Do subsidies to commercial R&D reduce market failures? …bhhall/others/... · 2012. 2. 3. ·...

Ž .Research Policy 29 2000 471–495www.elsevier.nlrlocatereconbase

Do subsidies to commercial R&D reduce market failures?Microeconometric evaluation studies 1

Tor Jakob Klette a,), Jarle Møen b, Zvi Griliches c

a Department of Economics, UniÕersity of Oslo, P.O. Box 1095 Blindern, N-0317 Oslo, Norwayb Norwegian School of Economics and Business Administration, Norway

c HarÕard UniÕersity and NBER, USA

Abstract

A number of market failures have been associated with R&D investments and significant amounts of public money havebeen spent on programs to stimulate innovative activities. In this paper, we review some recent microeconometric studiesevaluating effects of government-sponsored commercial R&D. We pay particular attention to the conceptual problemsinvolved. Neither the firms receiving support, nor those not applying, constitute random samples. Furthermore, those notreceiving support may be affected by the programs due to spillover effects which often are the main justification for R&Dsubsidies. Constructing a valid control group under these circumstances is challenging, and we relate our discussion to recentadvances in econometric methods for evaluation studies based on non-experimental data. We also discuss some analyticalquestions, beyond these estimation problems, that need to be addressed in order to assess whether R&D support schemescan be justified. For instance, what are the implications of firms’ R&D investments being complementary to each other, andto what extent are potential R&D spillovers internalized in the market? q 2000 Elsevier Science B.V. All rights reserved.

JEL classification: O30; O40; L10Keywords: R&D investment; High-tech policy; Evaluation studies; Microeconometrics

1. Introduction

The theoretical literature on market failures asso-ciated with R&D and technological innovations is

) Corresponding author. Tel.: q47-22-85-51-34; fax: q47-22-85-50-35; e-mail: [email protected], http:rrwww.uio.nor; torjkr

1 We have benefited from comments by Tore Nilssen, John vanReenen and participants at the NBER productivity meeting inDecember 1998. This project has received partial financial supportfrom the Research Council of Norway.

vast, and there is also a steadily growing empiricalliterature verifying the importance of spillovers inR&D and innovative activities. There is conse-quently little controversy among economists aboutthe desirability of governmental support to theseactivities 2, and all OECD countries have over sev-

2 See, however, the heated and wide-ranging debate on thisissue in Research Policy, starting with the review by DavidŽ .1997 of Kealey’s book on economic issues of scientific researchŽ . Ž .Kealey, 1997 . See also the exchange between Friedman 1994

Ž .and Griliches 1994 .

0048-7333r00r$ - see front matter q 2000 Elsevier Science B.V. All rights reserved.Ž .PII: S0048-7333 99 00086-4

( )T.J. Klette et al.rResearch Policy 29 2000 471–495472

eral decades spent significant amounts of publicmoney on programs intended to stimulate innovativeactivities. However, compared to the size of theprograms and the emphasis put on technology policyby politicians, the effort to evaluate in quantitativeterms the economic benefits and costs of R&Dsubsidies has been rather modest.

In this paper, we review some recent contribu-tions to this evaluation literature that use economet-ric techniques based on microdata, in particularfirm-level data. More specifically, we review themicroeconometric literature evaluating the effects ofgovernment sponsored commercial R&D. This kindof government support to commercial R&D projectsis supposed to target projects with large expectedsocial benefits, but with inadequate expected returnsto private investors. An important question is whetherthe government agencies are able to choose projectswith high social returns that the private sector wouldnot undertake on its own. 3

Evaluating the effects of government sponsoredprojects, one has to face the question of what wouldhave taken place without the subsidies, and it isimportant to realize that evaluating large scale sub-sidy programs is an exercise in counter factual analy-sis. Neither the firms receiving support, nor those notapplying, can be considered random draws. Con-structing a valid control group in this setting is quitechallenging and we relate our discussion to the re-cent advances in econometric methods for evaluationstudies based on non-experimental data.

Most of the available evaluation studies of R&Dprograms have not been based on microeconometrictechniques, but instead on case studies and inter-views with program and project managers. 4 Thesekey persons are typically asked to report the payofffrom the projects, and similar questions might beasked also to downstream users of innovationsemerging from the R&D program in question. 5 It is

3 Ž .See, e.g., Yager and Schmidt 1997 for a detailed andskeptical discussion of the government’s ability to reduce marketfailures in R&D activities.

4 Ž .Mansfield 1996 surveys this methodology and gives refer-ences to the previous literature.

5 Ž . Ž .Cf., e.g., Link 1996 and Link et al. 1996 . The 1996 bookŽ .by Link is reviewed by Averch 1997 .

easy to conceive an upward bias in the payoff re-ported by project managers, not least because a highestimate typically increases the chances that the R&Dprogram will be considered successful and continuedor replaced by a similar program. Also, one shouldnot underestimate the problems for the project man-agers in constructing an estimate of the payoff fromindividual projects, since such estimates are based oncounter factual questions similar to those faced bythe econometrician. 6 Another disadvantage of thecase studies is that they have high costs per caseŽ .project considered, and case studies consequentlytend to be quite selective and suffer from the objec-tion that they may not be representative. Finally,evaluation studies not based on ‘objective data’ maymore easily be biased, e.g., by prior beliefs, which isa problem because evaluation studies typically aredone by ‘professional evaluators’ who are part of thepolitical process that formulates the programs, andwho ‘‘are dependent on those commissioning theevaluation studies for further projects and studies,and risk losing future clients if they voice strong

Ž .criticism’’ Luukkonen, 1998 .It is outside the scope of this paper to discuss in

detail evaluation studies based on interviews andcase studies. Our study focuses on microeconometricstudies of firm level data or similar data sources, aswe pointed out above. It is also narrowly focused onthe impact on manufacturing performance of directgovernment support to commercial R&D-projects,and it largely ignores closely related issues such asthe impact of research in governmental labs, defenserelated R&D-contracts, support to basic research inuniversities and tax-breaks for R&D. 7 Furthermore,we do not review the literature that exclusivelyconsiders to what extent R&D subsidies crowd outprivately financed R&D investments 8, but our dis-cussion addresses this issue in the context of themore wide-ranging studies that we consider.

6 Notice that the project manager might have less informationthan the econometrician about economic results of competingprojects or firms.

7 Ž .See the survey by Hall and van Reenen 1999 on taxes andŽ .R&D, and Mowery and Rosenberg 1998 for a wide ranging

discussion of the other issues and further references.8 Ž .See David et al. 1999 for a survey.

