When Do Incumbents Learn From Entrepreneurial Ventures

8/10/2019 When Do Incumbents Learn From Entrepreneurial Ventures

1/25

Research Policy 34 (2005) 615639

When do incumbents learn from entrepreneurial ventures?Corporate venture capital and investing rm innovation rates

Gary Dushnitsky a,, Michael J. Lenox b

a The Wharton School, University of Pennsylvania, 2031 Steinberg HallDietrich Hall, Philadelphia, PA 19104, USAb Fuqua School of Business, Duke University, P.O. Box 90210, Durham, NC 27708, USA

Received 7 August 2003; received in revised form 20 June 2004; accepted 10 January 2005Available online 28 April 2005

Abstract

In this paper, we focus on the potential innovative benets to corporate venture capital (CVC), i.e. equity investments inentrepreneurial ventures by incumbent rms. We propose that corporate venture capital programs may be instrumental inharvesting innovations from entrepreneurial ventures and thus an important part of a rms overall innovation strategy. Wehypothesize that these programs are especially effective in weak intellectual property (IP) regimes and when the rm hassufcient absorptive capacity. We analyze a large panel of public rms over a 20-year period and nd that increases in corporateventure capital investments are associated with subsequent increases in rm patenting. 2005 Elsevier B.V. All rights reserved.

Keywords: Technology entrepreneurship; Corporate venture capital; Innovation; Appropriability

1. Introduction

A rms dynamic capability, i.e., its ability to de-velop and acquire valuable resources and capabilities,is largely related to nding knowledge external to therm and integrating it with internal knowledge ( Teece

et al., 1997; Henderson and Cockburn, 1994 ). In thisvein, scholars have examined a number of ways inwhich rms may source external knowledge includ-ing regional learning, university research, recruitment

Corresponding author. Tel.: +1 215 898 6386;fax: +1 215 898 0401.

E-mail address: [email protected] (G. Dushnitsky).

of high-human capital personnel, mergers and acquisi-tion, and alliances. However, little attention has beenpaid to the potential innovative benets of investingcorporate venture capital (CVC). Corporate venturecapital is equity investment by incumbent rms in in-dependent entrepreneurial ventures, i.e.,relatively new,

not-publicly-traded companies that are seeking capitalto continue operation ( Gompers and Lerner, 1998 ).At its peak in 2000, established rms invested more

than $16 billion, or 15% of all venture capital invest-ments, through nearly 400 CVC programs ( VentureEconomics, 2001 ). Despite a signicant drop-off incorporate venture capital investment corresponding tothe decline in the nancial markets in 2001, a number

0048-7333/$ see front matter 2005 Elsevier B.V. All rights reserved.doi:10.1016/j.respol.2005.01.017


2/25

616 G. Dushnitsky, M.J. Lenox / Research Policy 34 (2005) 615639

of leading corporations including Intel, Microsoft,Merck and others have continued to invest in newventures ( Chesbrough, 2002 ). Ostensibly, these in-

vestments are aimed to generate a substantial nancialreturn on investment through the sale of ownershipstakes post initial public offering (IPO) or througheventual acquisition. However, rms often assert thatone of the major motivations for investing is to gaina window onto valuable, novel technologies so as toimprove rm innovative efforts. Many rms report thatthis is their primary motivation for investing corporateventure capital ( Siegel et al., 1988 ).

In this paper, we explore the prospects for CVCto provide a window on technology. We investigatewhether CVC investment is associatedwith knowledgegains to incumbentrms. Do rms that invest corporateventure capital learn about and appropriate new tech-nologies and practices from those ventures in whichthey invest? Our investigation builds on two theoreti-cal pillars. First, the knowledge necessary to generateinnovations may likely reside outside the boundary of incumbent rms ( Arrow, 1974; Cohen and Levinthal,1990 ). Second, entrepreneurial startups may be a valu-able source of such knowledge ( Aghion and Tirole,1994; Kortum and Lerner, 2000; Shane, 2001 ).

Placed together, these two pillars suggest that cor-

porate venture capital programs may be an importantinstrument for accessing valuable knowledge fromentrepreneurial ventures. We propose that corporateventure capital investments increase incumbent rminnovation rates. We expect this effect to be especiallystrong when rms possess the internal capabilities toleverage venture knowledge and operate in an externalcontext that permits capitalization on that knowledge.In particular, we propose that in weak intellectualproperty (IP) regimes, CVC will provide privilegedaccess to venture specic knowledge and permit

greater opportunities to learn from that knowledge.Furthermore, we propose that the greater a rmsabsorptive capacity, the greater the marginal impact of CVC investment on rm innovation rates.

Empirically, we analyze a large unbalanced panel of U.S. publicrms duringthetime period19751995. Of those rms in our dataset, approximately 250 rms en-gaged at least in some level ofCVCinvestingduring thetime period. Data is pulled from VentureXpert, Com-pustat, and the U.S. Patent databases. Discrete choice,xed-effect and random-effect models are adopted to

establish a relationship between CVC investment andpatenting. We adopt a number of controls to addresspossible unobserved, time-variant heterogeneity. We

nd that increases in CVC investment are associatedwith subsequent increases in future citation-weightedpatenting rates. Furthermore, we nd that the magni-tude of this effect depends on the rms absorptive ca-pacity and the strength of intellectual property protec-tion.

The paper contributes to the growing literatureson sourcing external knowledge, broadly, and invest-ing corporate venture capital, narrowly. Unlike otherknowledge sourcing arrangements, CVC investment isa capital expenditure that is easily observed and mea-sured. The deployment of other external innovative in-puts is often difcult to observe and even more chal-lenging to determine their cost. For example, the costof establishing personal ties with university scientistsis difcult to estimate ( Cohen et al., 2002 ). Data onthe cost of maintaining R&D alliances is typically notavailable to researchers and may not even be calcu-lated by alliance members. The ability to measure thedollar amount of corporate venture capital investmentsenables us to better capture its elasticity with respectto incumbent rm patenting output. More importantly,theseinvestments areobservedirrespective of theirsuc-

cess or contribution to rm innovation rates.The research presented in this paper complements

the existing corporate venture capital literature in anumber of ways. Previous empirical research on ven-ture capital investments by corporations has focusedprimarily on the narrow, nancial returns to investingin new ventures (e.g., Gompers and Lerner). While ahandful of empirical studies have looked at the broaderstrategic benets of venturing activity (e.g., Siegel etal., 1988; Chesbrough, 2002 ), none of these studies di-rectly evaluate the effect of external corporate venture

capital investment on incumbent rm innovation usingpatenting.Our extensive longitudinal dataset allows formore sophisticated econometric models and allows usto address issues of unobserved heterogeneity and tem-poral precedence.

2. Theory and hypotheses

Researchers have long postulated that incumbentrms operating in competitive markets are inclined to


3/25

G. Dushnitsky, M.J. Lenox / Research Policy 34 (2005) 615639 617

innovate in order to sustain protability ( Schumpeter,1942; Arrow, 1962 ). Yet, despite this economic incli-nation to innovate, numerous researchers have high-

lighted the organizational limits of established rmsto generate innovations internally ( Henderson, 1993 ).The view that incumbent rms face difculties in gen-erating ground breaking, radical innovations is well es-tablished ( Tushman and Anderson, 1986; Henderson,1993 ). Innovation in large part requires the integrationof diverse knowledge sets ( Arrow, 1974 ). To the ex-tent there are constraints on the creation and sharingof knowledge within a single organization, incumbentrms may nd that they lack the knowledge necessaryto innovate.

To overcome these constraints, incumbent rmsmay seek to exploit knowledge external to the rm(Cohen and Levinthal, 1990 ). This has been the fo-cus of numerous studies that investigate the abil-ity to create new knowledge through the recombina-tion of knowledge across organizational boundaries(Henderson and Cockburn, 1994 ). Potential sourcesof knowledge include regional networks of employ-ees and rms ( Saxenian, 1990; Almeida and Kogut,1999 ), academic and government labs ( Cohen et al.,2002 ), and other established rms accessed eitherthrough technology alliances or mergers and acqui-

sitions (Hagedoorn and Schakenraad, 1994; Gulati,1995; Powell et al., 1996; Capron et al., 1998; Ahujaand Katila, 2001 ).

Entrepreneurial ventures may be a particularlyimportant source of knowledge. A number of scholarshave advanced the idea that entrepreneurial venturesare likely to be the source of highly valuable andinnovative ideas ( Kortum and Lerner, 2000; Zingales,2000 ). According to this line of reasoning, starscientists/innovators will opt away from xed salary(i.e., as an employee in a corporate R&D lab) and

towards prot sharing (i.e., founding their own newventure) when they have an idea that is highly lucrative(Anton and Yao, 1995; Gans and Stern, 2003 ). Thus,we expect to observe the formation of new venturesonly when entrepreneurs have highly innovativeideas (Aghion and Tirole, 199 4). Indeed, Kortumand Lerner (2000 ) observe that entrepreneurial,human-capital intensive ventures generate higherlevels of patenting output than established rms.Shane (2001) provides further empirical evidence thatnew venture formation is associated with underlying

entrepreneurial inventions that are of high economicvalue.

