rol

l or

Rol

ler

492

270

0.00

14.1

Rai

lway

Sw

itch

es a

nd S

igna

ls24

6

167

016

0.96

Tota

l for

sel

ecte

d cl

asse

s10

662

625

90.

24A

ll cl

asse

s

405

Pat

ents

inH

ighl

y ci

ted

pate

nts1

Ann

ual g

row

thC

lass

des

crip

tion

Cla

ss19

8397

5–83

1975

–83

(%)

Num

ber

Num

ber

Shar

e (%

)(%

)

420

0.00

18.8

Info

rmat

ion

Pro

cess

ing

Syst

em O

rgan

isat

ion

395

960

0.00

13.7

Dat

a P

roce

ssin

g:Sp

eech

Sig

nal P

roce

ssin

g,70

4L

ingu

isti

cs,L

angu

age

Tra

nsla

tion

,and

381

2.63

13.0

Ele

ctri

cal C

ompu

ters

and

Dig

ital

71

3P

roce

ssin

g Sy

stem

s:Su

ppor

t75

11.

3312

.1E

lect

rica

l Com

pute

rs a

nd D

igit

al P

roce

ssin

g 71

2Sy

stem

s:P

roce

ssin

g A

rchi

tect

ures

and

126

10.

7911

.6D

ata

Pro

cess

ing:

Veh

icle

s,N

avig

atio

n,70

1an

d R

elat

ive

Loc

atio

n22

30

0.00

11.3

Dat

a P

roce

ssin

g:G

ener

ic C

ontr

ol S

yste

ms

700

or S

peci

fic A

pplic

atio

ns54

35.

5611

.6D

ata

Pro

cess

ing:

Fin

anci

al,B

usin

ess

705

Pra

ctic

e,M

anag

emen

t,or

Cos

t/P

rice

D

eter

min

atio

n13

50

0.00

10.9

Dat

a P

roce

ssin

g:M

easu

ring

,Cal

ibra

ting

,70

2or

Tes

ting

350

0.00

10.7

Ele

ctri

cal C

ompu

ters

and

Dig

ital

Pro

cess

ing

709

Syst

ems:

Mul

tipl

e C

ompu

ter

or P

roce

ss25

70

0.00

10.5

Err

or D

etec

tion

/Cor

rect

ion

and

Fau

lt

714

Det

ecti

on/R

ecov

ery

108

16

0.56

Tota

l for

sel

ecte

d cl

asse

s63

383

322

0.51

All

clas

ses

Not

e:P

aten

t cl

asse

s w

ith

few

er t

han

10 p

aten

ts a

t th

e en

d of

each

per

iod

have

bee

n om

itte

d fr

om t

he t

able

.

406

Tab

le 1

4.4

Pat

ent

clas

ses

who

se c

ited

cla

sses

hav

e hi

gh g

row

th r

ates

*

HJT

sub

-U

S pa

tent

Cla

ss d

escr

ipti

onH

ighl

y ci

ted

All

pate

nts

excl

.ca

tego

rycl

ass

pate

nts

high

ly c

ited

TO

TA

L C

HE

MIC

AL

S0

021

Com

mun

icat

ions

437

0M

ulti

plex

Com

mun

icat

ions

337

9T

elep

honi

c C

omm

unic

atio

ns1

22C

ompu

ter

Har

dwar

e an

d So

ftw

are

1118

380

Cry

ptog

raph

y1

395

Info

rmat

ion

Pro

cess

ing

Syst

em O

rgan

isat

ion

1670

4D

ata

Pro

cess

ing:

Spee

ch S

igna

l Pro

cess

ing,

Lin

guis

tics

,Lan

guag

e1

705

Dat

a P

roce

ssin

g:F

inan

cial

,Bus

ines

s P

ract

ice,

Man

agem

ent,

or2

707

Dat

a P

roce

ssin

g:D

atab

ase

and

File

Man

agem

ent,

Dat

a St

ruct

ure

170

9E

lect

rica

l Com

pute

rs a

nd D

igit

al P

roce

ssin

g Sy

stem

s:M

ulti

ple

371

2E

lect

rica

l Com

pute

rs a

nd D

igit

al P

roce

ssin

g Sy

stem

s:P

roce

ssin

g1

713

Ele

ctri

cal C

ompu

ters

and

Dig

ital

Pro

cess

ing

Syst

ems:

Supp

ort

371

7D

ata

Pro

cess

ing:

Soft

war

e D

evel

opm

ent,

Inst

alla

tion

,or

Man

agem

ent

123

345

Sele

ctiv

e V

isua

l Dis

play

Sys

tem

s3

1T

OT

AL

CO

MP

UT

ING

1819

3343

5C

hem

istr

y:M

olec

ular

Bio

logy

and

Mic

robi

olog

y1

TO

TA

L D

RU

GS

AN

D M

ED

ICA

L I

NST

RU

ME

NT

S1

049

Mis

cella

neou

s el

ectr

ical

348

Tel

evis

ion

250

5Se

mic

ondu

ctor

tec

hnol

ogy:

appa

ratu

s,et

c1

TO

TA

L E

LE

CT

RIC

AL

21

TO

TA

L M

EC

HA

NIC

AL

00

TO

TA

L M

ISC

EL

LA

NE

OU

S0

0

Not

e:*C

lass

es fo

r th

e to

p 20

pat

ents

in e

ach

cate

gory

are

sho

wn.

The

ave

rage

gro

wth

rat

e of

the

clas

ses

ofth

e hi

ghly

cit

ed p

aten

ts is

abo

ve 2

8pe

rce

nt p

er a

nnum

;tho

se fo

r al

l pat

ents

abo

ve 5

2 pe

r ce

nt p

er a

nnum

.

those that are highly cited. The message is clear: using this measure, almostall the patent classes identified are in computing and communications tech-nology, and most are in data processing technologies more narrowly defined.

3.3 Citation Lags

Finally, observation (4) suggested that the average citation lag to GPTpatents might be longer. Of course, citations to patent data for a fixed timeperiod such as ours (1967–99) are always subject to truncation. For thisreason, we look at mean citation lags that are large relative to the averagecitation lag for patents applied for in the same year.

Table 14.5 shows that the 20 highly cited patents with long lags(greater than 70 per cent of the average citation lag) are typically in older


Table 14.5 Patent classes with highly cited patents that have long cite lags

HJT sub- US patent Class description Number category class of patents

TOTAL CHEMICALS 0

24 365 Static Information Storage and Retrieval 1TOTAL COMPUTING 1

32 604, 606 Surgery and Med. Instruments 339 623 Prosthesis (i.e., Artificial Body 1

Members), Parts Thereof, or ATOTAL DRUGS AND MEDICAL 4

INSTRUMENTS

TOTAL ELECTRICAL 0

51 Mat. Proc. and Handling 2264 Plastic and Nonmetallic Article 1

Shaping or Treating: Processes425 Plastic Article or Earthenware Shaping 1

or Treating: Apparatus54 359 Optics 2

TOTAL MECHANICAL 4

61 47 Agriculture, Husbandry, Food 368 Receptacles 7

53 Package Making 1206 Special Receptacle or Package 3383 Flexible Bags 3

69 428 Stock Material or Miscellaneous Articles 1TOTAL MISCELLANEOUS 11

technologies. It is noteworthy that there are none in the chemicals or elec-trical industries and only five in computing and drugs, most of which areto surgical innovations. The only highly cited computing patent with longcitation lags is a patent on an aspect of computer architecture taken out bySiemens in 1976; this patent has a mean citation lag of 23 years and is note-worthy because it is has essentially no citations until after it expired in 1994.It now has over 200. In general, given the fact that long lags by themselvesoften simply identify older and slower-moving technologies such as pack-aging, we will want to use this indicator in combination with our other indi-cators when looking for GPTs.

3.4 Summary

The GPT measures we have identified (the five generality indexes, the gen-erality of citing patents, within class growth in patenting, growth in citingpatent classes and the average citation lag) are promising, but clearly givecontradictory messages when taken separately. The goal is to combine themin a reasonable way to give an indication of the types of evidence GPTsleave in the patent statistics. We explore solutions to this problem insection 5, but first we summarise the relationship between them and theprobability that a patent is highly cited.

4 HIGHLY CITED PATENTS

Table 14.6 shows that the highly cited patents differ in almost all respectsfrom the population of all patents, and also from a 4 per cent sample ofpatents with at least one cite that we will use later as a control sample. Theytake longer to be issued, they have about twice as many claims, they aremore likely to have a US origin, and more likely to be assigned to a UScorporation, more likely to have multiple assignees and have higher cita-tion lags on average.14 They also have higher generality, no matter howgenerality is measured, and are in patent classes that are growing fasterthan average. Although the patents that cite them are more likely to becited themselves, they have only slightly higher generality than citingpatents in general.

More than half of the highly cited patents are in two of our six main tech-nology classes: computing hardware and software, and drugs and medicalinstruments. Of course, these are indeed the technology classes where weexpect to find modern-day GPTs. In Appendix Table A2, we broke thisdown, in order to identify the important technologies more precisely.Highly cited patents are more than twice as likely to be found in computer



Table 14.6 US patents granted 1967–99

Statistic All patents 4 % sample of Highly citedpatents patents

(�1 cite)

Number of patents 2 768 011 100 634 780Year applied for 1983.1 1982.8 1981.6Year granted 1984.6 1984.8 1983.9Average grant lag (years) 1.5 2.0 2.3Number of claims 12.1 12.5 23.6

Number of forward 6.72 8.85 204.71citations (to 2002)

Average citation lag 8.73 9.98 13.48Average class growth NA 2.98% 7.28%

(5 years)Share US origin 59.90% 61.20% 88.30%Share assigned to US 46.50% 47.70% 75.90%

corporationsShare multiple assignees 0.50% 0.60% 1.00%

Generality 1 (US class) 0.3417 0.5261 0.6416Generality 2 (IPC) 0.3548 0.5484 0.5716Generality 3 (US 0.2711 0.4167 0.4569

subcategory)Generality 4 (SIC of NA NA 0.5856

mfg-IPC)Generality 5 (SIC of NA NA 0.6444

use-IPC)

Average cites to citing 4.02 4.7 12.92patents

Total cites to citing patents 46.5 55.7 2663.9Average growth of citing NA 3.57% 7.69%

patent classes*Average generality of citing 0.3094 0.3487 0.3887

patents

Broad technology classes Chemicals 20.80% 19.20% 18.00%Computing 10.20% 11.40% 23.90%Drugs and medical 7.30% 7.00% 32.60%Electrical 17.10% 17.80% 9.70%Mechanical 23.00% 22.90% 6.10%Other 21.60% 21.80% 9.70%

Type of assigneeUS corporation 1 247 030 47 975 596Non-US corporation 885 533 31 798 75US individual 378 394 14 347 98Non-US individual 135 756 4 645 9US government 43 048 1 499 6Non-US government 9 845 370 1

hardware and software, computer peripherals, surgery and medical instru-ments, genetic technologies, miscellaneous drugs and semiconductors.

Table 14.7 shows a series of probit estimations for the probability thata patent with at least one cite will be highly cited, in order to provide amultivariate summary of the data in Table 14.6.15 Table 14.7 shows thederivative of the probability with respect to the independent variable thatare implied by the coefficient estimates. In the case of dummy variables, itshows the change in probability when the variable changes from zero toone. Because the probability of being one of the highly cited patents in thesample is very small (0.77 per cent), the values in Table 14.7 are small.Taking the grant lag as an example, the interpretation is that an additionalyear between application and grant increases the probability of beinghighly cited by 0.06 per cent, or from 0.77 per cent to 0.83 per cent at themean. Being a patent in the drugs and medical category increases the prob-ability by 2.9 per cent, which is a very large change at the mean probability.

Table 14.7 confirms the univariate differences between highly cited andall patents. In addition, this table shows that variations over time in theprobability of high citation do not greatly affect the coefficients (comparecolumn 4 with column 2). The only generality measure that enters signifi-cantly and positively in this regression is that based on the US class; theothers were all insignificant (IPC, technology sub-category and industry ofuse) or slightly negative (industry of manufacture). Also note that highlycited patents are far more likely to be cited by patents that are themselvescited by patents in many technology classes, once we control for the otherdifferences between highly cited and other patents.

5 IDENTIFYING GPT PATENTS

It is not obvious how to combine these measures to choose a sample of GPTpatents. In this first investigation of the topic, we have chosen simply to lookfor patents that are outliers in several of the categories, on the grounds thatsuch patents are likely to give us an idea of the technologies that have givenbirth to the most subsequent inventive activity in the largest number oftechnological areas. Accordingly, we began with the 780 highly cited patentsand then we chose a set of patents that fell in the top 20 per cent of thesepatents according to generality, citing patent generality and the subsequentfive-year growth of the patent’s class. We performed this exercise for eachof the five generality measures in turn.

Table 14. 8 shows the result: 20 patents out of the 780 were selected, manyby several of the different criteria. Selection by each of the five generalitymeasures is indicated by the presence of the measure in the table. Of these


411

Tab

le 1

4.7

Pro

bit

regr

essi

on fo

r hi

ghly

cit

ed p

aten

ts (

101

414

obse

rvat

ions

;78

0 hi

ghly

cit

ed)*

Var

iabl

eC

ited

pat

ent

Cit

ed a

nd c

itin

gC

ited

and

cit

ing

Incl

udin

g ye

arch

arac

teri

stic

spa

tent

cha

r.pa

tent

cha

r.du

mm

ies

dp/d

x***

Std.

err

dp/d

x***

Std.

err

dp/d

x***

Std.

err

dp/d

x***

Std.

err

Num

ber

ofcl

aim

s/10

0.

124%

0.01

1%0.

062%

0.00

6%0.

076%

0.00

7%0.

055%

0.00

6%D

(cl

aim

s m

issi

ng)�

0.40

0%0.

063%

0.03

3%0.

025%

0.03

9%0.

029%

0.50

8%0.

452%

Ave

rage

gra

nt la

g (y

ears

) 0.

062%

0.01

5%0.

057%

0.00

8%0.

066%

0.00

9%0.

054%

0.00

8%D

umm

y fo

r U

S or

igin

�0.

397%

0.04

7%0.

146%

0.02

5%0.

163%

0.02

8%0.

137%

0.02

3%D

umm

y fo

r U

S 0.

210%

0.04

6%0.

147%

0.02

6%0.

166%

0.02

9%0.

132%

0.02

4%co

rpor

atio

n�G

ener

alit

y 1

(US

clas

s)

0.22

6%0.

032%

0.20

2%0.

030%

Gen

eral

ity

5 (S

IC o

fus

e)

�0.

096%

0.06

3%

Ave

rage

cit

atio

n la

g 0.

068%

0.00

5%0.

078%

0.00

5%0.

066%

0.00

5%(r

elat

ive

to y

ear

aver

age)

A

vera

ge g

ener

alit

y of

citi

ng

0.33

1%0.

044%

0.47

0%

0.05

0%0.

334%

0.04

4%pa

tent

s

412

Tab

le 1

4.7

(con

tinu

ed)

Var

iabl

eC

ited

pat

ent

Cit

ed a

nd c

itin

gC

ited

and

cit

ing

Incl

udin

g ye

arch

arac

teri

stic

spa

tent

cha

r.pa

tent

cha

r.du

mm

ies

dp/d

x***

Std.

err

dp/d

x***

Std.

err

dp/d

x***

Std.

err

dp/d

x***

Std.

err

Dum

mie

s fo

r te

chno

logy

cla

sses

**

Che

mic

als�

0.30

8%

0.08

5%

0.20

1%

0.05

2%

0.23

9%

0.06

0%

0.19

1%

0.04

9%C

ompu

ting

�1.

085%

0.

162%

1.

007%

0.

147%

1.

005%

0.

158%

0.

904%

0.

138%

Dru

gs a

nd m

edic

al�

2.88

2%

0.31

8%

2.43

2%

0.29

1%

2.45

4%

0.29

2%

2.22

2%

0.27

5%E

lect

rica

l�0.

079%

0.

072%

0.

118%

0.

048%

0.

119%

0.

052%

0.

106%

0.

044%

Mec

hani

cal�

�0.

157%

0.05

7%

�0.

077%

0.

029%

�

0.08

5%

0.03

3%

�0.

069%

0.

027%

Yea

r du

mm

ies

No

No

No

Yes

Scal

ed R

-squ

ared

0.13

00.

222

0.21

70.

229

Log

like

lihoo

d�

3980

.69

�35

56.9

6�

3583

.02

�35

28.1

5

Not

es:

Coe

ffici

ent

esti

mat

es in

ital

ics

are

not

sign

ifica

nt a

t th

e 1

per

cent

leve

l.*

The

sam

ple

ofno

n-hi

ghly

-cit

ed p

aten

ts is

a 1

0 pe

r ce

nt s

ampl

e of

all p

aten

ts t

hat

have

tw

o or

mor

e ci

tati

ons.

The

ave

rage

pro

babi

lity

ofbe

ing

high

ly c

ited

in t

he s

ampl

e is

0.7

per

cen

t.**

The

exc

lude

d cl

ass

is o

ther

tec

hnol

ogie

s.**

* E

stim

ated

der

ivat

ive

ofpr

obab

ility

wit

h re

spec

t to

inde

pend

ent

vari

able

.For

dum

my

vari

able

s (�

),th

e di

scre

te c

hang

e in

pro

babi

lity

from

0to

1.

413

Tab

le 1

4.8

Hig

hly

cite

d pa

tent

s w

ith

high

gen

eral

ity,

clas

s gr

owth

and

cit

ing

pate

nt g

ener

alit

y

Pat

ent

Num

ber

App

lica-

In

vent

or

Ass

igne

eP

aten

t de

scri

ptio

nB

y U

S B

y IP

CG

ener

alit

ynu

mbe

rof

Cit

esti

on y

ear

stat

e,cl

ass

By

tech

B

y in

d.of

By

coun

try

sub-

man

u-in

dust

ryca

tego

ryfa

ctur

eof

use

3624

019

129

1970

IL,U

S N

alco

P

roce

ss fo

r R

apid

ily

0.84

60.

907

0.65

90.

856

0.79

8C

hem

ical

Dis

solv

ing

Com

pay

Wat

er-s

olub

lePo

lym

ers

3636

956

181

1970

DE

,US

Eth

icon

,Inc

.Po

lyac

tide

sut

ures

0.84

10.

696

0.82

5(a

bsor

babl

e)38

4219

412

519

71N

J,U

SR

CA

Info

rmat

ion

reco

rds

0.84

30.

730

0.83

0C

orpo

ra-

and

reco

rdin

g ti

onpl

ayba

ck s

yste

m

ther

efor

e (v

ideo

di

sc)

3956

615

178

1974

CA

,US

IBM

Tra

nsac

tion

exe

cuti

on0.

801

syst

em w

ith

secu

re

data

sto

rage

and

co

mm

unic

atio

ns45

2864

318

619

83T

X,U

SF

PD

CSy

stem

for

0.88

00.

797

0.69

6(F

reen

yre

prod

ucin

g pa

tent

)in

form

atio

n in

m

ater

ial o

bjec

ts i

na

poin

t of

sale

loca

tion

4558

413

377

1983

CA

,US

Xer

oxSo

ftw

are

vers

ion

0.82

6m

anag

emen

t sy

stem

414

Tab

le 1

4.8

(con

tinu

ed)

Pat

ent

Num

ber

App

lica-

In

vent

or

Ass

igne

eP

aten

t de

scri

ptio

nB

y U

S B

y IP

CG

ener

alit

ynu

mbe

rof

Cit

esti

on y

ear

stat

e,cl

ass

By

tech

B

y in

d of

By

coun

try

sub-

man

u-in

dust

ryca

tego

ryfa

ctur

eof

use

4575

621

186

1984

PA,U

S C

orpr

a Po

rtab

le e

lect

roni

c 0.

