Research Policy Contents lists available at ScienceDirect Research Policy jo - PDF document

Research Policy 43 (2014) 121… 137 lists available at ScienceDirect Research Policy funds and local biotechnology “rm creationChristos Kolympiris Nicholas Kalaitzandonakes Douglas Miller Management Studies, Wageningen University and Research Center, Hollandseweg 1, 6706 KN, Wageningen (Building 201), De Leeuwenborch, Room 5063,The NetherlandsbDepartment of Agricultural and Economics, University of Missouri, 125 A Mumford Hall, Columbia, MO 65211, USAcNew Analytics LLC and Hutton School of Business, University of the Cumberlands, Williamsburg, KY 40769, USA a r t i c l e i n f o Article history:Received 24 June 2012Received in revised form 3 July 2013Accepted 24 July 2013Available online 24 August 2013 JEL classi“cation:C23Keywords:Biotechnology birthFederal fundingNIH b s t r a c t A long stream of academic literature has established that public funding towards research and devel-opment matters for economic growth because it relates to increases in innovation, productivity andthe like. The impact of public funding on the creation of new “rms has received less attention in thisliterature despite theoretical constructs that support such association. In the present paper we studywhether indeed there is a relationship between public research funds and local “rm births in the con-text of the U.S. biotechnology industry. In doing so, we introduce a number of changes that strengthenthe robustness of our “ndings when compared with existing literature. These changes include a directmeasure of research expenditures and a considerably longitudinal dataset which allows us tocapture a structural relationship and not a chance event. We empirically demonstrate that increases inthe level of research funding from the National Institutes of Health towards biotechnology associate withincreases in the number of biotechnology “rm births at the Metropolitan Statistical Area level. Further,we reveal that public funds towards established “rms associate with local “rm births considerably morestrongly when compared with funds towards universities and research institutes/hospitals. We concludethe paper with academic and policy implications of the present work that highlight the complexity offactors that underlie the creation of local “rms in high technology industries. © 2013 Elsevier B.V. All rights reserved. 1. Introduction Following the Great Recession and recognizing the robustnessof the U.S. knowledge economy, which sustained high employmentand wages amid broad economic weakness, 2009 the U.S. gov-ernment made a strategic decision to substantially increase federalfunding for research and development (R&D) in high technologyindustries such as biotechnology ( 2009; Mervis, 2009 In2011 Ben Bernanke, the chairman of the Federal Reserve, pub-licly outlined the rationale and the in”uential role of government Corresponding author at: Management Studies, University andResearch Center, Hollandseweg 1, 6706 KN, Wageningen (Building 201), DeLeeuwenborch, Room 5063, The Netherlands. Tel.: +31 0317 482372.E-mail addresses: christos.kolympiris@wur.nl (C. Kolympiris), (N. Kalaitzandonakes), (D. Miller).1 subscribe to the de“nition of the knowledge economy in Powell andSnellman (2004) production and services based on knowledge-intensive activi-ties that contribute to an accelerated pace of technological and scienti“c advance aswell as equally rapid obsolescenceŽ. investments in R&D ( 2011 Not surprisingly, interest inmeasuring the returns to public R&D investments quickly followed,not only in the U.S. ( but also, in Europe, Australia,New Zealand and elsewhere (pg. 137 Stephan, 2012 These recentdevelopments have revitalized a general interest towards the rela-tionship between government funded R&D and economic growth.This relationship has been the focus of a long stream of researchthat has stressed the contribution of public R&D to increased inno-vation and productivity and has concluded that the social rateof return to public R&D investment is typically high (e.g. Beiseand Stahl, 1999; Mans“eld, 1991, 1995, 1997; Narin et al., 1997;Salter and Martin, 2001; Tijssen, 2002; Toole, 2012 Such “nd-ings have, in turn, supported continuing public R&D spending overtime.The conceptual underpinnings of such work are also strong.Investments in R&D tend to be risky, mainly due to limited knowl-edge appropriability and uncertainty of outcomes ( 1971 Such characteristics can discourage private parties from investingin R&D because the expected private rate of return is low. In fact,the social rate of return from R&D investments often outweighsthe private rate ( 1992; Hall, 1996 Consequently, gov-ernments be able to correct this market failure by funding 0048-7333/$ … see front matter © 2013 Elsevier B.V. All rights reserved. 122C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137 R&D and increase the odds of socially desirable outcomes ( 1962; Nelson, 1959 sort of argument for government intervention relies heavilyon a complete understanding and accounting of the bene“ts frompublic R&D funding. Yet, one potential bene“t from public R&Dfunding, “rm creation, has received relatively limited attentionin the academic literature despite the strong link between “rmcreation and economic growth ( Praag and Versloot, 2007;Wennekers and Thurik, 1999 Indeed, there are good theoreticalreasons to expect that public R&D funding encourage “rmcreation. For instance, increased R&D expenditures can expand theknowledge base developed in universities and other research insti-tutions and a part of it can be commercially exploited through“rm spinoffs ( and Logothetidis, 2008; Lockett andWright, 2005 “rms also be formed to capitalize onnon-appropriated knowledge (e.g. Acs et al., 2009; Audretsch andKeilbach, 2007 In this study we focus on the question of whether publiclyfunded R&D expenditures lead to “rm births in knowledge-intensive industries and in particular in biotechnology. A number ofstudies have examined the relationship between R&D expendituresand “rm births but most have not delineated the sources of fundsthat support R&D ( and Nerlinger, 2000; Goetz and Morgan,1995; Karlsson and Nyström, 2011; Kim et al., 2011; Kirchhoff et al.,2007; Woodward et al., 2006 Accordingly, our knowledge on theimpact of public R&D funding on “rm creation is limited.In our review of the literature we have identi“ed only that have focused on the impact of public R&D fundingon the creation of biotechnology startups: Chen and Marchioni(2008) and Zucker et al. (1998) Both studies “nd a positive rela-tionship between indicators of publicly funded R&D activity andlocal biotechnology “rm births. Our study adds to the “ndingsof these two studies and introduces a number of methodologi-cal and measurement improvements. For instance, these previousstudies do not distinguish between the type of organization thatreceives the public funding and performs the R&D. Here, we rec-ognize the potential for differential ef“ciencies between industrialand academic R&D organizations on the rate of “rm creation ( and Nerlinger, 2000; Karlsson and Nyström, 2011 and examinethe impacts of public R&D funds directed to universities, private“rms, research institutes and research hospitals separately. Thetwo previous studies have also measured the impact of federalR&D outlays on “rm creation in the biotechnology industry forrather short periods of time (up to two years). Here, extend theperiod of analysis to 18 years (1992…2010) recognizing the inherentlong cycles involved in R&D funding, knowledge development andpotential “rm creation from such new knowledge. As well, insteadof proxies of R&D intensity employed in the two previous studies(a life sciences index and a count of faculty members with grants)we use a more direct and sharper measure of R&D activity, namely,the dollar amount of R&D funding awarded to universities, private“rms, research institutes and research hospitals. focus on “rm births in the biotechnology industry for severalreasons. First and foremost, because the biotechnology industry isa core part of the knowledge economy and understanding how itgrows is important. Second, because the industry is a heavy recip-ient of federal research funds ( and Tulum, 2011 it is a 2 and Martin (2001) and Chaminade and Edquist (2006) elaborate that addi-tional considerations, besides the market failure arguments, are often in place beforethe government intervenes in the market place.3Data from the Association of University Technology Managers (AUTM) suggestthat over the last twenty years more than 9000 university spinoff “rms were cre-ated based on knowledge and intellectual property developed at major researchuniversities in the U.S. fertile ground for our investigation. because of the closelinkage between basic biotechnology research and commercialapplications, there is potential for a strong relationship betweenthe level of R&D activity and “rm births ( and Liebeskind,1998; McMillan and Narin, 2000 Fourth, because the biotech-nology industry exhibits a strong tendency to cluster in narrowgeographies ( and Stephan, 1996; Powell et al., 2002;Zucker et al., 1998 biotechnology “rm births tend to concentratein regions with large venture capital pools, specialized labor pools,and anchor institutions, like large biotechnology “rms and uni-versities ( and Stephan, 1996; Powell et al., 1996, 2012;Zucker et al., 1998 These are exactly the types of institutions andgeographies that a large share of public biotechnology R&D invest-ment is typically directed to. For these and other reasons, we expectthat if a relationship between public funding of R&D and “rm cre-ation exists, it should be possible to detect in the biotechnologyindustry.For our empirical analysis construct a rich dataset thatincludes all R&D funds from the largest funding source, the NationalInstitutes of Health (NIH), directed towards biotechnology researchfrom 1992 up to 2010. complement this dataset with informa-tion about biotechnology “rm births, venture capital investmentsand other relevant variables from Thomsons Financial SDC Plat-inum Database and other sources. organize the rest of the paper as follows: In the next sec-tion brie”y discuss the biotechnology industry and some of itscharacteristics that make it attractive for our analysis. In Sections 3and 4 review the relevant literature and develop our theoreticalexpectations on the effects of federal R&D monies on biotechnol-ogy “rm births. In Section 5 describe our econometric modeland estimation procedures, and in Section 6 we review the data In Section 7 present the estimation results and in Section we discuss how test the robustness of those results. Finally,in Section 9 we offer concluding comments, implications for policyand suggestions for further research. 