Both practitioners and scholars have proposed that

corporate venture capital may provide a valuable av-enue to access this pool of knowledge ( Chesbrough,2003; Gans and Stern, 2003; Poser, 2003; Roberts andBerry, 1985 ). By investing in new ventures, incumbentrms may gain a window onto the ventures technolo-gies and practices ( Chesbrough and Tucci, 2002 ). Ex-posure to novel, pioneering technologies may in turnincrease the likelihood that established rms wouldcreate breakthrough innovations ( Ahuja and Lampert,2001 ). Siegel et al. (1988) in a survey of 52 corporateventure programs, report that corporations rank ex-posure to new technologies and markets as the lead-ing objective for engaging in corporate venture capi-tal programs. Corporate venture capital may also con-tribute to rminnovativeness by increasing thedemandfor corporate innovations. Investment in ventures withcomplementary product and services may increase thedemand for current and future corporate products andconsequently encourage further corporate innovations(Brandenburger and Nalebuff, 1996 ).

A similar point has been made in the related tech-nology alliances literature. Hagedoorn and Schaken-raad (1994) and Ahuja (2000) indicate that knowledge

from innovative alliance partners may spillover andpositively affect the innovativeness of a rm. Stuart(2000) shows that rms patenting rates increase themore technologically advanced are its alliance part-ners. Rothaermal (2001) argues that incumbent rmsthat pursue alliances improve their new product devel-opment efforts when faced with radical technologicalchanges.To the extent thatentrepreneurial startupspos-sessvaluable knowledge,corporate venturecapital mayprovide a rm a window onto the operations of portfo-lio companies that can generate similar advantages to

those associated with having innovative alliance part-ners.

2.1. CVC investment

We propose that CVC investment provides a uniqueopportunity to learn from new ventures. In many in-stances, entrepreneurs desire and actively seek corpo-rate investment. Entrepreneurial ventures are willingto accept incumbent access to their operations in ex-change for the benets they receive from having a cor-


4/25


porate partner. In addition to funding, incumbent rmsprovide substantial value added services ( Block andMacMillan, 1993 ) as well as facilitate new ventures

access to complementary assets ( Teece, 1986; Pisano,1991 ), endorsements ( Stuart et al., 1999 ), and interna-tional product markets ( Acs et al., 1997 ). Indeed, em-pirical evidence indicates that CVC-backed venturesfair better than independent venture capitalist-backedones (Gompers and Lerner, 1998; Maula and Murray,2001 ).

There are at least three channels through which cor-porate venture capital activity facilitates rm learningfrom entrepreneurial ventures. First, the due-diligenceprocess provides the rm a unique opportunity to learnabout entrepreneurial inventions even prior to commit-ting capital. Post investment, an investor may learnabout novel technologies by maintaining board seats(or board observation rights) as well as utilizing dedi-cated liaisons. Finally, a failing venture may also con-stitute a learning experience to the extent that it offerstechnological insights, or conversely points at marketunattractiveness.

Prior to funding a venture, investors conduct thor-oughdue diligenceof the ventures operations andbusi-ness plan. This generally includes a background check of the founders and key management team, discussions

with key customers, and a review of the product andtechnology. Corporate venture capitalists often lever-agecorporate R&Dpersonnelto gauge a ventures tech-nological feasibility and consult corporate executivesto determine business and market risks. Henderson andLeleux (2002) observed that a person or team from thebusiness unit was typically involved in the due dili-gence process. For example, Agilent Technologies cor-porate venture capital group works closely with thecompanys existing businesses to share information,qualify investmentopportunities,and connect portfolio

companies to Agilents own initivatives ( Chesbrough,2002 :6).As a result, a rms business units become privy

to the novel technologies that they review. In addition,the business units often gained insight into future tech-nologies and products ( Chesbrough, 2002 ). Consider,for example, Intels investment in Berkeley Networks:As Intel performed its due diligence on its investment,it began to see the outlines of a possible strategy shift. . . culminating in the Intel Internet Exchange Archi-tecture, launched in 1999. The investment in Berke-

ley Networks allowed Intel to identify a promising op-portunity more quickly than it might have otherwise(Chesbrough, 2002 :9).

Once the investment round has taken place, cor-porations employ various mechanisms to learn fromtheir portfolio ventures. Corporate investors often se-cure board seats, or at least board observation rights,which provide them with knowledge of ventures keyactivities and technologies. UPS actively pursued suchadvantages from its CVC activities: United ParcelService decreases the distance between business unitsandventureinvestmentsbyrequiringboard observationrights for its leading business managers. Board meet-ings help senior UPS business managers learn aboutstart-up operations, technologies and business models,increasing UPSs ability to capture strategic value fromits venture investments (Corporate Executive Board,2000). A study of 91 U.S.-based ventures that oper-ated mainly in the computer and communication in-dustries during the late 1990s, nds that in 31% of thecases the corporate investor held a board seat and in40% of the cases it did not have a seat but did hold ob-server rights ( Maula, 2001 ). These results are echoedin a recent survey of European venture capital prac-tices (Bottazzi et al., 2004 ) that reports CVC investorsserve on portfolio companies boards (68%) and con-

duct close monitoring and site visits (70%) almost asfrequently as independent venture capitalists do (78%,76%, respectively).

Incumbent rms also institute specic organiza-tional routines to encourage and funnel learning fromventures at the post-investment stage. It has long beenrecognized that successful learning is contingent uponclose interaction between rms personnel that accom-modates rich media and information ows ( Arrow,1974; Daft and Lengel, 1986 ). Sony Corporation cre-ated two parallel and distinct functions responsible for

knowledge transfer between its portfolio companiesand the corporation ( International Business Forum,2001 ). In addition to securing board seats to its CVCgroup (Sony Strategic Venture Investments), Sonysbusiness divisions established liaisons with the ven-tures. These liaisons specic goal was to learn aboutand source the portfolio companys technology. Mo-torola pursued a similar strategy as part of its own CVCactivities (Corporate Executive Board, 2000).

CVC investors can derive benets even from failedportfolio ventures when technologies remain viable af-


5/25


ter the originating venture has dissolved ( Hoetker andAgarwal, 2004 ) and failure itself carries informationalweight ( McGrath, 1999 ). Chesbrough (2002:6) echoes

this view of CVC investments: Agilent has recentlyinvested in a startup company making wireless radio-frequency devices, a product area Agilent plans to ex-plore in its own business. If this investment is success-ful, Agilents future business will benet; if it fails,Agilent will get a valuable early warning about pitfallsto avoid in that business.

Through these learning structures, incumbent rmsthat invest in entrepreneurial ventures may increase thestock of knowledge from which they may base innova-tion. The larger a rms equity investment in new ven-tures, the greater the stock of entrepreneurial knowl-edge a rm has access due to either (a) access to agreater number of new ventures (i.e., more opportuni-ties to conduct and learn from due-diligence, as well asboard observation rights and witnessing failure), or (b)greater access to their portfolio companies (i.e., greaterleverage vis- a-vis the venture and hence more chancesto secure board seats and deploy liaisons).

The greater the stock of entrepreneurial knowledgea rm has accessed, the greater the subsequent inno-vation output. This increase in a rms innovation ratemay result from at least two mechanisms ( Ahuja and

Katila, 2001 ). First, to the extent that innovation resultsfrom the novel combination of existing knowledge, theexpansion of the rms knowledge base will lead to agreater number of possible knowledge congurations(Kogut and Zander, 1992 ). Second, the exposure tonew technologies and practices through CVC invest-ment will increase a rms absorptive capacity, i.e., theability to absorbanduse additional external knowledge(Cohen and Levinthal, 1990 ). Thus, we hypothesize:

Hypothesis 1. All else being equal, greater rm in-

vestment in entrepreneurial ventures leads to increasesin the investing rms innovation rate.

2.2. Intellectual property regime

The magnitude of the innovative impact of CVC in-vestment is likely to vary across both industries andrms. The degree to which a rm may innovate asa result of its investment in entrepreneurial ventureslikely depends on the inability of ventures to shel-ter their intellectual property ( Anand and Galetovic,

2000; Gans and Stern, 2003 ). All else being equal, themarginal benet of CVC investment should be greaterthe weaker the intellectual property regime, i.e., when

ventures struggle to protect their innovations from im-itation through legal mechanisms such as patents. Thereasons for this are two-fold.

First, corporate venture capital may be a uniquelyadvantageous strategy for gaining a window on en-trepreneurial technologies in weak IP regimes. Ven-tures will likelyresortto secrecy as a way to appropriatethe gains from innovation in industries where the intel-lectual property regime is weak ( Cohen et al., 2001 ).In strong IP regimes, ventures may have the option topatent their technology and license it to prospectiveincumbents. For example, biotech rms often patentand license their technology to pharmaceutical rms.CVC investment in these environments is less likelyto provide privileged access to external knowledge,compared with other strategies (e.g. licensing). In con-trast, in weak IP regimes, the ability to conduct due-diligence, as well as the chance to participate on boardmeetings provide the corporation with a unique oppor-tunity to pierce the veil of secrecy and provide an ef-fective vehicle to learn about the ventures closely-heldtechnologies.

Second, even when patented, a ventures innovation

is more likely to be appropriated by investing rms un-der weak IP regimes. Defending the rights for a patentis costly ( Lerner, 1995 ). When the intellectual propertyregime is weak, a cash-constraint venture may not havethe means to prohibit investors from appropriating itsknowledge. In weak patent regimes, ventures may ndit too costly to receive and defend patents for their tech-nology. Similarly, survivors of a failed venture maynd it difcult to enforce patent rights on the formerventures technologies. The same technology may leadto patents for investing rms whopossess the resources

necessary to prot from patenting. Incumbent rms aremore likelyto have theresources necessaryto ght law-suits andotherchallenges to their patents. Furthermore,incumbents are more likely to possess complementarycapabilities in research, manufacturing, and distribu-tion, which they can leverage to their advantage.

