Kenny, Charles. 2000 What do we know about Economic Growth Or Why don´t we know very much

8/4/2019 Kenny, Charles. 2000 What do we know about Economic Growth Or Why dont we know very much

1/22

What Do We Know About Economic Growth?

Or, Why Don't We Know Very Much?

CHARLES KENNYThe World Bank, Washington, DC, USA

and

DAVID WILLIAMS *

University of Oxford, UKSummary. The last 10 years has seen an explosion in cross country econometric studies ofgrowth, driven by two factorsnew mathematical models of the growth process that lendthemselves to econometric testing, and new data sets that make such testing possible. This paperlooks at a selective review of these studies. It concludes that the results are disappointing in that nomodel has proven robust to trial by repeated regression. The paper suggests some reasons for thisincluding that the tested models tend to be ahistorical and over-simple in terms of their causalaccounts. It concludes with possible lessons for econometric work in this area. 2000 ElsevierScience Ltd. All rights reserved.

Key words economic growth, theory, cross country regressions

1. INTRODUCTION1

In June 1996, The Economist magazinepublished a piece based on results from a globalstatistical study concluding that had Africancountries followed better policies, such as thosefollowed in eight fast-growing economies overthe last few decades, the region would havegrown 4.6% per annum faster than its historicalgrowth rateindeed, faster than the compara-tor set of fast-growing economies. A year later,The Economist published another piece based

on results from a global statistical study thatconcluded ``for much of the world, badclimates, poor soils and physical isolation arelikely to hinder growth whatever happens topolicy.'' This study concluded that, even hadAfrica followed better policies, it would havegrown 2.3% slower per year than the countriesof South and South East Asia.

These two articles, with their markedlydierent conclusions, provide an illustration ofthe problems facing even the best developmenteconomistsindeed, both articles were written

by Jerey Sachs (Sachs, 1996, 1997a). 2 Overall,attempts to divine the cause or causes of long-term economic growth, testing a wide range of

possible determinants using statistical tech-niques, have produced results that (like the twoSachs articles) are frequently contradictory toresults reported elsewhere. That is, empiricalevidence is hardly unanimous in support of aparticular view of the growth process.

If this is right, it might pose a serious prob-lem for development practitioners. As othershave argued, many of those concerned withstimulating growth in developing countrieshave a more or less universally applicable set ofpolicy prescriptions for achieving that goal.

The empirical evidence, however, seems toprovide little rm guidance for the universalecacy of any particular policy prescriptions.Indeed, we will argue that if the evidence showsanything at all, it is that markedly dierentpolicies, and markedly dierent policy mixes,may be appropriate for dierent countries atdierent times.

World Development Vol. 29, No. 1, pp. 122, 2001 2000 Elsevier Science Ltd. All rights reserved

Printed in Great Britain0305-750X/00/$ - see front matter

PII: S0305-750X(00)00088-7www.elsevier.com/locate/worlddev

*The opinions expressed here are the authors' own and

do not necessarily reect those of the World Bank, its

executive directors, or the countries that they represent.

Final revision accepted: 8 July 2000.

1


2/22

In addition to suggesting that at present wedo not know all that much with any certaintyabout the causes of long-term economicgrowth, the paper tries to investigate thereasons for this. Our approach here is a focus

on the structure of the theorizing abouteconomic growth upon which these econo-metric studies have been based. In otherwords, we do not attempt to distill out ofthese studies the ``real'' cause or causes ofeconomic growth; rather, we ask if there arefeatures of the theorizing behind them whichwould explain why these studies have notproduced universally compelling accounts ofthe process of economic growth. The argu-ment of this paper is that there are goodtheoretical reasons to match the empirical

evidence for thinking that the growth processis highly heterogeneous.

We argue that there is a crucial andimportant mismatch between the actualeconomic world, and the ``picture'' or ``mod-el'' of the economic world which underpins thecurrent search for causally signicant vari-ables. Specically, we suggest that currentthinking about economic growth has oftenfailed to grasp the complex causal nature ofthe social world, assuming that the compo-nents and processes of the economy are the

same across countries. Following from this, wethink that any account which takes as anassumption that the process of economicgrowth works more or less unaltered acrossenough countries to be proved or disprovedthrough the statistical testing of variables inlarge cross country regressions is likely to beinadequate.

Section 2 seeks to sketch out what we taketo be some of the key features of theorizingabout economic growth. It argues that thecommitment to what we call here epistemo-

logical universalism entails a commitment towhat we call ontological univeralism, and thatthese commitments underpin the currentsearch for causally signicant variables in thestudy of economic growth. Section 3 is areview of the available statistical evidenceconcerning the causes of economic growth.Section 4 of the paper ties the discussiontogether by suggesting some reasons why keyfeatures of the theorizing do not produce goodempirical accounts of the process of economicgrowth. In Section 5 we suggest a number of

possible implications for the current practiceof development, and for the way we theorizeabout economic growth.

2. THEORETICAL AND PRESCRIPTIVEUNIVERSALISM

The discussion in this section hinges aroundthe suggestion that there is a connection

between the kind of knowledge we can hope togain of a set of phenomena and the nature ofthat set of phenomena. For example, we areusually fairly condent that we can get law-likeknowledge of processes in the natural world,whereas we are less sure we can do so of thesocial or political worlds. This is at least in partthe result of the two kinds of world beingdierent; we are fairly sure there are someuniversal causal mechanisms in nature (gravity,for example), but are less sure there are in, say,politics or in family life, or at least if there are

such laws, it is not clear we are very neardiscovering them. Moreover, of course, part ofthe reason for this is that political and familylife are inhabited and constructed by active,creative, and thinking persons, and it is not atall clear we have any universal causal lawscovering their activities.

As regards current thinking about economicgrowth, this connection between the kind ofknowledge we can hope for, and the nature ofthe phenomena under study, implies that thekind of knowledge we can hope to have of the

growth process is determined by what we thinkeconomies and economic process are like. Wesuggest that, among many economists whostudy growth, there is a necessary commitmentinherent in cross country econometric studiesto two kinds of universalism as regards theirview of the growth process.

First, there is a commitment to what mightbe called ``epistemological universalism.'' Thisis inherited from the scientic revolution andinvolves a commitment to the production of``true'' knowledge generated following the

``rules'' of science, which in turn provides thebasis for prediction. This kind of knowledge is``universal'' in the sense of not being a reec-tion of a particular time or place; it is, to useNagel's phrase, a ``view from nowhere'' (Saiedi,1993, p. 22; Nagel, 1986). As regards the studyof economic growth it means a commitment tothe idea that all economic processes everywhereare, in principle, knowable. That is, the aim ofeconomics has been to provide theories gener-ated following scientic rules, which explain thelargest possible class of phenomena. This has

been evident since at least Adam Smith, whowas searching for the causes and mechanismswhich would ``connect together the otherwise

WORLD DEVELOPMENT2


3/22

disjointed and discordant phenomena ofnature'' (Smith, 1980, p. 119).

This commitment to epistemological univer-salism entails a prior commitment to whatmight be termed ``ontological universalism.''

This has two closely related elements. First, the``components'' of all economies are in some waythe same, and hence that economies andeconomic processes are comparable. Second,these components of the economy interact withone another in the same kinds of ways, thusproducing certain economic ``laws'' or regular-ities which operate across all economiesregardless of time or space. These two relatedcommitments underpin the search for the causesof economic growth across countries, and thecross country statistical techniques used to

discover them. If the components of economieswere not in some way the same, or if they didnot interact with one another in the same kindsof ways across countries, then cross-countrycomparisons would not be valid; they would notbe a comparison of like with like. It is thiscommitment to ``sameness,'' combined withepistemological universalism, which underpinsthe value placed on theoretical parsimony seenin discussions of economic growth. Only if alleconomies are fundamentally the same, in theircomponents and their processes, does the search

for fewer and fewer universal explanatoryprinciples or ``laws'' make sense.

There is a good case to be made that thesecommitments have underpinned the search forthe causes of economic growth since thebeginning of economics, and have subsequentlyinformed prescriptions about how to achieve it.We have already suggested they were present inAdam Smith. A chronological review of somerecent models suggests that this belief has, ifanything, become even stronger.

The HarrodDomar model posited a

concrete and linear relationship betweeninvestment and growth rates worldwide. It wasused by many development economists (in away unintended by Domar) to forecast GDPgrowth rates in the developing world, andestimate the level of foreign aid required toreach some target rate of growth. The WorldBank's RMSM (Revised Minimum StandardModel), developed in 1972, was based onHarrodDomar. This model argued thatcountries with a capital output ratio of, say,ve, needed 5% of GDP going to investment for

each 1% increase in growth. Thus, if investmentreached 15% in that country, growth wouldclimb to 3%. For a long time a country's capital

output ratio was used by the Bank to calculatethe foreign aid ``requirement'' of a country inorder to reach a growth target. 3 Althoughcapital output ratios were allowed to varybetween countries, the simple mathematical

model linking changes in the rates of invest-ment with changes in growth rates weresupposed to apply to all economies.