804

0.71

4R

esea

rch

tran

sact

ion

devi

ceIn

can

d sy

stem

the

refo

r46

7265

820

019

86N

J,U

S A

T&

TSp

read

spe

ctru

m

0.84

4w

irel

ess

PB

X47

8369

519

519

86N

Y,U

SG

ener

al

Mul

tich

ip in

tegr

ated

0.82

40.

674

Ele

ctri

c ci

rcui

t pa

ckag

ing

Co

confi

gura

tion

an

d m

etho

d48

2122

018

019

86W

A,U

S T

ektr

onix

Syst

em fo

r an

imat

ing

0.79

6pr

ogra

m o

pera

tion

and

disp

layi

ng t

ime-

base

d re

lati

onsh

ips

4885

717

183

1986

OR

,US

Tek

tron

ixSy

stem

for

grap

hica

lly0.

816

repr

esen

ting

op

erat

ion

ofob

ject

-or

ient

ed p

rogr

ams

4916

441

286

1988

CO

,US

Clin

icom

Inc

Port

able

han

dhel

d0.

912

0.82

40.

827

term

inal

4953

080

181

1988

CA

,US

Hew

lett

-O

bjec

t m

anag

emen

t 0.

794

Pac

kard

fa

cilit

y fo

r

Co

mai

ntai

ning

dat

a in

a co

mpu

ter

syst

em

415

5133

075

210

1988

CA

,US

Hew

lett

-M

etho

d of

mon

itor

ing

0.81

2P

acka

rd

chan

ges

in a

ttri

bute

Co

valu

es o

fob

ject

in

an o

bjec

t-or

ient

edda

taba

se51

5584

722

419

88O

N,C

AM

inic

omM

etho

d an

d

0.87

0D

ata

appa

ratu

s fo

rC

orp

upda

ting

sof

twar

e at

rem

ote

loca

tion

s50

9391

421

719

89IL

,US

AT

&T

Met

hod

ofco

ntro

lling

0.

832

the

exec

utio

n of

obje

ct-o

rien

ted

prog

ram

s53

4763

225

519

89N

J,U

SP

rodi

gy

Rec

epti

on s

yste

m

0.79

6Se

rvic

esfo

r an

inte

ract

ive

Co

com

pute

r ne

twor

kan

d m

etho

d of

oper

atio

n51

1947

520

019

91T

X,U

SSc

hlum

berg

er

Obj

ect-

orie

nted

0.81

0T

echn

olog

yfr

amew

ork

for

Cor

pm

enu

defin

itio

n51

3299

217

819

91N

Y,U

Sun

assi

gned

Aud

io a

nd v

ideo

0.85

60.

764

tran

smis

sion

and

rece

ivin

g sy

stem

(c

ompr

essi

on)

5307

456

173

1992

CA

,US

Sony

In

tegr

ated

mul

ti-

0.68

2E

lect

roni

csm

edia

pro

duct

ion

Inc

and

auth

orin

gsy

stem

416

Tab

le 1

4.8

(con

tinu

ed)

Cit

ing

Pat

ent

Mea

n ci

teM

ean

cite

lag

Sub-

cate

gory

US

clas

sG

row

th o

fC

lass

des

crip

tion

pate

nts

num

ber

lag

rela

tive

to

clas

s (

%)

gene

ralit

yav

erag

e

0.56

636

2401

917

.31.

915

523

15.2

Synt

heti

c R

esin

s or

N

atur

al R

ubbe

rs –

Par

t of

the

Cla

ss 5

20 S

erie

s0.

573

3636

956

21.5

6.1

3260

613

.0Su

rger

y0.

546

3842

194

8.6

�6.

024

369

12.9

Dyn

amic

Inf

orm

atio

n St

orag

e or

Ret

riev

al0.

658

3956

615

18.2

4.4

2270

520

.7D

ata

Pro

cess

ing:

Fin

anci

al,

Bus

ines

s P

ract

ice,

Man

agem

ent,

or C

ost

0.61

645

2864

314

.65.

022

705

14.7

Dat

a P

roce

ssin

g:F

inan

cial

,B

usin

ess

Pra

ctic

e,M

anag

emen

t,or

Cos

t0.

547

4558

413

13.1

3.5

2270

718

.9D

ata

Pro

cess

ing:

Dat

abas

e an

d F

ile M

anag

emen

t,D

ata

Stru

ctur

es0.

615

4575

621

10.4

1.2

5923

513

.3R

egis

ters

0.52

146

7265

89.

91.

421

455

13.2

Tel

ecom

mun

icat

ions

0.60

947

8369

59.

91.

446

257

20.1

Act

ive

Solid

-Sta

te D

evic

es(e

.g.,

Tra

nsis

tors

,Sol

id-

Stat

e D

iode

s)0.

570

4821

220

8.8

0.3

2239

519

.9In

form

atio

n P

roce

ssin

g Sy

stem

Org

anis

atio

n0.

581

4885

717

9.1

0.6

2239

519

.9In

form

atio

n P

roce

ssin

g Sy

stem

Org

anis

atio

n

417

0.53

949

1644

19.

21.

423

345

15.1

Sele

ctiv

e V

isua

l Dis

play

Sy

stem

s0.

593

4953

080

8.0

0.2

2270

718

.6D

ata

Pro

cess

ing:

Dat

abas

e an

d F

ile M

anag

emen

t,D

ata

Stru

ctur

es0.

557

5133

075

6.5

�1.

322

707

18.6

Dat

a P

roce

ssin

g:D

atab

ase

and

File

Man

agem

ent,

Dat

a St

ruct

ures

0.54

151

5584

78.

30.

522

709

30.2

Ele

ctri

cal C

ompu

ters

and

D

igit

al P

roce

ssin

g Sy

stem

s:M

ulti

ple

Com

pute

rs0.

593

5093

914

6.6

�0.

722

395

18.6

Info

rmat

ion

Pro

cess

ing

Syst

em O

rgan

isat

ion

0.54

253

4763

26.

5�

0.9

2270

924

.1E

lect

rica

l Com

pute

rs a

nd

Dig

ital

Pro

cess

ing

Syst

ems:

Mul

tipl

e C

ompu

ters

0.56

651

1947

56.

50.

023

345

26.5

Sele

ctiv

e V

isua

l Dis

play

Sy

stem

s0.

707

5132

992

6.6

0.1

2137

518

.5P

ulse

or

Dig

ital

C

omm

unic

atio

ns0.

539

5307

456

5.9

�0.

123

345

23.7

Sele

ctiv

e V

isua

l Dis

play

Sy

stem

s

patents, all but two were in technologies related to information and com-munication technology (ICT). The remaining two are the oldest (applied forin 1970) and cover a process that is useful in the making of paper, and insewage and waste treatment, and in absorbable sutures for surgery. All butone of the patents cover US inventions, five from California, three fromNew Jersey and the remainder from a number of other states. The sole excep-tion comes from a Toronto-based company. All but one of the patents wereassigned to corporations at the time they were taken out; the exception wasa patent for a method of compressing audio and video data for transmission.

The ICT-related patents cover a range of technologies: integrated circuitmanufacturing, handheld computers, spread spectrum technology, and soforth. What is noteworthy is the number of patents that relate to Internettransactions (e-commerce) and software development, especially object-oriented programming. Some of the e-commerce patents greatly precedethe actual use of the technology. For example, the celebrated Freeny patent(US4528643, shown in Figure 14.5) was applied for in 1983 and issued in


United States Patent 4,528,643Freeny, Jr. July 9, 1985

System for reproducing information in material objects at a point of salelocation

Abstract

The present invention contemplates a system for reproducing informa-tion in material objects at a point of sale location wherein the informa-tion to be reproduced is provided at the point of sale location from alocation remote with respect to the point of sale location, an ownerauthorization code is provided to the point of sale location in responseto receiving a request code from the point of sale location requesting toreproduce predetermined information in a material object, and the pre-determined information is reproduced in a material object at the point ofsale location in response to receiving the owner authorization code.

Inventors: Freeny, Jr.; Charles C. (Fort Worth, TX)

Assignee: FPDC, Inc. (Oklahoma City, OK)

Appl.No.: 456730

Filed: January 10, 1983

http://www.e-data.com/e-freeny.htm

Figure 14.5 The Freeny patent (4528643)

1985, but has been successfully asserted against such corporations asMicrosoft and Apple almost to the present day.16 The fact that the originaluse for which this patent was contemplated is unlikely to have beenInternet-based e-commerce reminds us to be cautious in our interpretationof the results in Table 14.8: we do not argue that the patents we identify arenecessarily the source of the GPT itself, but we do suggest that by identi-fying them via the subsequent growth and generality in their citing patents,we are observing the symptoms of the diffusion and development of ageneral purpose technology.

Looking at the actual ICT patents in Table 14.8 (rather than at theclasses in which they have been placed) yields the following summary: sevenare related to object-oriented and windows-based software, four to Internetcommerce and communication, three to audio-video applications, two tohandheld computing, and one each to telecommunications and semicon-ductor manufacturing. Thus the specific technologies identified as beingboth general and spawning rapid patenting growth are those related to theeffective use of the computer, especially for interacting and transactingover distance. That is, they are not computing hardware patents per se, butpatents on the technologies that allow a network of computers to operatetogether effectively and to interact with the users of those computers. Thisseems to us to characterise the GPT of the 1980s and 1990s, and we wouldtherefore declare our prospecting exercise a qualified success.17

6 CONCLUSIONS

Many empirical papers close with interpretive cautions and calls forfurther research. This chapter is no exception, but the caution and the callare stronger than usual. For reasons of limited time and computing power,we have not been able to explore the validity and use of the measures wehave constructed as much as we would like and we encourage further workin this area. In particular, all the generality measures suffer from the factthat they treat technologies that are closely related but not in the sameclass in the same way that they treat very distant technologies. Thisinevitably means that generality may be overestimated in some cases andunderestimated in others. One suggestion for future research would be toconstruct a weighted generality measure, where the weights are inverselyrelated to the overall probability that one class cites another class.

A second area of concern has to do with changes in the strategic usesof patents during the two decades we have studied. These changes are notunrelated to the growth in importance of ICT technologies but they mayalso have had an distorting impact on some of the measure we have used.


In particular, as Hall and Ziedonis (2001) have shown, one reason for rapidgrowth in semiconductor patenting after the mid-1980s is a conscious deci-sion on the part of some major firms to build up their patent portfolios inorder to fend off litigation and negotiate cross-licensing agreements. Thistype of strategy has spread throughout the industry and the consequencesfor patenting by firms such as IBM, Lucent and Hewlett-Packard has beenconfirmed by Bessen and Hunt (2004) and Hall (2005). The implication isthat citations to earlier patents in the ICT sector may be growing rapidlypartly because of a strategic shift as well as because the underlying tech-nology is growing in importance and diffusing throughout the economy.Sorting this out from our data will require more attention to the time seriesbehaviour of the indicators, improved generality measures and moredetailed investigation of the firms involved. In the interim, this chapter hasdemonstrated the potential validity of patent-based measures of GPTsand we hope it will spur further investigation into the use of patent data inthis way.


APPENDIX

Table 14.A1 SIC-industry correspondence for generality indices

Hall–Vopel quasi 2-digit industry SIC codes (1987)

01 Food and tobacco 20xx, 21xx 02 Textiles, apparel and footwear 22xx, 23xx, 31xx, 3021, 3961, 3965 03 Lumber and wood products 24xx 04 Furniture 25xx 05 Paper and paper products 26xx 06 Printing and publishing 27xx 07 Chemical products 28xx, excl. 283x, 284x 08 Petroleum refining and prods 13xx, 29xx 09 Plastics and rubber prods 30xx, excl. 3021 10 Stone, clay and glass 32xx 11 Primary metal products 33xx 12 Fabricated metal products 34xx 13 Machinery and engines 35xx, excl. 357x, 358x, 3524 14 Computers and comp. equip. 357x 15 Electrical machinery 358x, 3596, 360x, 361x, 362x, 363x,

364x, 3677, 369x, excl. 3690, 3695 16 Electronic inst. and comm. eq. 3651, 3652, 366x, 367x, excl. 3677, 3678;

3690, 3695, 381x, 382x, excl. 3827 17 Transportation equipment 372x, 373x, 374x, 376x–379x, excl.

3790, 3792, 3799 18 Motor vehicles 371x, excl. 3714; 375x, 3790, 3792, 3799


Table 14.A1 (continued)

Hall-Vopel quasi 2-digit industry SIC codes (1987)

19 Optical and medical instruments 3827, 384x, 386x 20 Pharmaceuticals 283x, 3851 21 Misc. manufacturing 387x, 39xx, excl. 3961, 3965 22 Soap and toiletries 284x 23 Auto parts 371424 Computing software 737x25 Telecommunications 48xx26 Wholesale trade 50xx27 Business services 73xx, excl. 737x28 Agriculture 01xx–09xx29 Mining 10xx, 11xx, 12xx, 14xx30 Construction 15xx–19xx31 Transportation services 40xx–47xx32 Utilities 49xx33 Trade 51xx-59xx34 Fire, Insurance, Real Estate 60xx–69xx35 Health services 80xx36 Engineering services 87xx37 Other services 70xx–99xx and not 73xx, 80xx, 87xx

Table 14.A2 Breakdown by technology sub-category

Sub-category All patents Highly cited patents Ratio

Number Share ( %) Number Share ( %)

Agriculture, Food, Textiles 24 134 0.9 8 1.0 1.17Coating 42 235 1.5 7 0.9 0.58Gas 13 614 0.5 4 0.5 1.04Organic Compounds 116 334 4.2 13 1.7 0.39Resins 96 948 3.5 40 5.1 1.45Miscellaneous 282 717 10.2 69 8.8 0.86Chemical technologies 575 982 20.8 141 18.0 0.86

total

Communications 118 316 4.3 51 6.5 1.52Computer Hardware 90 326 3.3 93 11.8 3.63

and SoftwareComputer Peripherials 24 147 0.9 28 3.6 4.09Information Storage 49 963 1.8 16 2.0 1.13Computer hardware 282 752 10.2 188 23.9 2.34

and software total




Number Share (%) Number Share (%)

Drugs 83 410 3.0 35 4.5 1.48Surgery and Med. 69 344 2.5 164 20.9 8.34

Instruments Genetics 31 794 1.1 24 3.1 2.66Miscellaneous 16 312 0.6 33 4.2 7.13Drugs and med. 200 860 7.3 256 32.6 4.49

instruments total

Electrical Devices 92 508 3.3 1 0.1 0.04Electrical Lighting 44 738 1.6 6 0.8 0.47Measuring and Testing 80 315 2.9 2 0.3 0.09Nuclear and X-rays 40 746 1.5 4 0.5 0.35Power Systems 97 739 3.5 4 0.5 0.14Semiconductor Devices 51 950 1.9 38 4.8 2.58Miscellaneous 66 440 2.4 21 2.7 1.11Electrical technologies 474 436 17.1 76 9.7 0.56

total

Mat. Proc. and 155 200 5.6 16 2.0 0.36Handling

Metal Working 88 661 3.2 11 1.4 0.44Motors and Engines� 102 504 3.7 1 0.1 0.03

PartsOptics 62 832 2.3 4 0.5 0.22Transportation 82 854 3.0 0 0.0 0.00Miscellaneous 143 849 5.2 16 2.0 0.39Mechanical 635 900 23.0 48 6.1 0.27

technologies total

Agriculture, Husbandry, 59 793 2.2 19 2.4 1.12Food

Amusement Devices 28 095 1.0 0 0.0 0.00Apparel and Textile 50 477 1.8 0 0.0 0.00Earth Working and 40 857 1.5 0 0.0 0.00

Wells Furniture, House 57 362 2.1 0 0.0 0.00

FixturesHeating 38 146 1.4 0 0.0 0.00Pipes and Joints 25 198 0.9 3 0.4 0.42Receptacles 58 616 2.1 32 4.1 1.93

NOTES

1. A preliminary version of this chapter was presented at the conference ‘New Frontiers inthe Economics of Innovation and New Technology,’ held in honour of Paul A. David atthe Accademia delle Scienze, Torino, 20–21 May, 2000. We are grateful to participantsin that conference, especially Paul David, John Cantwell, Giovanni Dosi, OveGranstrand and Ed Steinmueller, for comments on the earlier draft. The first authorthanks the Centre for Business Research, Judge Institute of Management, University ofCambridge for hospitality while this version was being written.

2. Griliches (1990); Pavitt (1988).3. In the US environment, this statement is increasingly less true, although the converse,

that not all patent subject matter is innovative, may be becoming more true.




Number Share (%) Number Share (%)

Miscellaneous 239 537 8.7 22 2.8 0.32

Other technologies 598 081 21.6 76 9.7 0.45total

All technologies 2 768 011 100.0 785 100.0 1.00total

Table 14.A3 Correlation matrix for generality indices (N780)

1 US class 2 IPC 3 US 4 Industry 5 Industry sub-category of of

manufacture use

Generality 1 1.000(US class)

Generality 2 0.555 1.000(IPC)

Generality 3 0.621 0.523 1.000(US subcategory)

Generality 4 0.238 0.590 0.599 1.000(industry ofmanufacture)

Generality 5 0.143 0.627 0.389 0.632 1.000 (industry of use)

4. As a general rule, the USPTO does not classify patents individually into IPC classes, butrelies on a map based on US classes and sub-classes to determine them. This is not ideal,but does mean that they incorporate some subclass information.

5. For recent evidence on this point, see Harhoff et al. (1999) for the results of a surveyof patent owners, and Hall et al. (2005) for results showing that the marketvalue–citation relationship is highly non-linear, with firms owning highly cited patentssubject to very large premia, as well as a graph showing the frequency distribution ofpatent citations.

6. This description of the meaning of patent citations is drawn from Hall et al. (2002).7. See Jaffe et al. (2000) for evidence from a survey of inventors on the role of citations in

both senses.8. The USPTO has developed over the years a highly elaborate classification system for the

technologies to which the patented inventions belong, consisting of about 400 main(three-digit) patent classes, and over 120 000 patent sub-classes. This system is beingupdated continuously, reflecting the rapid changes in the technologies themselves.Trajtenberg, Jaffe and Hall have developed a higher-level classification, by which the400 classes are aggregated into 36 two-digit technological sub-categories, and these inturn are further aggregated into six main categories: Chemical, Computers andCommunications (C&C), Drugs and Medical (D&M), Electrical and Electronics (E&E),Mechanical, and Others.

9. Note that generality is not defined if a patent receives no citations, and is zero by con-struction when a patent receives only one. We have omitted such patents in the tables andgraphs shown in this chapter. They comprise about one-quarter of all patents in oursample.

10. A typical citation lag distribution is shown in Figure 14.2. This curve was estimated fromthe observed data using the methodology of Trajtenberg et al. (1997). See Appendix Dof Hall et al. (2002) for details. From the graph it appears that over half the citations evermade are made in the first six years since the cited patent’s application date.

11. Slightly fewer than half the patents granted between 1967 and 1999 are assigned to UScorporations that we can identify (see Hall et al., 2002). However, many of these are inmultiple industries so the primary industry assignment may not be relevant for the par-ticular patent or citation that we are using.

12. The actual industry classification we use was developed by Hall and Vopel (1997) froman earlier classification used by Levin and Reiss (1984). It is based on four-digit SICsaggregated up to a level that is coarse enough to include most, but not all, whole firmsin the US manufacturing sector. We augmented this industry list with ten industries forthe non-manufacturing sector. See Appendix Table A1 for details.

13. We have omitted patent classes with fewer than 10 patents at the end of each period.14. Unlike the case of generality measures, the mean citation lag is linear in the citation

counts and therefore not a biased estimate, conditional on the total number of citations.It is, however, truncated at the end of the period, but this truncation affects both highlycited and non-highly cited patents equally.