Hypothesis 2. The weaker intellectual propertyregimes, the greater a rms investment in en-trepreneurial ventures will impact the rms innovationrate.


6/25


2.3. Absorptive capacity

We do not expect the marginal effect of CVC in-

vestment on innovation rates to be uniform across allrms. Within a focal industry, the degree to which arm may learn from its CVC investments will dependin part on the absorptive capacity of the rm. Cohenand Levinthal (1990) advance the view that internaland external sources of innovations are interdepen-dent. In line with their absorptive capacity argument,Kleinknecht and van Reijen (1992) report that havingan internal R&D department increases the likelihood of co-operative R&D with other rms. Others have alsoreported that rms with an expertise in a given researchdomain exhibit higher levels of knowledge absorptionfrom external sources ( Pisano, 1991; Veugelers, 1997 ).

We propose that the impact of investment in en-trepreneurial ventures on rm innovation rates will begreater for those rms who have a strong base in in-novation. The ability of an investing rm to transfer orcreate knowledge through its interaction with a venturelikely requires a rm to have sufcient technical un-derstanding to both grasp and capitalize on that knowl-edge. Internal research and development provides thefoundation upon which rms may learn from the ven-tures they invest. R&D personnel are often mobilized

to conduct technological due diligence, and some arelater chosentoserveas liaisons between corporate busi-ness units and the ventures. Indeed, Intels investmentexperience and high technological absorptive capacityallowed it to recognize, as early as the due diligencestage, the opportunity associated with Berkeley Net-works technology.

Hypothesis 3. The greater a rms absorptive capac-ity, the greater a rms investment in entrepreneurialventures will impact the rms innovation rate.

3. Data and method

In our analysis, we explore the relationship betweencorporate venture capital and incumbent rm innova-tion rates. Specically, we explore the varying levelsof rm CVC investment and compare their subsequentinnovation rates. We adopt patent data as a measure of innovation to be consistent with recent studies on tech-nology alliances ( Baum et al., 2000; Stuart, 2000 ) and

entrepreneurship ( Ahuja and Lampert, 2001; Almeidaet al., 2003 ). We cannot directly measure the gener-ation of patents as a result of investment in specic

ventures, because the data do not permit such attribu-tions. Following Hausman et al. (1984) , we adopt apatent production function (see methods below) to es-timate the sensitivity of a rms innovative output tovarious innovative-inputs in general (for a review seeGriliches,1990 ) andventure capital investments in par-ticular (Kortum and Lerner, 2000 ).

3.1. Sample

We constructed a large, unbalanced panel of U.S.public rms during the period 19691999. The panelincludes all public rms that invested corporate ven-ture capital or patented during this period. Theresultingdataset includes 2289 rms and 45,664 rm-year ob-servations. Thedatabase containsinformationon rmsventuring activity collected from Venture EconomicsVentureXpert database,patentingactivity fromthe Hallet al. (2001) datasetderivedfrom theU.S. Patent Ofce,and nancial data from Standard & Poors Compustatdatabase.

To constructour database,we rst identiedthe pop-ulation of rms engaging in corporate venturing activ-

ity through the VentureXpert database. 1 The databasecontains a comprehensive coverage of investment, exit,and performance activity in the private equity industryfrom 1969 to 1999. We searched the population of allprivate equity investments for any investments by rmsor their funds. 2 For these rms, we collected data onthe annual amount of venturing investments (disburse-

1 The database is offered by Venture Economics, a division of Thomson Financial. The data are collected through industry associ-ations (European Venture Capital Association, the National Venture

Capital Association,andotherkey associationsin Asiaand Australia)and the investment banking community. These data have been usedin several academic studies on the venture capital industry ( Bygrave,1989; Gompers, 1995 ).

2 We included the following VentureXpert categories: Non-Financial Corp Afliate or Subsidiary Partnership, Venture/PE Sub-sidiary of Non-Financial Corp., Venture/PE Subsidiary of OtherCompanies NEC, Venture/PE Subsidiary of Service Providers, Di-rect Investor/Non-Financial Corp, Direct Investor/Service Provider,SBIC Afliatewith non-nancial Corp.and Non-FinancialCorp.Af-liate or Subsidiary. We excluded investments by corporate pensionfunds because these investments are distinct and unlikely to result inlearning benets.


7/25


ments). Note, these amounts represent the actual dollarvalue invested during a given year, and should not beconfused with dollars committed to venturing activity

(usually reportedin theprofessional media), which rep-resent the total dollar amount a fund has committed toinvest over the funds life. We believe the former mea-sure is more appropriate given the focus of our studyon the benets of having a viable window onto innova-tive ventures. Similarly, we focus on CVC investmentmade directly in entrepreneurial ventures and excludecapital allocated to intermediaries such as traditionalventure capital funds (e.g., the practice of fund-of-funds).

We augmented our CVC data with patenting dataconstructed by Hall et al. (2001) (hereafter HJT) thatare based on the U.S. Patent Ofce database. We addedto our sample all U.S. rms within the same industries(based on four-digit standard industrial classication,SIC, classication) as those rms in our CVC datasetover the time period 19691999. The HJT datasetscontain detailed information on almost 3 million U.S.patents granted between 1963 and 1999, all citationsmade to these patents since 1974 (over 16 million),and a match of patients assignees to Compustat. Stan-dard & Poors Compustat database was used to provideannual rm-level accounting and nancial data thus

limiting our sample to publicly traded rms. An au-tomated, matching algorithm and hand-checking wereused to link the VentureXpert data with the HJT patent-ing dataset and Compustat. 3 Furthermore, while theHJT patenting dataset bases the ownership structureof rms circa 1989, we manually matched up rmsto U.S. Patent and Trademark Ofce (PTO) assigneecodes to ensure we capture the patenting activity of allrms.

The resulting sample included 2289 rms and45,664 observations. Two hundred and forty seven

rms of these rms invested corporate venture capi-tal some time during the period 19691999. For eachof the 2289 rms, we collected annual nancial data

3 The HJT database provides not only the name of the companyto which the patent was assigned but also the ultimate (parent) rmname. Therefore, a rm with a single entry in Compustat can havemultiple assignee-names in the HJT database. Since the knowledgegainedthrough a venturing activitycan be put to use in differentdivi-sionsand subsidiaries of thecorporation, we consider every assignee-name associated with a corporation,as long asit holds more than 50%of the original assignee.

such as research expenditures and book value of assetsfor the years 19691999 from Compustat. We estab-lished the annual patent output for each rm using the

HJT dataset. Approximately 80% of the public rmsfrom the VentureXpert database were issued a patentover the 19691999 time frame.

3.2. Measures

We capture the rate of innovation within rms usinga rms citation-weighted count of patents ( Citations ).There are a number of measures of a rms innovative-ness adopted in the empirical literature (see Hagedoornand Cloodt, 2003 , for a recent review). Some studiesemploy R&D expenditures as a proxy of a rms inno-vation competencies ( Henderson and Cockburn, 1994 )while others employ surveys of new product announce-ments (Acs and Auretsch, 1988 ). Patents (Griliches,1990 ) and patent citations ( Trajtenberg, 1990; Harhoff et al., 1999 ) are perhaps the most popular measuresof innovative performance. Based on an analysis of 1200 companies in high-tech industries, Hagedoornand Cloodt (2003: 1375) conclude the overlap be-tween each of these four indicators is that great . . .that in high-tech sectors any of these four indicatorscould be taken as a measure of innovative performance

in the broad sense. 4We opt to capture the rate of innovation within a rm

as the count of forward (future) citations to patents ap-plied for ina givenyearby a rm ( Citations ).5 In doingso, we build on previous studies that have found patentcitations is a good indicator of the value of the inven-tion (Trajtenberg, 1990; Harhoff et al., 1999 ). Previousstudies have employed patents and citation-weightedpatents to gauge innovative output in the chemical(Ahuja, 2000 ), pharmaceuticals ( Henderson and Cock-burn, 1994 ), information ( Stuart, 2000 ) and devices

(Brockhoff et al., 1999 ) sectors, among others.

4 We believe their conclusion applies to our study: Hagedoornand Cloodts (2003) sample and our sample share many character-istics. Specically, they studied companies operating in four high-tech industries (aerospace and defense, computer and ofce machin-ery, pharmaceuticals, electronics and communications), over the pe-riod 19921998. Their results hold for the sub-sample of U.S. rms(which constituted about 70% of their all sample).

5 To increase condence in our ndings, we also ran each of ourmodels with unweighted patent counts as well. We discuss thesealternative specications in the results.


8/25


Fig. 1. Total patents and citations in our sample (19691999).

Only patent applications that were later granted areincluded in our dataset. We chose to use applicationyear rather than grant year, because our main inter-est is to record changes to the focal rms knowledgebase rather than its ability and timing of appropriatingrents. Hall et al. (2001) report that the distribution of forward lag in citation (i.e., the number of years be-tween patents application date and later citing-patentsapplication date) is about 34 years. Thus, our citation-weighted patents variable is truncated on the right.Patents granted in recent years have not been availablelong enough to be cited by future patents (see Fig. 1).