While a range of other theoretical treatmentsof the investment-growth relationship aresomewhat more complex, investment wascentral to most early growth theories. Rostow'sstages of growth, Nurkse (1953) balancedgrowth, and Lewis' unlimited labor model allfeatured a central role for capital accumulation(Meier, 1999).

Domar himself argued that his model was

not appropriate for determining long-termgrowth rates, and supported instead the Solow-Swan neoclassical growth model which predic-ted growth rates were dependent, not oninvestment, but the rate of technologicalchange (Easterly, 1998). Crucially, this modelassumed that technology change occurred atthe same pace worldwide. It therefore had littleexplanatory power when looking at the hugedivergence in income growth rates acrosscountries, and so the neoclassical model waslargely abandoned in favor of what became

known as ``endogenous growth theory,'' inwhich the rate of technological change variedacross countries depending on other factors(Ruttan, 1998).

For example, an early endogenous growthmodel of Paul Romer's argued that output wasrelated to physical capital, labor, and knowl-edge, where the quantity of knowledge wasconnected with investment rates (Romer, 1993).In other words, investment in physical capitalreturned as the central proximate determinantof growth rates. Other models emphasized

investment in learning. Lucas' (1988) formalmodel argued that human capital was thesignicant factor determining growth ratesworldwide, while Rebelo (1990) suggested that,if countries invested in labor-augmentinghuman capital, returns to physical capitalwould not diminish, allowing growth tocontinue forever. This picked up on earliercross country work that had emphasized theimportance of education. In 1981, Easterlin hadlooked at growth experiences worldwide andconcluded that it was the late extension of

universal education in many countries thatanswered the question ``Why isn't the WholeWorld Developed?'' (Easterlin, 1981).

WHAT DO WE KNOW ABOUT ECONOMIC GROWTH? 3


4/22

But Romer and Lucas' formalizations havealso spurred an ever increasing number of long-term growth models that have moved beyondphysical and human capital. For example,Grossman and Helpman (1991) produced a

model that explored the role that trade mightplay in growth. These models produced a rangeof dierent conclusions, but many posited thatpolicies could have an important impact ongrowth rates even over the long term. Intandem with these developments in high theory,the 1980s saw the evolution of what came to beknown as the ``Washington Consensus'' ondevelopment policy. Among the policies linkedwith the consensus were devaluation, reductionof budget decits and ination, liberalization ofprices and interest rates, and privatization.

What connected these policies was a belief thatgovernments had played an overactive role inpromoting development, taking on tasks bestleft to the private sector, and abusing powersbest left unused. Widespread governmentintervention was not only unnecessary topromote growth, it was the chief barrier toachieving that growth. In addition, many policyprescriptions for developing countries revolvedaround reducing government spending whilecurbing ination (Polak, 1997; Walde & Wood,1999).

In a more explicit way than earlier theories,the ``Washington Consensus'' was underpinnedby a belief that all economies are similar andthus will respond in a similar manner to thesame policy change. Some of those working inthe eld went as far as to deny the need for aseparate discipline of ``development econom-ics'' arguing that ``once it is recognized thatindividuals respond to incentives, and that`market failure' is the result of inappropriateincentives rather than of non-responsiveness,the separateness of development economics as a

eld largely disappears'' (Kreuger, 1986).That belief remains strong among many

economists. But the diculty of pushingthrough reforms based on the WashingtonConsensus, and the limited impact these hadeven in those countries who succeeded inimplementing them, encouraged the professionto look for explanations in the realm ofpolitical economy. In particular, DouglasNorth's New Institutional Economics arguedthat underpinning economically successfulregimes was a strong network of property

rights, market structures and decentralized,democratic decision-making processes (North,1992). Some more recent oshoots of this

theory have suggested a large role foreconomic and social networks that promotetrust, thus suggesting the importance of thingssuch as equitable distributions of income andeorts to overcome ethnic divisions (Putnam,

Leonardi, & Nanetti, 1993; Fay, 1993;Svennson, 1994; Grenato, Inglehart, &Leblang, 1996; Seabright, 1997). Although notall of these positions have been formallymodeled to the same extent as earlier theories,their conclusions are clear enough to suggest,again, that these theories are seen to haveuniversal policy applicability.

A simplied chronology of ``what the main-stream considers is needed for developmentworldwide'' through to the early 1990s mightrun as follows, then: physical capital, human

capital, policy reform, institutional reform andsocial development. 4 It should be noted that,while these factors have been added to more orless formal models in approximately this order,they are not new to nonformal discussions ofdevelopmenta point we shall return to. Whilethe steady addition of dierent factors to thelist presented in formal models does suggest, inone way, an ever-greater appreciation of thecomplexity of the growth process, they tend tohave been added to basic models (especially intheir more formal incarnations) in a sequential

rather than parallel manner, and with thecontinuing assumption that the model isuniversally applicable.

Thus the current search for the cause orcauses of economic growth appears to befrequently informed by a commitment toproducing objective, scientic, and universalknowledge of economic growth, and this isunderpinned by the view that all economies aresubstantially similar in their components andprocessesthat there is but one basic produc-tion function driving all economies at all times

and in every time-frame.It should also be noted that universal models

of economic growth developed over the past 40years are, in their particularities and recom-mendations, in frequent contradiction with oneanother. For example, models have proposedboth limited intervention as opposed togovernment support for education and/or R&D(Rebelo, 1990; Romer, 1990), and openness aswell as trade controls (Romer, 1990; Young,1991). This alone might give pause to thosewho give advice based on model results, but it

would be a remediable problem if empiricalevidence clearly supported one model over theother.

WORLD DEVELOPMENT4


5/22

3. THE EVIDENCE

This section reviews how models of thegrowth process underpinned by universalistcommitments have held up to the empirical

evidence, following the widespread belief thatthe acceptability of an economic theory is to be

judged by its ability to explain and predictphenomena (Arrow, 1974; Friedman, 1953). 5

First, we will look at the predictive power ofvarious economic models when it comes toforecasting future growth. Then we turn to therecord of cross country econometric studiestrying to explain past growth using regressionanalysis. Finally, we look at a number of eventsand country experiences that are hard toexplain using some of the usual models and

explanations of growth processes.In econometric exercises, a comparison is

made between the growth performance of anumber of countries, and the country policiesand characteristics that are considered signif-icant determinants of growth. In eect, thesestudies test if, across a range of countries, andgiven certain other features of the economy,certain policies and characteristics are consis-tently associated with higher or lower growthrates. A review of these cross country regres-sions is therefore useful for testing views of

what causes growth embodied in phrases suchas ``the Washington Consensus'' (Williamson,1996), or in formal models that tend toassume that period, country and sectordierences are largely irrelevant. If multiplecross country studies over dierent time peri-ods using dierent samples and dierentregression techniques nearly all suggest that avariable considered signicant in a growthmodel is consistently and strongly related togrowth, this provides support for the univer-sal applicability of that model. Sadly, the

answer that appears to emerge from suchtesting seems to be that growth is a morecomplex process than can be captured byuniversal models.

Looking rst at economists ability to predictgrowth, Sherden (1998), among others, hascollected fairly damning evidence on the reli-ability of short-term forecasts of economicvariables, including economic growth rates. 6

His survey of recent studies suggests that justusing last year's growth as an estimate (the``nave forecast'') produces results as good as

the average forecaster. He also nds that noforecasters are consistently above average intheir predictive powers, that Keynesians and

non-Keynesians are equally goodor badand that the accuracy of economic predictionshas not improved over time.

It might be argued that predicting long-termgrowth should be easier than divining volatile

one-year growth outcomes. But the predictiveabilities of early development economistsregarding longer-term growth potential alsoappear to have been weak. Easterly and Levine(1997a) note that in the 1960s Africa waspredicted to grow faster than East Asia bymany if not most development economists. Atthe same time, the Soviet Union was beingtouted as a model for rapid development(Krugman, 1994).