15. The sample used is the 780 highly cited patents plus the four per cent random sample ofpatents with at least one cite shown in Table 14.2. The fact that we use a random samplerather than a population affects the constant term in this probit regression, so we do notreport it. The other coefficient estimates will not be affected by this procedure, althoughthe interpretation of the changes in probability will depend on the average probability inthe sample used.

16. This patent is currently owned by E-Data Corporation and was agressively asserted bythat company in the US beginning in 1996 (http://www.prpnet.net/7604.html).

17. Note that the industry of manufacture and industry of use measures do not identify thesoftware and Internet patents as GPTs, for reasons discussed earlier: they have beenobtained using a concordance that did not really admit these as patentable areas.


BIBLIOGRAPHY

Abramovitz, Moses and Paul A. David (1996), ‘Technological change, intangibleinvestments and growth in the knowledge-based economy: the US historical per-spective’, in D. Foray and B.A. Lundvall, Employment and Growth in theKnowledge-based Economy, Paris, OECD.

Bessen, James and Robert M. Hunt, (2004), ‘An empirical look at software patents’,Philadelphia: Federal Reserve Bank Working Paper No. 03–17/R.

Bresnahan, Timothy J. and Manuel Trajtenberg (1995), ‘General purpose tech-nologies: “Engines of growth”?’, Journal of Econometrics, 95, 83–108.

David, Paul A. (1990), ‘The dynamo and the computer: an historical perspective onthe modern productivity paradox’, American Economic Review, 80, 355–61.

David, Paul A. (1991), ‘Computer and dynamo: the modern productivity paradoxin a not-too-distant mirror’, in OECD (ed.), Technology and Productivity, Paris:OECD. Also Stanford University, CEPR Discussion Paper No. 172 (1989).

David, Paul A. and Gavin Wright (2003), ‘General purpose technologies and prod-uctivity surges: historical reflections on the future of the ICT revolution’, inP. David and M. Thomas (eds), The Economic Future in Historical Perspective,Oxford: Oxford University Press for the British Academy.

Griliches, Zvi (1990), ‘Patent statistics as economic indicators: a survey’, Journal ofEconomic Literature, 28, 1661–707.

Hall, Bronwyn H. (2005), ‘Exploring the patent explosion’, Journal of TechnologyTransfer, 30, 35–48.

Hall, Bronwyn H. (2002), ‘A note on the bias in the Herfindahl based on countdata’, in A. Jaffe and M. Trajtenberg (eds), Patents, Citations, and Innovation,Cambridge, MA, MIT Press.

Hall, Bronwyn H. and Katrin Vopel (1997), ‘Market value, market share, and innov-ation’, NBER, the University of California at Berkeley, and the University ofMannheim. Industry classification available at http://emlab.berkeley.edu/users/bhhall/index.html

Hall, Bronwyn H. and Rosemarie Ham Ziedonis (2001), ‘The determinants ofpatenting in the US semiconductor industry, 1980–94’, Rand Journal ofEconomics, 32, 101–28.

Hall, Bronwyn H., Adam Jaffe and Manuel Trajtenberg (2002), ‘The NBER patentcitations data file: lessons, insights and methodological tools’, in A. Jaffe andM. Trajtenberg (eds), Patents, Citations and Innovations, Cambridge, MA, MITPress. Also NBER Working Paper No. 8498 (October 2001).

Hall, Bronwyn H., Adam Jaffe and Manuel Trajtenberg (2005), ‘Market value andpatent citations’, Rand Journal of Economics, 36 (1), 16–38. Also NBER WorkingPaper No. 7741 (June 2000).

Harhoff, Dietmar, Francis Narin, F.M. Scherer and Katrin Vopel (1999), ‘Citationfrequency and the value of patented inventions’, Review of Economics andStatistics, 81 (3), 511–15.

Helpman, Elhanan and Manuel Trajtenberg (1998a), ‘Diffusion of general purposetechnologies’, in E. Helpman (ed.), General Purpose Technologies and EconomicGrowth, Cambridge: MIT Press, pp. 85–119.

Helpman, Elhanan and Manuel Trajtenberg (1998b), ‘A time to sow and a time toreap: growth based on general purpose technologies’, in E. Helpman (ed.),General Purpose Technologies and Economic Growth, Cambridge, MIT Press,Chapter 3.


Jaffe, Adam, Manuel Trajtenberg and Michael Fogarty (2000), ‘Knowledgespillovers and patent citations: evidence from a survey of inventors’, AmericanEconomic Review, Papers and Proceedings (May), 215–18.

Levin, Richard C. and Peter C. Reiss (1984), ‘Tests of a Schumpeterian model ofR&D and market structure’, in Z. Griliches (ed.), R&D, Patents and Productivity,Chicago, University of Chicago Press for the National Bureau of Economicresearch, pp. 175–208.

Office of Technology Assessment and Forecast (OTAF) (1976), US Department ofCommerce, Patent and Trademark Office, Sixth Report, Washington, DC, USGovernment Printing Office.

Pavitt, Keith (1988), ‘Uses and abuses of patent statistics’, in A.F.J. van Raan (ed.),Handbook of Quantitative Studies of Science and Technology, Amsterdam:Elsevier Science, pp. 509–36.

Rosenberg, Nathan (1976), ‘Technological change in the machine tool industry,1840–1910’, in N. Rosenberg (ed.), Perspectives in Technology, Cambridge,Cambridge University Press, pp. 9–31.

Sakakibara, Mariko and Lee G. Branstetter (2001), ‘Do stronger patents inducemore innovation? Evidence from the 1988 Japanese patent law reforms’, RandJournal of Economics, 32 (1), 77–100.

Silverman, Brian (2002), ‘International Patent Classification – US SIC concord-ance’, http://www.rotman.utoronto.ca/~silverman/ipcsic/ipcsicfiles.ZIP

Trajtenberg, Manuel, Adam Jaffe and Rebecca Henderson (1997), ‘Universityversus corporate patents: a window on the basicness of invention’, Economics ofInnovation and New Technology, 5 (1), 19–50.


15. Equilibrium, epidemic andcatastrophe: diffusion ofinnovations with network effectsLuís M.B. Cabral*

1 INTRODUCTION

It seems that important papers are characterised by long publication lags.Maskin’s famous mechanism design theorem and Holmstrom’s seminalpaper on managerial concerns each took about 20 years to get published.1

Prominent among the list of famous works that remained unpublished fora long time is Paul David’s ‘Contribution to the theory of diffusion’ (David,1969). In that paper, David develops an equilibrium model of new tech-nology adoption and shows how S-shaped diffusion paths reflect hetero-geneity among adopters.

In this chapter, I too focus on the issue of diffusion of innovations,specifically innovations subject to network effects. Like David and others, Istart from an equilibrium model of adopter heterogeneity. However, I willargue that, in the presence of strong network effects, the nature of the adop-tion process is quite different from what was previously characterised.In particular, I show that network effects imply discontinuous adoptionpaths – mathematically speaking, a catastrophe.

In a previous paper (Cabral, 1990), I noted how network effects may leadto discontinuous adoption paths. This chapter goes beyond Cabral (1990) intwo ways. First, I provide a more precise set of conditions under which a cat-astrophe takes place (section 3). Second, I suggest a possible test to distin-guish between alternative theories of new technology adoption (section 5).2

S-shaped diffusion paths, one of the most robust empirical regularitiesfound in the literature, are consistent with a number of theories. I considertwo types: (1) equilibrium diffusion theories based on adopter heterogen-eity, and (2) epidemic theories based on some form of imperfect informa-tion and/or word-of-mouth effects. I start from a model of the first type andadd network effects to it. I then compare it to a model of the second type,also allowing for the possibility of network effects.

427

2 AN EQUILIBRIUM MODEL OF NEWTECHNOLOGY ADOPTION

Consider a new technology available from some time t0. The cost of adopt-ingsuchinnovationisct perperiod.That is,uponadoption,aflowcostct mustbe paid. I assume that ct is decreasing, which reflects gradual, post-invention,technological development, as well as increased competition in supply.

I am particularly interested in the case when the benefit from an innov-ation can be measured by its use. For example, the benefit from having atelephone is proportional to the use that is made of such telephone (or, ifwe also want to consider the ‘stand-alone’ benefit from owning a telephone,then total benefit is a linear function of use). Formally,

Assumption 1 Each adopter’s benefit is proportional to use.

I am also interested in the case when the innovation is subject to networkeffects, that is, the case when adoption benefits are increasing in the numberof adopters. Specifically, suppose that each potential user derives a benefitfrom communicating with a set of other users. Such benefit can only begained if the other users are also hooked up to the network, that is, if theother users have adopted the innovation as well. Suppose, moreover, thatthe event of being part of the list of desirable links is independent of theuser’s type. Then the use of (and benefit from) the innovation is a linearfunction of the number of users.

Assumption 2 Each adopter’s willingness to pay is a linear function of thenumber of users.

I assume that potential adopters are different from each other.Specifically, each potential adopter is characterised by a parameter � � )that measures its willingness to pay for the innovation. Specifically, I assumethat

Assumption 3 Each adopter’s use of the innovation is proportional to theadopter’s type �.

Assumption 4 �N(�,�).

The above assumptions imply that, upon adoption (which I assume isirreversible), an adopter of type � receives a benefit flow given by

�((1�*)�*nt), (15.1)u�t


where nt is the measure of adopters at time t. The parameter * measures theimportance of network effects. In the extreme when *0, benefit is simplygiven by � (stand-alone utility), that is, independent of network size. In theopposite extreme, when *1, stand-alone utility is zero and benefit is pro-portional to network size.

It is straightforward to show that, if type � finds it optimal to adoptbefore time t, then the same is true for type �+��. It follows that, in equi-librium, the set of adopters at time t is given by all types with � greater thansome critical value. Let �+ be such critical value. The equilibriumconditions are then

where f is the density of �. The first equation guarantees that the marginaladopter (type �+) is just indifferent between adopting and not adopting attime t: the left-hand side is the (flow) cost of adoption, whereas the right-hand side is the benefit from adoption. The second equation is a ‘closure’condition: it implies that the network size is the measure of all adopters oftype greater than the marginal type.

The above equations can be combined to yield the following equilibriumcondition:

, (15.2)

where F is the cumulative distribution function of �.For each time t, and the corresponding value of ct, (15.2) can be solved

for . Each value of in turn corresponds to a value of n. Therefore,(15.2) induces an equilibrium correspondence E(t) giving the possibleequilibrium values nt for each t.3 Although the graph E(t) is a continuousand smooth manifold (Cabral, 1990), the equilibrium correspondence can,in principle, be multi-valued. In fact, in the presence of network effects thiswould not be a surprising feature.

3 NETWORK EXTERNALITIES ANDCATASTROPHIC ADOPTION PATHS

If network effects are non-existent or mild, then (15.2) induces a single-valued equilibrium correspondence E(t) and a continuous equilibriumadoption path (EAP). This is illustrated in Figure 15.1. The left-hand side

�+�+

1 � *F(�+) ct

�+

nt ��

�+

f (�)d�,

ct (1 � *)�+ � *�+Nt

�+(t)

Diffusion of innovations with network effects 429

depicts the two sides in (15.2). As can be seen, for each value of t (and ct),there exits a unique solution to the equation. This implies that the equilib-rium correspondence E(t) is single valued and there is a unique EAP,namely ntE(t).

Consider now the case when network externalities are significant. Thiscase is illustrated in Figure 15.2. The left-hand side of the figure shows that,for values of t slightly greater than the one corresponding to RHS2, severalsolutions exist to equation (15.2) (in this figure, equation (15.2) corres-ponds to LHSRHSi). This results in an equilibrium correspondence E(t)that is multi-valued for an interval of values of t. Such equilibrium corre-spondence is depicted in the right-hand side of the figure. A multi-valuedequilibrium correspondence means that there are multiple possible EAPs,in fact a continuum of them.


LHSRHS

�+

LHS

RHS1RHS2RHS3

1

nt

t

E(t)

Figure 15.1 Continuous adoption path with ‘mild’ network externalities(�5, �1, * .25)

Figure 15.2 Strong network externalities and catastrophe adoption path(�5, �1, * .5)

LHSRHS

�+

LHS

RHS1RHS2RHS3

nt

tt +

E(t)n**

n*

Despite this multiplicity of equilibria, it can be readily seen that everyEAP is discontinuous at least for some t. In fact, the most reasonable EAPconsists of following the lower branch of E(t) up to time t and thenjumping from ntn* to ntn**. In the jargon of topology, the point(t, n*) is a catastrophe point: although the equilibrium correspondence iscontinuous, a small increase in t implies a discontinuous change in thevalue of nt.

Figures 15.1 and 15.2 suggest that catastrophes are more likely whennetwork externalities are stronger. My first result formalises this intuition:

Proposition 1 For given � and �, a catastrophe point exists if and only if*�**. Conversely, for given � and *�0, a catastrophe point exists if andonly if ��*.

A formal proof may be found in the Appendix.In terms of actual behaviour, one would normally not expect to observe

a discontinuous EAP like the one suggested above. The shift from n* ton** would very likely occur over a period of time. In fact, if one assumesthat potential adopters make their decisions at time t based on theinstalled base at time t�� (that is, there is an observation lag �); then itcan be shown that, for small �, the adoption path follows closely E(t) upto time t and then moves gradually towards the upper portion of E(t)along a concave path. The result of this process is an S-shaped adop-tion path.4

4 ALTERNATIVE THEORIES

As I mentioned before, the economics literature has produced a largenumber of theoretical explanations consistent with the stylised fact of anS-shaped adoption path. Any claim for the worth of a new explanatorytheory has to be confronted with competing claims.

At the risk of over-generalising, we may classify the different theoriesinto two different categories.5 First, we have the equilibrium diffusion theo-ries based on adopter heterogeneity. These theories are similar to the modelI presented above (or vice versa), except for the inclusion of network effects.In words, these theories explain diffusion as a result of adopter hetero-geneity. Specifically, an S-shaped adoption path results from the shape ofthe cumulative distribution function of the adopters’ type. In particular, thesteep portion of the adoption path corresponds to a high density ofadopters around the relevant valuation parameter.

+

+


The epidemic theories, which are based on some form of imperfect infor-mation, provide an alternative explanation for S-shaped diffusion paths. Inits simplest form, the epidemic theory assumes that potential adoptersbecome aware of the existence of the innovation by word of mouth. Word-of-mouth dynamics are known to have the dynamics of medical epidemics,where the rate of change is proportional to the product of the number ofinfected and not-infected agents. This results in an S-shaped path verysimilar to the empirically observed paths.6

To summarize: an S-shaped adoption path does not require a catastrophe.In fact, it does not even require that there are network externalities at all orthat adopters are heterogeneous, so long as there is imperfect informationof some sort. Therefore, the simple observation of the aggregate rate ofadoption is not sufficient to validate any theory in particular. In the nextsection, I focus on empirical implications that separate the different theories.

5 TESTING BETWEEN THEORIES

As I have argued in the previous sections, there exist many theories that areconsistent with an S-shaped adoption path. Different theories must then bedistinguished by observables other than the diffusion path. My secondresult implies one such test for the case when the intensity of use can beeasily measured.

As I mentioned before, one natural motivation for the utility function(15.1) is the distinction between stand-alone and network-related benefit.Network benefit is proportional to total use, which in turn is proportionalto �nt. Based on this observation, different theories have different implica-tions with respect to the time path of average use, at, given by

where gt(�) is the density of � types who adopt by time t. Specifically, wehave the following results:

Proposition 2 Under equilibrium adoption with heterogeneous adopters,*0 implies that at is decreasing for all t; *1 implies that at is increasingfor low t and decreasing for high t.Proposition 3 Under epidemic diffusion, *0 implies that at is constantfor all t; *1 implies that at is increasing for all t.

Table 15.1 summarises Propositions 2 and 3.

at ��

�+ (1 � * � *nt)�gt(�)d(�)

��+ gt(�)d�

,


6 AN EXAMPLE: FAX MACHINES

In order to test the applicability of my theoretical results, I consider dataon the diffusion of fax machines in the USA. Figure 15.3 plots the value ofthe installed base of fax machines as well as the average use per machine(pages per machine), from the mid-1960s to 1990.

The data seem roughly consistent with the theory of diffusion withheterogeneous adopters and strong network externalities. First, around1987 there was a sharp increase in the installed base, which suggests acatastrophe point in the diffusion of fax machines. Moreover, the timepath of usage per machine seems consistent with the prediction ofProposition 2 for the case *1: the value of at is initially increasing andthen decreasing.


Table 15.1 Summary of Propositions 2 and 3

Theory No net effects Net effects

Heterogeneous adoptersEpidemic

→→ →

→→

Source: Farrell and Shapiro (1992); see also Economides and Himmelberg (1995).

Figure 15.3 Fax machines in the US: installed base and intensity of use

6m

3000

12000

3m

01970 1980 1990

Year

Installed base (IB) Pages per machine (PPM)

PPM

IB

Several qualifications are in order, however. First, the time series inFigure 15.3 is a bit too short to uncover a clear pattern in the evolutionof at. I am currently working on trying to extend this series, in the hope offinding stronger results. One problem with extending the series to the 1990sis that serious consideration must be given to the emergence of email as analternative to fax (including the emergence of electronic faxing).

Second, it should be noted that the value of pages per minute (PPM) isnot necessarily the best measure of at. In fact, it is not uncommon for faxmachines to be shared among several users. For this reason, using PPM asa measure of average use implicitly amounts to assuming that the numberof users per machine remained constant throughout the sample period.However, anecdotal evidence suggests that, as the price of fax machinesdropped over time, so did the number of users per machine. Figure 15.3 istherefore consistent with the epidemic-theory-cum-network-effects story.In other words, the time path of PPM is consistent with an ever increasingpath of at.

7 CONCLUDING REMARKS

Referring to the plethora of theories of S-shaped adoption paths, David(1969: II/13) argues that

It would be possible to find some pair of specifications [of F(�) and ct] that wouldgive rise to a diffusion path of the appropriate shape. Hence, meaningful effortsto distinguish between and verify alternative models of diffusion ought toinvolve some attempt at direct empirical validation of the component specifica-tions, including the postulated characteristics of the distribution f(X).

The analysis in this chapter suggests that the problem is deeper than that:not only there are multiple functional forms consistent with a giventime path; there are multiple theories that would produce identical out-comes. On the positive side, Propositions 2 and 3 suggest that, even withaggregate data only, there are ways of distinguishing between competingtheories.

To conclude, I should acknowledge that the model in this chapter isbased on a somewhat narrow class of innovations, namely communicationtechnologies, where benefits are derived from actual links between poten-tial users. However, as in much of the networks literature, results from thedirect network effects model can be extended to the case of indirect effectsas well. The crucial point is that Assumptions 1 and 2 (or a variationthereof) hold.


APPENDIX

Proof of Proposition 1 In a continuous equilibrium path (no catastrophe),all values ��0 will correspond to the marginal adopter at some time t. Thecondition determining the marginal adopter is

(15.3)

A necessary and sufficient condition for the EAP to be continuous is thatthe RHS cut the LHS from above for every t, that is,

or simply

for all ��0. Substiuting (15.3) for ct and simplifying, we get

(15.4)

The next step is to show that F��f is greater than 1 for some value of �.7

The derivative of F��f with respect to � is

It follows that, for � sufficiently large,

Since, in addition

it follows that there exists a � such that F(�)��f(�)�1. It follows that if *is sufficiently large, then the condition (15.4) is violated for some �.

The second part of the proposition is quite straightforward. For ��,(15.4) reduces to

*�12

� � 1

�√2$� � 1.

lim�→�

(F(�) � �f (�)) 1,

,

,� (F(�) � �f (�)) � 0.

,

,� (F(�) � �f (�)) f (�) � f (�) � ��

�2 �.

*(F(�) � �f(�)) � 1.