Variation in citing behavior across technological eldsand along years may also bias our citation-weightedpatents variable ( Hall et al., 2001 ). We address thesepotential biases in our methods discussion below. 6

Our primary independent variable is annual CVCinvestments in millions of dollars (CVC). This vari-able is calculated as the sum of all investments via allventuring funds bya rmin a year. Control variables in-clude annual rm research expenditures in millions of dollars ( Research ) and rm size measured as total rmassets in millions of dollars ( Assets ). We would expectthat larger in-house research expenditures would leadto greater patenting output ( Henderson and Cockburn,1996 ). Similarly, larger rms possess greater resourcesfor investing in research and thus are more likely topatent more ( Schumpeter, 1942; Cohen and Levinthal,

6 We also followed the citation-xed-effects approach of Hallet al. (2001) and rescaled our dependent variable by dividing theabsolute number of citations made to a given patent, by the averagenumber of citations made to all patents in the same technologicalgroupand application year. Ourresults are robust to thisspecication.See Hall et al. (2001) for a complete discussion.

1990 ). All three variables are adjusted to 1999 dollarsusing the Consumer Price Index.

In most models, we include a measure to capture the

technological opportunities available per industry. Wedene ( Industry Citations ) as the average number of citation-weighted patents applied to by rms in a givenyear in a given industry dened by each four-digit stan-dard industrial classication. This measure helps con-trol for time-variant,differences across industries, suchas citation ination in certain industries, that may bedriving an apparent relationship between CVC invest-ment and patenting. More importantly, by controllingfor Industry Citations , we attempt to address the factthat some industries at some points in time may ex-perience greater technological ferment that may driveboth the opportunities to invest in new ventures andthe opportunities to innovate internally ( Klevorick etal., 1995 ). We further address this potential confound-ing, alternative hypothesis by performing a separateanalysis for individual industry segments.

As anadditional control, we include a measure of thestock of patents a rm has been granted ( Patent Stock ).Since we analyze a relatively long period of time, it ispossible that unobserved, time-variant rm-level char-acteristics such as alliance activity or mergers and ac-quisitions affect patenting activity. Following Blundell

et al. (1995) , we calculate this measure by calculatingthe depreciated sum of all patents applied from 1963to the current year. 7

Patent Stock it

= ln(Patents it + (1 ) Patent Stock it 1)

We include this measure topartiallycontrol forpossiblerm-level unobserved heterogeneity that is not xedthroughout time. Previous studies examining patentingrates have used similar measures ( Ahuja and Katila,2001 ).

To testourhypothesis concerning themoderating ef-fect of intellectualproperty protection, wecreatea mea-sure derived from the Carnegie Mellon Survey (CMS)of Research and Development ( Cohen et al., 2001 ).

7 We adopted a depreciationrateof 30%as in Blundell etal. (1995) .We experimentedwith othervalues(e.g., 0%) andreceived consistentresults. At a depreciating rate of 30%, patents granted prior to 1963have little impact on 1969 Patent Stock especially given the 14 yearlag between patent application and granting.


9/25


The questionnaire was administered in 1994 to a ran-dom sample of U.S. manufacturing R&D labs drawnfrom the Directory of American Research and Tech-

nology. Overall, 1478 R&D unit managers answeredabout mechanisms they use in order to protect protsdueto inventions in their focal industry. These includeda variety of mechanisms such as patents andsecrecy. IPProtection reects the relative importance of patentingas a way to protect intellectual property. IP Protec-tion was measured as the mean percentage of innova-tions forwhich patentingwasconsidered effective. Thehigher the value of IP Protection , the greater the valueof patenting. Note, by construction, this variable variesacross industry but is time-invariant.

To test Hypothesis 3 , we capture the absorptivecapacity of the rm in a number of ways. It is acommon practice to capture rm absorptive capacitywith its contemporaneous R&D expenditure ( Cohenand Levinthal, 1990 ). Unfortunately, this may simplyadd noise to our results since CVC funds and R&Dlabs likely compete for corporate resources implyinga countervailing substitution effect. To avoid concernsover potential substitution between contemporaneouslevels of CVC and R&D, we use historical level of R&D.In particular, Past R & D measures the3-yearsumof past research and development outlays. That is, our

rst proxy for rm absorptive capacity is the sum of R&D in year t 2 to year t 4.

As an alternative measure for absorptive capacity,we employrm stockofpriorpatents (describedabove)as a proxy for absorptive capacity. On a theoreticallevel, we believe Patent Stock may be an attractive con-struct for rm absorptive capacity. Each dollar spenton internal R&D may not generate the same amountof knowledge stock. According to Hall et al. (2001) ,patents should be a good proxy for knowledge capitalbecause it represents the success of an R&D program

not just its input.Absorptive capacity, as originally conceived, is nota state variable rather it is domain specic. A rm maypossess an absorptivecapacity in onedomainof knowl-edge but lack absorptive capacity in another domain.To address the domain specic nature of absorptivecapacity, we adopt a third measure based on the tech-nological proximity between an investing rms do-main of expertise and the domain of expertise of therms portfolio ventures. In particular, we follow Jaffe(1986) and examine the extent of an overlap between

CVC and portfolio companies technological portfo-lios. Each rms technological portfolio is determinedby measuring the distribution across patent classica-

tions of thepatents associated with itsbusinesses,usingSilvermans (1996) concordance between SIC codes(i.e., line of business) and patent classes (i.e., domainof expertise). We believe this measure is more appro-priate than the cross-citation measure of technologicaloverlap ( Mowery et al., 1996 ) for the purpose of mea-suring a rms absorptive capacity. 8

Generating the Venture Proximity measure involvesa number of steps. First, we identify the technologicalportfolio that is associated with each corporate investorbased on its businesses ( Silverman, 1996 ). Next, foreach venture within an investing rms portfolio, weidentify the ventures domain of expertise using thesame methodology. Since these ventures are not yetpublicly traded companies, they are not typically as-sociated with a SIC code. Rather, we only know theirVenture Economics Industry Classication (VEIC)aVenture Economics proprietary industry classicationscheme. We manually assign a SIC to each venture us-ing a manual two-step process. 9 Third, the overlap intechnologicalportfolios (i.e., activityin the samepatentclasses) is then calculated for each CVC rm-portfolioventure pair. Finally, for each CVC investor, Venture

Proximity is dened as the average proximity acrossall of its portfolio companies.

8 The measure proposed by Mowery et al. (1996) is a strong indi-cator of a technological bond between two rms. However, it mayunderplay information about the similarities of rms that do not citeeachother (a likely event given thatentrepreneurial ventureshave notbeen around for a long period of time). The extent to which similarpatent classes are use by the two rms, irrespectively of whether ornot they cite each other, is consistent with the conceptual constructof rms absorptive capacity.

9

Coding each venture involved the following steps: (1) identifyall ventures in Venture Economics (VE) that ever pursued a publicoffering (IPO); (2) record their VEIC code as available through VE;(3) identify them through Compustat and record their SIC code; (4)generate an initial mapping of VEIC-to-SIC. (5) for a given VEICcode, identify all IPOed ventures and their SIC codes; (6) review rel-evant information about them from VE database. This includes thefollowing VE elds: company business description, company com-petitors,company customers, company internet techgroup, companyprimary customer type, company product keywords; (7) for each of the non-IPOed ventures, review thesame VE elds and assign an ap-propriateSIC code; (8) triangulate ventures line of business throughother databases (e.g., Dun and Bradstreet, Lexis-Nexis).


10/25


3.3. Method

Our dependent variable ( Citations ) is a count of patent citations for patents applied for by a rm in agiven year and as such is bounded at zero and assumesonly integer values. We address the discrete natureof Citations by adopting a negative binomial model.The negative binomial model is commonly used in thepatenting literature for over dispersed count data likeours (Griliches et al., 1987 ). The negative binomialmodel is a generalized form of a Poisson model wherean individual, unobserved effect is introduced in theconditional mean ( Greene, 2000 ). We do not adopta Poisson model because the assumption of constantdispersion appears violated, i.e. the mean and varianceof the event count are not proportional. A Lagrangemultiplier test of overdispersion proposed by Cameronand Trivedi (1998) was used to test this assumption. 10

In general, we assume a 1-year lag between our re-gressors and dependent variables. In other words, weexamine the association between last years value of our independent variables and this years patenting lev-els. In the case of Research , previous research lookingat the relationship between research expenditures andrm patenting have found high within rm correlationof R&D over time thus supporting the use of a one-

year lag (Hall and Zedonis, 2001 ). However, absentprior research on the impact of corporate venture capi-talonpatentingoutput, we take a cautiousapproach andemploy a distributed lag analysis for CVC ( Ahuja andKatila, 2001 ). There is likely to bea lag between invest-ment in new ventures and changes in innovative out-put for the investing rm. While the learning benetsfrom due-diligence may be immediate, more substan-tive learning is likely to take place only after ties havebeen established and trust is built. Furthermore, thelearning that mayderive from board seats and technicalliaisons may transpire over a longer period of time as aventures technology matures. Similarly, learning froma ventures mistakes and failures takes time as well.

Thus, we expect that the positive effect of CVC in-vestment on rm innovativeness will likely transpire

10 The LM test was operationalized using a likelihood ratio test inStata version 7.0.A signicant 2 statistic is interpreted as a rejectionof thebase hypothesis of constant dispersion. To increase condencein our results, we also ran our various models using a Poisson spec-ication and found consistent results. The Poisson models may beprovided by the authors upon request.

for years after the investment activity. In this spirit, weconsider both unrestricted and a geometric models of a distributed lag between CVC investment and inno-

vative output. In the unrestricted case, we consider theeffect CVC investment has on patenting output fromone to six years later: CVC t k where k = 16. We usea 2 test to test the joint hypothesis that the impact of all six lags is signicantly different from zero in com-bination. For the geometric distributed lag model, weassume that the joint impact of CVC activity over timehas the following standard functional form:

CVCdistributed lag = j (1 ) j (CVC) t j .