This would not, in itself, be a cause forconcern for recent theorists if early develop-

ment economics had few common elementswith contemporary development theory, butthat is not true (Srinivasan, 1993). While theemphasis might have changed, almost all themajor variable types that have at one time oranother been thought of as a major determi-nant of growth are present as chapters inArthur Lewis' The Theory of Economic Growth,rst published in 1955 (Lewis, 1965): the rightto reward, trade and specialization, economicfreedom, institutional change, the growth ofknowledge, the application of new ideas,

savings, investment, population and output, thepublic sector and power, and politics. 7 It doesnot appear that we have discovered a missing``silver bullet'' since then, at most it might besaid that we have re-arranged the bullets in themagazine.

It is not be surprising then, that tests ofmodern development theories based on pastperformance of countries also suggest a fairlypoor track record for theory. We will start withmodels that focused on investment which, as wehave noted, is seen by many development

economists (although not all: Morgan, 1967;Easterly, 1998) as a central determinant ofgrowth (Lewis, 1965, 1984).

The RMSM-type model that posits aconcrete, linear investment-growth relationshipperforms particularly badly. One recent esti-mate suggests that out of 138 countries, onlyve had a statistically signicant relationshipbetween one year's investment and the subse-quent year's growth with a reasonable capital-output ratio of between two and ve. Lookingat a particular country, if the HarrodDomar

model had been correct, Zambia (with areasonably high investment rate and largeinows of aid) would have a GDP per capita



6/22

income of perhaps $20,000 by 1994, instead ofthe actual level of around $600 (Easterly, 1998).

Even in studies that consider investment'simpact on growth in the context of many otherfactors, the evidence is far from reassuring for

HarrodDomar, as we will see from our reviewof the econometric evidence. At the same time,we will try to elaborate the reasons behind poorresults from regression analysis, and how thetechniques' foibles are useful for determiningthe cause of weak support for model-basedpredictions.

Figure 1 displays the simple relationshipbetween decade average investment as apercentage of GDP and decade average GDPgrowth across a number of countries over195090. Each point represents decade data for

one countryso that countries with 40 years ofinvestment and growth data will produce fourdata points. 8 As can be seen, there is clearlymore to explaining growth rates than invest-ment rates.

A simple regression analysis allows us tolearn that looking at the variation in investmentacross countries (without taking into accountany other inuences) can only account forabout 5% of the variation in growth acrosscountriesalthough the relationship is strongenough to suggest that we have found some-

thinghere, in that the variables are signicantlycorrelated. 9

At the risk of repeating a basic course ofeconometrics, what we have found depends ona number of assumptions, however. First, wehave assumed that it is investment that causesgrowth, not some other factor that is correlatedwith investmentor perhaps a factor that

explains both the rate of investment andgrowthrates. One example of such a factor is thestarting income per capita of a country in thesample. As we can see from Figure 2, there is apretty strong relationship between income per

capita at the start of a decade (measured on alog scale) and investment over the course ofthat decadericher countries invest more. 10

At the same time, there has been a divergencein incomes between rich and poor over the longterm (a phenomenon we will explore in greaterdetail later). Perhaps poorer countries grow lessfor some reason unconnected with investmentrates, and investment is only correlated withgrowth because it correlates with wealth. Weneed to test for the relationship betweeninvestment and growth allowing for initial

income, presented in Figure 3. The relationshipbetween income-adjusted investment andgrowth is still there. This suggests that invest-ment is not only related to growth because it isrelated to income per capita.

Again, at the risk of repeating the obvious,conditioning variables can have a signicantimpact on the apparent relationship betweenfactors in growth regressions. For example, weknow that, on average, richer countries aregrowing faster than poorer countries, so that ifwe estimated the equation

Growth C b initial income

b would take a positive value. But when weallow for investment rates, and estimate theequation

Growth C b initial income

v investment

Figure 1. The relationship between investment and growth.

WORLD DEVELOPMENT6


7/22

b has a negative value while v has a positivevalue (and one signicantly larger than in aregression of growth versus investmentalone). 11 This is the simplest form of theconditional convergence resultthat countrygrowth rates are inversely proportional toincomes (as would be predicted by the

neoclassical model) if various other conditionsare met.

Although not apparently a problem with therelationship between income, investment andgrowth, the above example does point up somepotential problems with multiple regressionanalysis. Explanatory variables are frequentlycorrelated themselves. For example, invest-ment, education and a number of variablesused to measure the quality of governmentinstitutions in a country are all positivelycorrelated with income per capita. There are a

number of possible explanations for suchcorrelation. There is a chance that the apparentrelationship is just luckin fact, there is no

causal relationship one way or the otherbetween the variables. Given the number ofvariables that have been thrown into crosscountry regressions, we would expect that somewould be correlated for just such a reason. Butit might be that one variable has a causalrelationship with the other (it is plausible to

assume, for example, that people in richercountries can aord more education). It mightbe that both variables, while not directly rela-ted, are caused by something else (higher edu-cation and investment might be correlatedpurely because both arein parta product ofhigher income).

It is in the nature of regression analysis that,if two variables that are correlated with eachother and with growth are entered into agrowth regression at the same time, theapparent strength of their individual relation-

ship with growth might appear much weakerthan if one or other is entered alone. Levineand Renelt (1992) performed an exhaustive test

Figure 3. The relationship between investment and growth corrected for initial income.

Figure 2. The relationship between income and investment.



8/22

to discover how widespread a phenomenon thisis. They took a number of variables commonlyused in econometric analyses of growth and ranthem in thousands of growth regressions withdierent conditioning sets of other variables, to

see if the test variable remained signicantlyrelated to growth in all of themwhich theydened as robust. Of the variables they tested,only investment survived as a robust vari-able. 12 This might suggest that all variablesbut investment have, at best, a relationship withgrowth dependent on a range of other condi-tionsor, at worst, they are only correlatedwith growth while the causal variable lies else-where.

In some ways, this might be an excessivelyrestrictive test, however. While the test will

fairly exclude variables that are only correlatedwith growth because they are caused by (orcorrelated with) another factor that itself has acausal relationship with growth, it will alsoexclude some factors that indirectly do have animpact on growth. For example, institutionalvariables are correlated with investment.Imagine that stronger government institutionsencourage higher investment in a country(because there is a lower risk of high inationor expropriation, perhaps), and that investmentencourages growth. If both variables are

entered into a regression at the same time, itmight be that the institutional variable doesnot show up as signicant, despite the factit does have an indirect impact on growth.Sala-i-Martin (1997) has designed a test forrobustness that allows for this problem withLevine-Renelt that we will discuss later.

In other ways, however, the Levine-Renelttest is not restrictive enough. Returning to thepotential problem that cross country regressionanalysis assumes that coecients and causalrelationships are the same across countries

and periods, the Levine-Renelt test is onlyperformed on a one-period global sample. On arelated note, the test does not allow for inter-actions between variables (for example, it mightbe that investment is only causally related togrowth in the presence of a strong institutionalstructure). To some extent, this problem can beovercome within the regression frameworkone can use interaction terms, or one variablemultiplied by another, to investigate thedierential impact of (as it might be) invest-ment in poor and good institutional environ-

ments. Later, we will look at splitting countrysamples as another method to look for changesin the strength and direction of causal rela-

tionships. But it should be noted that growthmodels rarely include the presence of interac-tion eects even at this simple level, and it issuch models that we are evaluating in thispaper.

A further problem with the Levine-Renelttest is that it measures the robustness ofcorrelation, but says little about causationbetween variables. Returning to the relation-ship between investment and growth, the evi-dence presented so far suggests that it is astrong (perhaps the strongest) correlate withgrowth that we have. Although it clearly missesthe great majority of the growth story, thismight suggest some support for a HarrodDomar type model. Correlation does not,however, prove causation. In fact, a dierent

statistical test performed by King and Levine(1994) suggests that, if anything, the relation-ship appears to operate the other waygrowthencourages companies and people to investmore, rather than investment speedinggrowth. 13 Again, it should be noted that thecausation issue can be examined using regres-sion analysis as well, and there are techniquesto (at least partially) overcome the eects ofbi-causal relationships of the type investment-causes-growth-causes-investment. 14 Nonethe-less, it suggests the need for yet more caution in

interpreting regression results in the absence ofsuch tests.

Turning to a broader analysis of the litera-ture on the econometric study of growth,dierent formulations, dierent proxies anddierent combinations of a wide range offactors have been subjected to many millions oftests in many thousands of papers producedusing cross country growth regressions. A listof variables that have been put into moderncross country growth regressions would includewell over 100 overlapping economic, policy,

structural, sociological, geographical andhistorical factors seen to inuence economicgrowth directly or indirectly. Yet a brief surveyof surveys appears to suggest that the results ofall this computation have been disappointingeven when holding to standards lower thanthose of Levine and Renelt.