* f (�) �ct

�2,

| ,

,� (1 � *F(�)) | � | ,

,� �ct

� � |,

1 � *F(�+) ct

�+


Clearly, for given � and *�0 this condition is violated for � sufficientlyclose to zero.

Proof of Proposition 2 Under heterogeneous adopter diffusion, per capitause is given by

Given that � is normally distributed, we have

If *0, then at�2fi(�+)/ (1 - F(�+)), which is increasing in �. Since �+ isdecreasing in t, it follows that at is decreasing in t. If *1, then at�2f(�+).Since � is decreasing in t, the value of at follows the value of f(�+t): increas-ing for low values of t, decreasing for high values of t.

Proof of Proposition 3 Under epidemic diffusion, the population ofadopters at time t is a representative sample from the population of poten-tial adopters. Moreover, type �’s use at time t is given by (1*)��*�nt.Together, these facts imply that

at(1*)��*�nt.

If *0, then at�, which is constant over time. If *1, then at�nt,which is increasing over time.

NOTES

* My first work in this area dates back to a second-year student paper at Stanford (in 1986)that eventually led to Cabral (1990). I am grateful to Paul David, Brian Arthur, and manyothers who encouraged me on this line of research. Regretfully, I alone remain respon-sible for all the shortcomings of this and previous related papers.

1. Both appeared in a recent issue of the Review of Economic Studies.2. Cowan (2005), also included in these proceedings, considers a model of cycles in art

appreciation and prices. Although I do not consider the possibility of cycles (I assumethat technology adoption is irreversible), our models share the prediction that, over time,consumers will shift between equilibria.

3. Specifically, E(t) is obtained by solving

,

where G(�) is the inverse of F(�).

1 � *(1 � nt) ct

G(1 � nt)

at ((1 � *) � *(1 � F(�+)))��2

f(�+)1 � F(�+)

.

at ��

�+((1 � *) � *(1 � F(�+)))�f (�)d�

��+f (�)d�


4. Cf Cabral (1990). Notice I do not call this an equilibrium adoption path since, for aperiod of time, the system is in disequilibrium, gradually moving from a low-adoptionto a high-adoption static equilibrium.

5. See Geroski (1999) for a recent survey.6. Jensen (1982) proposes an interesting variant of the epidemic theory based on imperfect

information about the value of the innovation. Specifically, he assumes that adoptersdiffer with respect to their prior beliefs that the innovation is profitable (other than this,he assumes adopters are identical). He shows that, starting from a uniform distributionof prior beliefs, an S-shaped equilibrium adoption path is obtained.

7. One side-result of the above condition is that, for given � and �, the EAP is continuousif * is sufficiently small.

REFERENCES

Cabral, Luís (1990), ‘On the adoption of innovations with “network” externalities’,Mathematical Social Sciences, 19, 299–308.

Cowan, Robin (2005), ‘Waves and cycles: explorations in the pure theory of pricefor fine art’, chapter 7 in this volume.

David, Paul A. (1969), ‘A contribution to the theory of diffusion’, Research Centerin Economic Growth Memorandum No. 71, Stanford University.

Economides, Nicholas and Charles Himmelberg (1995), ‘Critical mass and networksize with application to the US FAX market’, NYU Stern School of BusinessWorking Paper No. EC-95-11, August.

Farrell, Joseph and Carl Shapiro (1992), ‘Standard setting in high-definition televi-sion’, Brookings Papers on Economic Activity (Microeconomics), 1–77.

Geroski, Paul (2000), ‘Models of technological diffusion’, Research Policy, 29,603–25.

Jensen, Richard (1982), ‘Adoption and diffusion of an innovation of uncertain prof-itability’, Journal of Economic Theory, 27, 182–93.


16. Technological diffusion underuncertainty: a real options modelapplied to the comparativeinternational diffusion of robottechnologyPaul Stoneman and Otto Toivanen

INTRODUCTION

The purpose of this Festschrift is to celebrate the contribution of PaulDavid to various areas of Economics and especially the Economics ofInnovation. It is Paul’s work on technological diffusion that has had themost impact upon our work and especially Paul’s seminal contribution,David (1969), that introduced into the literature the probit-type model ofdiffusion. Prior to this work, diffusion was largely explained as the result ofan information-spreading process, whereas, here, for the first time,differences across firms in the returns that are to be realised from the use ofnew technology was argued to be the central element of the diffusionprocess. This not only emphasised the importance of user heterogeneity inthe process, but also, for the first time, provided the basis by which the stan-dard tools of rational choice, the key elements of economic analysis, couldbe brought to bear upon the analysis of diffusion.

The actual model presented in David (1969) is in many ways very basic.The principle argument is that firms (farms) make the decision whether toadopt a new technology or not on the basis of the cost of adoption relativeto the expected profit gain, the profit gain differing across firms (accordingto firm size). Changes in costs or benefits over time drive the diffusionprocess. Later modifications in the literature have extended the insight bymodifying the adoption condition to introduce an arbitrage as well as aprofitability criterion (not only is it profitable to adopt today but is it notmore profitable to wait until a later date), added expectations (on technol-ogy and prices) to the adoption criterion and also considered the competi-tive environment in which firms operate (for a literature review see

438

Stoneman, 2001). There is a further topic, however, that David did notaddress to any degree, and nor has much of the later literature, and that isthe impact of risk and uncertainty on the diffusion process. In this chapterthe diffusion of new technology in an uncertain environment is consideredvia a probit-type model of the David type.

Investment in general and the adoption of new technology in particularis a process inherently characterised by risk and uncertainty on both thecost and demand side. As an investment, the potential pay-off to the use ofnew technology is uncertain because future market revenues and produc-tion costs will not generally be known with certainty, while at the same timethe cost of adopting a new technology, for example, the price to be paid foracquiring new capital goods at different points of (future) time, is alsounlikely to be known with certainty. It may even be the case that new tech-nology investments are inherently more uncertain than investment ingeneral, for example, (1) new technology may well require additional invest-ments in human capital whereas investments in existing technology maynot and (2) new technology investments may have a higher degree of irre-versibility for there may only be very limited second-hand markets for suchtechnologies.

Although the modelling of technological diffusion has made considerableadvances in the last 10 years (Stoneman, 2001) uncertainty has played littlepart in this. The dominant treatment of uncertainty in the diffusion litera-ture goes back to Mansfield’s (1968) seminal but flawed (see Stoneman,1983) contribution, with a more suitable but dated means variance treat-ment in Stoneman (1981), and a further set of papers by Jensen (for example,Jensen, 1982; 1983). On the other hand, the investment literature has madegood use of models that place uncertainty and irreversibility as the centre-piece of the analysis (for a recent example see Caballero and Engel, 1999)although the empirical literature has placed little emphasis upon analysingthe effects of uncertainty directly (but see Driver and Moreton, 1991;Pindyck and Solimano, 1993; Leahy and Whited, 1996). It is thus somewhatsurprising that to date the diffusion literature has not followed this lead.

In continuing research, for which this chapter should be considered as aprogress report rather than a final outcome, we attempt to fill this gap bybuilding a model of technology diffusion that is based upon an optimisingfirm (project holder) facing an irreversible investment in a new technology ina world of uncertainty and then aggregating up to get an economy-wide levelexpression for the diffusion of the technology. The model is then applied toan inter-country data-set upon the diffusion of robot technologies.1 Ourapproach is largely based upon the real option approaches which were devel-oped and have become dominant in the investment literature (see, forexample, Dixit and Pindyck, 1994). To the best of our knowledge, real option

Technological diffusion under uncertainty 439

methods have rarely been used to analyse diffusion phenomena and, in par-ticular, we know of no empirical work into diffusion based upon a realoptions approach. The approach allows a relatively direct way of incorpo-rating well-defined measures of uncertainty into the estimating equation.

The exercise we undertake has five main potential pay-offs. Initially, ofcourse, it will provide insight in to the impact of uncertainty on thediffusion process, how diffusion may be modelled under uncertainty andthe utility of real option approaches for this purpose. Secondly, by explor-ing the real options approach we are able to provide a new and more rigor-ous theoretical foundation for some widely used empirical models oftechnology diffusion and show that these can be thought of as the aggre-gate implications of individual optimising behaviour where an ad hocstructure has been placed upon the adoption hazard. Third, although thereis some literature that compares diffusion across countries, that literature isquite sparse and our comparative international analysis also makes a rele-vant contribution in this area. Fourth, there has been a long-running argu-ment in macroeconomic policy discussion as to whether macroeconomicstability encourages investment (see, for example, Caballero, 1991; Driverand Morton, 1991; Pindyck and Solimano, 1993), and the work reportedupon here provides further evidence for this policy debate in that aby-product of the model that we construct gives some insight into whetherthe rate of technological change in an economy as measured by the speedof adoption of a particular technology is related to several indicators ofmacroeconomic stability. Fifth, on the basis of the empirical results we areable to conduct policy experiments to explore the effects of changing thevolatility of the environment in which firms make investment decisions.

In the next section we provide an introduction to robot technology. In thethird section we develop the real options model of technology choice andthen from that model construct an aggregate diffusion curve for an economy.In the fourth section we discuss the data used to test the model. In the fifthsection we discuss econometric issues. In the sixth section we present theresults of the estimation and discuss their implications. The work that wereport upon in this paper is still ongoing, thus to a considerable extent theresults that we present as this stage should be considered preliminary. Wethen use the estimated parameters in the sixth section to conduct the policyexperiments in the seventh section. The eighth section contains conclusions.

ROBOT TECHNOLOGY

TheInternationalFederationof Robotics (IFR)hascreatedan internationalstandard (ISO TR 8373) for robot technology with robots customarily


classified into standard and advanced robots and by the application area.International statistics2 on the use of industrial robots are compiled by theUnited Nations (UN) and the IFR, however, for the period 1981–93 onlyaggregate data is available. We use this data to measure the inter-countrystockof robots.These statistics cover28countries,16of whichwere includedin our actual sample,3 although due to gaps in the data our panel is unbal-anced, yielding a total of 161 observations. The 16 countries are listedtogether with summary statistics in Tables 16.2a and 16.2b (see below); thetables also report the observation period for each country. It is estimated(the statistics are not water-tight) that there were some 610 000 robots in useworldwide at the end of 1993. The rate of growth of this stock has been fast,ranging from the 16–23 per cent per annum. recorded at the turn of thedecade to latest levels of 6–8 per cent per annum. The average yearly changein the robot stock varies between almost 30 000 in Japan to 41 in Norway.

As evidence that robot adoption is significant, but different from aggre-gate manufacturing investment, in 1993 robot investment amounted to12 per cent of total machine tool investments in the USA, to 11 per cent inboth Germany and the UK, and to 6 per cent in France. Japan is by far thelargest user of robots whether measured by absolute (some 60 per cent ofworld stock) or relative numbers (in 1993 Japan had 264 advanced robotsper 10 000 employees in manufacturing when the country with the secondhighest density (Singapore) had 61). Robots are used in several industries,and perform a variety of tasks. Worldwide, the traditional ‘vehicle’ fordiffusion of robots has been the transport equipment industry (especiallythe motor vehicle industry), but lately, for example in Japan, the electricalmachinery industry has adopted more robots. The major application areasare welding, machining and assembly, with the leading application areavarying over countries. Although it would certainly be beneficial to havemore detailed country-level data on the composition of the robot stock,and its use, such country-level idiosyncrasies are to a great extent constantover time, and can be captured by country-level controls in the economet-ric model.

Intuition suggests that investments into a new technology such as robotsmay be more volatile than aggregate manufacturing investment. The reasonis that as the degree of technological uncertainty is greater (leading to agreater variance of future revenue streams), such investments are more sen-sitive to changes in other variables (such as prices and interest rates) thataffect investment decisions. To check whether this is the case, we comparedmanufacturing investment volatility to that of robot diffusion volatility inthe OECD countries of our sample (thereby excluding Singapore,Switzerland and Taiwan, for which it proved difficult to obtain a compara-ble investment series at this point). To be able to compare two different


forms of investment that are measured differently (manufacturing invest-ment in monetary terms, robots in units), we employed the coefficient ofvariation. Comparing for each of the 13 countries the two series over therobot observation period (see Table 16.1) we found that for all countries,the coefficient of variation of robot diffusion was larger than that of aggre-gate manufacturing investment (which includes robots, thereby biasingthese figures upwards). The mean ratio of robot coefficient of variationto that of manufacturing investment was 3.60, with a minimum of 1.26(Sweden) and a maximum of 5.74 (France). We read this as evidence thatrobot diffusion is indeed more volatile than aggregate manufacturinginvestment, as discussed in the Introduction.

THE MODELLING FRAMEWORK

In this section we develop a theoretical model of technology adoptionunder uncertainty that may be applied to our data upon robot diffusion.The initial unit of analysis is the project (as opposed to the more usual unitof the firm) where a project corresponds to investment in one unit of robottechnology. As our data does not contain observations upon individualprojects, we aggregate up from the project level to the country level, themodel predicting the number of robots installed in each country at timet which corresponds to the data available.


Table 16.1 Volatility of robot adoption in comparison to manufacturinginvestment

Country Ratio of robot diffusion coefficient of variation to manufacturing investment coefficient of variation

Australia 4.15677Austria 5.43715Denmark 3.340727Finland 1.91579France 5.724699Germany 2.729275Italy 4.548866Japan 1.73117Norway 2.944321Spain 2.780212Sweden 1.256557UK 1.422169USA 5.592314

The Adoption Timing Decision

We employ a real options approach adapted from a model proposed byDixit and Pindyck (1994: 207–11). Defining Pjt as the cost of a unit of robottechnology in country j in time t, assumed the same for all projects i, andRijt as the annual gross profit increase generated by project i in country j intime t, we assume that both Pjt and Rijt are uncertain but exhibit geometricBrownian motion such that,

dRijt/Rijt�Rijdt��Rijdz (16.1)

and that

dPjt/Pjt�Pjdt��Pjdz (16.2)

where dz is the increment of a standard Wiener Process. Note that the �smeasure the expected rates of growth or drift of the variables over timewhile the �s show the volatility or uncertainty attached to the variable. It isthrough the � terms that in this chapter we explore the effects of uncer-tainty on the diffusion process.

We assume that Pjt and Rijt are independent covariates with zero covari-ance. We also assume that Pjt and Rijt are independent of the (existing)number of robots in use and actual dates of adoption. In terms of thediffusion literature, this is the same as assuming that there are no stockand order effects in the diffusion process (see Karshenas and Stoneman,1993). The resulting model thus falls within the class of probit or rankmodels of diffusion introduced by Paul David and in which different ratesof adoption across countries will reflect the different characteristics ofthose countries, the characteristics being exogenous to the diffusionprocess.

Defining rjt as the riskless real interest rate in country j in time t, a projecti, which has not previously been undertaken, will be undertaken (started)in time t if

Rijt/Pjt�"ijt��ij(rjt��Rij)/(�ij�1) (16.3)

where "ijt is the threshold ratio of profit gains relative to the cost of acqui-sition above which project i will be undertaken in time t and below which itwill not (this being allowed to be time dependent for generality), and �ij isthe larger root of the quadratic equation

0.5 �(�Rij2 ,��Pj

2 ) � �ij�(�ij�1)�(�Rij��Pj) ��ij�(�Pj�rjt)0 (16.4)


enabling us to write (16.3) more generally as (16.5).

Rijt/Pjt�"ijt�F(rjt, �Rij2 , �Pj

2 , �Rij, �Pj) (16.5)

Following Dixit and Pindyck we assume that rjt��Rij�0 and rjt��Pj�0which implies that �ij�1, and then using (16.3) and (16.4) we may deducethat F2�0, F3�0, F5�0 but the signs of F1 and F4 are indeterminate. Thus,increases in uncertainty (�Rij

2 , �Pj2 ) will lead to increases in the threshold

value "ijt and an increase in the drift rate of increase of robot prices (�Pj)will reduce the threshold, however, the impact of an increase in the interestrate and the drift (rate of increase) of profit gains (�Rij) are indeterminate.Basically, although the direct impact of increases in the latter two parame-ters on (rjt��Rij) is clear, their impact on �ij/(�ij�1) is of the opposite sign.If there is no uncertainty so that �Rij�Pj0 then (16.3) collapses to

Rijt/Pjt�"ijtrjt��Pj (16.6)

which is the standard intertemporal arbitrage condition for adoption of anew technology in time t (see Ireland and Stoneman, 1986). In this case theimpact of changes in rjt and �Rij can be clearly signed as positive and zerorespectively.

The main result thus generated is that if �Rij2 and/or �Pj

2 increase, then thethreshold value for the ratio of profit gains relative to the cost of acquisi-tion, "ijt, will also increase. This does not necessarily imply however thatincreases in uncertainty will lead to a slower diffusion process. To illustratethe point we consider the probability that a randomly chosen project i willbe undertaken in time t given the adoption rule (16.3) and the paths of Rijtand Pjt as given by (16.1) and (16.2). The basis of the argument is that giventhe geometric Brownian motion assumed for P and R, the probability intime t that the realised or observed ratio of R to P will exceed any givenvalue will also be dependent upon on �Rij

2 and �Pj2 and in such a way that

one cannot predict a priori whether the probability that (16.5) will be metwill be increasing or decreasing in �Rij

2 and �Pj2 .

From Dixit and Pindyck (1994: 71–2) we know that if a variable x followsgeometric Brownian motion with drift at the rate � and volatility � then xat any time will be lognormally distributed with mean x0exp(�t) and vari-ance x0

2exp(2�t)[exp(�2t�1] where x0 is the value of x at time zero.Aitchinson and Brown (1957) show that the ratio of two lognormally dis-tributed variables is also lognormally distributed with a mean equal to thedifference of the two means and a variance equal to the sum of the two vari-ances. Given (16.1) and (16.2) above, we may then state that Rijt/Pjt in timet will be lognormally distributed with mean �ijt(R0/P0) �exp[(�Rij��Pj)t]


and variance �ijt2 �ijt

2 �[exp(�Rij2 ��Pj

2 )�1]. Clearly, the greater is �Rij2 and/or

�Pj2 the larger will be the variance of Rijt/Pjt in time t.Using L to represent the lognormal distribution function, we may then

state that the probability that a project i will be initiated in time t is givenby the probability that Rijt/Pjt�"ijt, written as Pr{Rijt/Pjt�"ijt} where

Pr{Rijt/Pjt�"ijt}1�L("ijt |�ijt, �ijt2 ) (16.7)

From the properties of the lognormal we may immediately state, ceteresparibus, that Pr{Rijt/Pjt�"ijt} is decreasing in "ijt, thus (1) the lower isthe threshold return to adoption the greater is the probability of adoption,(2) the greater is �ijtthe greater is the probability of adoption, but (3) theimpact of changes in �ijt

2 depends upon the value of "ijt and cannot besigned a priori.

The implications for the date of adoption of changes in �Rij2 and �Pj

2 arethus unclear in that an increase in these volatility parameters leads to anincrease in "ijt which reduces the probability of meeting the adoption criter-ion in time t, but it also leads to an increase in the variance of Rijt/Pjt whichmay impact positively or negatively upon the probability of adoption intime t. One cannot predict a priori the overall effect. One cannot, thus,make any predictions of the impact of increases in uncertainty upon thedate of adoption of a new technology. In addition, although for a given "ijtone may predict that the greater is �ijt the greater is the probability of adop-tion, this does not imply that increases in (�Rij��Pj) will lead to increasedadoption. This is because an increase in (�Rij��Pj) will change both "ijt and�ijt

2 as well as �ijt and we are unable to predict the overall impact of suchchanges.

Given that the modelling framework cannot predict a priori whetherincreases in volatility (or drift) of the cost and returns of adopting newtechnology will have a positive or negative impact upon the date of adop-tion, there is little advantage in attempting to empirically explore the rela-tionship between diffusion and uncertainty in terms of the impact ofuncertainty on a diffusion curve expressed as a function of time. In theabsence of any clear predictions from the model on the impact of uncer-tainty upon timing, one can never use the empirical results to rejecthypotheses relating timing to uncertainty, in that positive, negative or norelationships are all equally possible.