The geometric lag has the appealing attribute of havingone coefcient to capture the impact of CVC activity.

Since our sample traces multiple industries over along period of time, our ndings might be biased due tounobservable heterogeneity related to macroeconomicfactors, industry attributes, or rm characteristics. Inour analysis, we adopt a number of strategies to ad-dress possible unobserved heterogeneity in addition tothe use of Industry Citations and Patent Stock . First,year xed-effects are included to control for macroeco-nomic trends such as economic downturns and periodsof general technological ferment that may ultimatelyaffect overall patenting levels. Year xed-effects also

control for yearly variations in patent citation rates as aresult of right truncation of the data. To further addressbiases fromthe truncation of citation-weighted patents,we restrict our sample to the period 19751995. 11 Byexcluding the most recent years data, we also removeany bias introduced by the internet bubble economy of the late 1990s. By excluding the years prior to 1975,we improve the accuracy of the HJT matching of rmsto patents. 12

Second, we include rm-effects to address possiblerm-specic unobserved heterogeneity. Firm-effects

are a common practice in panel data analysis and serveas a control for unobserved, time-invariant rm-levelcharacteristics that may in part be driving patent activ-ity within the rm. We employ both rm xed-effects

11 While the number of rms in the sample remains unchanged(2289), limiting the time-period reduces the number of rm-yearobservations to 31,876.12 We estimated our models over a number of different time pe-

riods and found consistent results. Estimations based on the full19691999 panel improved our results. Tables for these estimationsmay be provided by the authors upon request.


11/25


and rm random-effects. In the negative binomial con-text, rm xed-effects are estimated using a condi-tional maximum likelihood estimator. The random-

effects specication assumes that the regressors andthe rm-specic effects are uncorrelated. As explainedin theanalysis below, each of themethods hasstrengthsandweaknesses andthe literature does nothave a strongpreference for one method. We also include sectordum-mies to control for possible xed industry effects thatare not captured by the rm-effects.

In summary, we adopt a negative binomial spec-ication with rm, sector, and year xed and randomeffects and lagged independent variables. The expectednumberof citation-weightedpatents given a setof inde-pendent variables may be given for the random-effectsspecication by,E [Citations it | X it 1] = it = exp( X it 1 + it + i)

where is the coefcient vector, X it 1 represents ourset of time-variant rm characteristics including CVC,and i and it are independent random variables.

4. Analysis and results

Fig. 2 (Panel A) presents a summary of total an-nual investment in new ventures during the period

19692003 (in 2003 dollars). We observe that corpo-rate venturing activity has gone through three wavesduring the last 30 years that closely parallel the invest-ment patterns of independent VC funds. The rst wavepeaked in the early seventies. Activity declined untilapproximately 1978, when changes in legislation ledto an increase in venturing investments by independentventure capitalists as well as established rms. Thissecond wave peaked around 1986 with total annual in-vestment at approximately $250 million per year (inoriginal dollars). Investments declined sharply after the

stock crash of 1987 to a level of $25 million in 1993.The third wave began with the rise of the Internet inthe mid-1990s. By 2000, the total amount of rounds inwhich established rms participated reached a recordlevel of nearly $16 billion (in original dollars).

Table 1 presents a summary of the investment ac-tivity of the 20 largest venturing rms in our sample interms of total cumulative dollars invested. Intel leadsthe list with total investments approaching $1.5 bil-lion since 1992. The top 20 is dominated by the largestelectronics and computer concerns such as Microsoft,

Sony, Motorola, AOL, and Dell. Johnson & Johnsonis the rst pharmaceutical rm to appear in the top 20at the 13th position with cumulative investments ap-

proximating $196 million. The relatively lower invest-ment levels by pharmaceutical rms may simply cap-ture the well documented inclination of these rms to-wards other forms of inter-organizational arrangements(Pisano, 1991; Baum et al., 2000; Rothaermal, 2001 ).For example, the relative incidence of licensing, incomparison to other channels for external knowledgeacquisition, may be particularly high in the pharma-ceutical industry ( Anand and Khanna, 2000 ).13 A ma- jority of the top 20 rms have started their corporateventuring funds after 1993. A few have been engagedin corporate venturing since the sixties including Xe-rox, Johnson & Johnson, and Motorola. A handful of rms including Intel, Sony and Xerox have numerousfunds with which to distribute funds.

Fig. 2 (Panel B) also presents a summary of total an-nualcorporateventurecapital activityby sectors duringthe period 19691997. 14 We exclude data from 1998and 1999 in the graphs since investment levels in theseyears were orders of magnitude greater for a number of sectors (inparticular, the information technology sectorcomprised of semiconductors, computers and telecom-munications). Consistent with Table 1 , we observe that

rms in the information and devices sectors, and tosome extent the pharmaceutical industry, actively en-gage in corporate venture capital. The chemical indus-try and the metals industry had lively CVC investmentduring the 1980s that was not matched during the fol-lowing wave of general CVC activity in the 1990s.We may speculate that much of the run up in CVC in-vestment in the late 1990s was driven by the Internet.For example, the publishing industries surge in CVCinvestment starting in 1995 was a result of establishedmedia moguls such as News Corp. investing in inter-

net start-ups that provided news and other informationon-line.

13 Anand and Khanna (2000:126) Pharmaceutical patents, in par-ticular, can be made very strong . . . . Hence, both the aggregate inci-dence of licensing, as well as the ratio of licenses to joint ventures(i.e., the relative incidence), is likely to be greater when propertyrights over technologies are strong, ceteris paribus. The implica-tions for rms innovation rates are discussed in the results.14 Sectors were denedby SICcode as follows:chemicals (28 ** ex-

cluding 2834 and 2836, 29 ** , 3080), pharmaceuticals (2834, 2836),devices (38 **), and information (357 * , 367* , 48** , 3663).


12/25


Fig. 2. Annual corporate venture capital investments (total and by sector). Panel A: total annual round amount by CVC and independent VC(IVC) investors. Panel B: annual CVC round amount by corporate sector.


13/25


Table 1Activity summary of 20 largest venturing rms (19691999)

Firm Year beganinvesting

Maximum annualventures

Total dollarsinvested a

Maximum annualinvested a

Average annualrounds

Total venturefunds

Intel 1992 179 1486 771 57 5Cisco 1995 55 1056 730 19 1Microsoft 1983 29 713 436 7 1Comdisco 1992 70 554 334 24 2Dell 1995 48 502 395 26 1MCI Worldcom 1996 11 495 410 8 1AOL 1993 39 333 169 10 2Motorola 1963 33 315 177 11 1Sony 1984 30 313 169 7 6Qualcomm 1999 5 262 207 5 1Safeguard 1983 21 231 118 6 3Sun Micro 1999 31 204 180 19 1J & J 1961 21 196 80 5 2Global-Tech 1999 13 188 122 11 1Yahoo 1997 5 186 163 3 1Xerox 1960 30 184 24 13 6Compaq 1992 21 182 113 5 3Citigroup 1999 11 156 93 9 1Ford Motor 1951 22 146 125 8 4Comcast 1996 16 144 84 11 1

a In millions of dollars.

Table 2 presents denitions, descriptive statistics,andthecorrelation matrix for thevariables in our patentcount model. We use the natural log of each of our

primary independent variables. We observe that rms,on average, are granted 13.5 patents per year and re-ceive approximately seven citations per patent. How-ever, the relatively few rms who patent in great num-bers skew these numbers. The main independent vari-able, CVC, varies from $0 to $52.5M in absolute terms.The correlations between independents variables werenot deemed high enough to warrant series concern of multicollinearity. To allayfearsof multicollinearity be-tween rm size and CVC, we estimated each of ourmodels using intensities and found consistent results.

Table 3 presents estimates from various specica-tions of our model. In Model 1, we present a base spec-ication where CVC is not included as an independentvariable. We estimate a rm xed-effect,negative bino-mial specication with Citations as our dependentvari-able using a conditional maximum likelihood (CML)estimator. 15 For simplicity, we do not present the coef-cient estimates for the year and sector dummies. We

15 Our overdispersion test statisticis presented in Models 4 through8 of Table 3 . The signicance of these estimates indicates a rejection

nd not surprisingly that larger rms ( Assets ) who en-gage in more research ( Research ) have higher citation-weighted patenting rates in the next year.

In Model 2, we introduce Patent Stock and Indus-try Citations as controls for time-variant unobservedheterogeneity. Recall that Patent Stock helps controlfor rm-specic heterogeneity such as the underlyinginnovativeness of the rm at different points in time. Industry Citations proxies for the technological oppor-tunities present in the industry at a particular time. Wend that both Patent Stock and Industry Citations havea signicant, positive effect on future patent levels. Incontrast with Model 1, we nd that large rms are lesslikely to generate citation-weighted patents. This result

indicates the importance of controlling for time-variantunobserved heterogeneity when studying a sample thatcovers a long period of time. In essence, the results sug-gests that themarginal output of quality patents is lowerfor larger rms, once we control for uctuation in in-dustry innovativeness (i.e., Industry Citations ) and theintrinsic and most importantly time-variant, or evolv-ing innovativeness of each rm (i.e., Patent Stock ).

of thehypothesis of constant dispersion anda preference for negativebinomial over Poisson.