Education is a staple of almost all morerecent discussions of economic growth. It is,however, found to be a particularly fragilecausal variable by Pritchett (1996b), who uses arange of regression tests to nd a relationship

that is sometimes signicantly negative. Ofcourse, a range of arguments that we have seenabove concerning the reasons why the invest-

WORLD DEVELOPMENT8


9/22

ment to growth relationship might not bepositive apply as much to signicant negativeresultsour only point here is to argue that therelationship is certainly not robustly positive. 15

Research and development expenditure

(another staple of some new growth theories)does little better at explaining growthit seemsto have no relationship with economic growthin poor countries and suer from reversecausation in OECD economies (Birdsall &Rhee, 1993).

Looking at the relationship between tradeand growth, both Edwards' (1993) survey andMaurer's (1994) survey found no strong evi-dence of a trade to growth linkage. A morerecent survey by Rodriguez and Rodrik (1999)nds considerable weaknesses in the econo-

metric literature on trade and growth sincethen, and concludes that ``(w)e nd little evi-dence that open trade policies in the sense oflower tari and non-tari barriers to trade aresignicantly associated with economic growth''(see also Walde & Wood, 1999). 16

Looking at scal and monetary variables,Tanzi and Zee (1996) conclude that evidence onthe relationships between tax policy, publicexpenditure, and debt neutrality and growth isinconclusive, 17 while Bruno and Easterly(1995) were unable to nd a long-term rela-

tionship between ination and growth. Simi-larly, institutional variables have weakpredictive powers when it comes to explaininggrowth. Przeworski and Limongi (1993) andBrunetti (1996) both nd that political regimetype (democratic versus authoritarian) has verylittle explanatory power in growth regressions,and Brunetti also provides evidence on theweakness of government instability, politicalviolence and policy volatility as determinants.

A range of other studies that have not reliedon advanced statistical techniques have also

pointed up simple problems with some recenttheories that emphasize the importance ofspecic policy or institutional variables. Overthe long term, US growth rates have beenremarkably stable (we can predict US GDP in1987 to within 5% simply by forecasting usingthe growth data from 1929 and the averagegrowth rate from 18801929). Further, between1870 and 1989, two-thirds of the present high-income countries have GDP growth rateswithin 0.2% of the US rate (Pritchett, 1996a). 18

It appears, then, that policy changes over time

and policy dierences across countries withinthe OECD and in wider samples of countrieshave had almost no eect on very long-term

growth rates. A linked problem faced by policy-based explanations of growth is the predict-ability of income based on past income. Ifreasonably unconstrained policy choiceexplains the rate of growth, and some countries

have followed better policies than others, thenthe rank ordering of country incomes shouldchange dramatically over time, and past incomeshould be little related to present income overthe long term. In fact, neither of these state-ments is true, and past income is a very strongpredictor of present income even over the verylong term (Kenny, 1999).

Over the shorter term, Jones (1995) studiesOECD countries in the postwar period. Hedocuments that trade openness, durablesinvestment, average years of education and

literacy rates have all risen. According totheories of growth that suggest these aresignicant determinants of economic perfor-mance, the OECD should have seen increasinggrowth rates. But, what change there has beenin these rates has been downward. 19

At even shorter, 15-year periods, Easterly,Kramer, Pritchett, and Summers (1993) notethat growth rates in a global sample of coun-tries are in fact highly volatile while many ofthe causal factors we usually look at, such aspolicies and institutional structures tend to

change rather slowly over this time frame.Figure 4 shows the relationship of growth ratesover two of the decade periods commonly usedin regression analysisas can be seen, therelationship is fairly weak. Compare this to thecorrelation of an education variable (the log ofone plus average years of schooling in thepopulation) across time in Figure 5. 20 In other

Figure 4. The relationship between growth over decades.



10/22

words, policies vary too slowly to explain thevolatility of shorter-term growth rates, and toofast to explain the stability of longer-termgrowth rates.

In addition, Stiglitz (1999) notes the problemposed for models which tie growth to policies(or, indeed, institutions) by the Chinese andRussian growth experience over the last decade.Since 1989, China's GDP nearly doubled, while

Russia's almost halved. This suggests that themore rapid reform path taken in Russiashock therapywas not a success. Growthmodels based on the Washington Consensus orNorth's market institutionalism would suggestthat Russia's growth rate should have increasedas it moved toward more market-friendlypolicies and institutions. On the other hand,drawing any strong conclusion against shocktherapy comes up against the problem thatPoland, the Czech Republic and Hungary allintroduced stronger reforms than Russia and

have since faired much better. Another prob-lem faced by theories based on a primacy ofnational policy is the weak response of Africancountries to adjustment. Several Africancountries that have conscientiously imple-mented sound macro policies have still onlyaveraged a growth of 0.5% per year over thelast few years (Buckley, 1999).

The huge regional inequality of incomewithin countries that face the same macropolicy framework is a nal reason to suggestthat there are factors at work other than those

macro conditions and institutions usually(more or less formally) modeled. The ruralpoverty rate in Guizhou province, China, is 7-

10 times higher than that in Guangdong prov-ince, for example. 21

As of yet, then, we do not appear knowwhether policy or institutional changes haveany major eect on long-term income levels,

and if they do, whether they cause a one-oincrease or decrease in output, or whether theylead to a permanent change in long-termgrowth rates. 22 Given that many theories ofeconomic growth provide such rm conclusionson the universal relationship between variablesand growth, it may come as a surprise that theeconometric evidence is so disappointing. Thisis especially so given that nearly all econometricwork on growth has been ``playing tennis withthe net down'' (Blaug, 1992): attempting simplyverication rather than falsicationproving

that the evidence does not contradict thetheory, rather than that the evidence proves thetheory. 23

Why is this? First, our relationships mayappear weak because so much of our data is soweak (Kamarck, 1983; Leontief, 1971). Deatonand Miller (1995) nd that, for a sample ofAfrican countries, GDP gures from twodierent sources over time have a correlationcoecient of less than 0.5 (the sources areSummers and Heston and the IMF). Killick(1992) nds that, between 198689, the IMF

estimates that African exports grew 0.3%, whilethe World Bank estimated the gure at 3.2%.Second, some of our results will be weakbecause many of our regression results arebased on faulty econometric techniquesal-though it is not always clear what the righttechniques are (Evans, 1996; Caselli, Esquivel,& Lefort, 1996; Levine & Renelt, 1991; Mayer,1992; Blaug, 1992; Pritchett, 1998). 24 Third,there are things that some are fairly condentare linked to the growth process, but that wehave struggled to nd a measure or proxy for

(knowledge and innovation, for example).Linked to this, weaknesses might persistbecause we are unclear with our terms, or ourterms are inherently unclear. Kamarck (1983)points out that African mines that feed minersa proper diet for several weeks after recruit-ment are increasing consumption expendi-turesbut in a way that denitely increaseshuman capital. Quite possibly, indeed, it raisescapital in a way that has a higher return than``investment'' expenditures on added equip-ment. In a regression-related example, Pritchett

(1996c) notes that the six trade variablesfrequently used in the growth literature arelargely uncorrelated across countries. 25

Figure 5. The relationship between schooling over

decades.

WORLD DEVELOPMENT10


11/22

Finally, and fundamentally, we have seenthat the relationships might appear weakbecause when we run large-scale cross countryregressions, we are assuming a global commonsystem of causal relationships between vari-

ables and output that might not exist at anylevel that can be caught using a cross countryregression. Indeed, Stewart (1991) and Grierand Tullock (1989) both nd strong support forthe contention that global samples should besplit by region, because relationships betweenvariables and output change sign and strengthof correlation across groups. For example,Stewart found that when running the sameregression in Latin America and Africa, thepopulation growth variable appeared to be apositive and signicant cause of economic

growth in Latin America while being a negativeand signicant cause of growth in Africa.Campos and Nugent (1999) nd that the insti-tutional variables most closely associated withgrowth vary between an East Asian and a LatinAmerican sample of countries. Limiting studiesto one region does not appear to solve theproblem, however. Four studies at the start ofthis decade performed regressions on a purelyLatin America sample (De Gregorio, 1992;Stewart, 1991; Cardoso & Fishlow, 1989; Grier& Tullock, 1989). Four variables were used by

more than one study, and of these, only one(export growth) was found signicant in all ofthe studies in which it was used (in this case,

just two studies). Depending on time frame orwhat else we take account of in our studies,policies or economic characteristics of a coun-try that look terribly important and positive inthe growth process in one study frequently looklargely irrelevant or negative in another.