2. The biotechnology industry The scienti“c origins of biotechnology can be traced back to theadvancements of molecular biology and related “elds in the 1950s( 1986 However, biotechnology as an industry began todevelop after the discovery of the basic technique for recombinantDNA in 1973 from Stanley Cohen of Stanford University and HerbertBoyer of University of California … San Francisco.The fundamental discoveries in genetic engineering led to anever-increasing rate of innovation. By the mid-1980s, a large num-ber of novel products and processes were being pursued in avariety of industries ( and Nelson, 1999 For instance,in the pharmaceutical industry regulatory proteins (e.g. humaninsulin and growth hormone), vaccines, antibiotics and monoclonalantibodies for diagnostic and therapeutic uses, were early targets.In agriculture, animal health products and growth promotants,and genetically engineered plants (e.g. plants resistant to herbi-cides, insects, diseases, and drought), were also broadly pursued.Improved amino acids, enzymes, vitamins, lipids, were the maintargets in the specialty chemicals industry. And, R&D activitiesextended in various other industries, from food processing to envi-ronmental remediation.The development of waves of biotechnology innovations andassociated competencies, such as genetic engineering, bioprocess-ing, genomics, proteomics, metabolomics and others, has sincecontinued and has led to an ever-expanding range of potential 4Biotechnology is not a heavy recipient of public R&D investments only in the US,but across the world (for instance see Dohse (2000) C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137123 commercial applications and in the scope of the industry ( et al., 2001 In parallel, scores of scholarly works have documentedthe industrys growth, its reliance on scienti“c talent and newknowledge development and transmission as well as its emergenceat the core of the knowledge economy (e.g. Audretsch and Stephan,1996; Liebeskind et al., 1996; Whittington et al., 2009; Zucker et al.,1998 Since its initial emergence, an important feature of biotech-nology has been that new knowledge development in the lab hasfound immediate commercial applications in the market place. Assuch, biotechnology has continued to blur the distinction betweenbasic and applied research. This feature, has led to its lasting suc-cess in attracting risk capital and in “rm creation ( & Young,2012 It has also proved a good “t and a catalyst for the emergenceof the entrepreneurial university ( 2006 With regard toscience, biotechnology has promoted interdisciplinary research,largely because it draws upon a number of diverse knowledgebases and housing all relevant knowledge under one roof is increas-ingly dif“cult ( 2006 Furthermore, it has promoted intenseuniversity-industry collaboration and commercialization of uni-versity research as well as “rm creation. sum, the biotechnology industry is at the center of the knowl-edge economy, it is heavily supported through government R&Dfunds, and many of its features make it an important industry toanalyze the relationship between public funding of R&D activities,knowledge development and “rm creation. 3. How can federal R&D expenditures lead to “rm births? Except, perhaps, in exceptional cases, there are inherent mea-surement dif“culties with establishing a direct causal relationshipbetween speci“c government R&D funds, ensuing new knowledgeoutcomes and resulting “rm births by tracing some sort of lineage.In this study, we instead ask the question whether an increase in theamount of public R&D investment in some location leads to a par-allel and measurable increase in “rm births within some distancefrom the location where the R&D activity takes place. In princi-ple, there are, at least, two mechanisms that public R&D spending lead to “rm creation. First, government R&D expenditures canlead to “rm births because they can expand the level of knowl-edge that can be pursued commercially by research organizationswith voluntary “rm spinoffs and other forms of ventures that haveformal ties to these organizations (e.g. joint ventures, subsidiaries,etc.). they can strengthen localization economies, whichcan attract and support the creation of new “rms in a given region.The “rst mechanism is straightforward. Federal funds canexpand R&D outlays which, in turn, would tend to yield moreknowledge ( and Levinthal, 1989 As Klevorick et al. (1995) public R&D expenditures can expand the scope ofopportunities that can be pursued commercially. As such, volun-tary spinoffs, which rely on the research and knowledge ”ows ofresearch organizations, could materialize ( 1983; Ndonzuauet al., 2002 Indeed, in knowledge industries, such as biotech-nology, voluntarily spinoffs from incumbent “rms as well asfrom public research institutions are common ( and It is not clear how many of the university spinoffs in the U.S. are biotechnology“rms as AUTM does not provide details about the industrial focus of these new“rms. However, from the few universities that do provide such details as well asfrom survey results, biotechnology startups appear to constitute a large majority( et al., 2012; Zhang, 2009 6Often, in the literature on the determinants of “rm births there is a distinctionbetween startups which are de“ned as new “rms without a speci“c scienti“c ori-gin, and between spinoffs/spinouts which refer to “rms that spawn from particularinstitutions such as universities and private “rms. In order to develop our theoreticalexpectations consult with research that uses both terms. 2008; Lockett and Wright, 2005 For example, the“rst ever biotechnology “rm, Genentech, a spinoff fromthe University of California, San Francisco. Similarly, IntermunePharmaceuticals was a spin out of Connetics Corporation andGuidant of Eli Lilly ( and Zipkin, 2002 The second mechanism that can encourage “rm births throughthe strengthening of localization economies is more indirect.Localization economies are typically de“ned as gains from the co-location of similar “rms. They can be partitioned to gains in theknowledge base of “rms (knowledge spillovers, network exter-nalities and the like) and gains from ef“ciencies in the costs ofdoing business (access to a skilled labor pool and specialized suppli-ers, availability of “rms in complementary industries) ( andSchnellenbach, 2006 An increase in the level of R&D expendituresin a particular location can lead to improvements in localizationeconomies and, hence, in the odds of local “rm creation. speci“cally, confronted with substantial research expend-itures, long research cycles, scienti“c complexities and a strictregulatory environment ( and Grabowski, 2007; Hausslerand Zademach, 2007 “rms in high technology industries oftenleverage the imperfect appropriability of knowledge ( 1962; Nelson, 1959 and its inherent dif“culty of transfer overphysical space ( 1998 by sourcing knowledge andknow-how from nearby research-intensive institutions. This hap-pens via interpersonal interactions of economic actors working insimilar problems, collaboration between nearby “rms, participa-tion in local professional networks and labor mobility of highlytrained employees ( et al., 2004; Dahl and Pedersen, 2004;Kolympiris and Kalaitzandonakes, 2013; Liebeskind et al., 1996;Saxenian, 1991 As the knowledge base of research organizations increases withtheir R&D investment ( and Levinthal, 1989 the opportunityfor knowledge acquisition by closely located “rms should increaseas well. Indeed, especially young “rms located in proximity toresearch intensive institutions often achieve higher innovative out-comes potentially due to such knowledge acquisition effects ( et al., 1994; Anselin et al., 1997, 2000; Fischer and Varga, 2003;Jaffe, 1989 Accordingly, the potential for knowledge acquisitioncan provide strong incentives for newly founded “rms to locate inproximity to sources of knowledge that can be exploited ( and Helpman, 1992 A somewhat different form of localization economies that isalmost unique to the biotechnology industry, has been described bya stream of research developed by Powell and his colleagues (e.g. 1996; Powell et al., 1996, 2002, 2012; Whittington et al.,2009 they explain, the diversity of local organizations (e.g.“rms, universities, etc.) in a region can boost the local biotechnol-ogy birth rate. When there is diversity in the local environment,communities can more easily overcome declines and increasedcompetition can create different standards, practices, rules andstrategies that are most relevant for success. This is particularly 7As we discuss in this section, a signi“cant body of research demonstrates theexistence and strength of knowledge spillovers between closely located actors. Nev-ertheless, other contributions, including Breschi and Lissoni (2001) have questionedthe relevance of knowledge spillovers. Similarly, physical proximity among actors also increase the chances that ones ideas can be appropriated by others ( and Flyer, 2000 perhaps due to increased local competition. For an opposing viewsee Feldman and Kelley (2006) and Audretsch and Stephan (1999) 8These studies have also noted that networks across relevant actors are an impor-tant source of innovation in biotechnology. Such networks can potentially in”uencethe startup rate of a region as long as new startups are attracted to them. To capturesuch ties we would need access to the formal and, potentially, informal linkagesof the biotechnology “rms, universities and research institutes in our dataset. Suchinformation is not available to us. However, insofar as organizations that receivemore funds tend to engage more heavily in networks, concerns of the impact of thisdata limitation on our work should be alleviated. 124C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137 relevant for biotechnology perhaps because its broad knowl-edge base favors numerous competing experimentations beforereaching the most desired outcome. Among the diverse organiza-tions, Powell and his colleagues highlighted the importance of alocal anchor entity (e.g. large “rm, research university, etc.) that. . .becomes a scaffolding that, either intentionally or unexpect-edly, assists subsequent connections. . .