14/25

Table 2Descriptive statistics for patenting models

Variable Description Mean StandardDeviation

Min Max 1 2 3 4 5

1. Patents Count of new patentsapplied for by a rm ineach year

13.544 62 .107 0 2405 1 .00

2. Citations Citation-weighted newpatents applied for by arm in each year

93.310 458 .381 0 12795 0 .89* 1.00

3. ln(CVC) Log of total annual CVCdollars invested ($M)

0.019 0.187 0 3.98 0.12* 0.13* 1.00

4. ln(Research) Log of total annualresearch expenditures($M)

1.677 1.814 0 9.45 0.45* 0.43* 0.16* 1.00

5. ln(Assets) Log of total assets of rm($M)

5.650 2.010 0.14 12 .61 0.34* 0.31* 0.13* 0.53* 1.00

6. Patent Stock Log of the depreciatedcount of patents issued toa rm

41.318 181 .321 0 5358 0 .98* 0.88* 0.12* 0.46* 0.35*

7. IndustryCitations

Average citations within afour-digit SICclassication in each year

88.377 185 .538 0 3304 0 .38* 0.42* 0.09* 0.35* 0.22*

8. IP Protection Relative effectiveness of patenting within eachindustry

0.418 0.075 0.233 0 .520 0.01 0.01 0.02 0.06* 0.12*

9. VentureProximity

Average technologicalproximity between a rm

and its ventures

0.530 0.379 0.061 1 .000 0.26* 0.30* 0.13* 0.16* 0.53*

n = 31,876 except for IP Protection where n = 10,469 and Venture Proximity where n = 1941. p < 0.01.


15/25


Table 3Firm patent citation levels across all industries (19751995)

Specicationsample model

Firm xed-effects negative binomial Firm random-effects negative binomial

Full sample a , 1 Full sample a , 2 Full sample a , 3 Full sample a , 4 Full sample, 5 CVC Investors b , 6ln(CVC) t i

w /geometric lag 0.065 *** (0.011) 0.066 *** (0.010) 0.043 *** (0.011)w /joint 2 test 81.73 ***

ln(CVC) t 1 0.040** (0.013)ln(CVC) t 2 0.047*** (0.013)ln(CVC) t 3 0.047*** (0.013)ln(CVC) t 4 0.039** (0.013)ln(CVC) t 5 0.028*(0.013)ln(CVC) t 6 0.021(0.014)

ln(Research) t 1 0.328*** (0.008) 0.069 *** (0.008) 0.067 *** (0.008) 0.067 *** (0.007) 0.077 *** (0.007) 0.131 *** (0.018)ln(Assets) t 1 0.058*** (0.007) 0.034*** (0.007) 0.033 *** (0.007) 0.033*** (0.007) 0.059 *** (0.006) 0.067*** (0.019)Patent Stock t 1 0.755*** (0.008) 0.759 *** (0.008) 0.757 *** (0.008) 0.868 *** (0.007) 0.904 *** (0.014)Industry Citations t 1 0.0003*** (0.000) 0.0003 *** (0.000) 0.0003 *** (0.000) 0.0002 *** (0.000) 0.0001 *** (0.000)Year xed effects Included Included Included Included Included IncludedSector xed effects Included c Included c Included c Included c Included IncludedObservations 28321 a 28321 a 28321 a 28321 a 31876 4473 b

Firms 1916 1916 1916 1916 2275 247Wald 2 5607*** 20758 *** 20848 *** 20877 *** 36814 *** 10155 ***

Overdispersion test 5010 *** 655***

Standard errors are in parentheses. p < 0.05. p < 0.01. p < 0.001.

a Firms who never patent are removed from the sample due to the xed-effects negative-binomial specication.b Sample is reduced to only those rms who make CVC investments.c The negative-binomial specication does not have rm-level dummies, thus sector xed effects are possible.

In Model 3, we add CVC investments into ourmodel using the unrestricted distributed lagged struc-ture. Consistent with Hypothesis 1 , we nd evidencethat CVC investment impacts rms patenting output.The coefcients of CVC are highly signicant for veout of the six lagged variables individually, and our joint 2 test suggests that the total effect of CVC in-vestments across the 6 years is statistically signi-cant. Looking at the coefcients, we see the impact

of CVC investment on quality patenting levels beinggreatest 23 years out. Over time, the impact of agiven CVC investment declines until eventually it hasno further effect. We estimated longer lag structuresand found consistently that after 5 years there was nodiscernable impact of CVC investment on patentingrates.

In Model 4, we substitute our geometric distributedlag for the unrestricted distributed lag structure of Model 3. The coefcient for the geometric distributedlag is positive and statistically signicant. As was men-

tioned above, this variable proxies for the joint impactof CVC activity over time on rm innovation rates.Similar to previous models, the coefcient estimatesfor ln(Research), Patent Stock, and Industry Citationsarepositive andhighly signicant andln(Assets) is neg-ative and signicant.

One of the tradeoffs in using the CML xed-effects,negative binomial specication is that rms who neverpatented over the time period are removed from the

sample including a number of rms who engage inCVC but did not patent in the 19751995 time frame.The exclusion of rms who engage in CVC but neverpatent may bias our results. To address this potentialbias, we adopt a random-effects specication of ournegative binomial model. The random-effects speci-cationhastheadvantagethatwemayusethefullsamplein our estimation. By adopting a random-effects spec-ication, we assume that the regressors and the rm-specic effects are uncorrelated. The literature exhibitsno strong preference for random-effects or xed-effects


16/25


Table 4Variable means and standard deviations by major sectors (19751995)

Sector Chemicals Devices Information Pharma Vehicles

ln(CVC) t iw /geometric lag ( = 0.6) 0 .037 (0.290) 0 .022 (0.197) 0 .069 (0.379) 0 .081 (0.354) 0 .021 (0.182)

ln(Research) 2 .291 (1.947) 2 .016 (1.520) 2 .406 (1.932) 3 .718 (1.995) 2 .304 (2.220)ln(Assets) 6 .299 (2.156) 4 .542 (1.778) 5 .340 (2.106) 6 .189 (2.166) 6 .086 (2.116)Patent Stock 2 .331 (2.009) 1 .963 (1.557) 1 .511 (1.743) 2 .851 (2.029) 2 .269 (1.829)Industry Citations 143 .323 (144.092) 106 .672 (177.973) 179 .630 (309.438) 169 .966 (76.189) 127 .805 (152.943)IP Protection 0 .345 (0.047) 0 .324 (0.107) 0 .248 0.087) 0 .432 (0.000) 0 .326 (0.031)Observations 2817 2815 4607 969 1830Firms 198 216 337 63 135

Standard deviations in parentheses.

in limited dependent variable models such as the neg-ative binomial ( Greene, 2000 ).

In Model 5, we estimate our model using Citationsas the dependent variable and a random-effects spec-ication. We nd positive, signicant coefcients forthe geometric lag calculation of ln(CVC), as well as forln(Research), Patent Stock, and Industry Citations.Thecoefcient for ln(Assets) remains negative and signif-icant. Our results for the larger sample are consistentwith the reduced sample from the xed-effects speci-cation. This gives us greater condence in our resultsgiven that we now include rms who never patented

throughout their life.One possible driver of our results is that rms are

self-selecting to invest CVC based on the attributesof the industry sector they operate in as well as vari-ous rm-level factors. In the presence of self-selection,our coefcient estimates may be overstating the directimpact of CVC investment and reecting underlyingdifferences between rms who invest and those whodo not. In Models 35, we control for stable rm dif-ferences with a rm-effect specication and controlfor time-variant differences by the inclusion of Patent

Stock . However, there might be additional time-variantfactors that affect the desire to pursue CVC investmentand the desire to increase patenting resulting in self-selection bias of our estimates. A more stringent testof the impact of rms CVC on its innovative outputshouldexplorewhether increasesin thelevels ofqualitypatenting are driven by prior increases in rms corpo-rate venture capital.

In Model 6, we follow this reasoning and limit oursample to only those rms that invest corporate ven-ture capital during the 20-year time period 19751995.

Here we directly estimate the impact greater CVC in-vestment has on quality patenting, given that a rmchose to make CVC investments. We once again con-trol for rm size, R&D expenditures, patent stock andindustry citations and estimate a random-effect, nega-tive binomial specication using Citations as our de-pendent variable. We nd a positive, signicant coef-cient estimate for ln(CVC). In other words, amongthose rms who invest corporate venture capital, wend that greater CVC investment increases the futureoutput of quality patents. Thus, a rms output of qual-ity patenting is associated not only with the existence

of CVC investment effort, but also with the magnitudeof the investment. 16

As a further analysis, we split the sample by sectorsand for each sub-sample we estimate a specicationsimilar to the one in Model 5 in Table 3 . Analyzingthe sector sub-samples allows us to further address anypotential industry-level confounding effects not cap-tured by Industry Citations and our sector dummies.We separately analyzed a negative binomial random-effects specication, using Citations as the dependentvariable, for the ve highest CVC investing sectors:

16 We have also addressed the self-selection issue more explicitlyby adopting a two-stage specication (2SLS). In the rst stage, weestimate the magnitude of CVC investments. The model is based onDushnitskyand Lenox (2003) who explore not only thepropensity toinvest CVC, but also the magnitude of those investments. The mainindependent variables in the rst stage are rms cash ow, rmshistory of innovation, and industry-level measure of technologicalferment. Our estimates in the second stage are consistent with pre-vious models; once again we nd a signicant, positive relationshipbetween CVC investment and rm quality patenting. The results arenot included for parsimony sake.