Perhaps econometric samples should notonly be split by region, subregion or even to theindividual country level, but perhaps they

should also be split by time. For example, itmight be that the oil crisis had a signicantstructural eect on the causal processes ofeconomic growth. In the short term, it doesappear that the strength of relationships andgrowth is very dependent on the period onestudies. 26 Longer term evidence for theimportance of period is provided by studies of19th and early 20th century growth, whereHarley and Crafts (1998), Morris and Adelman(1988) and Taylor (1996) nd that inuences ongrowth posited by modern theories have very

little explanatory power in explaining 19thcentury growth (further, Adelman, 1999, pointsout that there is not one model that can explain

all 19th century growth, either). This might notbe very surprising. The economy of present-dayPakistan looks nothing like the 1820s economyof the United Kingdomdespite the fact thattheir GDP per capita is broadly similar ($1,642

for Pakistan in 1992 as opposed to $1,756 forthe UK in 1820 Maddison, 1997). Two obviousreasons are that much of Pakistan's capitalstock involves goods not yet invented in 1820and that the relative world prices of goods havechanged dramatically over that time. 27

Samples could perhaps be split by sector aswell. Bernard and Jones (1996) nd evidencethat convergence results in OECD countries aredriven by services, while manufacturing showsno sign of convergence. This suggests the twosectors might work very dierently. Again,

however, splitting by sector alone again mightnot be enough. Mundlack, Larson, and Butzer(1997) report strong evidence against a globalproduction function for agriculture, for exam-ple. But if the only way to get a ``robust'' resulton the causes of economic growth is to split ourstudies by country, period and sector, thissuggests that the economics of an advancedindustrial country such as the United States in1960 can tell us little about the economics of anagricultural country such as Kenya's in 1960and even less about the economics of Kenya

today. Not only does the evidence suggest littlesupport for an assumption of homogeneity,then, but to the contrary, strong evidence infavor of heterogeneous processes at work.

All of this is not to say we know nothingabout the growth process. First, we have someevidence that institutions and/or structuralfactors matter in explaining the ``mean'' aroundwhich growth rates perform their violentrandom walk. In one of the most ambitiousrecent attempts to nd some ``robust'' vari-ables, Sala-i-Martin, in his (1997) four-million

regression test paper, nds 21 variables that heconsiders robust to a less rigorous version ofLevine and Renelt (1992) analysis. The test isbased on the idea that if correlations are nearlyalways (95% of the time) of the same sign evenallowing for dierent ``conditioning sets,'' itsuggests that variables have a strong positive(or negative) relationship with growth even ifthe correlations are not always statisticallysignicantly related to growth. 28 It should benoted that Sala-i-Martin's variables are onlyshown to be weakly robust in a one period

regression using a largely constant set ofcountries. Despite this weakness in the test, itmight be interesting to look at what survives.



12/22

The variables that survive the process includethose linked with religious aliation, region (acountry being located in Africa, for example),primary goods orientation (exporting agricul-tural and mineral products), latitude (being

near the equator), political and civil rights, therule of law, war, revolutions and coups,investment, foreign exchange variables, acountry's ``degree of capitalism'' and numbersof years the country can be classied an openeconomy. Fully half of these variables can beconsidered ``structural,'' and only two can beconsidered ``true'' policy variables. Thesepolicy variables are both general measures thatcapture a multitude of sinsdegree of capital-ism and number of years an open economyand are open to dispute as fair measures of

policy orientation.29

This might suggest that structural variablesthat governments have trouble changing, suchas religious beliefs, or geographic location, orthe quality of land, play an important role inthe long-term growth process, something alsosupported by a number of recent studiesincluding Sachs (1997a,b), Easterly and Levine(1997a), Kenny (1999) and Hall and Jones(1996). Again, the results also t with evidencepresented earlier suggesting that a country'sgrowth performance tends to look like a

random walk around a mean the mean is,perhaps, set by structural factors. 30 This is notto argue that a structural constraint is and willremain always a hurdle to growth whatever theperiod, sector, level of development or otherconstraints present. It is merely to note thatfactors beyond the power of man (or govern-ment) to change could be important elements indetermining growth rates over the long term,and that, given the importance of structuralvariables, it is quite possible that dierentpolicy mixes might be optimal in the presence

of dierent structural constraints.There are a number of other factors that

appear fairly robustly associated with growth.It is a near-universal phenomenon that theproportion of labor in agriculture declines ascountries become richer and that manufactur-ing and services contribute more to GDP(Syrquin & Chenery, 1989). We know thatconsumption of food as a percentage of totalexpenditure drops rapidly as people becomewealthier. 31 We have weaker evidence ofchanges in attitudes (Inglehart, 1995). What we

do not appear to know, and what global samplecross country econometric studies have notbeen able to tell us, are the causal relationships

at work and the processes that link factors tooutcomes.

4. DISCUSSION

It is a familiar (and justied) refrain thattheories or models of economic growth mustsimplify the world they attempt to depict.Indeed, most if not all growth models are onlydesigned to be partial rather than completemodels of the operation of economic change. Itwould seem to be a condition of good modelsor theories, however, that they do in some wayactually represent, albeit in a simplied form,that part of the world they are concerned with,capturing enough of the important causal real

world relationships to have explanatory andpredictive power.

While there are problems and inadequacies inthe statistical techniques frequently used toassess these theories, the universal failure toproduce robust, causally secure relationspredicted by models might suggest a broaderproblem than statistical methodological weak-nesses. 32 The evidence appears to suggest thatcountry growth experiences have been extre-mely heterogeneous, and heterogeneous in away that is dicult to explain using any one

model of economic growth. In this section wereturn to the connections between epistemologyand ontology and discuss two related problemswith the way in which economic growth hasbeen theorisedtheir ahistoricism and theiraccounts of causality.

First, despite the fact that economic growthis a process which takes place over oftenextended periods of time, most models ortheories of economic growth are in an impor-tant sense ahistorical. Apart from a fewexceptions they tend to assume that the process

of economic growth is the same the world overand through time (because the components andprocess of all economies are the same).

Why might this ahistoricism be a problem?Easterly (1997a) points out that there areprobably virtuous and vicious cycles at work indevelopment, connected with ``thresholdeects'' (an idea that dates back at least toRostow, and was revisited in Chenery, Elking-ton, & Sims, 1970). For example, if investmenthas a higher return in the presence of largestocks of capital because of various externali-

ties to capital accumulation (a central elementof Romer's thesis, looked at earlier), then if acountry has a certain critical mass of human

WORLD DEVELOPMENT12


13/22

and physical capital to start with, growth willbe a virtuous circle of capital accumulationattracting yet more capital. But a terms of tradeshock, for example, that rendered part of acountry's capital stock useless might shift some

countries near the critical mass level of capitalbelow the critical line, so shifting it from stronggrowth to decline. On the other hand, the sameshock would have little long-term impact oncountries distant from the threshold.

Ravallion and Jalan (1999) note anotherpossible cause of a vicious cycle: ``mixed''goods involving both public and private bene-ts (such as health care) will be less wellsupplied in poor areas. Agriculture is also fullof examples: returns to insecticide spraying arehighly dependent on water availability, early

planting, weeding and fertilizer. African eldtrials suggest that returns to insecticide spray-ing are eight times higher in the presence ofthese other factors than in their absence (Carr,1993). In order to escape a vicious cycle of lowreturns to agriculture, then, farmers need to beable to implement a range of new practices allat once.

The presence of virtuous and vicious cyclesitself raises questions of long-term pathdependency (and so historically-situated inter-pretations of the growth experience). A small

dierence in the initial capabilities or condi-tioning sets of a country could send it o onvery dierent growth trajectories. But this alsoraises the question of interaction eects acrossinternational borders. We have seen that theglobal and neighborhood environment some-times appears to have a signicant impact ongrowth outcomesand it might even be thatthese eects dominate long-term developmentoutcomes. While the dependency view ofhistory tends to be fairly linear and universal, incommon with the standard views of economic

growth, it does suggest the importance oflooking at the dynamics of growth at levelsother than the national. Indeed, it is possiblethat the global economy is a complex systemone that is impossible to dissect into compo-nent parts because the system itself arises fromthe interactions among those parts, involvingany number of feedback loops that producecounterintuitive results. This would bring intoquestion the validity of cross country growthregressions where the country is the unit ofobservation. 33 If this is true, it would also

make it impossible to produce a satisfactorypredictive model of a particular sector of aparticular type of economy.