Ž ( et al., 2012 In thiscontext of the biotechnology industry, federal R&D funds are allo-cated to different types of research institutions and might increasethe diversity of local research organizations. At the same time, thebulk of such funding goes to large research universities that as anchor organizations in various regions. As such, public R&Dfunding might increase “rm births in the biotechnology industrythrough the location effects described by Powell and his colleagues.Indeed, there is signi“cant evidence that most early biotechnologystartups were founded around universities, where major break-throughs took place ( and Stephan, 1996; Owen-Smithand Powell, 2004 Investments in R&D can augment knowledge spillovers amongproximate “rms but can also induce ef“ciencies in the costs ofdoing business of knowledge industries by enhancing specializedand localized labor pools and encouraging the presence of “rmsin complementary industries. Speci“cally in biotechnology, serviceproviders and other suppliers such as “rms with expertise in biolog-ical materials and advanced laboratory equipment tend to locate inregions with high research intensity ( and Sorenson, 2003a Insofar as newly founded biotechnology “rms are attracted bythe resource endowment of a given region ( and Sorenson,2003b an increase in the availability of research dollars can boostthe attractiveness of the region as a potential startup location.Along the same lines, an increase in R&D spending, especiallythrough government grants to universities that train new scien-tists through research, can enhance the talent in the local labor pooland accordingly boost the local availability of highly skilled labor( and St. John, 1996 Employee turnover from incumbent“rms can also enhance the local labor pool ( and Marschke,2005 A region with ample highly skilled labor could thereforeattract and encourage the creation of new local “rms.Regions rich in knowledge base also lead to “rm birthsthrough a mechanism described under the Knowledge SpilloverTheory of Entrepreneurship (KSTE). Knowledge originally devel-oped at incumbent organizations be commercially pursuedby alert individuals, often previous employees, who recognize thepotential of these projects and mobilize local networks and othermeans towards the development of their “rm ( et al., 2009;Audretsch and Keilbach, 2007 Empirical evidence on federal R&D expenditures and“rm births Empirical evidence in the literature is generally consistent withthe arguments in the previous section and suggests a positive rela-tionship between R&D expenditures and “rm creation. A stream ofresearch has documented a consistent positive association betweenproxies of R&D intensity, such as the number of R&D employees ina region, and the rate of local “rm births ( and Nerlinger, 2000; Previous employers can guard against the use of knowledge acquired internally( and Marschke, 2005 but signi“cant evidence of “rm spinoffs without formalties with the parent company ( et al., 2004; Klepper and Sleeper, 2005 suggests that such protection schemes are often bypassed.10 (1994) provides indirect evidence towards such effects by reporting that71 percent of “rm founders in his sample stated that their business originating ideacame either through replication or modi“cation of an idea encountered throughprevious employment. and Morgan, 1995; Karlsson and Nyström, 2011; Kim et al.,2011; Kirchhoff et al., 2007; Woodward et al., 2006 several studies in this stream of research havesought to distinguish the relative impact of academic and industrialR&D on “rm creation. These studies have found that conductsthe R&D matters to the rate of “rm creation and there appearsto be a stronger linkage between the creation of new “rms withR&D activity occurring in “rms rather than with R&D taking placein academic institutions ( and Nerlinger, 2000; Karlsson andNyström, 2011 What is dif“cult to infer from all these previousstudies, however, is the impact of federally funded R&D on “rmcreation because they measure R&D activity without any referenceto its funding source.A handful of studies have examined directly the relationshipbetween government R&D funding and “rm creation and here theevidence is more limited and nuanced. Kim et al. (2011) examinedthe factors that in”uence the annual rate of “rm births and deathsin the U.S. (without any speci“c industry focus) and found thatindustrial R&D expenditures had a positive effect on “rm births butgovernment R&D investments had no distinguishable effect. Samilaand Sorenson (2010) also studied the relationship between feder-ally funded R&D grants to academic institutions and the annual rateof “rm births in the U.S. (again without a speci“c industry focus). and Sorenson (2010) concluded that while in isolation fed-eral R&D funding to academic institutions did not affect the “rmbirth rate in the regions where the R&D activity occurs, in regionsrich in venture capital it did exhibit a positive effect. The authorsthen concluded that the local availability of venture capital actedas a catalyst to “rm creation.Two more studies, Chen and Marchioni (2008) and Zucker et al.(1998) have examined the linkage between federally funded R&Dand biotechnology “rm births and both have documented a positiverelationship. More speci“cally, Chen and Marchioni (2008) exam-ined the impact of federal research funds given to an MSA overthe 2003…2005 period on the number of venture capital-backedbiotechnology “rms in an in 2006. The level of federal fundsgiven to an was used in a principal component analysis alongwith other indicators, such as the number of life scientists, uni-versities, institutes and hospitals in an MSA, in order to constructa composite index of biotechnology research intensity in an MSA.In turn, this index was found to have a strong explanatory powerin the number of biotechnology “rm births. Hence, Chen andMarchioni (2008) provide indirect empirical evidence for a positiveimpact of public R&D funding on biotechnology “rm creation butthe marginal effect of such funding could not be computed. Zuckeret al. (1998) studied the births of biotechnology “rms in a givenU.S. Functional Economic Area in 1990. They evaluated the impactof federal funding on “rm creation by examining the relationshipbetween the number of faculty members in local universities thathave received federal grants between 1979 and 1980 and biotech-nology “rms ten years later. They found that federal R&D fundinghad a positive impact on biotechnology “rm births but, as in thecase of Chen and Marchioni (2008) the marginal impact of suchfederal investments on local “rm creation was not part of the anal-ysis.In broad strokes then, existing empirical evidence indicates that:(a) more R&D spending tends to increase “rm creation; (b) whoperforms the R&D matters and “rm R&D activities appear to have 11Indirect evidence is also provided by studies that examine the relationshipbetween the density of institutions that conduct R&D in a region and the rate ofnew “rm births at that region as well as from studies that analyze the rate thatacademic institutions and private organizations spawn startups (e.g. Baptista andMendonc¸ a, 2010; Klepper and Sleeper, 2005; Steffensen and Rogers, 2000; Stuartand Sorenson, 2003a C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137125 a higher marginal impact than university R&D activities on “rmbirths; and (c) increased government R&D spending have apositive effect on “rm creation under some conditions, especiallyin the biotechnology industry. The existing literature and empiri-cal evidence, however, leave a number of questions unanswered,including the most basic one … just how many more “rm birthsmight one expect from, say, an additional $1 million in public R&Dfunding in a given region? This is the principle question we addressin this study within the context of the U.S. biotechnology indus-try. To answer this question, we make a number of methodologicalimprovements in what we consider to be potential limitations inprevious studies.Speci“cally, the cycles of R&D grant acquisition, performance ofR&D, creation of new knowledge, and application of the new knowl-edge through the creation of new “rms can be long and variable inlength from one year to another. As such, we propose that the rela-tionship between public R&D funding and “rm creation must beexamined over a long period of time in order to allow for suf“-cient lags in the process and year-to-year variations in the ”ows offunds and “rms. the capacity of different types of researchinstitutions to translate knowledge into “rm births might be differ-ent, we also propose that the effects of public R&D expenditures on“rm creation should be considered separately by the type of recip-ient. Finally, we propose that instead of various indirect indicatorsof publicly funded R&D intensity, a direct measure of public R&Dfunding (dollars spent) should be used in this analysis so that themarginal effects on such spending can be calculated. implement the proposed improvements in the analysis thatfollows. More speci“cally, we use a direct measure of public R&Dspending; we extend the period of analysis to 18 years (1992…2010)to ensure adequate consideration of lags and year to year variationsin R&D investments and “rm births; and we partition public R&Dfunds to those directed to universities, private “rms and researchinstitutes and hospitals.Consistent with previous work that “nds the impact of R&Dactivities on “rm creation to materialize within a narrow geo-graphic scope, we measure the number of “rm births at the same where federal funds are allocated ( and Nyström,2011; Samila and Sorenson, 2010 The is used as the unit ofanalysis because it is small enough to capture the spatially boundednature of localization economies and the tendency of spinoffs tolocate close to parent organizations but it is large enough to exhibitindependent economic activity ( and Deitz, 2012; Samila andSorenson, 2011 Further, MSAs are generally more homogeneousacross U.S. states cities or other geographic units whichallows us to provide more meaningful comparisons across MSAsof rural and urban regions. 12For instance, in Chen and Marchioni (2008) and Zucker et al. (1998) the timespan of public R&D outlays measured and the lags between outlays and the new“rm counts are rather limited. Chen and Marchioni (2008) measure federal R&Dspending for a two year period and Zucker et al. (1998) for a one year period. Itis therefore possible, that the results could weaken or strengthen if their periodof analyses were lengthened allowing for multiple funding cycles, potential longerlags between funding and “rm creation as well as natural variations in externalconditions that also affect “rm births over time (e.g. overall business climateand conditions in “nancial markets).13Previous measures of public R&D activity have been somewhat crude. In additionto making dif“cult the derivation of a direct measure of the marginal effect of publicR&D spending on “rm creation, they also distort the potential relationship.For example, the count of faculty members supported by federal grants used in et al. (1998) mask important differences across the size of grants andaccordingly the level of knowledge generated and “rm created from public R&Dfunding.14The MSAs are U.S. population centers and exhibit less economic and geographicheterogeneity than the component spatial units (e.g., cities, towns, suburbs, villages,neighborhoods, and boroughs). 