17/25


chemicals, devices, information, pharmaceuticals, andvehicles. 17 Table 4 contains the summary statistics forourprimary independent variables broken down by sec-

tor. Because the means and standard deviations are re-ported for all rms within each sector (i.e., investing aswell as non-investing rms), the readershouldinterpretthe values of ln(CVC) with care. 18

Consistent with previous models, our results forModels 7 through 11 in Table 5 show that within eachof the sectors, rms who engage in more research ( Re-search ) and have been innovative in the past ( Patent Stock ) have higher citation-weighted patenting output.We nd that the coefcient of ( Industry Citations ) isstatistically insignicant in most models. This is notsurprising since the information contribution of thisvariable decreases once we split the sample by industrysector: the cross-sector variance does not enter the sub-samples and within sector temporal effects are mainlyabsorbed by the year effects. More importantly, we ndthat the coefcient for the geometric lag of ln(CVC)is positive and highly statistically signicant withinthe information and devices sectors, but insignicantwithin chemicals, pharmaceuticals, and vehicles. Fur-thermore, for the Information and devices sectors, thereturn to CVC investment is similar to R&D expendi-tures.

What may explain these ndings across sectors?Earlier, we hypothesized that the intellectual propertyregimes may drive cross-industryheterogeneity. Refer-ring to Table 4 , we observe that the two sectors withthe weakest IP protection are information and devicesfollowed by vehicles, chemicals, and nally pharma-ceuticals. Thus, we nd that the impact of investmentin entrepreneurial ventures on rm innovation rates issignicant only in those sectors where the intellectualproperty regime is weak.

17 We did not include additional sectors for parsimony sake. Theresults across these industry sectors and may be provided by theauthors upon request.18 The highest mean value for ln(CVC distributedlag ) is reported for

the pharmaceutical sector, which may be misleading. Overall CVClevels are higher in the information and devices sectors (see Fig. 1).However, the number of rms in the pharmaceutical sector is lower.Because the mean is calculated across all rms within a sector, thedenominator in the case of the pharmaceutical sector is lower. Con-sequently, the average CVC level across investing and non-investingrms is highest for that industry.

To explore this result further, we estimate a modelusing our intellectual property protection measure. Re-callthatthismeasureisonlyavailableforasubsetofour

industries thus reducing our sample size. In Model 12,we add IP Protection and interact it with CVC invest-ment. Not surprisingly, we nd that rms in industrieswith strong IP regimes, patent more. In addition, wend as hypothesized that the effect of CVC investmenton innovation is reduced the stronger intellectual prop-erty protection. Given the coefcients of Model 12, weestimate that only forsectors where IP Protection is lowwill we expect to see a positive relationship betweenCVC and innovation.

This raises an interesting question, if CVC invest-ment has positive innovative benets only in industrieswith weak IP regimes, why do rms in industries withstrongIP regimes invest corporateventure capital? Pre-vious work has found that rms invest in new venturesfor a number of reasons. While many rms are inter-ested in CVC as a window on novel technology, othersare interested in CVC purely as a nancial investmentor as a complement to other knowledge acquisitionstrategies ( Arora and Gambardella, 1990 ). Interviewswith CVC fund managers at pharmaceutical rms re-veal that these rms often pursue CVC as a vehiclefor identifying future alliance partners and takeover

targets. Indeed, a recent study nds that a pharma-ceutical rms patenting-output increases in the rmsR&D contracts, licenses, and acquisitions activity, butis insignicantly associated with the rms minority-equity investments ( Nicholls-Nixon and Woo, 2003 ).19

In addition, rms may fund ventures that develop com-plementary product or services, irrespective of the IPregime, as a strategy to increase the demand for corpo-rate innovations ( Brandenburger and Nalebuff, 1996;Dushnitsky, 2004 ).

Next,we raise thepossibility that theimpactof CVC

investment on rm innovative output is moderated byrm level capabilities, in particular, a rms absorptivecapacity. Table 6 analyzes the effect of a rms absorp-tive capacity on the innovation benets due to CVCactivity ( Hypothesis 3 ). We study only those sectorswhere CVC is associated with subsequent patenting

19 Nicholls-Nixon and Woo (2003) report signicant and high cor-relations between pharmaceutical rms investment and its R&Dcontracts and acquisition activity. This is consistent with the viewof CVC as a precursor to future alliance or takeover activity.


18/25


Table 5Firm patent citation levels by sector (19751995)

Sector, model Chemicals, 7 Devices, 8 Information, 9 Pharma, 10 Vehicles, 11 All Sectors, 12

ln(CVC) t iw /geometric lag ( =0.6) 0.032 (0.026) 0.091 *** (0.024) 0.061 ** (0.022) 0.018 (0.040) 0.141 (0.082) 0.362 * (0.176)

CVC IP Protection 0.756 * (0.411)IP Protection 2.424 *** (0.301)ln(Research) t 1 0.142*** (0.023) 0.085 * (0.034) 0.117 *** (0.021) 0.244 *** (0.048) 0.086 *** (0.024) 0.223 *** (0.016)ln(Assets) t 1 0.077 *** (0.019) 0.024 (0.030) 0.089 *** (0.020) 0.233 *** (0.043) 0.006 (0.024) 0.089 *** (0.015)Patent Stock t 1 0.860*** (0.019) 0.800 *** (0.022) 0.861 *** (0.019) 0.839 *** (0.032) 0.849 *** (0.022) 0.001 *** (0.000)Industry Citations t 1 0.0005*** (0.000) 0.0002 (0.000) 0.0001 *** (0.000) 0.0008 (0.001) 0.0001 (0.000) 0.001 ** (0.000)Year xed effects Included Included Included Included Included IncludedSector xed effects IncludedObservations 2817 2815 4607 969 1830 10469Firms 198 216 337 63 135 749Wald 2 5512*** 2807 *** 5255*** 1850 *** 3348*** 2872***

Overdispersion test 667***

505***

610***

234***

71***

5217***

A random-effects negative binomial specication is adopted in all models. Standard errors in parentheses. p < 0.05. p < 0.01. p < 0.001.

output, i.e., the information and devices sectors (wepool the sectors together and introduce sector dum-mies). InModel 13, wesee that for this reduced sample,we continue to nda positive, signicant coefcient fora geometric lag specication of CVC investment.

As a rst test of Hypothesis 3 , we split our samplebased on the median level of our rst proxy for ab-sorptive capacity (AC) Past R & D. We observe thatthose rms above the median level of Past R & D dif-fer considerably from those below the median level.High absorptive capacity rms engage in more R&Don average ($118Mversus $1.8M)andgenerate greatercitation-weighted patents on average (204 versus 7.2).This is driven in part by differences in rm size($2593M versus $1198M in assets on average). Av-erage CVC activity is substantially greater for high

AC rms versus low AC rms (ln(CVC)= 0.063 versus0.023).Estimating our model, we nd that for those rms

with low absorptive capacity (i.e. below the industrymedian) we are not condent that the marginal effectof CVC is greater than zero (Model 14). In contrast,for rms with greater than industry median absorp-tive capacity, we nd a signicant, positive relation-ship between CVC and quality patenting (Model 15).These results allude to an interaction effect betweenCVC and a rms absorptive capacity, consistent with

Hypothesis 3 . To see why, recall that each sub-sampleincludes CVC-investing rms as well as non-investingrms. We nd that the patenting output of low absorp-tive capacity CVC-rms is no better than that of theirnon-CVC, low absorptive capacity peers (Model 14).

However, high absorptive capacity CVC-rms experi-ence signicantly higher innovation rates compared tosimilar companies of high absorptive capacity that donot pursue corporate venture capital (Model 15).

As a more direct test, we pooled these two sub-samples together and interacted CVC investment withdummy variables indicating whether a rm has highabsorptive capacity or low absorptive capacity (usingPast R & D once again). We nd once again a signi-cant, positive coefcient on CVC investment for onlyrms with high absorptive capacity (Model 16).

As a robustness test, we substituted Patent Stock forPast R & D as our measure of absorptive capacity. Split-ting the sample into high and low patent stock usingthe four-digit SIC industry median, we nd a signi-cant positive effect only in the high AC sample. 20 InModel 17, we present the estimates from pooling highand low patent stock rms. We nd, once again, thatonly high absorptive capacity rms realize innovativegains from CVC investment.

20 We do not present these two models for parsimony sake.


19/25

Table 6Firm patent citation levels by absorptive capacity (AC) level for the information & device sectors (19751995)

AC measure,sample, model

All rms,13

Past R&D Past R&D, allrms, 16

Patent Stock,all rms, 17

Past R&D,CVC-rms, 18

V

Low AC rms, 14 High AC rms, 15 C

ln(CVC) t iw /geometric lag ( = 0.6) 0.081 *** (0.016) 0.257 (0.306) 0.078 *** (0.017)

CVC High AC 0.081 *** (0.016) 0.084 *** (0.015) 0.067 *** (0.017) 0CVC Low AC 0.131 (0.247) 0.284 (0.206) 0.305 (0.246) 0CVC AC CVC AC2

CVC unknown AC

ln(Research) t 1 0.118***

(0.018) 0.273***

(0.067) 0.073***

(0.022) 0.118***

(0.018) 0.119***

(0.018) 0.180***

(0.035) 0ln(Assets) t 1 0.078 *** (0.016) 0.018 (0.031) 0.085 *** (0.020) 0.079 *** (0.016) 0.078 *** (0.016) 0.114 *** (0.036) Patent Stock t 1 0.825*** (0.013) 1.075 *** (0.041) 0.809 *** (0.015) 0.825 *** (0.013) 0.823 *** (0.014) 0.911 *** (0.022) 0Industry Citations t 1 0.001*** (0.000) 0.0001 (0.000) 0.0001 *** (0.000) 0.001 *** (0.000) 0.0001 *** (0.000) 0.0001 (0.000) 0Year xed effects Included Included included included Included included inclSector xed effects Included Included included included Included included inclObservations 7422 2217 5205 7422 7422 1941 19Firms 553 175 378 553 553 113 1Wald 2 8306*** 1001*** 6189*** 8307 *** 8350*** 4402*** 4Overdispersion test 1294 *** 20.23 *** 1217*** 1294 *** 1288*** 237*** 2

A random-effects negative binomial specication is adopted in all models. Standard errors in parentheses. * p < 0.05. p < 0.01. p < 0.001.