( )T.J. Klette et al.rResearch Policy 29 2000 471–495 473

We start in Section 2 by considering five microe-conometric studies that directly try to evaluate theeffects of government sponsored commercial R&D,and we refer to these studies at several points in therest of the paper. Section 3 discusses some generalissues considered in the recent econometric literatureon evaluation studies when only non-experimentaldata are available, which is typically the case forR&D programs. Section 4 discusses more narrowlyhow the five studies and related studies address theessential issue of R&D spillovers. In Section 5, wediscuss some analytical questions related to marketimperfections and spillovers that need to be ad-dressed to decide whether the R&D support schemescan be justified. Our suggestions for future researchare summarized in the last section.

2. Five microeconometric studies of government-sponsored R&D

2.1. The SEMATECH research consortium in the US

Ž .Irwin and Klenow 1996 evaluated the SEMAT-ECH program in the US, which was a researchconsortium established in 1987. SEMATECH wasset up to promote US manufacturing’s role in thedevelopment of technology for production of semi-conductor products. 9 The consortium was initiatedwith fourteen firms but has since been somewhatrestructured with a few of the initial firms pullingout. About half of the consortium’s annual budgetŽ .about US$200 Mill. was financed through govern-ment subsidies in the period 1987–1996. 10

In their study based on annual firm-level data forŽ .the period 1970–1993, Irwin and Klenow 1996

found that SEMATECH was successful in eliminat-

9 Ž .Link et al. 1996 evaluate the returns to SEMATECH pro-jects through interview studies.

10 The government support to SEMATECH was ended in 1996,but the consortium has continued with private funding only. InAugust, 1998, it was announced that SEMATECH and the gov-ernment jointly will sponsor new university-based research centersto ‘‘study new methodologies in designing, testing and connectingmicrochip components’’.

ing excessive duplication of R&D, which was amajor objective of the consortium. At the same time,the SEMATECH firms had on average a more rapidgrowth in sales than non-member firms. Irwin andKlenow also compared the SEMATECH firms’ per-formance in terms of physical investment, returns onassets and sales, and labor productivity growth, butfound no systematic difference from non-memberfirms for these variables. Their analysis was basedon running a set of similar regressions of the form

Y sa qb Y qb DSMT qDummiesqe ,i t i 1 i , ty1 2 i i t

1Ž .

where Y is the performance measure of interest,i t

e.g., private R&D to sales ratio, for firm i in year t,while DSMT is a dummy which is one if the firm wasi

a member of SEMATECH and zero otherwise. Theirregressions include firm-specific parameters, a ,i

Žwhich are treated as so-called firm fixed or corre-. 11lated effects. The dummies include time dummies

and firm age dummies, while e is an error term.i t

The ‘experiment’ in the data allowing Irwin andKlenow to identify the interest parameter b , is the2

observations for non-member firms in the same in-Ždustry as the SEMATECH members i.e., the elec-

.tronic components industry, SIC 367 . The presenceof observations prior to the establishment of SE-MATECH is useful to add precision to the estimatesof the auxiliary parameters.

Irwin and Klenow focus on their estimate of b ,2

which, according to their computations, suggests sav-ings in R&D around US$300 Mill. But this estimatedoes not account for the dynamic effects captured bythe lagged dependent variable in their model. Thelong run effect of R&D membership is given by

Ž .b r 1yb , which is about 75% higher, and their2 1

estimated model consequently indicates that the R&D saving from SEMATECH was substantially higherthan US$300 Mill.

The study by Irwin and Klenow convincinglysuggests that SEMATECH has been a profitable

11 However, their regressions do not account for the bias createdby the presence of these fixed effects in combination with a

Ž .lagged dependent variable. See Nickell 1981 .


project in terms of social costs and benefits, as theconsortium has managed to eliminate wasteful dupli-cation of R&D, while preserving the same or per-haps even better R&D output despite the cut inR&D spending. It would seem useful to repeat theirexercise with a sample covering also the period after1993. As recognized by Irwin and Klenow, the mostimportant reservation one could raise against theiranalysis is probably the validity of the control group.Comparing the list of SEMATECH member firms tothe non-member US firms, it is clear that the SE-MATECH members are the leading US manufactur-ers in the electronic components industry, and thiswas true also when SEMATECH started. Irwin andKlenow try to account for the differences by incor-porating the fixed effects, but even when they condi-tion their analysis on such permanent differences, itremains questionable whether the non-members ofSEMATECH in the same industry reveal what themembers would have experienced without SEMAT-ECH in place. We will return to this issue when wediscuss methodological questions in evaluation stud-ies based on non-experimental data in Section 3below.

2.2. The Small Business InnoÕation Research pro-gram in the US

While SEMATECH largely targeted large andleading high-tech firms, the Small Business Innova-

Ž .tion Research SBIR program was intended to stim-ulate innovation in small, high-tech firms. TheSBIR-program was initiated in 1982 and the programmandated all federal agencies spending more thanUS$100 million annually on external research, to setaside 1.25% of these funds for awards to smallbusinesses. The percentage was increased to 2.5 in1992, and this amounted to US$1.1 billion in 1997.SBIR awardees must be independently owned, for-profit firms with less than 500 employees and amajority of shares must be owned by US citizens.

Ž .A recent study by Lerner 1998 has evaluated theperformance of the firms receiving SBIR awards inthe period 1983 to 1985. His study shows that SBIRawardees grew significantly faster, both in terms ofsales and employment, than similar, non-supportedfirms from 1985 to 1995. These findings are based

on an econometric analysis of a sample where theSBIR awardees are matched with similar firms. 12

The first part of Lerner’s analysis presents variousŽ .statistics mean, variance and various percentiles

from the distribution of growth rates separately forthe supported firms and the non-supported firms. The

Ž .second part considers regressions similar to Eq. 1 ,but without firm fixed effects and with observations

Žfor only two years that are ten years apart i.e., with.1985 corresponding to ty1 and 1995 to t . Lerner

explicitly stresses the need to assess the long-termimpact of the awards, and he also states that heideally would have liked to examine the relationshipbetween program participation and firms’ valuation.Using firms’ valuation as the dependent variable wasnot possible, however, because only a small fractionof the SBIR awardees was publicly held. He choseinstead growth in sales and growth in employment as

Ž .proxies, referring to Gompers and Lerner 1997 whohave shown that these measures are highly correlatedwith the valuation that venture capitalists assign toprivate firms.