5. Methods and procedures specify a two-way “xed effects model in which the depend-ent variable yitis the number of biotechnology “rm births inMetropolitan Statistical Area i and year t. Given that the depend-ent variable is an observed count, the general form of the expectedcount is formulated as follows: E(yit|Xit, Ai, t) = mXit +iaiAi+ttt(1) where Xitis the 1 × 13 row vector that contains thirteen non-constant explanatory variables, which we describe later in thissection. Function m is a link function that maps the linear com-bination of the explanatory variables into an expected count thatis non-negative. two-way unobserved components in our model are repre-sented with the second and third terms in Eq. (1) In particular, Aiequals 1 for i and is 0 otherwise, and tequals 1 for year tand is zero otherwise. The year dummies can capture time-varyingeffects, such as favorable or unfavorable environments in “nan-cial markets on “rm births (e.g. a hot IPO marketŽ ( andSchwert, 2002 The variables are used to capturetime-constant factors that affect “rm location choice for a widerange of industries, including biotechnology. For example, the can approximate the local effect of taxes ( 1985,1989; Rathelot and Sillard, 2008 and economic initiatives ( and Rottner, 2008 which despite potential deviations from year toyear are expected to be largely time constant. set of explanatory variables in Eq. (1) includes the laggeddependent variable (i.e. the “rm birth rate in each at time tŠ1)in order to capture potentially dynamic relationships in “rm births.Regions conducive to new “rm creation are expected to show ahistorical pattern of “rm births ( and Koster, 2011 sowe expect a positive sign for the coef“cient of the lagged dependentvariable. also use the estimate of this coef“cient to distinguishbetween the short and long run effects of federal funds on the local“rm birth rate ( 2009 Since the lagged count variable in Xitis a function of thecross-sectional unobserved component (represented by the time-invariant variable Aiin (1) the presence of laggeddependent variables in panel models generally causes bias and 15To estimate the model in (1) we use the Poisson maximum likelihood estima-tor. Following earlier contributions to this literature, the standard Poisson varianceassumption of equal conditional means and variances be too restrictive formuch of the economic count data encountered in practice, so we relax this con-dition in our estimated model. When relaxed, the computed standard errors arenot sensitive to any conditional variance assumption and allow for arbitrary serialcorrelation ( 1999 16 should note that because some MSAs cross state borders, local taxes andeconomic initiatives may vary within an MSA. Accordingly, the interpretation of theregion-speci“c dummy variable as a true “xed effect may be limited. As a robustnesscheck, whenever the analysis is limited to the MSAs that do not cross state bordersthe results remain qualitatively similar to those reported in Table 3 17 have also run “xed effects models using procedures that are included inpopular statistical software, such as SAS and STATA, (e.g. TCOUNTREG). Unfortu-nately, some of these procedures are still experimental and as such they only reportlimited “t statistics. Furthermore, the available routines are suitable mainly for one-way count panel models but in our application need to estimate a restricted orpartial-two-way model (some cross-sectional effects are restricted to zero) becausesome cross-sectional observations have limited variability. In addition, in all the pro-cedures used, the number of observations drops drastically because the standardFE Poisson model as described in Wooldridge (2002) drops all observed zeros fromthe log-likelihood. As a result, the sample of such models is quite different fromthe base model and as such the results are not directly comparable. Given all theseproblems as well as the “ndings of Greene (2002) that the “xed effects estimatorin different families of nonlinear models is biased, opted to not present “xedeffects estimates here. 126C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137 inconsistency problems in the standard estimators for panel mod-els ( 2002 For one-way panel models of count data,the potential inconsistency problem be resolved by usingquasi-differenced data or an instrumental variables estimator.Given that our data have two-way unobserved components,we use the variable approach, which is roughly compara-ble to the within estimator. Although the within estimator is alsobiased for data sets with a “xed temporal dimension, the bias con-verges to zero as the number of time series observations becomeslarge. Our data set has 18 years of time-series observations, so thepotential bias in the two-way estimator of Eq. (1) be smallfor this relatively large data set, and we present alternative ver-sions of the estimated model in order to evaluate the evidence ofpotential bias in the estimator.To test whether federal funds relate to local “rm births andwhether the type of the recipient mediates that effect, we includethe average total amount of federal grants awarded at is uni-versities, private “rms and institutes/hospitals from tŠ1 to tŠ5 linear and quadratic form. The amount of funds that each from NIH within a “ve year window is fairly stable dueto the common multiyear nature of grants awards. Due to thesimilarity across adjacent observations of the lagged variables,strong correlations exist among yearly lagged observations, and an average value versus separate year lags in the empiricalmodel. use a “ve year lag average because the period up to “veyears before “rm birth is perhaps the most relevant in capturingthe true effects of federal funds because a large number of grantsexpire after the “ve year window. quadratic form of the vari-ables in question is included in the analysis in order to account forpotential nonlinearities in the relationship between federal R&Dmonies and “rm births at the level.In line with our theoretical expectation, we anticipate an overallpositive contribution of NIH funds to “rm births, and the rate increasing at an increasing rate (i.e. the quadratic coef“cientis positive) or increasing at a decreasing rate (i.e. the quadraticcoef“cient is negative). Also in line with our theoretical expec-tations, we expect the magnitude of each variable that measuresthe amount of funds provided to a given type of institution tovary among types of institutions which would signal that differ-ent types of recipients have distinct effect on the generation ofnew “rms from federally funded research. As well, we include threeinteraction terms (university funds * private “rm funds, universityfunds * research institute funds, research institute funds * private“rms funds) to account for potential synergies (positive sign of theinteraction terms) or congestion effects (negative sign of the inter-action terms) that arise when funds are allocated to differentrecipient types/organizational forms in the same MSA.To account for time-varying factors that can in”uence the yearlybirth rate of a given region we add three relevant control factors inthe analysis. The “rst control variable is the average GDP of the in the “ve years prior to a “rm birth. The variable is usedto capture the effect of overall economic conditions on the yearly“rm birth rate. These conditions are expected to arise from factorssuch as regional cost advantages, amenities and a regions prestige( and Gray, 2002; Frenkel, 2001; Stuart and Sorenson, 2003b 18In order to include early years in the empirical analysis, use a 5 year averagefor available observations. For example, the observation for 1995 is the averagevalue of years 1992…1994, which is a 3 year average. Thus, the empirical analysisonly omits data on the dependent variable for 1992. As show in section 8, thisempirical choice did not have a noticeable impact on our results.19Previous research has found 5-year windows for lag structures to be appropriate(e.g. Aharonson et al., 2008; Baum et al., 2000 As shown in Table 5 to test therobustness of our “ndings to the speci“cation of the year lag structure speci“edthe relevant variables with different lag structures and found nearly identical resultsto those reported in Table 3 Given the highly localized nature of venture capital investments inbiotechnology ( et al., 2011; Powell et al., 2002 and thecontribution of venture capital to regional “rm birth rate ( and Sorenson, 2010, 2011 the second control variable is the aver-age total venture capital funds invested in biotechnology “rms inthe in the “ve years prior to a “rm birth. expect a positivesign for the variable in question. Note that the aforementionedcontrol variables are speci“ed as 5-year averages so that we couldcompare them with the corresponding variables that measure the5-year average in”ow of NIH dollars at a given MSA. To ensure thatour empirical analysis is robust to the size of the MSA, alsoinclude a control variable for the population of the MSA at year t. 6. Data sources and presentation The data used to construct the variables that measure the yearlyfederal funds allocated to universities, private “rms, and researchinstitutes/hospitals (and the associated interaction and quadraticterms) were obtained from NIHs Research Portfolio Online Repor-ting Tools (RePORT). collected data from 1992 2010 onthe amount awarded by NIH to every principal investigator (PI)as well as each funded projects title and each PIs af“liation atthe time the project was funded. order to identify biotech-nology grants, a keyword search performed for all projecttitles. we sorted out the biotechnology grants, adjustedthe nominal award money to 2007 values using the CPI and clas-si“ed each projects PI af“liation to universities, private “rmsand research institutes/hospitals after consulting with the catego-rization of each institution as private “rm, university and so onprovided by RePORT. Whenever in doubt, visited each institu-tions website. Then, we constructed the MSA-speci“c explanatoryvariables by adding the in”ation adjusted award monies for eachtype of institution. The effect of NIH money on local “rm births could have been larger in the earlyyears because the industry have not attained complete maturity during theseyears. Accordingly, “rm births prior to 1992 would also be of interest to the presentstudy. Unfortunately, 1992 is the “rst year for which NIH data are available. Further,the boom in the biotechnology industry occurred some years later than 1992, so1992 is still among the early years.21NIHs RePORT reports the PI and the institution that each project is awarded. Itis possible that the PI may have more than one af“liation. Nevertheless, this doesnot present a problem as we allocate the funds to the institution reported in theNIH grant. NIH grants are made, principally, to an institution rather than a PI andas such the research is expected to occur in the recipient institution listed in theaward. While no of“cial statistics exist, information provided by the Of“ce of Statis-tical Analysis and Reporting (OSAR) of NIH, indicates that the majority of NIH grantsare individually allocated to a single institution and a single PI. As such, we allocateall awarded monies in each grant to the location of the primary recipient insti-tution and PI. For the small number of NIH grants involving multiple institutions,this attribution scheme might lead to misallocation of funding among institutionsand locations but do not expect this type of error to be signi“cant. indi-rectly tested for the signi“cance of such a potential error by taking advantage of arecent change in NIH grant awards. Speci“cally, since the beginning of 2007, NIHhas allowed grants to have multiple investigators ( 2006 When re-estimated our base model on a limited data set for the 1992…2007 period, we foundthe results to be qualitatively similar to the results of Table 3 where the analysisextended to 2010.22The list of biotechnology keywords was constructed after consulting withbiotechnology researchers employed at the authors institutions. Almost 400 key-words were used but 155 of them characterized 