However, we do have one clear prediction from the theory – that uncer-tainty as measured by the volatility of the costs and returns impacts posi-tively upon the critical adoption threshold ("ijt). It is around this predictionthat the rest of this chapter is constructed. Essentially, we concentrate uponadoption condition (16.3), that is, Rijt/Pjt�"ijt, but rather than replacing


Rijt/Pjt by a stochastic function of time and looking at the probability thatat any time t a project will be started, we instead explore whether thediffusion patterns observed are consistent with the prediction of (16.3) thata project i will be adopted at the first point in time at which the realised ratioof Rijt/Pjt is greater than or equal to "ijt. In essence, therefore, we estimate adiffusion curve not as function of time but as a function of realised Rijt/Pjt.In doing this we model "ijt inter alia as a positive function of volatility andjointly test this hypothesis. It should be noticed, however, that even if wecannot reject that the threshold is positively related to volatility, this will notnecessarily imply that uncertainty slows diffusion when diffusion is consid-ered solely as a function of time. We would instead only be able to say that,given the realised path for Rijt/Pjt, diffusion has been slowed by uncertainty.

The basic model of firm behaviour we therefore have in mind is that for aproject i the costs and returns follow geometric Brownian motion as detailedabove. The decision-maker knows this and is aware of size of the drift andvolatility indicators. From this he/she is able to calculate a threshold valueof the ratio of returns to cost. As time proceeds a realised time path forRijt/Pjt is mapped out. Eventually (or not, for some projects) Rijt/Pjtwillequal or exceed the critical value (for the first time) and the decision-makerwill, upon that criterion being met, commit to the project.

In pursuing this basic framework the obvious issue arises as to how thedecision-maker obtains estimates of drift and volatility. It would, of course,be possible to allow some sort of Bayesian updating process of assumedpriors but this is not really consistent with the modelling detailed above. Inour empirical explorations we have explored the properties of the costs ofand (determinants of) returns to robot technologies in each of the coun-tries in our sample and are unable to reject the hypothesis that these costsand returns follow geometric Brownian motion processes with constantdrift and volatility. Given the constancy of drift and volatility, very fewrealised observations would be required by a decision-maker to get a trueestimate of the drift and volatility or, on average, decision-makers wouldgenerate true estimates at an early date. We thus assume that decision-makers from the earliest days of the diffusion process have such estimatesand these estimates are those that are apparent when estimated from dataupon the diffusion process from its start date to the end date of our sample.

The Hazard Rate

Equation (16.5) above is the condition to be met if project i is to be initi-ated in country j in time t. This condition may be written as (16.8)

Rijt/Pjt�"ijt�0. (16.8)


To move to a model that is appropriate to the macro data available to us,and as we are unable to measure Rijt directly, we assume that Rijt may bewritten as

Rijtexp (aij�&kbkXkjt) (16.9)

where the term &kbkXkjt is a weighted sum of a number k of macroeco-nomic variables that are the determinants of the return to robots on averageacross projects, and the aij term picks up the heterogeneity of returns todifferent projects arising from various real-life factors that, although theymay be known with certainty to individual decision-makers, cannot beincorporated into the model. This may include such factors as firm-specificrisk, expectations and characteristics. If the Xkjt terms all follow absoluteBrownian motion (with drift and volatility �Xkj and �Xkj), as we can empir-ically show that the variables that we include in Xkjt do, then Rijt will followgeometric Brownian motion so that �Rij and �Rij

2 will also be weighted sumsover k of �Xkj and �Xkj

2 and independent of i. "ijt will thus also be indepen-dent of i and may be written as (16.10)

"ijt�F(rjt, &kbk2�Xkj

2 , �Pj2 , &kbk�Xkj, �Pj). (16.10)

Substituting from (9) into (8) and taking logs then yields the adoption con-dition that

&kbkXkjt� lnPjt� ln"jt�aij�0 (16.11)

Following Karshenas and Stoneman (1993) for each country j, assumingthat the distribution of ai, V(ai), remains invariant across the projects overtime and is distributed independently of &kbkXkjt� lnPjt� ln"jt, one maywrite that the probability of a project i, that has not been started previously,being undertaken in the small time interval {t, t�dt} that is, the hazardrate, is given by (16.12)

hijtProb{&kbkXkjt� lnPjt� ln"jt�aij�0}

V(&kbkXkjt� lnPjt� ln"jt) (16.12)

which is independent of i, and thus may be written as hjt.

The Aggregate Diffusion Curve

Define Sjt as the number of robots installed or projects being undertakenin country j in time t, S*jt as the number of robots that would be installed if


all projects in country j for which robot use were feasible in a technologicalas opposed to an economic sense were being pursued and �j as the rate atwhich past robot investments disappear from the stock on account,perhaps, of the churning of firms (a consequence of our assumption thatR follows a geometric Brownian motion is that it is bounded below to benon-negative and a robot will be operated forever and thus physical obso-lescence is ruled out). Then immediately we may write (16.13)

Sjt(1��j)Sjt�1�hjt{S*jt�Sjt�1} (16.13)

where, as defined above, hjt is the hazard rate, that is, essentially the proba-bility that any project in country j not started by time t will be initiated attime t.

Two issues with respect to this specification merit discussion. First,usually in diffusion models the attrition effect is implicitly allowed to bezero, probably on the basis of an unspecified assumption that either sucheffects do not exist or that any such effects are immediately counteractedthrough greater gross investment in robots. Here, instead, we rigorously testas to whether this ‘depreciation term’ is non-zero. Secondly, S*jt plays arather different role than is often assumed in diffusion analysis. It is oftenassumed in such models that as t tends to infinity that Sjt tends to S*jt. Here,however, this is not necessarily so. Here the diffusion process will only con-tinue as long as hjt is positive and hjt may tend to zero well before SjtS*jt.

Equation (16.13) is a standard logistic diffusion curve (see Stoneman,2001) with the rate of diffusion (setting �j equal to zero) given by the hazardrate, hjt, as determined by (16.12). A useful variant on (16.13) can beobtained using the approximation that (x�y)/y logx� logy which yieldsthe Gompertz diffusion curve (16.14)

lnSjt� lnSjt�1��j�hjt{lnS*jt� lnSjt�1}. (16.14)

It is the Gompertz curve upon which we concentrate below.4 One may notethat diffusion curves such as (16.13) and (16.14) are usually based uponinformation spreading or epidemic arguments (see Stoneman, 2001) but inthis case that is not so. When they are so based, the lagged stock usuallyappears on the on the right-hand side as a proxy for current stock. In thecurrent model the lagged stock appears in its own right.

Detailed Model Specification

There are three remaining issues that need to be settled before the modelcan be operationalised. The first is the modelling of S*jt. We take two


different approaches to this. The simple approach is to assume, given thatrobots are generally only used in the manufacturing sector, that S*jt is pro-portional to manufacturing output, INDjt, in country j such that

S*jtIND�jt. (16.15)

However, it is very likely that, for example, the optimal robot stock variesover countries in a manner that is country specific, time-invariant andunobservable to us.5 This suggests the use of a random effects specificationsuch as

S*jt (16.16)

where, E [-j]0, var(-j)�2�-2 and � is a measure of the variance share of

-j. Such an approach, however, implies that in the estimating equation thecountry specific time-invariant error term is multiplied by hjt. This createsa random coefficients – type (equi-correlated) model that cannot be esti-mated with standard methods. It is thus necessary to resort to a simulationestimator (see below) to estimate this model.

The second issue to address is to list the macro variables that are includedas determinants of Rijt. We initially assume that the potential profit gainfrom installing robots is determined by real gross domestic product, GDP,and the share of investment in GDP. The logic is that demand conditionsas measured by GDP will be a main determinant of profit gains from adop-tion, whereas I/GDP will reflect the general investment climate and, thus,also the profitability of robot adoption. Both variables are expected tocarry positive coefficients. In addition, because the rate of inflation appearsin a number of the investment and uncertainty studies referred to above asa key measure of uncertainty or stability, for the purposes of empiricalexploration we have further included the growth rate of the price level ofmanufactured products, i, as a proxy for the rate of inflation, into theempirical model (although not the actual price level due to difficulties incross-country comparison).

Finally, we need to be specific as to the functional form of the hazardfunction. Given our limited sample size, we decided in favour of parame-terised models and have assumed that the hazard function is exponential(we test the assumption of an exponential functional form against that ofa Weibull hazard). We also assume that the baseline hazard takes an expo-nential functional form (to guarantee non-negative hazards) and is linearin its arguments. Additionally, to simplify the model we assume that F(�)in (16.10), which specifies the threshold level for investment, is linear in itsarguments.

IND�jte

�-j


Bringing together all the parts of the model with these assumptions leadsto the estimating equation (16.17) into which the more complex version ofthe expression for S*jt is incorporated.

(16.17)

The analysis above suggests that d�0, �1�0, �12�0, �2�0, �22�0,�3 � 0, �31�0, �32�0, and ��0 with other parameters unsigned.

DATA SOURCES

In the Appendix we discuss data sources in more detail. Here we provide anoverview. The sources and nature of the international data upon the robotstock was discussed in the second section above. Robot price data were notavailable on an individual country basis, but we located such data forGermany (and Italy). The results that we present are based on the Germandata. The prices were converted into real US dollars using the respectiveconsumer price index and the yearly (average) exchange rates as reportedin the International Monetary Fund’s (IMF’s) International FinancialStatistics Yearbook (IFS).

The data on macro variables (see Tables 16.2a and 16.2b) comes mainly(but not solely, see the Appendix) from Penn World Tables Mark 6(Summers and Heston, 1991). The relevant variables are GDP, the invest-ment share of GDP, the real interest rate (taken from IFS statistics) and therate of inflation measured using the Penn World Tables’ index on prices inmanufacturing. Tables 16.2a and 16.2b reveal that there is considerablevariation in the macro variables over countries. The mean of GDP growthvaries between 11.72 per cent (Taiwan) and 4.4 per cent (Norway). Themean of the real interest rate varies between 0.69 per cent (Switzerland) and7.45 per cent (Taiwan) whereas the lowest mean inflation rate (note that thisis the price level of manufactures, not the consumer price index) is found inUSA (�2.37 per cent) and the highest in Spain (5.3 per cent). There are alsoconsiderable inter-country differences in the volatility measures, but thedifferent measures do not necessarily move in parallel.

The structure of the model requires for estimation purposes not onlyestimates of variables in levels but also estimates of the �js and �js for

ln� Sjt

Sjit�1� �

� d � exp(�0 � �1GDPjt � �11�GDPj � �12�GDPj

� �2I�GDPjt � �21�IGDPj � �22�IGDPj

� �3lnPjt � �31�Pj � �32�Pj � �4rjt � �5i)(�lnINDjt � -j � lnSjt�1) � (1 � �)�jt


451

Tab

le 1

6.2a

Cou

ntry

-lev

el d

escr

ipti

ve s

tati

stic

s

Cou

ntry

GD

PG

DP

GD

P

GD

PIn

vest

men

tP

erce

ntag

e I/

GD

P

I/G

DP

In

flati

on(m

illio

nsgr

owth

gr

owth

grow

th

shar

e of

chan

ge in

grow

thgr

owth

of19

85

perc

enta

ges.

d.G

DP

I/

GD

Ps.

d.U

S do

llars

)(I

/GD

P)

perc

enta

ge

Aus

tral

ia27

1.73

820.

0594

0.08

120.

0347

25.6

375

�0.

0183

�0.

0064

0.08

02�

0.01

2419

85–9

141

.198

00.

0310

2.27

720.

0757

0.08

99A

ustr

ia98

.665

50.

0644

0.08

220.

0279

24.7

909

0.00

350.

0073

0.06

410.

0211

1982

–92

21.5

569

0.01

631.

7335

0.05

400.

1334

Den

mar

k74

.667

40.

0633

0.07

490.

0242

21.2

909

�0.

0015

�0.

0096

0.09

140.

0196

1982

–92

14.6

161

0.01

912.

0364

0.08

430.

1302

Fin

land

67.5

622

0.05

350.

0785

0.04

6730

.518

2�

0.03

15�

0.01

270.

0976

0.00

7219

82–9

213

.252

00.

0506

3.47

130.

0959

0.12

35F

ranc

e94

8.40

070.

0622

0.08

210.

0226

26.8

200

0.00

400.

0021

0.05

490.

0084

1988

–92

86.0

395

0.01

911.

0208

0.05

120.

0742

Ger

man

y93

3.32

970.

0685

0.08

010.

0258

23.9

727

�0.

0011

�0.

0077

0.05

220.

0216

1982

–92

228.

9880

0.01

960.

8855

0.03

320.

1283

Ital

y72

9.10

650.

0629

0.08

430.

0291

24.1

364

�0.

0029

�0.

0083

0.06

550.

0177

1982

–92

158.

3957

0.01

410.

6772

0.03

410.

1044

Japa

n1

739.

0154

0.07

870.

1080

0.03

1934

.427

30.

0094

0.01

550.

0591

0.01

5819

82–9

246

8.72

600.

0166

3.31

420.

0443

0.12

92N

orw

ay61

.633

00.

0461

0.08

020.

0457

27.7

000

�0.

0247

�0.

0114

0.09

650.

0110

1983

–92

7.76

650.

0391

3.72

050.

1011

0.10

29Si

ngap

ore

28.9

984

0.10

640.

1284

0.05

9535

.800

0�

0.00

490.

0121

0.08

11�

0.00

4219

82–9

210

.460

00.

0460

3.69

270.

0808

0.05

40

452

Tab

le 1

6.2a

(con

tinu

ed)

Cou

ntry

GD

PG

DP

GD

P

GD

PIn

vest

men

tP

erce

ntag

e I/

GD

P

I/G

DP

In

flati

on(m

illio

nsgr

owth

gr

owth

grow

th

shar

e of

chan

ge in

grow

thgr

owth

of19

85

perc

enta

ges.

d.G

DP

I/

GD

Ps.

d.U

S do

llars

)(I

/GD

P)

perc

enta

ge

Spai

n38

7.24

300.

0737

0.09

220.

0262

25.3

778

0.02

400.

0117

0.05

460.

0531

1984

–92

87.2

258

0.02

423.

5643

0.06

070.

0948

Swed

en12

8.59

690.

0545

0.07

050.

0273

21.4

909

0.00

45�

0.00

360.

0833

0.00

2719

82–9

224

.377

50.

0210

2.17

410.

0771

0.11

29Sw

itze

rlan

d11

7.10

190.

0606

0.07

500.

0242

30.9

800

0.00

640.

0035

0.08

000.

0226

1983

–92

23.1

568

0.01

902.

6599

0.05

420.

1322

Tai

wan

130.

6258

0.11

720.

1306

0.03

4922

.566

7�

0.02

050.

0214

0.09

550.

0135

1982

–90

44.1

117

0.03

022.

1575

0.09

470.

0831

UK

742.

3100

0.06

050.

0699

0.02

5217

.872

70.

0189

0.00

270.

0771

�0.

0136

1982

–92

153.

1283

0.02

271.

6298

0.06

420.

1082

USA

425

0.05

720.

0577

0.07

490.

0251

22.0

833

�0.

0038

�0.

0023

0.06

38�

0.02

37

1982

–92

137

0.01

400.

0224

0.00

161.

8004

0.07

060.

0016

0.00

420.

0162

Not

es:

Fir

st c

olum

n gi

ves

the

coun

try

nam

e an

d pe

riod

use

d in

est

imat

ion

(whi

ch is

due

to

diff

eren

cing

one

yea

r le

ss t

han

the

obse

rvat

ion

peri

od).

In m

ost

case

s,ro

bot

data

lim

its

the

obse

rvat

ion

peri

od.E

ach

colu

mn

give

s th

e co

untr

y le

vel m

ean

and

s.d.

ofth

e re

spec

tive

var

iabl

e.N

ote

that

,for

exa

mpl

e,th

e G

DP

gro

wth

per

cent

age

and

GD

P g

row

th d

iffer

bec

ause

the

form

er is

cal

cula

ted

over

the

rob

ot s

tock

obs

erva

tion

per

iod

only

,the

latt

er o

ver

a lo

nger

per

iod

(defi

ned

in t

he A

ppen

dix)

.For

the

dri

ft (

grow

th)

and

vola

tilit

y va

riab

les,

no s.

ds a

re r

epor

ted

as t

hey

are

byde

finit

ion

zero

.

453

Tab

le 1

6.2b

Cou

ntry

-lev

el d

escr

ipti

ve s

tati

stic

s

Cou

ntry

Rob

ot s

tock

Cha

nge

inP

erce

ntag

e R

obot

pri

ce

RP

dri

ftR

P s.

d.R

eal i

nter

est

Indu

stri

al(N

o.of

robo

t ch

ange

RP

(19

85

rate

prod

ucti

onro

bots

)st

ock

US

dol

lars

)(p

erce

ntag

e(m

illio

ns

poin

ts)

of19

85

US

dolla

rs)

Aus

tral

ia1

229.

3750

154.

2500

0.15

0698

.951

00.

0832

0.21

396.

3238

59.6

261

1985

–91

402.

7757

50.1

960

0.06

8331

.094

21.

3662

2.11

26A

ustr

ia66

7.63

6415

0.09

090.

3091

90.4

676

0.04

970.

2007

4.66

2913

.979

819

82–9

256

6.18

7510

1.24

960.

0808

29.8

562

0.95

284.

4845

Den

mar

k30

1.45

4548

.454

50.

2216

90.4

676

0.07

600.

2102

7.09

8911

.415

619

82–9

219

3.69

5829

.709

80.

1095

29.8

562

1.42

061.

0931

Fin

land

494.

2727

92.3

636

0.30

9390

.467

60.

0756

0.21

826.

1285

13.8

356

1982

–92

342.

9624

37.4

039

0.17

9729

.856

23.

0889

1.12

85F

ranc

e8

380.

2000

128

9.00

000.

1811

114.

1126

0.08

100.

2144

6.03

7717

6.15

8919

88–9

22

071.

8759

180.

4896

0.06

3829

.469

90.

4904

3.88

55G

erm

any

1753

3.18

183

371.

8182

0.25

8290

.467

60.

0516

0.20

163.

8156

282.

7428

1982

–92

1216

5.35

331

782.

6613

0.08

5629

.856

21.

1037

30.7

243

Ital

y7

573.

3636

151

3.36

360.

3307

90.4

676

0.09

620.

2194

4.66

622.

5276

1982

–92

545

1.81

9068

5.07

990.

1996

29.8

562

1.85

230.

2000

Japa

n16

729

6.63

6429

859.

8182

0.25

5690

.467

60.

0333

0.22

383.

7269

472.

7758

1982

–92

111

333.

3099

1430

6.09

710.

1040

29.8

562

0.74

6465

.712

8N

orw

ay42

2.40

0042

.600

00.

1345

93.0

191

0.08

170.

2150

5.58

719.

1056

1983

–92

128.

8825

19.0

916

0.09

7030

.180

61.

4845

1.21

37Si

ngap

ore

807.

4545

189.

5455

0.54

8790

.467

60.

0457

0.36

365.

7492

7.04

2619

82–9

279

8.02

7023

8.40

230.

4845

29.8

562

5.61

782.

9696

454

Tab

le 1

6.2b

(con

tinu

ed)

Cou

ntry

Rob

ot s

tock

Cha

nge

inP

erce

ntag

e R

obot

pri

ce

RP

dri

ftR

P s.

d.R

eal i

nter

est

Indu

stri

al(N

o.of

robo

t ch

ange

RP

(19

85

rate

prod

ucti

onro

bots

)st

ock

US

dol

lars

)(p

erce

ntag

e(m

illio

ns

poin

ts)

of19

85

US

dolla

rs)

Spai

n1

642.

1111

342.