Lerner considers different interpretations of hisresults including capital market imperfections andregulatory capture. It is well known that small R&Dintensive firms may have difficulties rising capitaldue to informational assymmetries, and there is alsoan extensive literature suggesting that governmentinvolvement may be affected by distorted incentivesfor politicians and government decision-makers. Withrespect to the latter group, program managers maytarget anticipated winners so that the SBIR ‘‘canclaim credit for the firms’ ultimate success, even ifthe marginal contribution of the public funds wasvery low’’. Lerner argues that picking winners shouldbe easier in low tech than in high tech industries,whereas a signal to investors about project qualityshould be particularly valuable in high tech indus-

12 To be more specific, Lerner analyses two samples, one whereeach of the SBIR awardees is matched with a firm of similar sizefrom the same industry and another where each of the awardees ismatched with a firm of similar size from the same region. Note,that even though Lerner uses matching to construct the compari-son group, he does not proceed using a formal matching estimatorin the analysis. We will discuss various aspects of matching as amethod for constructing the counter factual outcome in Section 3.


tries where traditional financial measures are of littleuse. He finds that the superior performance of SBIRawardees is particularly significant in high-tech in-dustries, and furthermore that the first award to afirm plays a significant role, while the marginalvalue of subsequent awards declines sharply. Basedon these findings, Lerner concludes that the SBIRprogram seems to have played an important role incertifying firms’ quality and the technological meritsof the firms’ projects, thereby alleviating capitalmarket imperfections. However, he also finds evi-dence of distortions in the award process. Interviewswith program managers revealed that they had facedpolitical pressure to make geographically diverseawards, and this may explain why the SBIR programseems to have been less effective in regions with fewhigh-technology firms.

2.3. Japanese research consortia

The SEMATECH program was inspired by thesuccess of Japanese research consortia in the semi-conductor industry and other high-tech industries.

Ž .Branstetter and Sakakibara 1998 have examinedthe performance of the Japanese research consortiain these industries, combining econometric tech-niques with an interview study. The Japanese re-search consortia were heavily subsidized by theJapanese government; government subsidies coveredon average two thirds of the research costs for theprojects carried out within the consortia. Branstetterand Sakakibara argue that the Japanese researchconsortia were primarily aimed at bringing togethercomplementary R&D projects, thereby making theR&D projects more productive and also more prof-itable. 13 In this view, the research consortia haveraised the learning opportunities and thereby stimu-lated to more R&D. Notice that this situation isdifferent from the SEMATECH case discussed above,where the consortium eliminated excessive duplica-tion of parallel research rather than promoted com-plementary research. Branstetter and Sakakibara’s

13 They motivate this focus with a reference to Cohen andŽ .Levinthal 1989 , who emphasize that firms typically undertake

R&D to learn about competitors’ innovative activities. See Section5.1 for further remarks on this issue.

econometric results show that a membership in theJapanese research consortia typically stimulated pri-vate R&D spending, and also made the researcheffort more productive.

Branstetter and Sakakibara’s result on R&Dspending is obtained by estimating a model slightlydifferent from Irwin and Klenow’s non-structural

Ž Ž . .model cf. Eq. 1 above :

log R&D sa qb log Capital qb SŽ . Ž .i t i 1 i t 2 i t

qDummiesqe 2Ž .i t

The left hand side variable is private R&D spendingin firm i in year t, while the first explanatoryvariable on the right hand side is physical capitaladded to control for size effects. S is the number ofi t

research consortia in which the firm is involved inyear t and b is the parameter of interest. Branstet-2

ter and Sakakibara present estimates where theymake different assumptions concerning a , assumingi

a to be either random effects or firm fixed effects.i

The dummies include both time and industry dum-mies as their sample covers several high-tech indus-tries. The equation is estimated on an unbalancedsample of 226 firms over the period 1983–1989,with 141 firms participating in at least one researchconsortium during the sample period, while the re-maining 85 firms did not. As pointed out above, theirresults revealed a positive and statistically significantvalue for the interest parameter b .2

To examine whether the research consortia cre-ated spillovers and thereby made the research effortmore productive, Branstetter and Sakakibara esti-mated several patenting equations. Using data onpatents granted to Japanese firms in the US,Branstetter and Sakakibara started by estimating anequation with the log of patents as dependent vari-able 14 :

log P q1 sa qb log R&DŽ . Ž .i t i 1 i t

qb log Capital qb SŽ .2 i t 3 i t

qDummiesqe . 3Ž .i t

Their point estimate of the consortia coefficientb suggests that membership in an additional consor-3

14 Branstetter and Sakakibara follow the patent literature bydating each patent according to the patent’s year of application.


tium tends to raise patenting by 5%, and this effect isstatistically significant irrespective of whether a isi

treated as a random firm specific effect, or as a firmfixed effect. Branstetter and Sakakibara consider sev-

Ž .eral alternative specifications of Eq. 3 and concludethat the positive effect of membership in researchconsortia is robust.

The final part of Branstetter and Sakakibara’sanalysis focuses more closely on the R&D spilloversassociated with membership in a research consor-tium. This analysis is carried out by augmenting Eq.Ž .3 with two additional terms representing spillovers.The basic spillover variable is constructed as aweighted sum of other firms’ R&D, where theweights reflect the ‘technological distance’ betweenthe firm in question and each of the other firms. Theprimary additional term representing spillovers inBranstetter and Sakakibara’s analysis is this spillovervariable interacted with a dummy variable reflectingmembership in research consortia. Their parameterestimate for this interaction term is positive andstatistically significant when a is treated as a ran-i

dom effect, and Branstetter and Sakakibara concludethat membership in research consortia augmentsknowledge spillovers.

One final, interesting aspect of Branstetter andSakakibara’s study is their use of interviews to sup-plement the econometric analysis. The responses intheir interview study are consistent with their findingthat government funds did not substitute for privateR&D spending, but rather tended to increase thefirms’ own R&D spending. Interestingly, the inter-views also suggested that selection into the researchconsortia was not biased towards the best projects;firms which are technology leaders in a field tend tobe reluctant to participate in projects which willspread their superior knowledge and where they havelittle to gain. We will discuss how this selectionissue affects the interpretation of the estimated pa-rameters in Section 3.

2.4. GoÕernment support to commercial R&D pro-jects in Israeli firms

Ž .The study by Griliches and Regev 1998 illus-trates how the production function framework widelyused to study returns on R&D, can easily be adaptedto study the effects on private firm performance of

government-funded R&D. 15 Their study covers theoverall effort by the Israeli government to promoteR&D related to manufacturing activities, incorporat-ing a number of governmental programs. 16 Theyestimate the private returns accrued to the supportedmanufacturing firms, created by the government-funded R&D. Their preliminary results suggest thatthere are large private benefits to the firms carryingout these government-funded R&D projects, andtheir estimate of the rate of return on these R&Dinvestments is high. The social rate of returns is evenhigher if these R&D programs in addition generatedany spillovers as presumably was expected.