20/25


Finally, we adopt our measure of venture proximityas our measure of absorptive capacity. Because Ven-ture Proximity is missing for non-investing rms (who,

by denition, have no ventures), we limit our samplesolely to CVC-investing rms in the subsequent anal-yses. For comparison, in Model 18 we replicate Model16 for the limited sample of 113 CVC-investing rms.We nd once again a signicant, positive coefcienton CVC investment for only rms with high absorptivecapacity (using the Past R & D measure). The sign andsignicance of ln( Assets ), ln( Research ), Patent Stock and Industry Citations coefcients remain unchanged.

In Model 19, we present estimates splitting the sam-plebased on those with high venture proximity andlowventure proximity using the median venture proximityfor a given four-digit SIC industry. 21 Surprisingly, wend that those rms who invest in technologically dif-ferent ventures experience greater increases in qualitypatents than those whoinvest in technologically similarventures. Further inquiry revealed a concave relation-ship between Venture Proximity and Citations . 22 InModel 20, we estimate the interaction effect betweenCVC investment and Venture Proximity and VentureProximity squared. We nd a positive, signicant coef-cient for the interaction with the primary term and anegative, signicant coefcient for the interaction with

thesquared term. Thesimilar absolute magnitudeof thecoefcients suggests a concave relationship betweenzero and one with approximately a 0.5 Venture Prox-imity score having the largest marginal effect of CVCinvestment.

This concave relationship is consistent with twopotential explanations: substitution and competition.Both suggest that little learning occurs when rmsknowledge bases are diverse due to a lack of absorp-tive capacity, but offer different reasons for lack of

21

Due to data limitations, we were not able to calculate VentureProximity for all the rms in the Information & Devices sectors. Tomaintain comparable samples, we created a dummy variable Un-known AC to control for those rms for which we could not calculateVenture Proximity .22 While R& D Expenditures and Patent Stock are concerned with

the overall magnitude of an investing-rmsknowledge base, VentureProximity gauges the specic overlap in knowledge bases vis- a-visportfolio companies. The former measures should exhibit a positiverelationship with rm patenting output: the greater rms knowledgebase, the more it is able to learn from CVC. However, the lattermeasure should exhibit an inverted U-shape relationship, where amoderate level of overlap is associated with highest rm patenting.

learning when knowledge bases overlap. The rst viewsuggests that as two agents become closely aligned intheir knowledge sets, their knowledge becomes redun-

dant, andthus very little learningoccurs( Mowery et al.,1998; Ahuja and Katila, 2001; Lenox and King, 2004 ).In the context of CVC, ventures and investing rmswho are well-aligned in their technologicalknowledge,have little to learn from one another. As the divergencebetween knowledge sets grows, investing rms will beable to learn novel insights from their ventures. Eventu-ally, however, this learning will diminish as the invest-ing rms knowledge is so divergent from the venturesknowledge space that the investing rm fails to assim-ilate knowledge from the venture.

The second explanationassumes thatCVCinvestorsare interested in learning but have little opportunities todo so given theactions of ventures. 23 While thegreatestlearning potential may exist when two agents becomeclosely aligned, competitive pressures may lead highlyinnovative entrepreneurs to avoid CVC investors: Asmall high-technology company might be reluctant toapproach IBM or Sony directly for funding. Therefore,thevery companies in which these corporations wantedtoinvestwereusuallytheonesthatnevermadeittotheirdoorsteps ( Gompers, 2002 :1). Based on a large sam-ple analysis, Dushnitsky (2004) concludes that ven-

tures are most likely to be driven away from corpo-rate investors when the two operate within the sameindustry. 24 Consequently, CVC investors will not beexposed to novel technologies in their primary domainof expertise or will be privy only to lower quality en-trepreneurial inventions.

The results presented in Tables 3, 5 and 6 are robustto a variety of specications. To test the sensitivity of our results to construct measurement, we estimated allthe models specied above using simple patent countsrather than citation-weighted patent counts. Our coef-

cient estimates were consistent though often slightlyless signicant. In addition, we explored alternative

23 To the extent that entrepreneurial startups may be a valuablesource of innovations ( Aghion and Tirole, 1994; Kortum and Lerner,2000; Shane, 2001 ), CVC investors may be interested in learningfrom them. Moreover, the corporation may be particularly interestedin learning from ventures that operate in the same technological do-main. These ventures are likely to possess the cutting edge technol-ogy within that technological eld.24 Dushnitsky (2004) also reports that competitivepressures dimin-

ish as the level of complementarity between the two increases.


21/25


measures of corporate venture capital investment in-cluding the intensity of CVC spending and the numberof ventures invested. Once again, our results were con-

sistent with the specications presented. Finally, wetested alternative specications using negative bino-mial xed-effect and Poisson xed-effect estimatorsand found consistent results. The authors may providethese results upon request.

5. Discussion

Consistent with Hypothesis 1 , we nd strong evi-dence that increased corporate venture capital invest-ment is associated with higher future patent citationlevels. Consistent with Hypothesis 2 , we present evi-dence that the relationship between CVC and patent-ing is driven by CVC investments in industries wherepatent effectiveness is low. Consistent with Hypothesis3, we nd that the marginal contribution of CVC in-vestment to patenting is greater for rms with greaterabsorptive capacity. However, we present evidence thatwhen ventures andinvesting rms overlap in their tech-nological knowledge sets, the learning benets of CVCinvestments may be limited.

Our ndings echo recent work on inter-industry

differences in appropriability regimes ( Cohen et al.,2001 ). We nd evidence that corporate venture capi-tal serves as a window on technology only in the in-formation and devices sectors. These two sectors arecharacterized by relatively weak intellectual propertyregimes.Wedo notnd evidencethat rms in thechem-icals, vehicles, and pharmaceuticals sector (where theIP regime is stronger) receive an innovation benetfrom CVC investment. We speculate that rms in thesesectors are pursuing CVC primarily for identicationof future partners or direct nancial returns. However,

we should be careful not to infer too much from thelack of statistical nding in these industries.Our ndings are also consistent with recent work

on absorptive capacity. To the extent that absorptivecapacity is domain specic, we expect to observe anon-monotonic relationship between absorptive capac-ity and learning. On one hand, a rm will learn littleif it lacks expertise related to the knowledge the rmwishes to acquire ( Cohen and Levinthal, 1990 ). On theother hand, to the extent the desired knowledge over-laps with the rms existing knowledge base, coun-

tervailing substitution and competitive effects kick-in(Lenox and King, 2004 ). In particular, rms may ndthat the have little to learn from ventures with simi-

lar technological expertise. In addition, ventures mayhavestrongincentives to eitheravoidor at thevery leastkeepat arms lengths,CVCinvestors whose knowledge(and likely whose products) greatly overlap their own(Dushnitsky, 2004 ).

Our results are robust to a number of controls andmodel specications. We adopt a lagged specication,examiningthe association between previous years val-ues of our independent variables and this years patent-ing levels, in an effort to mitigate concerns of reversecausality. Year xed-effects, annual industry citationlevels, and sector dummies help alleviate concerns thattechnological ferment is driving our ndings. The in-clusion of rm size, research expenditures, previousinnovation rates, and rm-effects help control for thepossibility that rms who pursue CVC are more inno-vative than those who do not.

Furthermore, our ndings are robust to inter-industry heterogeneity. The analysis includes a groupof industry peers all non-CVC rms who operatein the same industry that do not exhibit the sameincreases in patenting output as CVC rms do. Thissupports our view that it is the practice of corporate

venture capital, rather than industry attributes, that isassociated with enhancement in rm innovativeness.Moreover, our inter-industry ndings are conservative.If industry technological opportunity or rm innova-tiveness were driving our results, we would expect therelationship betweena rms CVCinvestmentlevel andits innovative output to disappear in weak IP regimes.All else being equal,more innovativerms with greaterinnovative opportunities will patent less in weak IP en-vironments. While we nd that rm citation-weightedpatentingdeclinesas patent effectiveness decreases,we

report a positive and signicant relationship betweenCVC and innovation within the segment of the popu-lation characterized by weak patent effectiveness.

While we can never completely rule out the possi-bility of some unobserved time-variant factor drivingour results, any potential confound to the observed re-lationship would have to vary contemporaneously withCVC investment. For example, their remains the pos-sibility that rms simultaneously increase their CVCinvestments as they pursue new technology alliancesor university research consortia. However, we believe


22/25


our inclusion of the stock of past patents goes alongway to control for these non-CVC activities aimed atacquiringexternal knowledge.Moreover, the results for

the Venture Proximity analysis are consistent with theview of CVC as an important venue for learning anddriving rms innovative output. Other measures of ab-sorptive capacity (such as R&D expenditure) may ac-tually signify the fact that rms decide to adopt a moreinnovation-focused strategy relative and pursues alter-native innovation efforts. Venture Proximity , however,is specic to rms CVC activity and is not confoundedwith other

Documents

When Do Incumbents Learn From Entrepreneurial Ventures