2222

0.23

2696

.027

20.

0902

0.23

086.

4693

85.6

590

1984

–92

100

6.87

9021

7.36

710.

0287

30.3

798

2.73

465.

5779

Swed

en2

781.

2727

311.

3636

0.12

7090

.467

60.

0877

0.21

995.

8330

26.0

212

1982

–92

110

6.02

9889

.770

00.

0318

29.8

562

3.95

391.

6903

Swit

zerl

and

868.

5000

197.

7000

0.33

3593

.019

10.

0524

0.19

240.

6964

28.3

199

1983

–92

695.

1109

132.

3976

0.13

3030

.180

61.

0631

2.83

59T

aiw

an46

2.44

4414

3.55

560.

7961

88.0

445

0.04

220.

2668

7.44

5853

.477

119

82–9

043

5.53

6211

1.32

170.

8907

32.7

886

1.83

4110

.284

2U

K4

301.

5455

625.

9091

0.21

5190

.467

60.

0576

0.20

454.

6691

146.

1418

1982

–92

215

0.90

5712

7.88

700.

1345

29.8

562

1.46

639.

3619

USA

25

849.

8333

346

8.66

670.

1787

91.6

561

0.06

860.

2447

4.88

6180

0.38

7419

82–9

21

465

9.53

931

890.

8775

0.13

7628

.763

00.

0035

0.01

271.

6758

221.

5590

Not

es:

Fir

st c

olum

n gi

ves

the

coun

try

nam

e an

d pe

riod

use

d in

est

imat

ion

(whi

ch is

due

to

diff

eren

cing

one

yea

r le

ss t

han

the

obse

rvat

ion

peri

od).

In m

ost

case

s,ro

bot

data

lim

its

the

obse

rvat

ion

peri

od.E

ach

colu

mn

give

s th

e co

untr

y le

vel m

ean

and

s.d.

ofth

e re

spec

tive

var

iabl

e.N

ote

that

,for

exa

mpl

e,th

e G

DP

gro

wth

per

cent

age

and

GD

P g

row

th d

iffer

bec

ause

the

form

er is

cal

cula

ted

over

the

rob

ot s

tock

obs

erva

tion

per

iod

only

,the

latt

er o

ver

a lo

nger

per

iod

(defi

ned

in t

he A

ppen

dix)

.For

the

dri

ft (

grow

th)

and

vola

tilit

y va

riab

les,

no s.

ds a

re r

epor

ted

as t

hey

are

byde

finit

ion

zero

.

several variables, in particular, GDP, the investment share of GDP(I/GDP), and robot prices. The measures of growth rates and volatilityfor each of these variables we get by assuming that potential adoptersknow the true growth rates and volatilities over identical periods in eachcountry (the length of the period is 1960–92, but modest changes in thismake little difference), and apply these in their decision-making. Then,relying on the assumption that each of the country-specific time series isa random walk, we estimate the country-specific growth rate and thevolatility using maximum likelihood (we have tested the null hypothesis ofa random walk, and are unable to reject it for any of the series. See theAppendix).

ECONOMETRIC ISSUES

In estimating (16.17) there are several relevant econometric issues. In par-ticular we note the following.

Spurious versus True State Dependence

It is well known that initial conditions (we do not observe robot diffusionfrom T0) may be correlated with time-invariant unobservables(Heckman, 1981). We control for this initially by allowing the lagged stockof robots to be correlated with -j. Note, however, that we do observecountry-wise robot stocks ‘almost’ from the beginning (apart from inFrance): this is likely to reduce the problem considerably. As can be seenfrom Table 16.3, for 12 out of 16 countries, the first observed stock accountsfor less than 13 per cent of final observed stock and for some, substantiallyless. For four countries, the first observed stock is over 20 per cent of thefinal observed stock. These countries are Australia (29.97 per cent, 1stobservation year 1984), France (40.44 per cent, 1987), Norway (26.04 percent, 1982), and Sweden 24.73 per cent, 1982). Together, these countriesaccount for 4 per cent of sample robot stock at the end of 1992. In futurework, it is our intention to use the correction suggested by Heckman, andregress the initial (first observed) stock of robots on observables, but wehave not done so here.

Endogeneity of Price Terms

Although it may be plausible to think that individual firms take prices asgiven when contemplating the adoption of robots, it is less plausible toassume that the price of robots is exogenous at the country level. This


would seem to be the case at least for the countries with higher adoptionrates (Japan being the prime example). We will test for the exogeneity ofrobot prices in future work but have not done so here.

Aggregation

Our current framework implicitly assumes homogenous firms up to the iid.error term in the hazard function. In future work we plan to exploit theknowledge that (1) large firms usually adopt earlier (for example, Rose andJoskow, 1990; Karshenas and Stoneman, 1993) and (2) the firm size dis-tribution is (close to) log-normal, and thus use a simulation estimator thatallows us to aggregate over unobserved firm size differences, where themass of the distribution is given by industrial output, but have not doneso here. This will affect both our hazard rate, and our estimate of S*jt.

Estimation Methods

We use a generalised method of moments estimator (GMM) and amethod of simulated moments estimator (MSM) respectively, for modelswith and without a country-specific time-invariant error term. Both


Table 16.3 Country-level initial stocks and final stocks of robots

Country Initial stock Final stock Ratio of initial 1st year ofof robots of robots to final stock observation/no.

of robots of observations

Australia 528 1 762 0.2997 1984/7Austria 57 1 708 0.0334 1981/11Denmark 51 584 0.0 873 1981/11Finland 35 1 051 0.0333 1981/11France 4 376 10 821 0.4044 1987/5Germany 2 300 39 390 0.0584 1981/11Italy 450 17 097 0.0263 1981/11Japan 21 000 3 49 458 0.0601 1981/11Norway 150 576 0.2604 1982/10Singapore 5 2 090 0.0024 1981/11Spain 433 3 513 0.1233 1983/9Sweden 1 125 4 550 0.2473 1981/11Switzerland 73 2 050 0.0356 1982/10Taiwan 1 2 217 0.0005 1981/9UK 713 7 598 0.0938 1981/11USA 6 000 47 000 0.1277 1981/11

estimators minimise #jt in equation (16.18), thereby matching the predic-tions as closely as possible with observations (ln(Sjt/Sjt�1)), where (16.18)is derived from (16.17) above.

(16.18)

A no equi-correlation assumption is identical to imposing �1 on the data.Currently we use S30, where S is the number of simulations, for theMSM estimator unless otherwise stated.

Instruments

Both estimators require instruments. Berndt et al. (1974) show that optimalinstruments are of the form

A*(xjt)GoD(xjt)+.�1 (16.19)

where � is a vector of parameters, xjt is a data vector,

D(xjt)E [,#jt/,�|xjt] (16.20)

and

.E [#jt#+jt |xjt]. (16.21)

These, however, suffer from reliance on functional form assumptions:if there is functional form misspecification, the resulting estimator is con-sistent but inefficient (Newey, 1990). Newey (1990) has proposed instru-ments that are asymptotically optimal even under functional formmisspecification. Currently we use (untransformed) explanatory variablesas instruments.

Generated Regressors

Our drift and volatility variables are generated, leading to biased standarderrors. In future work we intend to correct these as suggested by Pagan(1984) but have not done so here.

#jt ln� Sjt

Sjit�1� � �

� d � exp(�0 � �1GDPjt � �11�GDPj � �12�GDPj

� �2I�GDPjt � �21�IGDPj � �22�IGDPj

� �3lnPjt � �31�Pj � �32�Pj � �4rjt � �5i)(�lnINDjt � -j � lnSjt�1) � (1 � �)�jt


RESULTS

Our estimation results are presented in Table 16.4. We experimented withestimating models where the constant term (the depreciation rate timesminus one) was constrained to be non-positive. These generally failed toconverge, and the constraint was always binding. We therefore impose theconstraint of a zero depreciation rate in what follows. In column (1) wepresent the GMM estimates where no equi-correlation is assumed. Incolumn (2) we then present the MSM estimates that allow for country-specific unobserved component in the frictionless stock of robots.

From the GMM estimates, the (natural log of) S*jt is given by 0.62 timesmanufacturing output. Turning then to the determinants of the hazardrate, we find first of all a large negative constant term (point esti-mate �1.856 with s.e. 1.184) that together with other estimated coefficientsguarantees that the estimated (baseline) hazard is less than unity for allobservations. Of the level variables, the investment share of GDP (I/GDP)(we here omit country and time subscripts) and GDP obtain positivecoefficients as expected, but of these only the first one is significant. Thecoefficient of the real interest rate carries a negative sign, but is insignifi-cant. The log of robot prices carries the correct, negative, sign and is sig-nificant at the 5 per cent level. Of the drift (growth) variables, the growth inthe investment share and the growth of robot prices (incorrectly) obtainnegative coefficients, but neither is significant. The growth of GDP carriesa precisely measured positive coefficient. All four volatility variables (infla-tion, the volatility of the growth of the investment share of GDP, of GDPand of robot prices, respectively) carry negative coefficients in line with thepredictions of theory. However, of these, only one (volatility of GDPgrowth) is significant at even the 10 per cent level. This would suggest thatuncertainty does not play a major role in the diffusion of robots.

Turning to column (2) one immediately notices that our estimate of �,the variance share of the country-specific time-invariant error term, is stat-istically insignificantly different from unity but significantly different fromzero. This implies that all the variance of the error term is captured by thetime-invariant, country-specific part. A high variance share is not a sur-prise, given the large differences in stocks of robots across countries. Also,the zero variance share of the linear additive iid error carries the interpre-tation that the depreciation rate is uniform, constant, and zero. Note alsothat we find no sign of spurious state dependence being a problem, as theestimated correlation between lagged robot stock and the time-invarianterror term is very small (�0.000023) and very imprecisely measured.

Comparing the point estimates to those in column (2) it is clear that thecoefficients on the drift and volatility variables are most affected, together



Table 16.4 Estimation results

Variable/parameter (1) (2)GMM MSM

Baseline hazardConstant/�0 �1.855681 �6.004694***

1.188489 0.657220GDP/�1 0.000051 �0.003182***

0.000101 0.000248�GDP/�11 30.822349*** 109.904215***

8.895551 6.509909�GDP/�12 �24.749952* �59.492837***

14.490138 8.306346I/GDP/�2 0.050468*** 0.106185***

0.017385 0.012490�IGDP/�21 �19.602290 �97.052057***

12.600056 7.446970�IGDP/�22 �6.496848 �44.250598***

4.589849 3.948424lnP/�3 �0.551984** �0.944563***

0.234123 0.144453�P/�31 �3.953244 �0.003910

4.371913 0.004187�P/�32 �3.121895 �20.161940***

3.678035 2.646841r/�4 �0.011807 �0.008959

0.020639 0.006719i/�5 �0.817569 �1.219140***

0.540648 0.368441

Frictionless robot stocklnIND/� 0.623337*** 94.977696***

0.021108 9.993399Variance share of the country specific — 1.000001***

time-invariant error term/� 0.000022

Correlation coefficient between the — �0.000023country specific error term 0.000051and lagged robot stock/�

nobs 161 161R2 0.6755 —min. dist. 0.4791 31.2525s.e. 0.1789 0.1725

Notes:S30 (Sno. of simulation draws).***sign at the 1 per cent level, **sign at the 5 per cent level, *sign at the 10 per cent level.

with the constant which is now estimated to be �6.00 (p-value .00). We nowfind that of the levels variables, I/GDP and robot prices carry significantcoefficients of the right sign (I/GDP positive, robot price negative), and thatof GDP a coefficient of the wrong (negative) sign. The growth (drift) ofI/GDP carries a negative and significant coefficient. That of robot prices(contrary to expectations) obtains a negative coefficient but this is impre-cisely measured. The coefficient of the drift of GDP carries a positive signand is highly significant. Most importantly, all the volatility variables’coefficients carry coefficients of the expected sign (negative) and are pre-cisely measured. Finally, another large change resulting from allowingequi-correlation is that the coefficient of industrial output obtains a verylarge (and very precisely measured) coefficient.

To check the robustness of the results in column (2), we conducted thefollowing tests. First, we estimated the model (and that in column (1))assuming a Weibull hazard which is a generalisation of the exponential(Weibull generating the exponential when the Weibull parameter p1).Our estimate of the Weibull parameter was p1.2732 (s.e. 1.9769) and wetherefore cannot reject the null hypothesis of an exponential hazard func-tion. Secondly, we estimated the model in column (2) allowing for a non-positive constant (non-negative depreciation rate). Our point estimate ofthe depreciation rate was �.0000 (s.e. .01673), and we thus consider it jus-tified to impose a zero depreciation rate.

POLICY EXPERIMENTS

To quantify and illustrate the effects of policy on investment decisions, weconducted several policy experiments at country level. The central ques-tion, of course, is whether a government policy that yields a more stable(less volatile) environment will lead to substantially faster adoption (thisassertion being based on the assumption that [robot] diffusion is welfareenhancing). It should be noted, however, that the exercises we undertakeare only partial. We allow that changes in volatility impact upon the thresh-old rate of return above which adoption will take place and thus the hazardrate. We have not allowed for the second order effect which means that asvolatility changes the time paths of Rijt and Pjt will also be changing. It ispossible that these second order changes that we do not consider could out-weigh the effects we do consider. We report the results of our experimentsbelow in Table 16.5.

The first column gives the country in question and column (1) reports theestimated hazard rates of adoption (using the preferred equi-correlatedspecification) using the country-specific means of the explanatory variables,


461

Tab

le 1

6.5

Pol

icy

expe

rim

ents

Cou

ntry

Pre

dict

edIn

flati

onG

DP

gro

wth

GD

P v

olat

ility

I/G

DP

gro

wth

I/G

DP

vol

atili

tyR

obot

pri

ce

haza

rdvo

lati

lity

One

s.d.

One

s.d.

10 %

10 %

10 %

10 %

10 %

10 %

10 %

10 %

10 %

10 %

decr

ease

incr

ease

decr

ease

incr

ease

decr

ease

incr

ease

decr

ease

incr

ease

decr

ease

incr

ease

decr

ease

incr

ease

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

(12)

(13)

Aus

tral

ia0.

0047

1.11

580.

8962

0.40

962.

4417

1.42

410.

7022

0.83

771.

1938

1.42

410.

7022

1.53

910.

6497

Aus

tria

0.00

061.

1766

0.84

990.

4051

2.46

841.

4667

0.68

181.

0343

0.96

691.

4667

0.68

181.

4986

0.66

73D

enm

ark

0.00

061.

1720

0.85

320.

4393

2.27

661.

4569

0.68

640.

9854

1.01

481.

4569

0.68

641.

5277

0.65

46F

inla

nd0.

0752

1.16

250.

8602

0.42

202.

3697

1.37

440.

7276

0.73

641.

3580

1.37

440.

7276

1.55

260.

6441

Fra

nce

0.00

001.

0947

0.91

350.

4057

2.46

461.

4479

0.69

061.

0393

0.96

211.

4479

0.69

061.

5408

0.64

90G

erm

any

0.00

011.

1693

0.85

520.

4149

2.41

051.

5027

0.66

550.

9893

1.01

081.

5027

0.66

551.

5016

0.66

60It

aly

0.00

031.

1358

0.88

050.

3959

2.52

611.

4536

0.68

800.

9718

1.02

901.

4536

0.68

801.

5563

0.64

26Ja

pan

0.00

001.

1705

0.85

430.

3052

3.27

611.

5975

0.62

601.

0960

0.91

241.

5975

0.62

601.

5701

0.63

69N

orw

ay0.

0533

1.13

360.

8821

0.41

442.

4133

1.31

550.

7602

0.78

701.

2707

1.31

550.

7602

1.54

260.

6483

Sing

apor

e0.

0021

1.06

800.

9363

0.24

384.

1020

1.88

360.

5309

0.95

371.

0486

1.88

360.

5309

2.08

140.

4804

Spai

n0.

0000

1.12

250.

8909

0.36

302.

7545

1.55

070.

6449

1.26

230.

7922

1.55

070.

6449

1.59

260.

6279

Swed

en0.

0002

1.14

760.

8714

0.46

072.

1705

1.38

320.

7230

1.04

440.

9575

1.38

320.

7230

1.55

780.

6419

Swit

zerl

and

0.00

061.

1749

0.85

110.

4384

2.28

121.

4344

0.69

711.

0639

0.93

991.

4344

0.69

711.

4740

0.67

84T

aiw

an0.

0053

1.10

660.

9037

0.23

804.

2025

2.00

850.

4979

0.81

961.

2200

2.00

850.

4979

1.71

240.

5840

UK

0.00

001.

1409

0.87

650.

4636

2.15

681.

4330

0.69

781.

2011

0.83

261.

4330

0.69

781.

5102

0.66

22U

SA0.

0000

1.01

990.

9805

0.43

902.

2777

1.40

950.

7095

0.96

421.

0371

1.40

950.

7095

1.63

790.

6106

Ave

rage

0.00

891.

1320

0.88

470.

3912

2.66

201.

5089

0.67

060.

9866

1.03

421.

5089

0.67

061.

5872

0.63

40s.

d.0.

0220

0.04

370.

0358

0.06

980.

6390

0.18

450.

0688

0.14

200.

1550

0.18

450.

0688

0.14

370.

0468

Not

es:

The

gro

wth

var

iabl

es r

efer

to

thos

e us

ed in

est

imat

ion.

calculated over the estimation period. As can be seen, the predicted hazardrates vary substantially between countries (even those that are zero at fourdigits – UK, USA, Japan and France – are strictly positive at eight digits).Surprisingly, Japan’s predicted hazard rate is small: this reflects at least par-tially the fact that Japan’s high stock of robots is captured by a high esti-mate of S*jt. Columns (2) to (13) report the ratio of the predicted hazardafter the experiment to the predicted hazard before the experiment (that is,the hazard given in column (1)).

The first experiment was to increase and decrease inflation by one stan-dard deviation from its country mean. Column (2) shows that decreasinginflation by one standard deviation increases the hazard rate of robot adop-tion by between 2 and almost 18 per cent, and on average by 13 per cent.Increasing inflation by the same amount decreases the average hazard rateby over 11 per cent.

For all other variables, the experiment we conducted was a 10 per centincrease or decrease;6 the reason being that as these measures are constantfor each country, no standard deviations are available. We find that decreas-ing GDP growth by 10 per cent leads to a 60 per cent decrease in the hazardrate; mirroring this, an equivalent increase more than doubles the hazardrate in each country, with the highest increases being more than four-fold (Singapore and Taiwan). Changes in the volatility of GDP growthhad pronounced effects, too, with an over 50 per cent increase following a10 per cent decrease in GDP growth volatility. Increasing volatility by thesame amount results in an average hazard rate that is just two-thirds of thepredicted. Looking then at the effects of growth of the investment share ofGDP, we find the effects are moderate. A 10 per cent decrease (increase)only slightly lowers (raises) the hazard rate. The effects of similar relativechanges in the volatility of this growth rate are much more pronounced,leading to a 50 per cent increase and a one-third decrease respectively.Finally, we looked at the effects of changes in the volatility of the stochas-tic process determining robot prices. These are of the same order of mag-nitude as changes in other measures of volatility.

CONCLUSIONS

In this chapter a real options based model of the diffusion of a newprocess technology has been developed and applied to international dataupon the diffusion of robots. There are limits on the predictive power ofthis model, but it indicates that uncertainty will impact negatively uponthe threshold value of the ratio or returns to costs above which adoptionwill take place. Using the volatility of several macroeconomic indicators


as measures of uncertainty, the preliminary empirical analysis reportedhere confirms this prediction. In addition, it is found that there are sig-nificant country-specific effects, that variables in their levels generallyimpact as expected but that the results with the respect to the drift (orgrowth) of relevant variables are weaker. There is still considerable furtherwork to be undertaken upon the estimation, but the results as presentedhere are encouraging and (1) support the use of real options basedmethods, (2) generally confirm the hypothesis that uncertainty detersinvestment in new technology, and (3) support the view that, to somedegree at least, different rates of diffusion in different countries reflect theuncertainties in their macro environments.