Griliches and Regev estimate production func-Ž .tions incorporating R&D capital K , allowing fori t

a separate coefficient on the share of R&D capitalŽ . 17accumulated with government funding s :i t

ln QrL sa qb ln CrL qb ln MrLŽ . Ž . Ž .i t i t i ti 1 2

qb ln KrL q b d sŽ . Ž .i t3 3 i t

qDummiesqe , 4Ž .i t

where Q, L, C and M are output, labor, physicalcapital and materials. As above, the subscripts referto firm i in year t. The dummies include a numberof control variables in addition to year and industrydummies. The parameter of interest is d , which canbe interpreted as the effective premium or discounton government supported R&D. As mentionedabove, their preliminary results suggest quite a high,

15 Ž .See Griliches 1979 . The production function approach haspreviously been used to study the impact of government funded

Ž .R&D in US manufacturing firms in Griliches 1986 . A very largeshare of this R&D in the US has, however, been related to defensecontracts and there are a number of reasons why it is hard tomeasure the real effects of defense related R&D projects, as

Ž .discussed by Griliches 1979 .16 This governmental support to R&D includes commercial R&D

projects, support to consortia engaged in ‘generic’ technologiesŽ .Magnet-program , National S&T Infrastructure program, USA-Israel binational program, and defense related contracts.

17 This specification is based on the observation that

w xK q 1qd K s K q K qdKw xŽ .1 2 1 2 2

s ln K qln 1qd s , ln K qd s,Ž .where K and K are two types of capital with different effi-1 2

ciency, and K s K q K . d is the efficiency premium of the1 2

second type, while s is the share of this kind of capital. The lastapproximation is good if ds is small.


positive and statistically significant premium on gov-ernment supported R&D, based on a sample of morethan 11 000 firm-year observations covering the pe-riod 1990–1995. One suspects that the high premiumis due to the government picking the best firms, andthat the estimated rate of return therefore is upwardbiased. However, this does not seem to be the case.The premium is particularly high when fixed effectsare accounted for, suggesting a negatiÕe selectionbias where firms with a high share of R&D capitalaccumulated with government funding typically havelow average productivity levels. We will return tothis issue in Section 3.

Their finding of a high premium on R&D pro-jects funded by the government suggests that theseprojects should have been profitable also for thefirms themselves. 18 According to their estimates, thegovernment picks good projects in commercial terms,but the projects seem to be too profitable to justifygovernment support. The question then emerges inwhat way a study is useful for evaluation of govern-mental support to commercial R&D, given the ambi-guity of the interpretation of rate of return estimates.That is to say, a low rate of return estimate suggeststhat the projects might have been unsuccessful, whilea high rate of return estimate suggests that the firmsshould have been able to fund the R&D activitiesthemselves, unless there are significant capital mar-ket imperfections affecting R&D investments. Oneor two additional steps are consequently required todraw any conclusions about the social value of theseR&D programs. First, their study should be supple-mented by a study of how private R&D spendingtends to respond to R&D subsidies and, second,spillover benefits should be estimated, as we willdiscuss in Section 4.

2.5. GoÕernment support to commercial R&D pro-jects in Norwegian high-tech firms

Ž .Klette and Møen 1999 study the impact of aseries of governmental programs aimed at supporting

18 Ž .As discussed in Griliches and Regev 1998 , a high premiumdoes not necessarily imply a high marginal rate of return on thesupported projects, as the support typically went to R&D intensivefirms and their model assumes diminishing marginal returns toR&D capital.

commercial R&D projects in Norwegian manufac-turing related to information technology. These ITprograms were intended to stimulate complementaryR&D activities, especially in high-tech manufactur-ing, and the effort peaked in the four years 1987–1990. The econometric analysis reveals few signifi-cant differences between the supported firms and thenon-supported firms in the same industries, despitethe large amounts of R&D support provided. Simi-larly, at a more aggregated level, the study finds thattargeted industries did not show any outstandingperformance compared to the rest of the manufactur-ing sector in Norway, nor in comparison to the sameindustries in other OECD countries. 19 The studyconcludes that the effort to promote IT-related manu-facturing has been largely unsuccessful, and the studyproceeds by examining why the IT programs hadsuch a poor coordinating performance.

In terms of the performance measure used byŽ .Griliches and Regev 1998 , i.e., total factor produc-

Ž .tivity growth, Klette and Møen 1999 find that thesupported firms did significantly worse than thenon-supported firms. Considering this performancemeasure alone, one is led towards the conclusion thatgovernmental support is associated with significantlypoorer performance. However, the systematic differ-ence between supported and non-supported firmsdisappears when a broader set of performance mea-sures is considered. It is difficult to conceive thatthere is a causal relationship between governmentsupport and poor performance in terms of total factorproductivity growth, and it seems more plausible thatthe relationship runs the other way; the governmenttried to save some of the main high-tech firms asthey encountered problems when the IT industry wasrestructured towards the end of the 1980s. Thispossible interpretation illustrates why there might bea negatiÕe selection bias in the parameter estimatescapturing the effect of government support, and wewill discuss how this selection bias can be reducedor eliminated in Section 3.

The microeconometric part of the study by KletteŽ .and Møen is similar to Irwin and Klenow 1996 in

19 The Norwegian governmental support to R&D in the targetedindustries seems to have been high in relative terms also in aninternational perspective. See below.


that the estimating equations are reduced-form equa-tions with a number of different performance mea-sures as left hand side variables: private R&Dspending and physical investment, growth in sales,employment and productivity, and returns on assetsand sales. The estimating equations do not include

Ž .lagged dependent variables in contrast to Eq. 1 , butthe main results are based on models including fixed,firm level effects. The first, microeconometric partof the analysis is based on firm and plant level datafor the period 1982–1995.

As mentioned, the study also contains a moreaggregated analysis, based on industry-level data forNorway and other OECD countries. This part of theanalysis examines the overall performance of thetargeted high-tech industries. 20 The motivation forthis is that some of the benefits from the programcould spill over to non-supported firms with theresult that the comparison between the supportedfirms and the non-supported firms would underesti-mate the effect of the program. To the extent thatthese spillover effects were important, these effectsshould show up in the performance at a more aggre-gated level. At the more aggregated level, it is,however, difficult to identify a control group, i.e., asimilar non-supported industry, and Klette and Møenconsider two alternatives. The first comparison isbetween the targeted high-tech industries and the restof the manufacturing sector as a whole. This isclearly not a clean quasi-experiment, but it is never-theless interesting to compare, e.g., the profit rates

Ž .and the returns to investments R&D and physicalin the targeted industries to other industries in acost-benefit perspective. The second comparison atthe industry level is based on OECD data for thetargeted high-tech industries in Norway and in otherOECD countries. Once more, the contrast betweenindustry performance in Norway and the other OECDcountries is far from a clean quasi-experiment, as thesame high-tech industries also received considerablegovernmental support in the other OECD countries.As far as the OECD data go, they suggest that the

20 Two alternative definitions of the targeted industries wereconsidered: a widely defined group ISIC 382, 383 and 385, and amore narrowly defined group; ISIC 3825 and 3832.