Regarding causality, most growth modelscould be said to have perhaps an over-simplecausal account at their heart. John Stuart Millargued that the ``cause'' of any particular eventis the set of conditions or factors which, taken

together, are a sucient condition for the eventto occur (Mill, 1975, Book 3). That is, eventsrarely, if ever, have a single cause, but arerather the result of a conjuncture of severalfactors or conditions. Mackie (1980) hassuggested that even this is not a adequateaccount of the complex causal nature of events,and has instead suggested that we view causa-tion in `INUS' terms. For Mackie a cause is anyfactor or condition that is itself an insucient,but necessary, part of a conjuncture of factorsor conditions that are unnecessary but sucient

for the event to occur. To put it in terms ofeconomic growth, it may be that increasingrates of investment is an INUS factor. It maybe an insucient but necessary part of a set ofconditions which cause growth to occur, butthat growth is not necessarily always caused bythis conjuncture of conditions.

Given the complex causal nature of theworld, Mackie argues that scientic laws do notstate ``what always does happen,'' but rather``what would, failing interference, happen''(Mackie, 1980, p. 76). Given that, as far as we

know, the social world is more causallycomplex than the natural world, and hencethere are likely to be many more things whichcould ``interfere'' with the supposed causalconnection, and given the evident impossibilityof isolating causal relations in the social worldin a ` laboratory,'' it might be a mistake tothink that there are any small number of INUScauses of economic growth which will workunaltered across countries over time. AsFrancois Perroux has argued,

when a rst class statistician establishes a parrallel be-tween the growth of GNP and that of major aggre-gatesinvestment, consumption, savings, etc.thereis good reason to doubt whether what he is doingamounts to more than taking the rst step towards acausal analysis (Perroux, 1983, p. 27).

More than 40 years ago, Gunnar Myrdal saidmuch the same thing. He argued that econo-mists concerned with economic growth need toaccept not just that it may have a great numberof causes, but also that these do not work in

any ``linear'' manner. He suggested that prob-lems like economic growth should be examinedusing the concept of ``circular causation'' where



14/22

a change in one factor would aect a number ofother factors, and these changes would in turnfeedback on the rst factor (Myrdal, 1957, p.16). The essence of a problem such as economicgrowth is that ` it concerns a complex of

interlocking, circular, and cumulative changes''(p. 14). For Myrdal this had two implications.First, it was ``useless to look for one predomi-nant factor, a `basic factor'as everything iscause to everything else in an interlockingcircular manner'' (p. 19). Second, viewingeconomic growth in these terms meant aban-doning the search for neat econometric models:``the relevant variables, and the relevant rela-tions between them are too many to permit thatsort of heroic simplication'' (p. 101).

The causal issues involved in economic

growth are further complicated by the prob-lems of accounting for the causal signicance ofpeoples' beliefs about themselves, about others,and about the future. In contrast to atoms inchemical reactions, economic agents areconscious that they are part of an event, areconscious of others, and of past events, and thiscan eect their behavior. What people do in thesocial world is in part the result of what theythink about other agents, what they think otheragents think about them, and so on, and whatthey believe about the future. Again as Franc-

ois Perroux has argued, economics has yet toget to grips with the idea that individualeconomic agents are active, thinking persons,not simply through-puts in the working out oftimeless and spaceless economic laws and rela-tions (1983, pp. 6970). But this injects animportant element of indeterminacy intoprocesses of economic change, which is notgenerally captured by the current way ofthinking about growth. Sadly, there is littlereason to think we will ever be able to graspfully beforehand this kind of process. Over the

long term especially, it suggests that thecomplexity of predicting the eects of policy orinstitutional change on economic relations oroutput.

Finally, as the ongoing debates about theIndustrial Revolution in England show, evenhistorically-centered, nonuniversal, complexcausal models of growth cannot be conclusivelyproventhe historical record does notstraightforwardly provide a single plausibleaccount of any event.

Overall, while the work of the natural scien-

tist is underpinned by an assumption that therereally are a few general principles which struc-ture the natural world, and while economic

theory would suggest the same is true of theeconomic world, there appears to be littleconvincing empirical evidence that this isactually the case as regards growth. This mightmean that current models only capture a

supposed theoretical relationship that we havelittle reason to think will actually be manifestedin the real economic world.

The above argument might suggest thatmathematical modeling techniques have inva-ded territory to which they are ill-suited,imposing unrealistic criteria of rigor andprecision. 34 We do not want to embroilourselves too far in the debate over formalmodeling in economics, between those who seemodeling as the ``mental gymnastics of aparticularly depraved mind'' and those who see

non-mathematical treatments as ``fat, sloppyand vague'' (Samuelson & Klein, in order,quoted in Roy (1989, p. 140)). Many earlydevelopment economists did not use mathe-matical models (Lewis'book is almost comple-tely absent of equationsalthough he did havea role in starting the trend toward moremathematical treatments), but while some earlyand most later development economists did usemodels, there is apparently little dierencebetween them in terms of their predictive andexplanatory ability.

In judging the utility of complex models, wewould, however, argue with Krugman when heclaims that self-conscious formalization leadsto better results. As he points out formalmodeling cannot easily handle complex ideasonly recently has it been able to encompass theidea of increasing returns, and it still struggleswith imperfect competition. Krugman's argu-ment that the ``essential logic'' of high devel-opment stories can emerge in a simpliedsetting using models (Krugman, 1997) isperhaps close to a truism, and suggests that we

will lose the important nuances which appearvital in accounting for growth outcomes. Toargue that we are lost in the fog without modelsmight be to wish the all too real fog away.Because the formal models tend to be evenmore parsimonious than descriptive theoriesabout development, it is perhaps not unsur-prising that they have been, if anything, evenless helpful in illuminating what appears to be ahighly complex and varied phenomenon. 35

Perhaps one could develop a model thatallowed for the heterogeneity, and took into

account all of the interactions between vari-ables, but that model could not be testedbecause (as we have seen in regard to trade) our

WORLD DEVELOPMENT14


15/22

theoretical concepts are loose, because there arenot (and could not be) the undisputed econo-metric tools to declare such a model uniquelyand causally congruent with the data, andbecause there is certainly not the quantity of

accurate data required to test it. Further, ourexperience with short-term predictive models ofgrowth might give us pause in such an exercise.It appears that models that use thousands ofvariables to predict economic conditions a yearor two out are no better than those that useonly one or two variables (Sherden, 1998).

Heterogeneous growth experiences are only asurprise if we think that ``underneath,'' as itwere, the components and process of all econ-omies are the same, and that the complexinteraction of external conditions, noneco-

nomic factors and peoples' beliefs have no realcausal importance in explaining growth expe-riences. We have argued that there are goodempirical and theoretical reasons for thinkingthis is not the case. Instead,

we are faced at every turn with the problems of organ-ic unity, of discreteness, of discontinuitythe whole isnot equal to the sum of the parts, comparisons ofquantity fail us, small changes produce large eects,and the assumptions of a uniform and homogeneouscontinuum are not satised (Keynes, 1993).

5. CONCLUSION

There are, we think, a number of conclusionsand implications which follow from the analy-sis we have presented here. First, a review of theavailable evidence suggests that the currentstate of understanding about the causes ofeconomic growth is fairly poor. Clearly therehave been development ``successes'' just as

there have been development ``failures.'' Whatwe are arguing is that we are in a weak positionto explain why some countries have experi-enced economic growth and others not. Thisshould, we think, induce us all to be a littlemore cautious in the certainty with which wehold to present models and modes of thinking.

In part because of this (but also because thefuture is sure to be dierent), we do not thinkthat there is any very rm empirical basis forthe condence with which (any) developmentpractitioners advocate particular policy

prescriptions. The practice of development hasrelied upon theories which all purported toexplain why developing countries had not

experienced economic growth, and purportedto be able to oer cogent advice on what shouldbe done to overcome this. Yet the availableevidence suggests that none of these theorieshas so far been able to do this very well (even

when playing the far easier game of explainingpast growth rather than improving futuregrowth).

Even if some might argue our selective reviewhas been overly pessimistic in its evaluation ofthe output from cross country regressions, theconcerns we have raised over the leap betweencorrelation and causation, the quality of data,the looseness of concepts and the sheer numberof variables that have been put into regressionsshould at least suggest the fear that policyrecommendations are sometimes nothing more

than a mechanical output of the particularchoices made by the model-builder.