Preliminary policy experiments suggest, within their stated limitations,that modest changes in GDP growth and all our measures of volatilityhave pronounced effects on the hazard rate of adoption. As the effects ofsuch changes cumulate rapidly over time, our results suggest – under thehypothesis that more rapid diffusion of new technologies is welfareenhancing – that managing the stability (volatility) of the (macro) envi-ronment in which firms make investment decisions is of paramountimportance.

APPENDIX

Data Sources and Variable Definitions

Robot and robot price data: World industrial robots 1994, United Nations.For robot prices, we use the German prices on the unit value of robot pro-duction. Prices for 1979 and 1978 were calculated using the results of aregression of prices on a constant, years and squared years. The price serieswas deflated using the consumer price index reported in the IFS statistics(the only price series available for all countries). We use prices in constant1985 US dollars for all countries in the regressions. To calculate drift andvolatility for robot prices we use the whole series.

GDP, from Penn World Tables Mark 5.6. Defined as GDPreal GDP percapita (series name CGDP) times population in 000s (POP)/1 000 000. Tocalculate drift and volatility for GDP, we use 1960–92 data, as this was thelongest series available to all countries.

Manufacturing output. Manufacturing in 1986 defined as Man. index ofmanufacturing times GDP (defined as above) times the manufacturing shareof GDP. For other years, the level of manufacturing is calculated from the1986 (1992 for Taiwan, see below) figure using the index of manufacturing.


The index is from International Financial Statistics (for Taiwan fromFinancial Statistics, Taiwan District Republic of China: these are designedto conform to the IFS statistics) and GDP is derived from Penn World Tablesas described above. The share of manufacturing as a percentage of GDP in1986 is from the UN Statistical Yearbook 1995 for all other countries butItaly and Taiwan (not available for these two). For Italy, the manufacturingshare of GDP was calculated as the ratio between figures ‘industria in sensostretto’ (total industry output) and ‘totale’ in Table ‘Tavola 10 – Produzioneal costo dei fattori – Valori a prezzi 1995’ (‘Table 10 – Production at factorcost – 1995 values’). The table can be found on the web page of the NationalInstitute of Statistics of Italy (http://www.istat.it/) as file TAVOLE-PRO-DUZIONE.XLS. The manufacturing share of GDP was calculated for1986. For Taiwan, the source is the file VIGNOF4D.XLS to be found onthe web page of the Directorate General of Budget, Accounting andStatistics of Taiwan (http://www.dgbasey.gov.tw/). The file contains TableH-2 Structure of Domestic Production, in which the manufacturing shareof GDP is reported for 1992–95. We used the 1992 figure.

Investment share of GDP: Penn World Tables (CI). Drift and volatility cal-culations as with GDP.

Price level in manufacturing, Penn World Tables (PI). Inflation in country iin year t defined as inflit ln(PIit)� ln(PIit–1).

Real interest rate. Defined as the difference between the nominal interestrate and inflation, both calculated from indexes from the IFS. Inflation iscalculated from the consumer price index as reported in IFS (it is the onlyprice index available for all sample countries in IFS). We use the moneymarket rate from IFS statistics as the nominal interest rate.

Testing the Assumption of a Random Walk

Our theoretical model assumes that the relevant (country-level) time-seriesare random walks (with drift). In the estimations, we use as explanatoryvariables ML estimates of the drift (growth rate) and standard error ofGDP, investment share of GDP, and of robot prices. To test that our time-series are random walks, we regressed the difference in the series on a con-stant, time trend, and lagged level using data from the period 1960–92 (wealso estimated the model without the time trend, with similar results). Wethen calculated a Dickey-Fuller test. Table 16.A1 summarises our findings,presenting the D-F test value (p-value) of the GDP and I/GDP estimationsfor each country separately (32 degrees of freedom), and ‘finally’ the robotprice tests.


NOTES

1. In an earlier paper, Toivanen et al. (1999), we have also looked at the comparative inter-national diffusion paths of robot technology. That paper contains no formal modelling ofthe impact of uncertainty upon the diffusion process, instead some rather ad hoc theoris-ing was relied upon. Here, with the explicit theoretical model we are able to more preciselydefine appropriate measures and functional forms and also more reliably interpret theresults. In addition, we have undertaken considerably more estimation experiments thanwe had completed at the time the previous paper was published.

2. Our discussion on robots and robot statistics relies heavily on World Industrial Robots1994 (Geneva: United Nations).

3. Twelve countries were excluded for a variety of reasons, for example, for Russia andHungary there was no reliable data on required macrovariables (and even the robot datais suspicious) and the Benelux countries (Netherlands, Belgium and Luxembourg) weresummed together in the robot statistics.

4. The choice of the Gompertz as opposed to logistic formulation is to some degree condi-tioned by the success of this formulation in our earlier work, Toivanen et al. (1999), butexperiments also indicate that empirically the Gompertz model is to be preferred to thelogistic model.

5. Neither of these approaches take any note of changes in the nature of robot technologyover time. In principle this could be catered for by the introduction of trend terms or bythe definition of robot price Pjt as quality adjusted. In this chapter, however, neither routeis explored.

6. Halving the changes to one-half of a standard deviation and 5 per cent respectivelyled to effects that were, to a rough approximation, on average half of those reported here.


Table 16.A1 Dickey-Fuller tests

Country GDP I/GDP

Australia 1.3640 (1.0000) 4.4769 (1.0000)Austria 0.0234 (1.0000) 2.3260 (1.0000)Denmark 1.1187 (1.0000) 2.5019 (1.0000)Finland 2.0427 (1.0000) 2.8554 (1.0000)France 1.0780 (1.0000) 2.2507 (1.0000)Germany 1.5006 (1.0000) 2.9139 (1.0000)Italy 0.9978 (1.0000) 3.0360 (1.0000)Japan 1.1567 (1.0000) 2.0718 (1.0000)Norway 1.9845 (1.0000) 1.9929 (1.0000)Singapore 2.8244 (1.0000) 1.6285 (1.0000)Spain 0.8829 (1.0000) 1.8924 (1.0000)Sweden 2.0859 (1.0000) 2.9694 (1.0000)Switzerland 0.5560 (1.0000) 2.2222 (1.0000)Taiwan 4.6409 (1.0000) 1.3666 (1.0000)UK 1.8332 (1.0000) 2.5871 (1.0000)USA 1.3868 (1.0000) 4.3405 (1.0000)

REFERENCES

Aitchinson, J. and J.A.C. Brown (1957), The Lognormal Distribution, Cambridge,Cambridge University Press.

Berndt, E.R., B.H. Hall, R.E. Hall and J.A. Hausman (1974), ‘Estimation andinference in nonlinear structural models’, Analysis of Economic and SocialMeasurement, 3, 653–66.

Caballero, R-J. (1991), ‘On the sign of the investment–uncertainty relationship’,American Economic Review, 81, 279–88.

Caballero, R-J. and E.M.R.A. Engel (1999), ‘Explaining investment dynamicsin US manufacturing: a generalized (S,s) approach’, Econometrica, 67, (4),783–86.

David, P.A. (1969), A Contribution to the Theory of Diffusion, Stanford Centre forResearch in Economic Growth, Memorandum No. 71, Stanford University.

Dixit, A. and R. Pindyck (1994), Investment Under Uncertainty, Princeton,Princeton University Press.

Driver, C. and D. Moreton (1991), ‘The influence of uncertainty on UK manufac-turing investment’, Economic Journal, 101, 1452–9.

Heckman, J. (1981), ‘The incidental parameters problem and the problem of initialconditions in estimating a discrete time-discrete data stochastic process’, in C.F.Manski and D. McFadden (eds), Structural Analysis of Discrete Data withEconometric Applications, Cambridge, MA: MIT Press.

Ireland, N. and P. Stoneman (1986), ‘Technological diffusion, expectations andwelfare’, Oxford Economic Papers, 38, 283–304.

Jensen, R. (1982), ‘Adoption and diffusion of an innovation of uncertain prof-itability’, Journal of Economic Theory, 27 (1), 182–93.

Jensen, R. (1983), ‘Innovation adoption and diffusion when there are competinginnovations’, Journal of Economic Theory, 29 (1), 161–71.

Karshenas, K. and P. Stoneman (1993), ‘Rank, stock, order and epidemic effects inthe diffusion of new process technology’, Rand Journal of Economics, 24 (4),503–28.

Leahy, J. and T. Whited (1996), ‘The Effect of uncertainty on investment: some styl-ized facts’, Journal of Money, Credit and Banking, 28, 64–83.

Mansfield, E., (1968), Industrial Research and Technological Innovation, New York,Norton.

Newey, W.K. (1990), ‘Efficient instrumental variables estimation of nonlinearmodels’, Econometrica, 58 (4), 809–37.

Pagan, A.R. (1984), ‘Econometric issues in the analysis of regressions with gener-ated regressors’, International Economic Review, 25 (1), 221–47.

Pindyck R.S. and A. Solimano (1993), ‘Economic instability and aggregateinvestment’, NBER Macroeconomics Annual, Cambridge, MA: MIT Press,pp. 259–303.

Rose, N. and P. Joskow (1990), ‘The diffusion of new technologies: evidence fromthe electricity utility industry’, Rand Journal of Economics, 21, 354–73.

Stoneman, P. (1981), ‘Intra-firm diffusion, Bayesian learning and profitability’,Economic Journal, 91 (362), 375–88.

Stoneman, P. (1983), The Economic Analysis of Technological Change, Oxford:Oxford University Press.


Stoneman, P. (2001), The Economics of Technological Diffusion, Oxford, Blackwell.Summers, R. and A. Heston (1991), ‘The Penn World Table (Mark 5): an expanded

set of international comparisons, 1950–1988’, Quarterly Journal of Economics,106 (2), 327–68.

Toivanen, O., P. Stoneman and P. Diederen (1999), ‘Uncertainty, macroeconomicvolatility and investment in new technology’, in C. Driver (ed.), Investment,Growth and Employment: Perspectives for Policy, London, Routledge;pp. 136–60.


PART V

Postscript

17. An appreciation of Paul David’sworkDominique Foray*

Paul’s work is like an extraordinary theatrical performance. It is full ofheroes, including Galileo1 and Edison,2 you will meet primitive tribes – thePomo Indians and the !Kung Bushmen of the Kalihari,3 as well as nymphs4

and strange little monsters (our ‘less reliable American cousins’5). It is abody of work where amazing machines are driven by a frantic will tosurvive – QWERTY,6 the reaper, and the robot.7 There is sex (the estima-tion of ‘objective probabilities of conception from isolated coitus onvarious days of the menstrual cycle’8), blood (or at least weapons like theboomerang9 or near-weapons like nuclear power10), and weird, almostimaginary objects such as the marine chronometer11 or the so called‘N times 384 Kbps standard’.12 You can visit the 1900 Paris Exposition,13

as well as the Crystal Palace Great Exhibition of 1851,14 and you canexplore the secret Venice of 1332.15

Apart from the panda16 and a few horses, there are not really manyanimals, but there are all kinds of networks, made up of strange and dis-turbing people, like the secret society of snow-shovellers,17 or the mysteri-ous Hungarian sect of Zipernowsky, Blathy and Deri.18

It is a body of work with lots of drama and all sorts of accidents: nuclearones,19 a fire in Baltimore,20 even a train derailment.21 And, of course, themillennium bug.22 It is thrilling, fertile, evocative and exuberant, with fren-zied battles giving way to galleries of portraits. It makes one dizzy to con-template it all. Is that the primitive tribe that uses the digital boomerang tohunt with? Did the Grand Duke of Tuscany save us from the year 2000 bug?

But there is a key that allows us to understand the meaning of this greatspectacle, to put all the characters in the right place and to grasp whathappens to the machines and the institutions and why disasters happen.I found this key when I read Robert Musil’s great work.23 Musil writes:‘The course of history is shaped by the action of myriad little causes whichall operate in an unpredictable way.’ Musil labels this notion anti-hero andpetit bourgeois, while remarking that the philosophy of the history of great

471

causes, with its fine intellectual pathos is really only heroic in appearance,because it does not take the facts at face value.

Indeed, says Musil, sometimes a very small movement at the right time isall that is needed to change substantially the end result. The philosophy ofsmall causes, of sensitivity of effects to the very tiniest differences in initialconditions, and the huge contrast possible in the end between minutecauses and enormous effects, introduce vagueness, doubt and a basic ambi-guity into human activity. Musil defends this view of history, which hefirmly shares with Paul. He says it both respects the facts, which other the-ories do not, and leaves us free to take action, and perhaps even be left witha better outcome at the end of the day. Musil concludes:

people say a taste for this way of thinking betrays a crudely mechanical mindand a cynical Philistine attitude. I’d like to say that it contains enormous opti-mism. Because if we do not have to rely any longer on the likes of a spindly scare-crow in a field deciding our destiny, but are just covered with a lot of little weightstangled up like pendants, it’s up to us to tip the balance.24

Our steps are influenced by a tangle of weights, whose balance we maybe able to shift. Abundance of resources is up to us!25 And knowledge open-ness as well!26

This is a key to Paul’s work. ‘It’s up to us to tip the balance’ (ibid.). Thankyou, Paul, for giving us an economics of optimism!

NOTES

* I am grateful to Bronwyn and Ed for editorial assistance.1. ‘Patronage, reputation and common agency contracting in the scientific revolution’.2. ‘The hero and the herd’.3. ‘Information technology, social communication and the wealth and diversity of nations’.4. ‘From market magic to Calypso science policy’.5. ‘European nuclear power and their less reliable American cousins’.6. ‘Clio and the economics of QWERTY’.7. ‘The reaper and the robot’.8. ‘Making use of treacherous advice’.9. ‘Digital technology boomerang: new IPR protections’.

10. ‘Performance-based measures of nuclear reactor standardization’.11. ‘Keeping your bearings on public R&D policy’.12. ‘The ISDN bandwagon is coming’.13. ‘Computer and dynamo’.14. ‘The landscape and the machine’.15. ‘The evolution of intellectual property institutions’.16. ‘Intellectual property institutions and the panda’s thumb’.17. ‘Path-dependence and predictability’.18. ‘The economics of gateway technologies and network evolution’.19. ‘Learning from disasters’.

472 Postscript

20. ‘New standards for the economics of standardization’.21. ‘Transport innovations and economic growth’.22. ‘The millennium bug meets the economics of QWERTY’.23. Musil’s words come from R. Musil, ‘Der Man ohne Eigenschaften’ (French translation:

‘L’homme sans qualité) and from J. Bouveresse, ‘L’homme probable: R. Musil, le hasard,la moyenne et l’escargot de l’histoire’.

24. In L’Europe désemparée, Essais, Conferences Critiques, Aphorisme et Réflexions, l’Europedésemparée, Paris: Seuil, 1984, p. 142 (our translation).

25. ‘Resource abundance and American leadership’.26. ‘Knowledge, property, and the system dynamics of technological change’.

An appreciation of Paul David’s work 473

Abramovitz, M. 10, 13Abramson, A. 149adoption of new technology 16, 17,

434, 438adoption decisions 24–5, 29, 30catastrophe adoption path 430, 431,

435continuous adoption path 430epidemic adoption theory 432, 433,

436equilibrium adoption theory 428–9,

431, 432, 433, 436fax machines 433, 434increasing returns to adoption 23–4,

34network effects 428, 429, 430, 431uncertainty 439see also competing technologies; real

options model of technologyadoption; robots

Aghion, P. 302Aitchinson, J. 444Aitken, H. 141Akerlof, G. 189Allen, R. 243Alma-Tadema, L. 179, 180, 181, 182,

183Alston, L. 324American Telephone and Telegraph

(AT&T) 120, 122, 127, 141share of US patenting 141, 142, 143technological specialisation

path dependency 154profile of 46–8, 149, 150, 151, 152,

153Andersen, H. 128Antonelli, C. 158Aoki, M. 353Archibugi, D. 128, 365, 366Arora, A. 80Arrow, K. 13, 55, 241art

inter-painter price relationships 201,202

oil paintingsprices of 179

price measures of demand 165, 166,167–77, 190–93, 195–6

‘fad component’ 193–5inherently good painters 198–9,

200study data 177, 178study methodology 177, 178, 180

tastesavant-garde effect 188–9, 196, 197,

198, 199, 200–201characteristics approach 164–5conformity effect 188, 196, 197,

201path dependence 162, 163, 182,

183volatility of 160, 161, 162

trends in popularity of artists180–82, 183, 188, 189, 197, 198,201, 202–3

see also tasteArthur, W. 24, 25, 26, 30, 31, 36, 39,

60, 118Arundel, A. 363, 367, 369AT&T, see American Telephone and

TelegraphAtkinson, A. 58Audretsch, D. 256Ayers, F. 173

Bacharach, M. 164Balassa, B. 373Banerjee, A. 24Barnes, R. 161Barré, R. 258Barrera, M. 127, 137Basberg, B. 128Baumol, W. 158, 160Bayer 120

475

Index

share of US patenting 131technological specialisation

path dependency 154profile of 133, 134, 135, 136, 137,

150, 151, 152see also IG Farben

Bayma, T. 341Beaton, K. 122Beaver, D. 256Becker, G. 160, 189Beer, J. 119, 129, 136Bell, A. 141Bell, C. 161Bell, Q. 161Beniger, J. 213Berndt, E. 457Bernheim, B. 24Berthet, M. 245, 249Bessen, J. 420Bianchi, M. 160Bienz-Tadmor, B. 78Birdzell, L. 324, 327Bonnard, P. 181, 182, 183Bordo, M. 325Boucher, F. 181, 182, 183Bound, J. 126Bourdieu, P. 159Boyer, H. 77Braun, T. 365Bresnahan, T. 332, 389Brock, W. 24Brown, J. 308, 309, 311, 444Brynjolfsson, E. 24Burstall, M. 85

Caballero, R.-J. 439, 440Callon, M. 242, 255, 256Canaletto, A. 179, 181, 182, 183Cantwell, J. 365, 369, 374Carlsson, B. 365Casson, M. 297Centre National de la Recherche

Scientifique (CNRS) study ofcollaboration between researchers258–91

Cézanne, P. 179, 181, 182, 183Chandler, A. 71, 121, 122, 130, 153,

299Chanel, O. 165chemical industry 119–20

geographic origin of researchactivities 370, 371, 372

knowledge specialisation 373–5, 376,377, 378, 379, 380

shares of US patenting 131technological specialisation 151see also Bayer; Du Pont; IG Farben

Chien, R. 84Church, J. 24, 25Claude Gellée 181, 182, 183CNRS, see Centre National de la

Recherche Scientifiqueco-publication in scientific research 255

determinants of 255, 256future research on 285, 286see also Centre National de la

Recherche ScientifiqueCoase, R. 7Cockburn, I. 77, 79, 81, 82Cohen, M. 296Cohendet, P. 303Coleman, D. 122Collins, W. 180, 181, 182, 183communities of practice

interaction with epistemiccommunities 314

learning in 306–7, 309, 310, 311competing technologies 29

models of 27–9, 30–34, 35, 36–9proof of proposition 1 40–41proof of proposition 2 41–3proof of proposition 4 43–6

rate of convergence to technologicalmonopoly/market sharing 35–9

sequence of historical events 38time required 38, 39

relative impact of increasing returnsand degree of heterogeneity 38,39

competitive advantage in industry 209composition effects 59, 60, 61computer manufacturers 217

software systems 217, 218see also information and

communication technologyConant, J. 364condensed matter physics 258conspicuous consumption 159Constable, J. 181, 182, 183Cottereau, A. 248, 249, 250