99% of all grants in our dataset. Theabbreviated list is presented in Appendix Table A1 It should be noted that med-ical and biotechnology research often overlap. There are, however, many cases ofmedical research that is different from biotechnology. Examples include researchon clinical diagnostics, clinical test kits, infectious diseases and magnetic resonanceimaging (MRI).23Starting in 2007, NIH implemented a new system to measure the amount offunds for biotechnology. Our measure of biotechnology funds, which is comparableto the updated NIH system, is conservative when compared to the original NIHestimate; the 2007 real total biotechnology amount estimate of NIH is about 5 billiondollars while our estimate for the same year is about 3 billion dollars. C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137127 Fig. 1. Historical NIH “nancial outlays towards biotechnology. Table 1Correlation coef“cients between award amounts per year. IN PR UN 1.00 0.55 0.57IN 0.55 1.00 0.52PR 0.57 0.52 1.00 UN: average NIH funds from tŠ1 to tŠ5 towards universities per MSA year; IN,average NIH funds from tŠ1 to tŠ5 towards institutes and hospitals per MSA year;PR, average NIH funds from tŠ1 to tŠ5 towards private “rms per year. 1 presents the historical real NIH funding levels for biotech-nology by institution type. Biotechnology funds increased throughthe 1990s, ”attened-out between 2003 and 2004, and stabilizedstarting in 2005. The proportion of funds directed to the differ-ent types of institutions remained stable over time mainly becausemost grants are multiyear awards. Universities attract the largestshare of the funds while private “rms receive the least amount ofgrants from NIH. as seen in Table 1 the correlationcoef“cients between monies for the different types of institutionson a per-MSA-year base are moderate, ranging from 0.52 to 0.57.The magnitude of those correlations suggests that there is variationin funding levels across the different types of institutions within theMSAs, and this should help us to estimate the separate effects of thefunds on “rm births. used the Thomsons Financial SDC Platinum Database, theZoominfo web-based database, and the web-based Moneytreereport to identify biotechnology “rm births and to construct the 24NIH is currently required to set aside 2.5 percent of its extramural R&D bud-get exclusively for grants of the Small Business Innovation Research (SBIR) program( (Ed.), 2009 which mainly go to private “rms. The required percentage hasslightly ”uctuated over time but some of the NIH funds issued to private “rms areSBIR grants. Also note that the majority of funds for private “rms do not come fromthe federal government but from other sources like venture capital funds. Hence,the overall “rm-generation capacity of private “rms will be underestimated herebecause do not include the total research amount received by private “rms(besides NIH funds). dependent variable and the time lag variable. Each “rms loca-tion and founding date were generally available in all three datasources, but missing observations were gathered from the web-sites of the individual “rms. All three data-sources report “rmsthat during their lifetime received funds from venture capital “rms.Perhaps due to the highly selective nature of venture capital invest-ments, venture capital-backed “rms are often overperforming andgenerate substantial revenues and associated increases in valueadded and jobs; they are then precisely the type of “rms that (2009) argues public policy should focus on. exten-sion, venture capital-backed “rms appear a suitable sample for ourstudy. Fig. 2 presents the number of “rm births by year, whichexhibit an increasing trend over time even though year to year 25 (2009) even suggests that the criteria used by venture capitalists in choos-ing the “rms they invest in should be adopted by federal agencies, such as NIH, thatprovide funds.26Approximately 80 percent of the “rms were common in all three databases butSDC was more comprehensive in reporting the foundation date of each “rm as wellas its address, which were the two main pieces of information we sought from thesedata sources in order to assign the birth of each “rm to an at a certain year.Importantly, for about 90 percent of the “rms we used to construct our dependentvariable SDC provided the status of each “rm as of the end of 2010, which is theyear our analysis ends. The large majority of the “rms were in business for at least7 years after their births and some were merged or acquired.27Although venture capital…backed “rms often locate where venture capitalinvestments occur, because of our focus on “rms of this kind have over-looked some “rms in regions with less venture capital activity, which could leadto a potential bias in our estimates. In general, we do not expect this issue to beimportant as 60 percent of the “rms in our sample were founded outside the threetraditional venture capital hubs of San Francisco, Boston and San Diego. Further-more, existing empirical evidence from a broad set of industries indicates that theeffect of regional venture capital activity per se on “rm foundings is not particularlystrong ( and Sorenson, 2010, 2011 Empirical evidence from the biotechnol-ogy industry suggests that the impact of venture capital activity on “rm births iseither weakly positive and lessens even more when other factors (e.g. universitypresence) are explicitly considered ( and Sorenson, 2003a or it is negative,potentially due to the fact that venture capital activity favors the creation of a smallnumber of large “rms ( et al., 1998 128C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137 Fig. 2. Biotechnology “rm births from 1992 to 2010 (1298 “rm births). variations are substantial. Firm births peaked in 2000 with 102 newbiotechnology “rms and declined after 2005.To illustrate the spatial character of “rm births, Fig. 3 presentsthe cumulative amount of NIH grants collected from 1992 to 2010for all MSAs in the U.S. along with their cumulative “rm birthsfor the same period. Each is represented by its principal cityas de“ned by the U.S. Census Bureau; for those MSAs with cities, the more geographically central city in the isdepicted in the map. MSAs are classi“ed according to their NIH fundaccumulations, and larger symbols in the “gure indicate MSAs withmore biotechnology “rm births. The general pattern observed from 3 is that the MSAs that host institutions that have attractedlarge amounts from NIH have also experienced more “rm births.Only 12 percent of the MSAs (9 of 75) with the highest NIH fundaccumulations did not have any “rm births while the correspond-ing percentage for MSAs with lower or no NIH fund accumulationswas 76 percent (107 of the 140) and 97 percent (151 of the 156),respectively. For example, Bostons had 181 “rm births andthe largest funds accumulation of all MSAs with more than $4.1billion from 1992 to 2010. Also, San Francisco had 201 “rm birthswhile having the 6th highest total NIH fund accumulation. In con-trast, Los Angeles had only 27 “rm births while having received the3rd total largest amount from NIH with more than $2 billion from1992 to 2010. While Fig. 3 implies a positive association between“rm births and total NIH funds, it does not provide a comprehen-sive picture of the relationship under consideration because it doesnot account for temporal, spatial, and other structural effects thatmight shape “rm “rms in various locations. The estimated countdata model is therefore expected to provide more speci“c evidenceon the conditional impact of NIH funds on “rm births across the U.S.For the remaining variables, we used data from the U.S. Bureauof Economic Analysis to construct the GDP for each and the nominal GDP values to 2007 dollars using the CPI.Data available at the U.S. Census Bureau was employed to formthe population variable. Finally, the variable that measuredthe venture capital investments at a given was built with data from the SDC Platinum database. SDC provided the nominal venturecapital amounts awarded to biotechnology “rms per year. In orderto construct the variable in question, converted these amountsto 2007 dollars using the CPI and summed up the values for all “rmslocated in each MSA. 2 presents descriptive statistics for the dependent vari-able and selected independent variables. The average number of“rm births per year is 0.18 with a standard deviation of 1.The dependent variable is right-skewed because most of the MSAsdid not have any “rm births in a given year. On average, universi-ties receive about $3.69 million per year, research institutesreceive about $1 million per year, and private “rms attractabout $0.12 million per year. Note that regardless of the typeof the institution, the standard deviation of NIH funds is greaterthan the variable mean, which indicates the wide range of values inthe observed funding levels. As with the dependent variable, theseexplanatory variables exhibit strong right skewness because mostobservations in a given year equal zero. Most MSAs did not receivebiotechnology-related funds from venture capitalists but the rel-ative size of the standard deviation of the average venture capitalinvested as compared with its mean (36.8 versus 4.9) indicates thatwhen (and where) venture capital investments occurred, they wereof signi“cant magnitude. Finally, the average GDP for a given almost $19 billion with most MSAs having a GDP of more than$2.4 billion. 7. Estimation results The “t statistics reported at the bottom of Table 3 comefrom a maximum likelihood (ML) estimation of the previouslydescribed Poisson count model which based on an exponen-tial speci“cation for the conditional mean. however, that 28For the cross-sectional “xed effects, the estimators of these parameters areonly identi“ed when the associated dependent variable exhibits changes during the C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137129 Fig. 3. MSA NIH funds (2007$) and “rm births from 1992 to 2010. Table 2Descriptive statistics of variables used in the empirical analysis. Mean Median Mode Standard Number of biotechnology “rm births for each year 0.184 0.000 0.000 1.019Average federal funds amount towards universities in the 5 years proceeding “rm birth (2007 3.695 0.000 0.000 13.795Average federal funds amount towards private “rms in the 5 years proceeding “rm birth (2007 M$)0.119 0.000 0.000 0.692Average federal funds amount towards institutes in the 5 years proceeding “rm birth (2007 0.996 0.000 0.000 7.460Average total venture capital funds invested in the in the 5 years proceeding “rm birth (2007 M$) 4.881 0.000 0.000 36.817 population at time t (Million) 0.654 0.225 0.110 1.512Average GDP in the 5 years proceeding “rm birth (2007 18,955.818 4653.830 2473.650 55,877.052 there might be unobserved or dif“cult to measure time-varyingregional characteristics (for instance, the quality of support towardsentrepreneurship provided in the MSA) that can boost or hinderthe “rm birth rate in a given across years. If such unobservedfactors do exist, they lead to a violation in the assumptionof independence across observations ( and Schaffer, 2007;Stimson, 1985 For this reason we compute