The possible dangers inherent in the use ofregression analysis as a source of policy advicecan be illustrated by looking at the use ofsometime signicant coecient estimates. Forexample, dierent studies suggest that thefollowing will increase per capita growth by1%: increasing the average years of schooling inthe labor force by 1.2 years, increasing the ratioto GDP of public investment in transport andcommunication by 1.7%, a fall in ination of

26%, a reduction in the ratio of the governmentbudget decit to GDP of 4.3% (Easterly,1997a). But, rst, we know that these results arenot robustthat the size and signicance of thecoecients change dramatically across samples.Second, we know one of the reasons for thislack of robustness is that these variables arelinked with many others. It might be thatchanging policies to improve these variables isimportant to growth outcomes, it might be thatthey happen to be usually correlated with thefactors that really matter, and intentionally

trying to change them without altering the trulysignicant correlate will have no eect ongrowth. There might also be interactionsbetween the variables. For example, if thegovernment uses budget decits to nancetransport and communications, what happensto growth? It might also be that these policiesand outcomes are signicantly linked to thegrowth process in one country at one time, butnot signicantly linked in another country atanother time, or that dierent initial conditionscompletely reverse the eects of policy change

at dierent times in dierent places. The ina-tion variable shows up the importance ofaccounting for nonlinear relationships between



16/22

variables. If reducing ination by 26% reallyincreased growth by 1%, this suggests thatBrazil's GNP per capita growth rate wouldhave gone from 5% in 1994 when the countryhad an ination rate of 2,948% to a per capita

growth rate of somewhere in the region of118% in 1996 when the ination rate haddropped to 16%.

On the other hand, many of the variablesfound unrobust or insignicant in cross countryregressions might in fact have a large eect oncertain countries at certain times while beinginsignicantly or negatively related in othercountries or at other times. The danger ofrelying on such studies for policy advice is not

just one of commission, then.We would argue that this should encourage

development practitioners to exercise caution inusing coecients for anything more than themost tentative of illustrations. At the sametime, this study might suggest the need forcaution in assuming that the ``lessons'' from thedevelopment experience of one sector or onecountry at one time can be applied substantiallyunaltered in dierent sectors or dierent coun-tries at later times. To quote an early develop-ment economist:

all of this is not an argument against economic devel-

opment. But it shows the necessity of applying West-ern experiences and `rational'methods only with greatcaution and with a real eort to understand the cul-ture, value system and social structure of each popula-tion group that is to be exposed to a developmentpolicy (Lauterbach, 1957).

We know that in practice this is an eort farmore widespread than it is in theory (Cohen,

1995), but to the extent that theory drivespractice, it is an eort that should be mademore open and more central to theoreticaltreatments.

It should be noted that some recent treat-

ments of the development process are re-em-phasizing its complex and interlinked nature,perhaps moving the eld back to the moreholistic treatments of the 1950s and before(Stiglitz, 1998). While it can well be argued thatsaying everything is important has little prac-tical use, it might be a more accurate charac-terization of the development process thanfocusing on one or two keys to developmentsuccess. As we do not seem to know exactlywhat we are aiming at, it might be better to usea shotgun than a rie, then.

On the theoretical side, perhaps, more energyshould be directed toward understanding thecomplex and varied inner workings of actualeconomies rather than trying to assimilate theminto abstract universal models. Over 100 yearsago, Thorstein Veblen said substantially thesame thing. He argued then that economistsneeded to abandon their animistic bent, andinstead turn to a study of the complex causalprocesses of economic development (Veblen,1948). Even historical accounts of economicgrowth in particular places will not necessarily

lead to conclusive causal narratives, but itmight at least allow us to escape from some ofthe disabling commitments which characterizecrosscountry statistical investigations intogrowth. It would perhaps allow for the devel-opment of more appropriate interventions,tailored more closely to the particular circum-stances of particular countries.

NOTES

1. The authors would like to thank two anonymousreviewers for their helpful comments.

2. We mean no disrespect to Jerey Sachs, a widely

known, widely respected and highly prolic economist,

who has produced some of the most cited recent work

on economic growth (Asian Development Bank, 1997;

Sachs & Warner, 1995, 1996) and has played a vital

theoretical and practical role in our dealings with

economies in transition (Sachs, 1993). Regarding the

two articles, Sachs argues that, while point estimates

and some important implications changed from one

paper to the next with the addition of geographicvariables, the important benets of good policy remain

signicantthe articles are dierent, but not contra-

dictory, then (correspondence with authors, June 23,1998).

3. In defense of World Bank economists, they no

longer actually believe in the model as a helpful

representation of the long-term growth processif they

ever did. On the other hand, the model still crops up in

discussions of country growth prospects in World Bank

documents and elsewhere (noted in Easterly, 1998;

Deverajan, Easterly, & Pack, 1999).

4. A cynic might note that this list moves from the

relatively simple to overcome toward the impossible tochange (even more so if we take the story in to the later

1990s and add geographic factors).

WORLD DEVELOPMENT16


17/22

5. We do not want to enter into the detailed debates

about methodology in economics. Most economists tend

not concern themselves with such issues, but we think

that most would accept the rough and ready view that

economic theory is to be ``judged'' against the empirical

evidence.

6. Studies with similar ndings include Pons (1999)

and Greer (1999). A recent test of the World Bank's

projections for developing country GDP growth one

year ahead do suggest that they are better than the nave

forecastbut only marginally (Verbeek, 1999).

7. Another example of an early development economist

who seems to cover most if not all of our present

concerns is Lauterbach (1957). He specically mentions:

savings, population growth, barriers to trade, lack of

credit organizations, political traditions, ethnic and

religious division, civil disorder, ination, income distri-

bution, communications systems, crime, public health,

education, technology diusion, property rights,

exchange controls, types of government, corruption

and beliefs. Moreover, Lauterbach and Lewis both stood

on the shoulders of generations of earlier economists

(and others) who had exploited similar ideas. Sachs

(1997a,b) quotes Smith in support of both his contention

that a small government is the sucient condition for

growth and to back the idea that geographical conditions

can be a serious impediment to development. Smith

(1976) also argues that inequitable landholding is bad for

economic development. Heilbroner had not forgotten the

question of the development capacity of tropical coun-

tries in 1963, indeed he argued that: ``Not many years ago

the prima-facie `evidence' made the climate theory of

underdevelopment virtually the prevalent explanation of

economic backwardness.'' The idea that savings and

investment are important to growth goes back even

before Smith, to the Mercantilists (Mason, 1998). On

human and social capital, in Anna Karenina Tolstoy

writes about what was clearly an active debate in 19th

Century Russia over the reasons for the backward state

of the countrycovering topics such as education,

insucient market reforms to encourage peasant entre-

preneurship in the countryside, authoritarian versus

participatory methods of agricultural development.

8. The data are from Summers and Heston (1991).

These data and regressions presented in the following

notes are based on all available data from a 160-country

sample (that of Easterly & Levine, 1997a).

9. Regressing decade average GDP growth on a

constant and decade average investment, N 425,C 3X255, b 0X065, the t-statistics on investment is

4.83, Adjusted R2 is 0.05.

10. With average investment as the dependent variable,

N 439, C 27X9, b 5X9, the t-statistics on log

initial income is 17.4, Adjusted R2 is 0.41.

11. In a regression of growth against a constant, log

initial income and investment, N 423, C 11X

2,b 1X13 (t-stat. )8.3), v 0X15 (t-stat: 9.35), adjusted

R2 is 0.19.

12. Fuentes' (1995) selective review of the growth

literature also argues (in part based on the Levine and

Renelt results) that available estimates of the technology

and convergence parameters appear to be quite sensitive

to sample selection, econometric specication or even

the list of regressors included in the regression.

13. The arguments that rage over growth accounting

exercises, which look at the relationship between stocks

of human and physical capital and GDP levels, are a

further sign of the disputed link between investment and

growthand the role of human capital and technology.

Depending on the exercise one chooses to believe, the

East Asian growth miracle can be explained almost

entirely in terms of high rates of physical and human

capital accumulation (or investment in goods and

education), or largely on improvements in the produc-

tivity of human and physical capital due to technology

adaptation and advance (see, for example, Young, 1995

& Crafts, 1998on the side of capitalversus Klenow &

Rodriguez Clare, 1997on the side of total factor

productivity). Even using similar methodologies

(weights and procedures) the measured contribution of

factors varies greatly by country and period (Crafts,

1999). Felipe (1999) concludes his survey of TFP growth

and capital accumulation in East Asia by arguing that

``this work has become a war of gures. From the

crudest calculations to the most detailed studies F F F. In

many cases these are straight exercises in data mining

embedded in fancy empirical methodsF F F It seems that

re-working the data one can show almost anything.''

14. Techniques to test causation include the Granger

Causality Test, two simple (and somewhat problematic)

measures to overcome bi-causality include using initial

values of variables rather than period averages and the

(slightly more advanced) two-stage least-squares tech-

nique.

15. A non-econometric exercise can emphasize the

point. Worldwide, school enrollment and life expectancy

have shot up over the last 30 yearson average, life

expectancy has risen by four months each year since

1970, while adult literacy has risen from 46% to 70%over that same period. These increases have been felt on

every continent and in nearly every region (World Bank,



18/22

1999). Yet what movement we have seen in worldwide

growth rates over that period has been downward.