Žincrease, and perhaps also the level relative to pri-.vate R&D spending , of governmental support to

these industries was significantly larger in Norwaythan in most of the other countries in the second halfof the 1980s.

Ž .In a companion study, Klette and Møen 1998examine more closely the effect of the R&D subsi-dies on private R&D spending in the supportedfirms. The first part of their analysis uses a non-structural econometric approach similar to Branstet-

Ž . Ž .ter and Sakakibara 1998 , as specified in Eq. 2above. The analysis suggests that governmental R&Dsupport did not crowd out private R&D spending,but nor did the firms increase their own R&Dspending as was expected in the ‘matching grant’contract scheme that was widely used. In the secondhalf of their study, they introduce a structural modelfor R&D investment which incorporates a ‘learning-by-doing effect’ in R&D, where accumulated R&D

Ž .capital past R&D effort has a positive impact onthe productivity of current R&D. 21 This frameworksuggests that temporary R&D grants might have hada more lasting, positive effect on private R&Dspending after the support had expired, but the em-pirical results at this point are more suggestive thanconclusive.

3. Estimating counter factual outcomes fromnon-experimental data

As we will clarify below, the results in the studiespresented in Section 2 are based on the assumptionthat R&D subsidies to a large extent are allocatedrandomly to firms and projects. With enough ran-domness in the allocation process, data for the firmsreceiving R&D subsidies as well as for similarnon-supported firms provide us with quasi-experi-ments and a basis for causal, econometric analysis.Given the many factors involved in the politicaleconomy process that determines the allocation ofR&D subsidies, random allocation may not be too

21 Ž .The same framework is also used in Klette 1996 and KletteŽ .and Johansen 1998 .


misleading in some cases. However, assuming thatgovernments’ deliberate selection process is largelyrandom is clearly dubious and there might be asignificant bias involved in the estimated impactparameters. This section tries to clarify the potentialbiases involved and explain how the methodologycan be improved by drawing on some recent ad-vances in econometrics associated with evaluation oflabor market programs. 22

3.1. Selection and the problem of the counter factual

Ž . Ž .Both Irwin and Klenow 1996 , Lerner 1998 andŽ .Klette and Møen 1999 use the outcome of the

non-supported firms to estimate what the supportedfirms would have experienced had they not beensupported, and the two studies from Japan and Israeluse their econometric models as devices to generatesimilar counter factuals. 23 The difference in perfor-mance between supported and non-supported firms isthe estimated gross impact of the R&D supportschemes. The performance of the non-supported firmsmay, however, differ systematically from what thesupported firms would have experienced in the ab-sence of the support schemes, and this is the selec-tion bias problem that we referred to above. 24 As weshall argue below, such a systematic difference doesnot make the evaluation results uninteresting, but itlimits the kind of counter factual questions the evalu-ation results can answer.

To aid the discussion, let us address the selectionissue somewhat formally, and assume that the perfor-

22 Ž . Ž .See Angrist and Krueger 1998 and Heckman et al. 1998Ž .and references cited there. Blundell 1998 gives a simple intro-

duction to the econometric literature.23 We will not consider the evaluation literature on estimating

‘treatment effects’ with various levels of ‘treatment’, as thiscomplicates the analysis considerably. This literature is surveyed

Ž .in Angrist and Krueger 1998 .24 An interesting analysis of the choice of comparison groups

when evaluating technology innovation programs is Brown et al.Ž .1995 . They suggest that the counter factual outcome should beconstructed using the performance of firms with rejected applica-tions only, and not the performance of all non-supported firms.The rejected project applications are hardly a random group ofprojects, but they may in some settings be as close to a controlgroup as it is possible to get.

mance of a firm i in period t, denoted Y , is giveni t

by

Y sa ql qb D qu 5Ž .i t i t i i i t

where D is a dummy variable which is one if thei

firm has received R & D support and zerootherwise 25, a is a firm specific intercept, l re-i t

flects shocks common across firms, and u repre-i t

sents temporary fluctuations in unobservables. WeŽ .have abstracted from other observable regressors

for simplicity. To be concrete, a represents perma-i

nent differences in firm performance while u repre-i t

sents temporary fluctuations in performance aroundthe firm specific means, due to effects specific for

Ž .individual R&D-projects. Eq. 5 incorporates het-Ž .erogenous responses to the R&D support ex post

as indicated by the subscript i on the b-coefficient,and the distribution of these coefficients may differsystematically between the supported and the non-supported firms. Indeed, the agency allocating theR&D support might try to allocate their funds on thebasis of anticipated differences in the b ’s.i

Most of the studies above present estimates wherea is treated as a firm specific parameter, i.e., wherei

a is allowed to be correlated with D . In this way,i i

the estimated impact parameter is not biased even ifthe supported firms are non-randomly selected, aslong as the selection is based on firm characteristicsthat are largely invariant over time. Assuming thatdata are available before and after the supportedfirms have received their support, i.e., at times t0

and t , this gives the estimator1

s s n nb̂ s Y yY y Y yYž / ž /did t t t t1 0 1 0 ,s nsDY yDY

s nwhere DY and DY are the average changes inperformance from before to after the R&D supportscheme was operating, and the superscripts s and nrefer to the supported and the non-supported firms,respectively. In the econometric literature, this esti-

25 We have ignored the time subscript on D for simplicity, andi

we will focus the discussion on situations where the econometri-cian compares outcomes from before and after the program hastaken place.


mator is now commonly referred to as the ‘dif-ference-in-differences’ estimator. 26 Assuming thatD and u are uncorrelated, we have thati i t

ˆ S<plim b sE b D s1 'bŽ .did i in™`

which is a parameter of interest, representing theaÕerage impact of the R&D-support on the sup-ported firms. 27 This is the parameter of interest ifwe want to do a cost-benefit analysis of the R&Dsupport scheme. Notice, however, that this parametermay not be informative of what would happen if theR&D support scheme was extended to previouslynon-supported firms, when there are systematic dif-ferences in the responses to R&D support betweenthe supported and the non-supported firms. 28

As mentioned, most of the estimates presented inthe four studies discussed above are based on the‘difference-in-differences’ estimator or similar esti-

Ž .mators, and the study by Heckman et al. 1998suggests that this method is preferable to alternativessuch as the widely-used parametric selection-correc-

Ž .tion method introduced by Heckman 1979 and themore recent matching methods discussed in Heck-

Ž .man et al. 1998 .The econometric evaluation literature has noticed

that there may remain a serious problem due toŽ .correlation between the temporary shocks u andi t

the probability of being selected into the program. 29

Discussing the results from the study by Klette and

26 With observations for more than two years, a preferableestimator might be the ‘within’-estimator widely used in the paneldata literature. The ‘within’-estimator is closely related to the‘differences-in-difference’-estimator. It is possible to estimate thetime profile of the impact by considering a number of ‘difference-in-differences’ estimates when observations for more than two

Ž .years are available. See Heckman et al. 1998 .27 In the econometric literature, this parameter is often termed

the mean impact of the treatment on the treated. We have thatS < <b s E b D s1 s b q E b y b D s1 ,Ž . Ž .i i i i

where b is the population mean impact effect.28 One could, however, extrapolate the impact analysis to the

non-supported firms also in this case by adding assumptions aboutthe functional form for the b -distribution, along the lines ini

Ž .Heckman 1979 .29 In studies of training programs, it has been observed that

individuals tend to be selected into the program during periodswhen they perform particularly badly, i.e., have particularly lowincome. This is the so-called ‘Ashenfelter-dip’.