standard errors clus-tered at the level using generalized estimating equations(GEE). practice accounts for potential clustering effects andyields identical estimates to These last estimates of the “ttedPoisson count model are reported in the “rst column of Table 3 Aswe discuss in detail in Section 8 a number of robustness checkssuggest that our estimates are generally consistent across differentestimators, sets of regressors, lag structures, and model speci“ca-tions.To evaluate the potential bias due to the presence of the “xedeffects and the lagged dependent variables in (1) estimate analternative model and report the results in the second column of 3 The second model includes all the independent variablespreviously discussed except the temporal lag. The small differencesbetween the estimates of the two models suggest that the temporal sample period ( 2008 Accordingly, the set of MSA-speci“c dummies wasde“ned as the set of 104 MSAs that over the sample period had at least one “rmbirth, and use a partial or restricted form of the “xed effects speci“cation.29GEE is a method to estimate the standard errors which “rst estimates the vari-ability within the de“ned cluster and then sums across all clusters ( 2006 lag does not induce substantial bias in the parameter estima-tors. Accordingly, in the following discussion we refer only to theestimates of the full model in the “rst column.The joint signi“cance tests reported at the bottom of Table 3 strong signi“cance for the MSA-speci“c and the year-speci“c variables. well, the condition number for the setof explanatory variables (39.67) reduces inference concerns thatrelate to multicollinearity because it is within the range of the gen-erally regarded as safe level of 30 and well below the worrisomecondition number of 100 ( et al., 1980 The estimated coef“cients for the NIH variables are largelypositive and provide empirical support to the proposition that fed-eral R&D spending contributes to local biotechnology “rm births.Because the NIH funding variables for the recipient institutionsappear in multiple terms (i.e. linear, quadratic, and cross-productterms), have to combine the individual marginal effects to eval-uate the overall marginal effect from NIH funding on “rm births.The semi-elasticities for the three recipient types are reported forthe two estimated Poisson models in Table 4 For research universi- 30 also estimated models with one-way (only MSA-speci“c) “xed effects, andthese have largely similar results with those reported in Table 3 31Separate year-speci“c and MSA-speci“c “xed effects were mostly statisticallysigni“cant as well and are not reported in Table 3 for ease of exposition. Importantly,the signi“cance of the MSA-speci“c dummy variables suggests that time-invariantcharacteristics such as “scal policies at the regional level have an impact on thelocation patterns of biotechnology “rm births. 130C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137Table 3Count estimates of Poisson model with mean equal variance assumption relaxed. The dependent variable is the number of biotechnology “rm births at time t in i. Model 1: Poisson (without variancerestriction) full modelModel 2: Poisson (without variancerestriction) excluding temporal lag Estimate S.E. Estimate S.E. Intercept Š4.2939 Š4.2626 Biotech “rm births in in 10.0186 0.0235Average amount awarded ($2007 to MSA universitiesfrom tŠ1 to tŠ5 0.0612 Average amount awarded ($2007 to MSAinst./hospitals from tŠ1 to tŠ5 0.0538 Average amount awarded ($2007 to MSA private “rmsfrom tŠ1 to tŠ5 0.6245 (Average amount awarded ($2007 M.) to universitiesfrom tŠ1 to 0.00010.0002 0.00010.0002 amount awarded ($2007 M.) to from tŠ1 to tŠ5)20.0002 0.0002 0.0002 0.0002(Average amount awarded ($2007 to MSA private “rmsfrom tŠ1 to tŠ5)2Š0.0014 0.0117 Š0.0017 0.0119University funds * inst./hospitals funds Š0.0003 0.0004 Š0.0004 0.0004University funds * private “rms funds Š0.0065 Š0.0065 Private “rms funds * inst./hospitals funds Š0.0073 Š0.0076 Average amount of venture capital funds invested tobiotechnology “rms from tŠ1 to 0.00090.0007 Š0.0008 0.0007Population of the at t 0.0267 0.1093 0.0288 0.1110Average GDP observed at the MSA from tŠ1 to tŠ5 0.0000 0.0000 0.0000 0.0000Year “xed effects Yes Yes “xed effects included Yes YesScale 0.5613 0.5615Fit statistics (AIC and loglikelihood from maximumlikelihood estimation)Test of joint signi“cance of year “xed effects 3.660 of joint signi“cance of “xed effects 41.160 41.890 likelihood Š1890.148 Š1391.756AIC 3504.438 3504.477Multicollinearity condition number 39.673 38.395Number of observations 6480 6480 Notes: 1. The log link function was used for the Poisson model. 2. Standard errors reported are clustered at the level.a In order to include years 1992…1996 in the analysis, the averages are calculated as the average of available observations. For year 1996 for example, the average used inthe model is the average NIH$ from 1992 to 1995, which is a 4 and not 5 year average. b The omitted year is 2007.*0.10 signi“cance. ** 0.05 signi“cance. *** 0.001 signi“cance. Table 4Semi-elasticities for the variables that measure the rate at which federal funds towards different types of institutions associate with biotechnology “rm births at the Model 1: Poisson (withoutvariance restriction) full modelModel 2: Poisson (without variancerestriction) excluding temporal lag Percentage change in the expected number of “rm births in the focal at year t, given a1 million increase in the average NIH funding level of universities located in the tŠ1 to tŠ55.9388 5.9288Percentage change in the expected number of “rm births in the focal at year t, given a1 million increase in the average NIH funding level of research institutes/hospitalslocated in the from tŠ1 to tŠ54.9923 5.2518Percentage change in the expected number of “rm births in the focal at year t, given a1 million increase in the average NIH funding level of private “rms located in the tŠ1 to tŠ558.1078 59.2508 ties, the results of Table 4 indicate that, on average, an additional $1million of public R&D funding awarded to universities over a 5 yearperiod is expected to generate in the following year an increaseof 5.93 percent in local “rm births per MSA. Regarding the shortrun and long run effects of federal money on local “rm births, estimated long-run effect is close to 2 percent higher than theshort-run effect (0.0612 and 0.0623 respectively).With regard to the marginal effect of federal R&D funds awardedto private “rms on the regional “rm birth rate, the estimated 32The long run effect is estimated with /1 Š  where  is the short run effect and is the estimated coef“cient of the lagged births variable ( 2009 semi-elasticity reported in Table 4 implies that, on average, an addi-tional $1 million of federal R&D funds awarded to private “rms overa 5 year period is expected to increase the number of local “rmbirths in the following year by 58.11 percent per MSA. This “nd-ing provides empirical support to the proposition that the impactof federal R&D funding on “rm creation is sensitive to the recipi-ent type. What makes this “nding striking is the sheer differencein magnitude. Federal R&D funding directed to private “rms isfound to have an impact on “rm creation that is almost ten timeslarger than that directed to research universities. This differenceis consistent with the “ndings of previous studies that documentdifferential effects between industrial and academic R&D on “rmbirths ( and Nerlinger, 2000; Karlsson and Nyström, 2011 C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137131 Similar to the results for universities, the estimated short-run andlong-run multipliers for private “rms are nearly identical, and themarginal effects do not exhibit a higher order relationship that iseconomically relevant.With regards to federal R&D funds awarded to research insti-tutes and hospitals we “nd that the impact of such funds on localbiotechnology “rm creation is slightly lower than that of researchuniversities but substantially lower than that of private “rms. Inparticular, our results show that, on average, an additional $1 mil-lion of federal R&D funds awarded to research institutes/hospitalsin a “ve year period is expected to increase the number of biotech-nology “rm births in the following year by about 5 percent for the of interest. Given that unlike private “rms, universities andresearch institutes are not driven mainly by pro“t maximization, apossible impetus of the observed difference be the potentialtendency of private “rms to direct their efforts mainly towards theend of research with higher commercial value. It also be theresult of diminishing returns of public R&D funding to “rm creation(note the difference in the relative size of public R&D spending inuniversities and private “rms), or some other factor.The statistical insigni“cance of the lagged dependent variablesuggests that prior local “rm births do not have strong explanatorypower in the rate of local “rm formation and the same holds forventure capital investments, the size of the and its economicgrowth. 8. Sensitivity analysis In order to evaluate the robustness of our “ndings performeda number of tests and we present the results in Table 5 First, we tested for the case under which our results are in”u-enced by a virtuous cycleŽ where funds go to regions that alreadyhost many “rms and in turn (mainly due to the enhanced capacityof “rms to generate new “rms from federal funds) these regionsend up with more “rms. To address the issue, we computed theannual NIH funds towards private “rms in the same MSA measur-ing only those funds that went to “rms that were at least sevenyears old. shown in Model 1 in Table 5 the results once againindicate that federal funds towards incumbent “rms promote local“rm births and imply that the estimates reported in Table 3 a lower bound on the impact of such funds on the localstartup rate. The estimates of the university and institutes impactsare in the same range with the estimates of Table 3 but the instituteimpact is no longer statistically signi“cant.The estimated scale parameter for the Poisson model reported in 3 indicates under-dispersion in the variance of the dependentvariable (i.e. the observed variance is less than a standard Poissonrandom variable), which be due to the large number of obser-vations with a zero count. to this potential under-dispersion,we also constructed a Generalized Poisson model ( 1993;Famoye and Singh, 2003 which has properties that are appropri-ate for under-dispersed data ( 2011 The results of this modelare presented as Model 2 in Table 5 and are qualitatively similar the results presented in Table 3 In Model 3 we present estimates of a linear probability modelthat tests the sensitivity of our results to the exclusion of MSAswithout “rm births across years from the analysis that relates to 33 are grateful to two anonymous reviewers for pointing this potential issueand for suggesting ways to address it.34The Poisson model with the variance equal to the mean is not reported in Table 3 the Poisson variance assumption was not supported by the data.35The main difference is the statistical signi“cance