16. The dominant theories also suggest that periods of

open trade policies among countries should see conver-

gence in incomes between themso that the dispersal ofincome levels should decline over these periods. The

Rodriguez and Rodrik (1999) study suggests, instead,

that periods of high taris (the depression and the

interwar period) saw convergence in country incomes,

while times of freer trade (the period up to 1878 and the

postwar period) have seen divergence of income.

Further, the relatively open East Asian economies have

seen divergence since 1960 while the relatively closed

Latin American economies have seen convergence.

17. Rather worryingly, they nonetheless conclude that,

despite the lack of empirical support, trusted modelsshould still be used.

18. Clark noted this same phenomenon as early as

1944. He pointed out that even WWI had only a

``comparatively slight and temporary eect'' on long-

term trend growth rates. As we have seen, the same is

true of WWII, suggesting that, even when countries are

doing their best to have an impact on output, they have

only a fairly limited long-term eect (Clark, 1944, p. 4).

19. Pack (1994) also concludes that recent growthmodels do poorly when tested against events such as the

productivity slowdown in the OECD countries or the

Asian growth acceleration.

20. Data for graphs come from Easterly and Levine

(1997a).

21. The reforms introduced in China since the late

1970s have apparently slowed the rate of convergence

across regions of China, but cannot be blamed for the

extent of the initial disparity of incomes (Ravallion &

Jalan, 1999).

22. More broadly, the two most basic, and linked,

questions about the nature of growth remain disputed:

is there convergence and are there decreasing returns

to capital? The answers to these questions depend on

who you ask and how you phrase the question

(Kocherlakota & Yi, 1995; Quah, 1993; Evans, 1996;

Barro & Sala-i-Martin, 1992; Parente & Prescott,

1993). The evidence apparently does not support a

rm judgment between the conclusions of neoclassical

variants, new growth variants or structuralist variants

(see Young, 1995 versus Romer, 1993 on object versusidea gaps).

23. We would accept that there are huge problems with

falsication in economicsone reason that it cannot be

the type of science that some practitioners hope for it. It

is near impossible to impose the necessary controls to

solve identication problems or to ensure the use of the

``right'' model (McCloskey, 1983).

24. For example, Caselli et al. (1996) introduced a

generalized method of moments estimator to overcome

inconsistencies in previous crosscountry-regression work

connected with correlated individual eects and endog-

enous explanatory variables. Pritchett (1998) argues that

their technique is itself awed, however, due to the

volatility over time of growth in a country.

25. The variables are frequency of nontari barriers,

average taris, structure-adjusted trade intensity, an

openness index, a trade distortion index, and price

distortions.

26. This can be illustrated by studying Easterly and

Levine (1997b) Table IV, Column 4 regression (used

because the data were near to hand). This is a 1960-90

regression of decade growth in GDP per capita against

an Africa dummy, a Latin America dummy, log initial

income, log initial income squared, ethnolinguistic

fractionalization, log of years of schooling per capita,

assassinations per capita, nancial depth, the black

market premium and government surplus. Running theregression one decade at a time (without the decade

dummies) the variables that are signicantly related to

growth in each period are: 1960s the Latin America

dummy, the income terms, the black market rate; 1970s,

the Latin America dummy, the income terms, ethnolin-

guistic fractionalization, schooling, the black market

rate, the government surplus; 1980s, the Africa and

Latin America dummies, the income terms, the black

market rate, the government surplus. Across periods,

then, the only robust results from this regression are the

Latin America dummy, initial income and the black

market rate (which might be proxying for terms of tradechanges)although some loss of signicance should be

expected from the attenuation of t-stats given smaller

samples.

27. DeLong (1992) estimates that real incomes in the

United States have risen eight-fold over 18951990.

Incomes in terms of purchasing power of particular

goods have shown very dierent changes, however. In

terms of the number of mirrors the average American

can buy, US citizens are 60 times richer today than in

1895. In terms of the number of silver teaspoons, the

average American is actually poorer (by about 10%)than they were in 1895.

WORLD DEVELOPMENT18


19/22

28. It will be remembered that Levine and Renelt's test

looked at whether the variables remained siginicantly

positively or negatively related in the presence of

dierent conditioning sets.

29. One measures ` capitalism'' at period end, leadingto questions of causality, the other is a composite

measure that includes a variable from the mid-1980s, a

variable calculated only for Africa, and a black market

rate variable which is probably connected with terms of

trade shocks (see Easterly et al., 1993).

30. This also gives more credence to the explanations

of the present distribution of global wealth argued by

Landes (1998) and Diamond (1998)arguments that, in

Landes'case, involve factors going back to at least the

Renaissance, and in Diamond's, back to approximately

10,000 BC.

31. Having said that, there appears to be a surprisingly

weak relationship between economic growth and many

desirable outcomes that we have traditionally assumed

went along with higher incomes. For example, economic

growth does not appear to have been behind the

worldwide improvement in life expectancy, or improved

political rights, or access to sanitation and clean water

(Easterly, 1997b).

32. It is possible that we have missed the causally

signicant variablesbut if it is true that we have been

looking in all the wrong places for the development

silver bullet, this too suggests we need to re-examine our

theories.

33. Some economists are now trying to model such

global systems, spurred on by Krugman's work in the

eld of the economics of geography (Krugman, 1997).

Examples include Matsuyama (1996) and Baldwin,

Martin, and Ottaviano (1998).

34. As Frank Hahn, the president of the Econometric

Society in 1970, complained about econometric models

in general: ``there is something scandalous in the

spectacle of so many people rening the analysis of

economic states which they have no reason to suppose

will ever, or have ever, come about'' (Mayer, 1992, p. 2).

35. To take but one example, Romer's (1987) model

predicts that labor growth will be bad for long-run

growth because increased labor supply reduces rm's

incentives to invest. This might sound plausible for

manufacturing. In agriculture, however, there is plenty

of micro evidence to suggest that increasing labor

pressure is vital to spurring investment and advance

through land-saving technologies (Tomich, Kilby, &

Johnston, 1995, p. 385).

REFERENCES

Adelman, I. (1999). Fallacies in development theory andtheir implications for policy. Paper presented at `` TheFuture of Development Economics in Perspective.''Dubrovnik, May.

Arrow, K. (1974). Limited knowledge and economicanalysis. American Economic Review, 64(1), 110.

Asian Development Bank. (1997). Emerging Asia:Changes and challenges. Manilla: Asian DevelopmentBank.

Baldwin, R., Martin, P., & Ottaviano, G. (1998). Globaldivergence, trade and industrialization: the geographyof growth take-os. Center for Economic PolicyResearch Discussion Paper 1803, February.

Barro, R., & Sala-i-Martin, X. (1992). Convergence.Journal of Political Economy, 100(2), 223251.

Bernard, A., & Jones, C. (1996). Comparing apples tooranges: productivity convergence and measurementacross industries and countries. American EconomicReview, 86(5), 216238.

Birdsall, N., & Rhee, C. (1993). Does research anddevelopment contribute to economic growth in devel-oping countries? World Bank Policy Research Work-ing Paper, No. 1221.

Blaug, M. (1992). The methodology of economics: Or howeconomists explain. Cambridge: Cambridge Univer-sity Press.

Brunetti, A. (1996). Political variables in crosscountrygrowth analysis. Journal of Economic Surveys, 11(2),163190.

Bruno, M., & Easterly, W. (1995). Ination crises andlong-run growth. World Bank Policy Research Work-ing Paper, No. 1517.

Buckley, R. (1999). Annual review of development eec-tiveness. Washington, DC: World Bank.

Campos, N., & Nugent, J. (1999). Development perfor-mance and the institutions of governance: evidencefrom East Asia and Latin America. World Develop-ment, 27(3), 439452.

Cardoso, E., & Fishlow, A. (1989). Latin Americaneconomic development: 19501980. National Bu-reau of Economic Research Working Paper No.3161.

Carr, J. (1993). Improving cash crops in Africa. Wash-ington, DC: World Bank.

Caselli, F., Esquivel, G., & Lefort, F. (1996). Reopeningthe convergence debate: a new look at crosscountrygrowth empirics. Journal of Economic Growth, 1,363389.

Chenery, H. B., Elkington, H., & Sims, C. (1970). A

uniform analysis of development patterns. Economicdevelopment report 148. Cambridge, MA.: HarvardUniversity Center for International Aairs.



20/22

Clark, C. (1944). The economics of 1960. London:Ma

Documents

Kenny, Charles. 2000 What do we know about Economic Growth Or Why don´t we know very much