Ž .Møen 1999 , we observed above that the poorgrowth performance, in terms of total factor produc-tivity for the supported firms, might have been dueto the government supporting some large firms thatwere facing particularly severe problems when theIT industry was restructured towards the end of the1980s. In such a case, there is a positive relationshipbetween receiving R&D support and the prospect ofgrowing more slowly than the average, and thegrowth performance of the non-supported firms isnot very useful for estimating what the supportedfirms would have experienced had they not beensupported. Consequently, the ‘difference-in-dif-ferences’ estimator underestimates the impact of theR&D-support on the supported firms. Similarly, in

Ž .the Japanese case, Branstetter and Sakakibara 1998find from their survey study, that firms with the mostpromising projects in a technological field were re-luctant to participate in research consortia, whichcreates a similar downward bias in the ‘difference-in-differences’ estimator.

On the other hand, it is easy to conceive that thebias can go the other way in cases where firms applyfor support because they have discovered particularlypromising R&D projects. The screening of projectsin the government agencies will also tend to create aselection bias in the estimated impact. More pre-cisely, if there is a positive correlation between afirm hitting particularly promising projects that tendto generate above average performance growth insubsequent years, and the chance of the firm receiv-ing R&D support, the ‘difference-in-differences’ es-timator will overestimate the impact of the R&D-support on the performance of the supported firms.Previous studies of the effectiveness of R&D subsi-dies in stimulating private R&D spending have been

Ž .criticized by Kauko 1996 along these lines, andamong the studies we have reviewed in Section 2,this problem seems particularly relevant for the eval-

Ž .uation of the SBIR program by Lerner 1998 . Heconcludes that the subsidies awarded under that pro-gram seem to have played a certifying role, helpingthe selected firms to attract venture capital. If, how-ever, awards conveyed information to the marketabout the quality of recipient firms, this implies thatthe SBIR officials on average succeed in ‘pickingwinners’. If so, one would expect awardees to per-form better than non-supported firms even without


the awards, and the effect of the awards has to someextent been overestimated.

The econometric literature has suggested that suchbiases can be reduced or eliminated by augmentingthe ‘difference-in-differences’ estimator, incorporat-ing conditioning variables reflecting the pre-programperformance. 30 That is, differences in longitudinalchanges in performance between supported andnon-supported firms should control for pre-program,temporary shocks that influence the probability ofbeing supported, e.g., pre-program changes in R&Dor firm growth. Similarly, one would also like tocontrol for anticipated future temporary shocks thatinfluence the probability of being supported by con-ditioning on forward looking variables, in particularphysical and R&D investment and perhaps alsohiring or firing.

In the review of the study of SEMATECH inSection 2, we raised the issue that the members andnon-members in SEMATECH were to a large extentquite different firms in terms of size and closeness tothe technological frontier. As emphasized in Heck-

Ž .man et al. 1998 , such differences make the evalua-tion results critically dependent on assumptions aboutfunctional forms, both in terms of the performanceequation and the selection equation, and Heckman etal. find that this tends to generate substantial biasesin the case they examine. Exploring various match-ing-procedures as well as regression methods, Heck-man et al. conclude that evaluation results are only

Žreliable when they are based on ‘treated’ units cf..supported firms which are similar to some of the

Ž .‘non-treated’ units cf. non-supported firms . For thesupported firms that cannot be adequately ‘matched’,the comparison to non-supported firms can give quitemisleading inference of the impact.

3.2. SpilloÕers and the counter factual: ‘Catch-22’?

Using the non-supported firms to evaluate whatwould have happened to the supported firms if theyhad not been supported, assumes that there are nospillover effects of the R&D support scheme to the

30 Ž . Ž .See Angrist and Krueger 1998 , Blundell 1998 and Heck-Ž .man et al. 1998 for details.

non-supported firms, which is clearly a strong as-sumption. The question is whether the performanceof the non-supported firms can be considered inde-pendent of the support given to the supportedfirms. 31 One could argue both ways in terms of thebias this problem introduces in the estimated impactof the R&D program; the impact will be underesti-mated if the non-supported firms tend to benefit,e.g., from pure knowledge spillovers from the R&Din the supported firms, while the impact will beoÕerestimated if the non-supported firms are hurt asthey lose relative competitiveness to the supportedfirms.

This spillover issue is particularly problematicsince spillovers to technologically related firms areoften a major justification for such programs in thefirst place. This implies a ‘Catch-22’ problem: If theprogram is successful in creating innovations thatspill over to technologically related firms, it will bevery difficult to find similar non-supported firms thatcan identify the counter factual outcome for thesupported firms. This problem is particularly trans-parent if one tries to evaluate the performance of thesupported firms by means of the matching procedure

Ž .described in Blundell 1998, Section 5.4.2 . Thematching estimator suggested by Blundell is givenby

1b̂ s Y y v YÝ Ýmm i i j jž /NS igS jgN

where Y and Y are the post-program outcomes for ai j

supported and a non-supported firm, respectively,while N is the number of supported firms. S and NS

refer to the groups of supported and non-supportedfirms. v is a weight indicating the ‘similarity’i j

between the two firms before the R&D-support wasprovided. Our point is that similar weighting schemeshave been used to identify ‘technologically related’firms when estimating the impact of R&D spillovers,

31 Ž .Manski 1993 considers a closely related problem; the as-sumptions required for identification of spillover effects, when wewant to condition on regressors that tend to eliminate independentvariations in the spillover variable. This is largely the reverse of

Žthe question we discuss in this section. See also Griliches 1998,.ch. 12 .


Ž .as in studies by Jaffe 1986 , Branstetter and Sakak-Ž .ibara 1998 and others that we will discuss in the

next section. This suggests that the better a firmseems to satisfy the conditions required to identifythe counter factual outcome in the absence ofspillovers, the worse might this spillover problem be.