of the quadratic terms. But,their magnitude is so small that it suggests a nearly nonexistent economic impactfrom these higher-order terms. the Maximum Likelihood estimator used for the count mod-els. In Model 3, those MSAs are included in the analysis. AlthoughOLS ignores the count feature of the dependent variable, the esti-mates largely corroborate our previous “ndings that funds directedtowards private “rms have a much stronger marginal effect on local“rm births than funds directed towards universities and researchinstitutes/hospitals.Another estimator that ignores the count feature of our databut addresses potential endogeneity between the dependent andthe independent variables is the Arellano-Bond estimator ( and Bond, 1991 our application, in Model 4, we use it totest the possibility that our estimates are plagued by the poten-tial endogeneity of the lagged dependent variable. The estimatedcoef“cients do not provide evidence that such an issue is presentin our models.Models 5…8 test the sensitivity of our estimates to the modelspeci“cation. In particular, we build models that include each fund-ing category separately and then a model that does not includethe interaction terms. The estimates of these models are largelycomparable with the estimates presented in Table 3 Models 9 and 10 are built to test the robustness of our estimatesto the construction of the time lag we used to estimate the effect offederal funds (5 year average). In model 9, the 5 year average lag isreplaced by a one year lag and in model 10 the 5 year average lagis replaced by a 3 year average lag. Both models suggest that thechoice of the lag structure does not signi“cantly impact the results.Finally, as note at the bottom of Table 3 in order to includeearly years in the empirical analysis, we use a 5 year average foravailable observations where for example, the observation for 1995is the average value of years 1992…1994, which is a 3 year average.In model 11 we test the potential in”uence of that choice in ourestimates and build models that use observations only after 1997.The results are nearly identical to those presented in Table 3 In conclusion, the tests we conducted support the overall con-clusions have drawn from our main results reported in Table 3 indicate that our “ndings are robust to alternative model spec-i“cations and data constructions. 9. Summary, discussion and concluding comments Partly due to a strategic decision of the U.S. government to sub-stantially increase funding towards R&D in order to boost the U.S.economy during the severe downturn of the late 2000s, interestin the relationship between public R&D spending and economicgrowth has been revitalized. A longstanding academic literaturehas established that public funding matters for economic growthas it increases innovation, productivity and the like. Nevertheless,the impact of public R&D funding on the creation of new “rms hasreceived little attention despite strong theoretical constructs thatsupport the association. Indeed, empirical evidence on the relation-ship between public R&D funding and “rm creation has been scantand indirect.In this study, we have analyzed the relationship between federalR&D funds and local “rm births in the U.S. biotechnology indus-try and our “ndings suggest that government R&D spending has apositive impact on “rm creation. Our results, therefore, corrobo-rate the empirical evidence that has been provided by a handful of 36To construct the instruments we use a one year lag of the independent variables,which model as weakly exogenous implying that we allow for the case thatthey are correlated both with past and . . .possibly current realization of the errorŽ(pg. 86 Roodman, 2009 Different con“gurations of the set of variables we use toconstruct the instruments yield qualitatively similar results to those reported in 5 The estimates presented in Model 4 of Table 5 are derived from the Arellano-Bond system estimator, which, in our application, yields similar estimates to thedifference estimator. 132C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137Table 5Estimates of models testing the robustness of the results. The dependent variable is the number of biotechnology “rm births at time t in i. number 1 2 3 4Model description Model where the NIH funds towardsprivate “rms measures only fundstowards “rms that were at least 7 yearsold at the year in question. Standarderrors are clustered at the MSA levelGeneralized Poisson Linear probability model withheteroskedasticity … robuststandard errorsArellano … bond estimator Estimate S.E. Estimate S.E. Estimate S.E. Estimate S.E. Intercept Š4.3297 Š4.3212 0.0381 0.0934 Š0.2554 Biotech “rm births in MSA in 10.0525 0.0084 0.0161 0.1876 0.0620 Average amount awarded ($2007 to universities from tŠ1 to tŠ5 0.0676 0.0138 0.0129 0.0043Average amount awarded ($2007 to inst./hospitals from tŠ1 to tŠ5 0.0423 0.0546 0.0881 0.0942 Average amount awarded ($2007 to private “rms from tŠ1 to tŠ5 0.5855 0.2635 0.1638 (Average amount awarded ($2007 to universities from tŠ1 to tŠ5)20.0000 0.0001 Š0.0001 0.0001 0.0001 0.0000 0.0000(Average amount awarded ($2007 to inst./hospitals from tŠ1 to tŠ5)20.0000 0.0002 0.0002 Š0.0001 0.0002 Š0.0011 (Average amount awarded ($2007 to private “rms from tŠ1 to tŠ5)2Š0.0292 0.0301 0.0050 0.0062 0.0152 0.0187 0.0083 0.0054University funds * inst./hospitals funds Š0.0005 0.0006 Š0.0002 0.0002 Š0.0002 0.0004 Š0.0009 University funds * private “rms funds Š0.0095 0.0073 Š0.0074 Š0.0054 Š0.0017 0.0011Private “rms funds * inst./hospitals funds0.0081 0.0099 Š0.0080 0.0014 0.0080 0.0039 0.0034Average amount of venture capital fundsinvested to biotechnology “rms from tŠ1to tŠ5Š0.0016 0.0018 Š0.0008 0.0008 Š0.0023 0.0044 Š0.0002 0.0007Population of the at t 0.2593 0.0366 0.0494 0.1564 0.9392 Average GDP observed at the fromtŠ1 to tŠ50.0000 0.0000 0.0000 0.0000 0.0000 Year “xed effects Yes Yes Yes Yes “xed effects Yes Yes Yes YesScale 0.5849Fit statistics (for countmodels the log likelihoodand the AIC are frommodels estimated withmaximum likelihood)Test of joint signi“cance of year “xedeffects6.690 of joint signi“cance of “xedeffects31.360 likelihood Š1992.011AIC 3678.757 3379.300Adjusted R20.739Wald 2297.350 condition number 41.079 39.673 41.579 33.978Number of observations6480 6480 6480 5831 C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137133 Š4.4941 Š3.3875 Š3.5403 Š4.3855 Biotech “rm births in in tŠ1 0.1172 0.0824 0.1137 0.0398 0.0250Average amount awarded ($2007 to universities from tŠ1 to tŠ5 0.0692 Average amount awarded ($2007 to inst./hospitals from tŠ1 to tŠ5 0.0244 Average amount awarded ($2007 to private “rms from tŠ1 to tŠ5 0.5981 (Average amount awarded ($2007 to universities from tŠ1 to tŠ5)2Š0.0004 Š0.0003 (Average amount awarded ($2007 to inst./hospitals from tŠ1 to tŠ5)20.0003 0.0000 0.0001 0.0001(Average amount awarded ($2007 to private “rms from tŠ1 to tŠ5)2Š0.0668 Š0.0291 0.0184University funds * inst./hospitals fundsUniversity funds * private “rms fundsPrivate “rms funds * inst./hospitals fundsAverage amount of venture capital fundsinvested to biotechnology “rms from tŠ1to tŠ5Š0.0002 0.0009 Š0.0006 0.0004 0.0018 0.0008 Š0.0044 Population of the MSA at t Š0.1028 0.1475 0.2478 0.3224 Š0.0295 0.1134Average GDP observed at the MSA fromtŠ1 to tŠ50.0000 0.0000 0.0000 0.0000 0.0000 0.0000Year “xed effects Yes Yes Yes Yes “xed effects included Yes Yes Yes YesScale 0.6101 0.6418 0.6396 0.5802Fit statistics (for countmodels the log likelihoodand the AIC are frommodels estimated withmaximum likelihood)Test of joint signi“cance of year “xedeffects4.110 0.310 8.030 of joint signi“cance of “xedeffects10.100 likelihood Š2091.908 Š2197.114 Š2190.437 Š1974.693AIC 3356.823 4109.498 4091.873 3637.283Adjusted R2Wald 2Multicollinearity condition number 14.822 12.412 12.763 18.177Number of observations 6480 6480 6480 6480 134C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137Table 5 (continued ) number 9 10 11Model description Model where the 5 year lagaverage of the NIH funds isreplaced with a one year lag.Standard errors are clusteredat the levelModel where the 5 year lagaverage of the NIH funds isreplaced with a 3 year lagaverage. Standard errors areclustered at the levelModel where all observationshave a 5 year lag average.Standard errors are clusteredat the level Estimate S.E. Estimate S.E. Estimate S.E. Intercept Š4.3137 Š4.3607 Š4.2763 Biotech “rm births in in 10.0499 0.0323 0.0161 0.0277 0.00020.0232 amount awarded ($2007 M.) toMSA universities from tŠ1 to tŠ530.0598 0.0571 0.0596 Average amount awarded ($2007 M.) toMSA inst./hospitals from tŠ1 to tŠ530.0337 0.0488 0.0478 Average amount awarded ($2007 M.) toMSA Private Firms from tŠ1 to tŠ530.4764 0.6066 0.5731 (Average amount awarded ($2007 toMSA universities from tŠ1 to 0.00010.0001 0.00010.0001 0.0000 0.0002(Average amount awarded ($2007 toMSA inst./hospitals from tŠ1 to tŠ5)0.0000 0.0001 0.0002 0.0001 0.0002 0.0002(Average amount awarded ($2007 toMSA private “rms from tŠ1 to 0.00260.0017 0.00620.0067 0.00280.0119 funds * inst./hospitals funds Š0.0001 0.0002 Š0.0003 0.0003 Š0.0004 0.0003University funds * private “rms funds Š0.0046 Š0.0051 Š0.0074 Private “rms funds * inst./hospitals funds Š0.0037 Š0.0071 Š0.0069 Average amount of venture capital fundsinvested to biotechnology “rms from tŠ1to tŠ5Š0.0006 Š0.0008 0.0007 Š0.0006 0.0009Population of the MSA at 0.00670.1164 0.0226 0.1140 0.02820.1319 GDP observed at the MSA fromtŠ1 to tŠ50.0000 0.0000 0.0000 0.0000 0.0000 0.0000Year “xed effects included Yes Yes Yes “xed effects included Yes Yes YesScale 0.5637 0.5613 0.5559Fit statistics (for countmodels the log likelihoodand the AIC are frommodels estimated withmaximum likelihood)Test of joint signi“cance of year “xedeffects4.370 of joint signi“cance of MSA “xedeffects38.930 likelihood Š1901.148 Š1890.202 Š1160.255AIC 3521.668 Š1633.262 2760.192Adjusted R2Wald 2Multicollinearity condition number 34.609 37.300 39.569Number of observations6480 6480 5040 The log link function used for the Poisson model.a In order to include years 1992…1996 in the analysis, the averages are calculated as the average of available observations. For year 1996 for example, the average used in the model is the average NIH$ from 1992 to 1995, whichis a 4 and not 5 year average. In models 9 and 10 the estimates represent a one year lag and a three years average respectively. The square terms as well as the interaction effects are calculated accordingly. b The omitted year is 2007. * 0.10 signi“cance. ** 0.05 signi“cance. *** 0.001 signi“cance. C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137135 previous studies that have also found a positive relationshipbetween public R&D funds and regional “rm births, especially inthe biotechnology industry. Our analysis focused on biotechnologybecause it is central to the knowledge economy; it is a heavy recip-ient of federal research funds; it displays a close linkage betweenbasic research and commercial application; and it tends to clusteraround geographies and institutions that receive signi“cant gov-ernment R&D funding. As such, we presumed that if a positiverelationship between public