476 Index

Cowan, R. 24, 25, 30, 31, 160, 306Cozzi, G. 160creativity 62, 63, 64Crémer, J. 313cross-licensing agreements 340, 343Cusumano, M. 24, 29Cyert, R. 296

Dalum, B. 374Dasgupta, P. 7, 232, 240, 242, 251, 255,

323, 326, 328, 347, 348, 364data warehouses 229David, P. 5–18, 23, 25, 51–2, 56–8, 60,

62, 70, 118–20, 141, 143, 158, 163,182, 190, 207, 218, 232, 240–42,251–2, 255, 286, 307, 323, 325–8,344, 347–8, 353, 361, 364, 367, 379,389–90, 427, 438–9, 443, 471–2

de-centred and distributed learning310, 311, 312

communication 312–13de Gennes, P.-G. 258de Lasalle, P. 244, 246–7, 249de-localisation of knowledge 212–14De Marchi, N. 160, 164De Piles, R. 164, 165Deane, P. 177, 178, 179Debreu, G. 13decentralised system of knowledge

management 231–2decision-making structures 64Degas, E. 181, 182, 183Deng, Z. 333Diamond, A. 255diffusion of innovations and new

technology 16, 17fax machines 433, 434see also adoption of new technology;

real options model oftechnology adoption; robots

dissonance 311distant past historicism 167distributed information technology

217, 218Dixit, A. 439, 443, 444Dornseifer, B. 130Dosi, G. 160, 296, 299Driver, C. 439, 440Du Pont 120, 122, 127

share of US patenting 131

technological development 129–30technological specialisation

path dependency 154profile of 137–40, 150, 151, 152

Duguid, P. 308, 309, 311Durlauf, S. 24dynamic efficiency of economic

systems, conditions for 61–6 , 67

Eckhardt, S. 82economic theory 3, 4economics of science 255, 256Economides, N. 24, 433economies of learning 208Edison, T. 141, 143Egidi, M. 299Eisenberg, R. 325, 338, 340electrical equipment industry 120

development of 141shares of US patenting 141–3technological specialisation 151vertically integrated systems 141, 149see also American Telephone and

Telegraph; General ElectricEliasson, G. 296, 308Engel, E. 439enterprise management software 229epistemic communities

interactions with communities ofpractice 314

production of knowledge 306ergodic processes 52, 53Ernst and Young 81ETAN 332, 333, 334European Commission 365, 369European Technology Assessment

Network (ETAN) 332, 333, 334experimental learning 307, 308, 311–12

Fagerberg, J. 366Fai, F. 118, 119Falcon, J.-P. 244, 249Farrell, J. 24, 34, 433Favereau, O. 298fax machines 433, 434Federal Reserve Bank of Dallas 223Filene, E. 214, 217firms

building of a common knowledgespecific to the firm 313–14

Index 477

core competences 300exchange of knowledge through

networks 301, 302governance 304, 305knowledge formation in the firm

309, 310management of collectively

distributed knowledge withinthe organisation 308, 309

non-core competences 301peripheral activities 302ranking of activities 302, 303, 304,

314declassifying routines 303–4structure of governance 304

theories of 296, 297competence 300, 304, 305principal/agent theory 297–8processor of information, as 297–8processor of knowledge, as 298–9transaction cost 298, 300

flexible production 216, 217, 222; seealso product variety

Foray, D. 158, 255, 258, 361, 367Ford, H. 213, 216, 339Ford Motor Company 216forgetting 311Foss, N. 296, 302Frank, R. 160, 189Fransman, M. 296Freeman, C. 133Freeny Jr., C. 418Frey, B. 160Frost, R. 8

Galambos, L. 71Gambardella, A. 71, 77, 79, 80, 81, 370Gandal, N. 24, 25General Electric 120, 122, 127, 141

share of US patenting 142, 143technological specialisation

path dependency 146, 154profile of 143–6, 150, 151, 152,

153General Motors 216general purpose technologies (GPTs)

analysis ofdata 390, 391generality measurement 393,

395–400, 419, 420, 421, 423

identifying GPT patents 410,413–17, 418, 419

definition of 390ICT-related patents 418, 419patent characteristics 392see also patent citations; patents

geographic origins of researchactivities 370, 371, 372

Ghoshal, S. 305, 311, 312Gibbons, M. 255, 308Gilbert, R. 336Ginsburgh, V. 160globalisation as cause of technological

change 65Godin, B. 365, 373Gogh, V. van 160, 181, 182, 183, 188Gombrich, E. 161, 162Gomperts, P. 331Goodwin, C. 160Gorman, W. 164Gould, S. 57GPTs, see general purpose technologiesGrabowski, H. 78, 84Grampp, W. 160Granstrand, O. 366, 379Green, J. 337Greenstein, S. 328Griliches, Z. 16, 123, 127, 128Grindley, P. 340, 343Grossman, S. 335growth 66

conditions for 61–6Guerzoni, G. 177

Haber, L. 119, 129Hadley, W. 340Hall, B. 329, 340Hals, F. 180, 181, 182, 183Hand, J. 333Hart, O. 335Hayek, F. von 299Heckman, J. 455Heller, M. 325, 338, 340Helpman, E. 389, 390Henderson, R. 71, 77, 79, 81, 82, 85,

346, 393, 397Heston, A. 450higher education, see university-based

researchHill, B. 41, 43

478 Index

Himmelberg, C. 433Hounshell, D. 122, 126, 130, 137, 140,

213, 214, 215, 216Hughes, T. 141, 213Hunt, R. 420

Iansiti, M. 366ICT, see information and

communication technologyIG Farben 120, 122, 127, 133

share of US patenting 131technological development 129, 130technological specialisation

path dependency 154profile of 133, 134, 135, 136, 137,

150, 151, 152see also Bayer

IMF 450incentives and institutional standards

224, 225, 226, 249, 328, 329–30increasing returns to adoption 23–4,

25, 26, 34, 39 individual knowledge 208individual learning 208information and communication

technology (ICT) 11construction of integrated systems

218–19decentralising information

processing 218, 219distributed technology 217, 218incentives and institutional

standards 226interpersonal communication

226–8minicomputers 218modelling business processes 228–9patents 418, 419supporting local learning 219, 220see also computer manufacturers;

softwareinformation search costs 334innovation economics 3, 4, 5, 6innovative capacity 4, 5intellectual property 12, 335

security interests in 333–4see also patents

intensive use of knowledge 8International Monetary Fund (IMF)

450

interpersonal communicationexchange of knowledge 226–8

interrelatedness of technology 150Ireland, N. 444Ironmonger, D. 164irreversibility 59Islas, J. 24

Jacquard, J.-M. 248, 249, 250Jaffe, A. 393, 395, 397Janson, A. 188Janson, H. 188Jensen, R. 439Jones, R. 122Joskow, P. 456

Kahneman, D. 55Karshenas, K. 443, 447, 456Katz, J. 256Katz, M. 23, 24, 25, 27, 29, 31Kemerer, C. 24Kenney, M. 349Kirman, A. 190Klemperer, P. 336Klevorick, A. 83knowledge 6

circulation 210codification 10de-localisation of 12–14individual 208intensive use of 8management 230, 231, 232–3

decentralised system of231–2

meaning of 230–31organisational 208

value of 209production of 8public domain 7, 12

financing of knowledgeproduction 240–41

public–private interactions 12role of, in industry 211, 212tacit 10, 11transfers of 12see also knowledge commons;

knowledge integration;knowledge openness;knowledge persistence;knowledge specialisation

Index 479

knowledge commons 8, 9localised 58, 60

knowledge integration 363, 366, 367,368

chemical and pharmaceuticalindustries 375–6, 377, 378, 379,380

future research 381policies for 380

knowledge openness 239, 240, 241,242

collective ethos 246, 247, 248efficiency of 248–50establishment of technical standards

249reward system 248, 250, 251see also open science; open source

software; open technologyknowledge persistence 362, 364, 365,

366chemical and pharmaceutical

industries 373–5, 377, 378, 379,380

future research 381see also knowledge integration

knowledge specialisationRSI index 381–2specialisation profiles of chemical

and pharmaceutical industries377–8, 380

see also knowledge integration;knowledge persistence;technological specialisation

Konno, N. 304Kortum, S. 342Koski, H. 37Kremer, M. 344Krugman, P. 26

Lamoreaux, N. 71, 334Lancaster, K. 164Landau, R. 366Landseer, E. 180, 181, 182, 183Lane, D. 24Langlois, R. 296, 302, 328, 339Lasdon, L. 180Leahy, J. 439Leamer, E. 278learning 308

by doing 9, 10

de-centred and distributed 310, 311,312, 313

economies of 208experimental 307, 308, 311–12governance for 313individual 208organisational 208‘technology of 208–9through error production 310see also communities of practice

Leibig, J. 82Lerner, J. 342Lev, B. 333Liebowitz, S. 24, 29Linden, G. 329, 330Llerena, P. 314Loasby, B. 296localised introduction of new

technologies 58factors affecting 65

localised knowledge commons 58, 60localised problem-solving 229lock-in effects 14, 15, 61, 119Lundvall, B. 307, 361Lyons silk industry

diffusion of new technology 248–50

invention in 243–4sharing of knowledge 244–8, 250,

251

Maclaurin, W. 143Madison, J. 16Magalhães, R. 309Malerba, F. 365Malo, S. 367, 370Malraux, A. 162Manet, E. 181, 182, 183, 188Mansfield, E. 364, 439March, J. 296, 299, 305Marcus, G. 162Marengo, L. 296, 299Margolis, S. 24, 29market sharing 23, 26, 34; see also

competing technologies markets for technology

financial institutions, role of 331global market 349–53information search costs 334institutional settings 327

480 Index

intellectual property rights 335security interests in 333–4

limitation of liability 334patent offices, role of 341–3patent-pooling agreements 343patents 335, 336, 337–8

‘efficient breach’ 344extension of ‘eminent domain’ 344fragmentation 338, 339, 340, 343,

344legal costs 340

research and development tax credits332–3

standards 328, 329–30startup firms, government support

for 332technology suppliers, role of 331,

344university research 344valuation of technology 333venture capitalists, role of 331–2

Marriot, O. 122Martin, B. 380mass production system 213, 214, 216,

217information goods 220–21

Matraves, C. 74Maxwell, R. 82McCain, R. 160McCormick, C. 214, 215McCormick, L. 214, 215McCormick Reaper Works 214, 215

production system 214–16McDermott, C. 162McPherson, M. 160Meissonier, E. 180, 181, 182, 183Meliciani, V. 365memory 230, 311Menger, P.-M. 160Merges, R. 78, 335, 337, 340Merton, R. 364Metcalfe, J. 65Metcalfe, S. 308minicomputers 218Mitchell, B. 177, 178, 179Monet, C. 164, 181, 182, 183, 188Moore, J. 335moral property rights 241, 242Moreton, D. 439, 440Mowery, D. 345, 346, 347, 349

Mullins, N. 256Musil, R. 471, 472

Narin, F. 364, 367Nash, L. 364national competitiveness, scientific and

technological specialisation, roleof 362

National Research Council 340national systems of innovation (NSI)

361Nattier, J.-M. 180, 181, 182, 183Nelson, R. 78, 120, 136, 299, 312, 337,

340, 361network effects 15, 24, 29, 428, 429,

430, 431Newey, W. 457Nijkamp, P. 37Nobel, D. 143Nohria, N. 305, 311, 312non-ergodic processes 52, 53; see also

past dependence; path dependenceNonaka, I. 304Nooteboom, B. 308, 312North, D. 324NSI 361Nuvolari, A. 243

OECD 307Office of Science and Technology

(OST) 365Office of Technology Assessment and

Forecast (OTAF) 392oil firms, shares of US patenting 132open science 7–8; see also knowledge

opennessopen source software 227open technology 243, 251; see also

knowledge openness; Lyons silkindustry

Oren, S. 24Organisation for Economic

Cooperation and Development(OECD) 307

organisational capability 208organisational knowledge 208

value of 209organisational learning 208organisational memory 230Orsenigo, L. 365, 370

Index 481

OST 365OTAF 392

Pagan, A. 457Pareto, V. 13, 14past dependence 52, 53, 54, 56

role of internal factors 57Patel, P. 128, 362, 366, 369patent citations 391–2

citation lags 407, 408, 409highly cited patents 393, 394

characteristics of 408, 409, 410generality measures 400, 401–2probability of 410, 411–12technology sub-categories of 421,

422, 423see also general purpose

technologies; patentspatents

characteristics of 409cross-licensing agreements 340, 343growth of patent classes 400, 403–5,

406, 407, 409legal costs 340pharmaceutical industry 78, 83, 87technology sub-categories of 421,

422, 423university-based research 345, 346see also general purpose

technologies; markets fortechnology; patent citations

path dependence 51, 52, 53, 54, 56,60–61, 66, 67, 118

characteristics of 52, 54definitions of 163external factors, role of 58, 59, 60feedbacks 54, 55, 56, 119internal factors, role of 58, 59, 60, 61irreversibility 54, 55local externalities 54, 55, 56, 58lock-in 61sequence of steps 54strength of 119theory of 13–15

path independence 163Pavitt, K. 123, 128, 362, 365, 366, 369,

373Peltzman, S. 84Penrose, E. 299personal computers

construction of integrated systems218–19

decentralising informationprocessing 218, 219

supporting local learning 219, 220pharmaceutical industry 70, 71

biotechnologydevelopment of 81impact of 77, 78, 79

collaborative research 80, 81commercialisation of penicillin 72competition 75development of 71–2, 113geographic origin of research

activities 371, 372health-care systems, structure of 84innovation 85–6

economic benefits from 73, 74, 75forms of 75imitator firms 93innovative firms 93levels of 73, 74

knowledge specialisation 373–5, 376,377, 378, 379, 380

levels of concentration 75–6, 86model of new drug development

87–92, 93, 114extension of time of patent

protection, effect of 108, 113firms’ activity in different

therapeutic categories 106imitative products, number of 101,

102, 103, 104increase in number of firms, effect

of 108, 113increase in stringency of approval

procedures, effect of 108, 113 innovative products, number of

101, 102, 103, 104innovative products, share of 105market concentration 94, 95, 96,

108number of firms in each

therapeutic area 100number of innovative and

imitative products in eachtherapeutic area 108, 109, 110,111, 112

number of products in therapeuticarea 99

482 Index

number of therapeutic areasdiscovered 98

performance index 107surviving firms 97

new firm entrants 75, 78patents 78, 83, 87price regulation 84, 85product approval 83–4, 87publicly funded research 76, 82–3random screening 71, 73, 74, 75, 76

source of first-mover advantage 75rate of technological change 81rational drug design 76, 77research approach 70, 72, 73transforming research into

successful products 372university research 82university spin-offs 77vertical integration 81

Pharmaceutical ManufacturersAssociation 78

Pianta, M. 365, 366Pindyck, R. 439, 440, 443, 444Pisano, G. 71, 362, 366, 370Pissarro, C. 181, 182, 183Plumpe, G. 122, 126, 127, 129Pommerehne, W. 160Porter, M. 362Prencipe, A. 366Price, D. de S. 365Price, R. 16producer–user relationship in industry

212product diversification, link to

technological diversification 153product selection decisions 25, 29, 30,

37product variety 221, 222, 223

decentralisation, need for 223, 224,225

see also flexible productionproductivity of scientific research,

determinants of 255prospect theory 55public and quasi-public databases 341

Quillen, C. 340

Rallet, A. 258Rauch, J. 26

real options model of technologyadoption 439–40, 442–50

data sources 450, 463–4descriptive statistics 451–4methodology 455–7, 463, 464, 465see also robots

recent past historicism 167Reich, L. 120, 122, 126, 127, 141, 143,

146, 149Reitlinger, G. 166, 177, 179Rembrandt van Ryn 181, 182, 183Renoir, P. 181, 182, 183, 188reputation capital 241research and development (R&D) tax

credits 332–3Rheims, M. 161Richardson, G. 299Robertson, P. 328robots 440–41

adoption of 441determinants of 458, 459, 460government policy changes, effect

of 460–62numbers 453, 454, 456uncertainty, impact of 462–3

application areas 441investment in 441

volatility of 441, 442prices 450, 453, 454, 455, 456, 458,

460, 463see also real options model of

technology adoptionRohlfs, J. 24Roos, J. 309Rose, N. 456Rosen, R. 256Rosenberg, N. 24, 136, 213, 324, 327,

362, 366, 389Rostoker, M. 340Ruskin, J. 158, 160, 164, 166Ruttan, V. 55

Saloner, G. 24, 34Sanderson, W. 9, 307Santangelo, G. 153Saviotti, P. 24Scherer, F. 123, 126Schmookler, J. 121, 123, 126Schwartzman, D. 71Schwerin, J. 243

Index 483

Science Citation Index (SCI) 279, 281,282

Scotchmer, S. 337sectoral knowledge bases 367, 368; see

also knowledge specialisationsecurity interests in intellectual

property rights 333–4self-sustaining process of growth and

innovation 62, 63Sewell, J. 71Shapiro, C. 23, 24, 25, 27, 29, 31, 336,

433Sharp, M. 372Shi, Y. 255Shrum, W. 256Silverman, B. 395, 398Simon, H. 299Sisley, A. 181, 182, 183slack 311Sloan, A. 217, 219Smith, A. 164, 189, 212Smith, J. 122, 126, 130, 137, 140Smith, S. 24social referral networks 229–30Soete, L. 123, 365, 373software

enterprise management 229open source 227systems 217, 218

Sokoloff, K. 334Solimano, A. 439, 440Somaya, D. 329, 330startup firms 332Stephan, P. 255, 256Sternberg, R. 378Stigler, G. 212Stiglitz, J. 58Stocking, G. 122Storper, M. 278Sturchio, J. 71Summers, R. 450Sutton, J. 74Swann, G. 198, 201, 203Swanson, R. 77systematisation 213

tacit knowledge 10, 11taste

aspiration 160association 160

bandwagons of 190conformity 189distinction 159, 160, 189path dependence of 162, 163,

182price as a measure of 165, 166volatility of 160, 161–2see also art

technological commons 57technological disparities between firms

210technological diversification

interrelatedness of technologicalactivities 153–4

link to product diversification 153

motives for 153see also technological specialisation

technological knowledge 4, 5collective activity, as 56–7

technological monopolies 23different monopolies in different

markets 39increasing returns to adoption 25,

26, 39see also competing technologies

‘technological opportunity’ 209technological specialisation 362

research studydata 121–2measure of specialisation 122–8

see also American Telephone andTelegraph; Bayer; Du Pont;General Electric; IG Farben;knowledge specialisation;technological diversification

technology adoption, see adoption ofnew technology

‘technology of learning’ 208–9technology suppliers 331, 344technology transfer 208Teece, D. 299, 340, 343Thomas, L. 84, 85Throsby, D. 160Tijessen, R. 369Tirole, J. 302Toniolo, G. 309, 310Torre, A. 258Trickett, A. 24Tversky, A. 55

484 Index

university-based research 82, 241, 344,345, 346, 347

commercialisation 345impact on academic norms

347–9licensing 345, 346patents 345, 346

university spin-offs 77

valuation of technology 333van Gogh, V. 160, 181, 182, 183,

188van Wijk, E. 369Vaucanson, J. 244, 249Veblen, T. 159Venables, A. 26venture capitalists 331–2Vernon, J. 78, 84Vicari, S. 309, 310Vincenti, W. 347, 366von Hippell, E. 243

von Krogh, G. 309, 313Vopel, K. 395, 420, 421

Walras, M.-E.L. 13, 14Walsh, J. 341Waren, A. 180Watkins, M. 122Wenger, E. 306,311White, M. 165Whited, T. 439Wilkins, M. 122Wilkinson, L. 215Wilson, G. 249Winter, S. 26, 118, 120, 136, 299Witt, U. 55

Young, A. 212

Ziedonis, R. 329, 340, 346, 347, 420Zucker, L. 278Zuscovitch, E. 301

Index 485