The motivation for introducing the matching esti-mator into the econometric tool box is that it requiresonly weak assumptions about functional forms, as

Ž .we noted above see Heckman et al., 1998 . Thisargument suggests therefore that it might be difficultto identify the impact of R&D programs more gen-erally, without imposing strong functional form as-sumptions. 32 As is so often the case in economics,one does not get very far in causal inference withnon-experimental data unless a significant amount ofstructure is imposed on the analysis. To conclude,we face the paradoxical situation that if an evalua-tion study finds little difference between the sup-ported firms and the non-supported firms it couldeither be because the R&D program was unsuccess-ful and generated little innovation, or because theR&D program was highly successful in generatingnew innovations which created large positivespillovers to the non-supported firms.

3.3. Focus on a few successes?

w xT he economic value of one great industrial ge-nius is sufficient to cover the expense of the

Žeducation of a whole town Marshall, 1920, p..179 .

It has been widely recognized that the economicbenefits from research projects tend to have a highlyskewed distribution, with a median return whichmight not be very high, but a few projects generate ahigh mean return; see, e.g., Scherer and HarhoffŽ . 331999 . This represents a further challenge to re-gression analysis of the impact of R&D subsidies,

32 Ž .See Manski 1993 for a formal analysis of the functionalform assumptions required for identification in a closely relatedcontext, and within a regression framework.

33 This observation is closely related to the finding that thedistribution of the value of patents is highly skewed with a long

Ž .right tail, see, e.g., Pakes 1986 .

and such skewness might be particularly pronouncedfor the outcome of government sponsored R&Dprojects to the extent that governments tend to sup-port high-risk R&D. This observation raises thequestion of whether the main parameter of interest isthe aÕerage impact of the R&D-support on thesupported firms. More precisely, we might be inter-ested in the average rate of return to the whole R&Dsubsidy program, but the weighted average estimatesprovided by the ‘difference-in-differences’ estimatoror similar estimators will typically not apply theeconomically relevant weights to the individual ob-servations, and we may want to pay more attentionto the economically interesting outliers than suchestimation procedures tend to encourage.

To the extent that the estimated impact parameteris driven by a few high-return observations, theconfidence intervals for the impact parameter will belarge and poorly approximated by the routinely re-ported intervals based on asymptotic normal distribu-tions. Even if calculated correctly,34 the confidenceinterval obtained will be large, reflecting the substan-tial uncertainty that prevails in trying to infer theimpact parameter when the outcomes are character-ized by a highly skewed distribution with long righttails. This suggests that we might need to consider anumber of independent evaluation studies, saythrough meta-analysis, before we can provide anestimate of the impact of the R&D subsidies withmuch precision.

Recent econometric advances suggest that it mightbe possible to estimate the distribution of the subsidyimpacts across firms, 35 but we believe that thesemethods should only provide a first step in a closerinvestigation of the economic benefits of the mostimportant innovations generated by the R&D sub-sidy programs. It would be useful to merge econo-metric studies of the kind discussed in this paperwith more detailed case studies of the most success-ful projects, and perhaps also some of the less suc-cessful projects.

34 Whatever that means, but say from bootstrap estimates for theargument’s sake.

35 Ž . Ž .See Heckman et al. 1997 and Abadie et al. 1998 .


4. Identifying spillovers and the social benefits ofR&D projects

We noted above that spillovers tend to invalidatethe non-participants as a control group. However,measuring the magnitude of the spillovers generatedis by itself a crucial part of evaluating the programs.The studies discussed in Section 2 covered quite wellthe benefits to the private firms receiving the sup-port, while in most of the studies the spillovers tonon-supported firms and pecuniary externalities tocustomers and consumers were not extensively ad-dressed.

A full cost benefit analysis of an R&D supportscheme would involve estimating the expression

˜ ˜ ˜w s s Dp s ,s q Dp s q Dp sŽ . Ž . Ž .Ž .Ý Ý Ýi i j ligS jgN lgR

˜qÝD CS yd s 6Ž . Ž . Ž .˜ Ž .where D is used to indicate the counter factual shift

in the various variables as follows 36 : The first sumcovers the change in profits in the group of sup-ported firms, S, due to the R&D support scheme.Note that the benefit for each firm belonging to Scan be decomposed into a direct effect capturing theincrease in profits due to the support they havereceived themselves, s , and an indirect effect captur-i

ing the change in profits due to the support receivedby other firms, s. The latter component may bepositive, negative or zero, depending on what kind ofspillovers are present. The second summation term

Ž .on the right hand side of 6 captures the change inprofits in the group of non-supported firms, N, in thesame industry as the supported firms. 37 The sign ofthis term is also ambiguous. We will refer to it as theindirect effect on the non-supported firms, and itmay be a mixture of knowledge and rent spillovers.The next two terms represent rent spillovers alone.That is, the third sum captures the change in profitsin firms in the rest of the economy, R, due, e.g., topecuniary externalities as inputs become cheaper or

36 Ž .All variables on the right hand side of 6 should be inter-preted in terms of present values of current and future benefits.

37 Our concept of industry is at this point loosely defined asfirms which are technologically related.

better. The fourth term is the increased consumersurplus in the economy. The last term represents thedeadweight loss associated with the funding of theprogram.

4.1. The treatment of spilloÕers in the eÕaluationstudies

Ž .Using Eq. 6 to fix ideas, we will now brieflydiscuss how far the various evaluation studies re-viewed in Section 2 go towards incorporating the fullwelfare effects of the programs they examine. Withrespect to estimation techniques, the most ambitiousattempt to estimate spillovers among these studies is

Ž .Branstetter and Sakakibara 1998 . Still, this studyexplores only the effect of the programs, i.e., thesubsidized research consortia, on the participatingfirms and other firms in the same industry. In otherwords, they deal roughly with the first two sums inŽ .6 , while ignoring pecuniary externalities to firms inother industries and to consumers. With respect tothe first sum, the change in profits for the supportedfirms, it is not relevant to distinguish between thedirect and the indirect effect of subsidies, as thesubsidies are given to consortia and not to individualfirms.

Branstetter and Sakakibara find evidence that par-ticipation in research consortia rises research output,even after controlling for research input and firmspecific effects. It seems reasonable to interpret thisas a pure knowledge spillover, but comparing their

Ž .framework to Eq. 6 , we should note that they havenot considered how the increased research outputaffects the consortia participants’ profits. Dependingon how close competitors the participants are in theoutput markets, there may be negative rent spilloversbetween them, and the innovative gains may partlyaccrue to customers and suppliers. However, focus-ing on innovative output seems like a reasonablestrategy when the total welfare effect cannot bemeasured, since increased research efficiency neces-sarily increases total welfare. Turning next to theindirect effects of the program on non-supported buttechnologically related firms, they find clear evi-dence of general R&D spillovers, but they do notidentify the extent of spillovers from the subsidizedconsortia to the non-members.