R&D spending and “rm creation exists,it should be evident in the biotechnology industry.By developing separate measures of the effect of federal R&Dfunds awarded to different types of recipient research institutions,we “nd that public funds dispensed for R&D to existing “rmshave, proportionally, a much stronger positive impact on local“rm births than funds directed to universities and research insti-tutes/hospitals. also “nd that temporal and spatial factors inour analysis (two-way “xed effects) explain part of the variation inthe birth rate of biotechnology “rms across different MetropolitanStatistical Areas in the U.S. Hence, year-to-year variation in marketconditions (e.g. hot IPOŽ markets) as well as regional differencesin infrastructure, business climate, taxes, initiatives and other fac-tors tend to condition the overall impact of public R&D funding on“rm creation. These “ndings reinforce our view that analysis of therelationship between public R&D spending and “rm creation mustbe done over a long period of time and across large geographies to allow for lags as well as spatial and temporal variation that areinherent in such a relationship.The speci“c quantitative results in our study are of particu-lar interest. “nd that, depending on the recipient of funds, a$1 million increase in the average amount of federal R&D fund-ing associates with an increase of 5…58 percent in the number oflocal biotechnology “rm births a few years later. Therefore, themagnitude of the effect has a considerable range. The speci“c sizeof this marginal effect is important because it relates to the job-creation process and other direct economic bene“ts associated withentrepreneurship and “rm creation ( Praag and Versloot, 2007 Given that such marginal effects have not been produced in otherstudies, it is important that future studies con“rm and re“ne them.Having a deeper understanding of such marginal effects across sec-tors and geographies as well as of the conditions that determinethem could boost the overall impact of public R&D funding oneconomic growth and employment. It could also provide a moreaccurate accounting of the extent public R&D funding corrects mar-ket failures in certain knowledge sectors.Measurement of the marginal effects of public R&D on “rmcreation, like those we provide in this study, also offer use-ful insights in the debate about the gradual emergence of theentrepreneurial university, which in addition to its teaching andresearch mission, promotes local “rm births and economic devel-opment ( 1998 Our empirical estimates suggest that Table A1Keywords used to identify biotechnology related grants. chromatography Chinese Hamster Ovary gene silencing oligonucleotideagarose cho cellsgene targeting oligo-nucleotideAgarose gel electrophoresis chromatid Gene therapy oligonucleotide ligation assayallele chromatin gene transfer oligonucleotide microarrayampli“ed fragment lengthpolymorphismChromatography gene”ow oligonucleotide probesanticodon chromosomal fragmentation genetic PCRAntigen chromosomal mapping genetic discoveries pcr testAntisense chromosomal mutation genetic disorders peptideArabidopsis Chromosome genetic engineering plasmidascites chromosome walking genetic map Polyacrylamide gel electrophoresisassay chrondocyte differentiation genetic marker polyclonal antibodiesbacillus Cistron genetic modi“cation polymerase chain reactionBacillus thuringiensisclone genetic parametersPolymerase chain reaction (PCR)bacteriophage coli genomics recombinant adenovirus technologyBeta-glucans collagen glycopeptides recombinant allergensbeta-glucuronidase combinatorial biocatalysis glycoprotein recombinant antigensbioassay Combinatorial chemistry glycosaminoglycan recombinant collagensbiocomputing Cyclic (cyclic adenosinemonophosphate)glycosidase recombinant enzymesBio“ltration Cytogenetic glycosylation recombinant genesbiolistics cytokines glycosyltransferases recombinant proteinsbiomass cytokinesis hormone Restriction-fragment-length polymorphism (RFLP)biomedicine cytosine immunoaf“nityReverse transcriptasebiopharma Directional cloning immunoassay Southern blotBiopharming dna Interferon Southern hybridization (Southern blotting)bioplastics dna biosensor Introgression stem cellsbiopolymers dna chips kinase tissue engineeringbioprocess dna detection knockout mice transcription factorsBioprocessing dna inihibitors ligase chain reaction Transgenic plantsBioreactor dna modi“cation Microarray Western blotBioremediation dna polymerases microbial biotechnologybiosensor dna-chip mitosisbiosynthesis elisa monoclonal antibodiesBiotelemetry embryoandgenetic Mouse modelBioticstress enzyme Mutagenesisblastocyst Enzyme-linked assay (ELISA):mutagenic substancebovine somatotropin fab mutationCdna Field trial nanobiotechnologycell culture functional genomics neuroncellular assays gene neuropithelial stem cellscellular signaling Gene (DNA) sequencing Northern blotcentromere gene ampli“cation nucleotidechimeraplasty Gene mapping d-glycolsylation 136C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137 publicly funded R&D in universities and research institutes has apositive impact on local “rm creation. Characterizing the differ-ences in such marginal effects between universities and private“rms as well as across locations and sectors help identify theirsources and improve the ef“ciency of the entrepreneurial univer-sities and their impact on local economies through the creation ofnew ventures.Before closing, we note that our empirical models contain someunexplained variance. This variance can be reduced in a numberof ways. have discussed a number of mechanisms that havebeen identi“ed in the literature and could explain the differencesin the marginal effects of public R&D spending we found in thisstudy. However, we have not tested the relevance of such mech-anisms in our analysis. Explicitly accounting for the mechanismsthat shape the differential impact of government R&D spending on“rm creation across recipient institutions, is an important area forfurther investigation. Also, we treat all institutions in our three fundrecipient categories (university, private “rm, and research instituteor hospital) as homogeneous (aside from MSA-speci“c difference),but there are distinct types of academic institutions ( Gregorioand Shane, 2003 or private “rms that affect the local rate ofassociated “rm births differently. Future research can determineif these within-group differences are signi“cant and how the dif-ferent types of subgroups perform. Finally, data limitations do notallow us to directly account for the network ties in the biotech-nology industry that have been noted in previous work. Testingwhether federal funds in”uence the creation of such ties, whichcan promote local “rm creation, could yield important insights.AcknowledgementsResearch funding provided by the Ewing Marion KauffmanFoundation Strategic Grant #20050176 is gratefully acknowledged.An earlier version of the paper was presented at the 35th DRUID Cel-ebration Conference at ESADE Business School in Barcelona, Spain. thank Peter Klein and James Kaufman for valuable inputs dur-ing the development of the paper and Sebastian Hoenen and LifengTu for assistance in data collection. Appendix A. See Table A1 ReferencesAbel, J., Deitz, R., 2012. capital? Z.J., Audretsch, D.B., Feldman, M.P., 1994. Managerial Z.J., Braunerhjelm, P., Audretsch, D.B., Carlsson, B., 2009. spillover R., Echambadi, R., Franco, A.M., Sarkar, M.B., 2004. throughof B.S., Baum, J.A.C., Plunket, A., 2008. the P.D., 2008. Koster, S., 2011. rates„evidence L., Varga, A., Acs, Z.J., 1997. sity422…448. L., Varga, A., Acs, Z.J., 2000. demic Bond, S., 1991. evidenceStudies N.S., Liebeskind, J.P., 1998. ties& K., 1962. In:University K., 1971. K. D.B., 1998. Review D.B., Keilbach, 2007. ship. D.B., Stephan, P.E., 1996. biotechnology. D.B., Stephan, P.E., 1999. and F.-J., Nerlinger, E.A., 2000. “rms:155…176. R., Mendonc¸ a, J., 2010. knowledge-based T.J., 1985. effectsness T.J., 1989. effects T.J., Gray, W.B., 2002. locationmentalAldershot, P., 2012. The H., Malmberg, A., Maskell, P., 2004. globalGeography, J.A.C., Calabrese, T., Silverman, B.S., 2000. compositionManagement M., Stahl, H., 1999. Research D.A., Kuh, E., Welsch, R.E., 1980. Library. B., 2011. In:GrowthŽ, A., 1994. Review J., 2006. compiling S., Lissoni, F., 2001. milieux:255…273. Logothetidis, S., 2008. andJournal C., Edquist, C., 2006. vationShapira,Researchpp. K., Marchioni, 2008. nology W.M., Levinthal, D.A., 1989. Economic M.S., Pedersen, C.Ø.R., 2004. industrial Gregorio, D., Shane, S., 2003. than J.A., Grabowski, H.G., 2007. different? D., 2000. Research T., Schnellenbach, J., 2006. spillovers375…395. 2012. H., 1998. new F., 1993. in F., Singh, K., 2003. Advances M.P., Kelley, M.R., 2006. governmentResearch Varga, A., 2003. evidence A., 2001. tan C. Kolympiris et al. / Research Policy 43 (2014) 121 … 137137 Garvin, D.A., 1983. agement S.J., Morgan, R.S., 1995. “rms. 2002. lished Z., 1992. 94, G.M., Helpman, E., 1992. MIT B.H., 1996. Smith,Institution E., 2009. C., Zademach, H.-M., 2007. ture,Schmalenbach J.M., 2011. A.B., 1989. 957…970. C., Nyström, K., 2011. Desai,Development. 1986. Press. J., Marschke, G., 2005. the Y., Kim, Yang, T., 2011. regional154…166. B.A., Newbert, S.L., Hasan, I., Armington, C., 2007. sityEntrepreneurship: S., Sleeper, S., 2005. A.K., Levin, R.C., Nelson, R.R., Winter, S.G., 1995. ni“cancePolicy C., Kalaitzandonakes, N., 2013. among C., Kalaitzandonakes, N., Miller, D., 2011. capital Tulum, Ö., 2011. of A., Zipkin, I., 2002. Journal J.P., Oliver, A.L., Zucker, L.G., Brewer, M.B., 1996. ing,Organization A., Wright, 2005. university Schwert, G., 2002. Journal E., 1991. 20, E., 1995. characteristics, E., 1997. empirical G.S., Narin, F., 2000. innovation: J., 2009. 323, D.C., Nelson, R.R. (Eds.), 1999. Seven F., Hamilton, K.S., Olivastro, D., 1997. technology F.N., Pirnay, F., Surlemont, B., 2002. creation. R.R., 1959. Political A., Schaffer, M.E., 2007. Stata L., Pammolli, F., Riccaboni, 2001. dynamics485…508. J., Powell, 2004. theScience G., 2006. R., John St., C.H., 1996. “rms 1996. try. Koput, K.W., Bowie, J.I., Smith-Doerr, L., 2002. scienceRegional Koput, K.W., Smith-Doerr, L., 1996. andtive Packalen, K.A., Whittington, K., 2012. genesis:Powell,University Snellman, K., 2004. ology, R., Sillard, P., 2008. location499…514. C.S., 2006. entrepreneurship. D., 2009. A.J., Martin, B.R., 2001. research: S., Sorenson, O., 2010. Research S., Sorenson, O., 2011. growth. A.L., 1991. Valley. S., 2009. public J.M., Flyer, F., 2000. eign1175…1193. Rogers, E.M., 2000. university. P., 2012. J.A., 1985. nal T.E., Sorenson, O., 2003a. in32, T.E., Sorenson, O., 2003b. entrepreneurial R.J.W., 2002. and A.A., 2012. dence Praag, C.M., Versloot, P.H., 2007. review S., Thurik, R., 1999. Small W.C. (Ed.), 2009. tutes K.B., Owen-Smith, J., Powell, 2009. innovation54, R., Lewandowski, B., DiPersio, C., Ford, R., Streit, R., 2012. Impacts D., Figueiredo, O., Guimaraes, P., 2006. sity J.M., 1999. models. J.M., 2002. J.M., 2009. Western J.L., Rottner, R.M., 2008. entrepreneurship. J., 2009. using C., 2006. ent L.G., Darby, M.R., Brewer, M.B., 1998. of