Ž .Irwin and Klenow 1996 resemble Branstetterand Sakakibara in that they consider membership ina subsidized research consortium and not individu-ally received R&D subsidies under a program. Re-

Ž .ferring back to Eq. 6 , one could say that Irwin andKlenow sign the first term on the right-hand side,i.e., the effect of the program on the participants, asthey find increased profitability for SEMATECHmembers. The level of profitability is obviously af-fected by other factors than SEMATECH member-ship, but their results suggest that members haveincreased profitability relative to non-members alsowhen they control for such factors. The differencecould in principle be due to non-members facingstronger competition after the introduction of SE-MATECH, i.e., a negative pecuniary externality be-longing to the second term on the right hand side ofŽ .6 , but the authors’ interpretation is that it is mostlikely due to increased research efficiency within theconsortium. 38

Ž .Branstetter and Sakakibara 1998 , like Klette andŽ .Møen 1999 , try to capture both the effect on the

supported firms and the effect on the non-supportedfirms in the same industry, but they ignore possiblerent spillovers to other industries and to consumers.The empirical part of the study starts out estimatingthe effect of the support on the supported firms, 39

but find no significant impact neither in the short norin the long run. This could, as mentioned in Section2, be due to strong spillovers from the supported tothe non-supported firms, and they investigate thisissue by comparing the growth of the supported high

Ž .tech industry including the non-supported firmsboth to growth in overall manufacturing, and togrowth in similarly defined high tech industries inother OECD countries.

38 Their interpretation is based on the finding that SEMATECHmembers significantly reduced their R&D-intensity relative tonon-members, and the assumption that non-members’ R&Dspending was not affected by SEMATECH. This assumption is ofcourse crucial, as we discussed in Section 3 under the heading of‘Selection and the problem of the counter factual.’

39 Since the subsidies in the programs they evaluate are given toindividual firms, they could, in principle, have distinguishedbetween direct and indirect effects of the support on the supported

Ž .firms, but their focus is implicitly on the sum of the two effects.

Ž .Lerner 1998 concerns himself only with theeffect of subsidies on the subsidized firms, but hestates explicitly that his inability to assess the socialreturn to the program is the most critical limitationof the study. Furthermore, he is well aware of theestimating problem that positive spillovers representin that they, if present, reduce the difference inperformance between subsidized and non-subsidizedfirms.

Ž . Ž .Like Lerner 1998 , Griliches and Regev 1998do not explicitly deal with spillovers. However, theframework Griliches and Regev use is one whicheasily lends itself to incorporating such effects theway they are usually treated in the more generalliterature on R&D spillovers. We will now turn tothis larger literature as it is obviously of great rele-vance with respect both to methodology and to R&Dpolicy. First, we take a closer look at the frameworksavailable to study R&D spillovers and then webriefly review the main findings and raise someconcerns.

4.2. Traditional approaches to the study of R&DspilloÕers

There are two main strands of literature investi-gating the empirical importance of R&D spillovers.First, there are case studies that try to estimate thesocial return to particular research projects by exten-sively tracing the effects of the resulting innovations.This approach was first used to evaluate publicinvestments in agricultural research, but private R&Dinvestments have also been studied. The most fa-

Ž .mous example of the latter is Mansfield et al. 1977finding a median social rate of return of 56%, morethan twice the comparable median private rate ofreturn. 40 The detailed information provided by casestudies has been extremely valuable for understand-ing the mechanisms at work in technologically ad-vanced industries and markets. However, as pointedout in the introduction, case studies always sufferfrom the objection that they may not be representa-

40 Ž .More recent studies include Bresnahan 1986 on computers,Ž .Trajtenberg 1983; 1989 on CT scanners and the Bureau of

Ž .Industry Economics 1994 on 16 innovations in Australia.


tive. This has motivated econometric work, which isthe other main approach.

Most econometric studies have been performedwithin a production function framework where a‘pool’ of outside knowledge is included in the pro-duction function of a firm or an industry. This is theidea utilized in the study by Branstetter and Sakak-

Ž . 41ibara 1998 . The R&D pool is constructed as aŽ .weighted sum with weights ideally representing the

relevance of R&D undertaken elsewhere in theeconomy, i.e.

S s w K 7Ž .Ýi t i jt jtj

where S is the spillover pool, and w is thei t i jt

effective fraction of knowledge in firm j which isfreely available to firm i at time t. The weights areusually considered a measure of the proximity be-tween the firms, and have been constructed in anumber of ways. According to the survey by MohnenŽ .1996 , both product fields, types of R&D, patentclasses, input-output flows, investment flows andpatent flows have been utilized, 42 and he suggestsother possibilities such as flows of R&D personnel,qualifications of R&D personnel, and R&D cooper-ation agreements.

Ž .As pointed out by Griliches 1979 , there are twodifferent concepts of spillovers behind these mea-sures. First, a firm may benefit from research under-taken elsewhere to the extent that changes in themarket prices of its inputs do not fully reflect thevalue of the innovations. From a production functionpoint of view it is not really a spillover, but ameasurement problem. If price indexes fully reflectquality adjustments, R&D embodied in inputs willnot be relevant as a separate variable. Lacking qual-ity-adjusted price indexes, however, one can try to

41 The basic idea is most completely spelled out in GrilichesŽ . Ž .1979 and Griliches 1995 , but was first applied by Brown and

Ž .Conrad 1967 at industry level data. Branstetter and SakakibaraŽ . Ž .1998 build on Jaffe 1986 in their particular implementation ofthe framework.

42 Ž .Many of the studies reviewed by Mohnen 1996 use industrylevel data rather than firm data.

trace the effect of R&D rents not appropriatedthrough the product prices by including the R&Dinvestments of the producers in the production func-tion of the buyers in proportion to the purchasesdone. 43 We follow several previous writers and usethe term ‘rent spillovers’ for this effect. True knowl-edge spillovers, however, are ideas borrowed fromother researchers, and one would think that thesespillovers increase with the technical relatedness andgeographical closeness of firms. According to thisview, measures based on product fields, patentclasses, types of R&D, R&D cooperation, or quali-fications of R&D personnel seem most suited to

Ž .constitute the weights in Eq. 7 , maybe augmentedwith geographical distance. 44

4.3. Estimating knowledge and rent spilloÕers

It is widely acknowledged in the empirical litera-ture that it is hard to distinguish knowledge spillovers

Ž .from rent spillovers. The methodology of Jaffe 1986is probably the one that comes closest to looking forthe former type. Following suggestions in GrilichesŽ .1979 , he links an outside pool of R&D to firmperformance. Jaffe also extends the basic frameworkby controlling for differences in technological oppor-tunities across different sectors and by allowing forthe amount of spillovers received to depend on thefirms’ own R&D investments. His key contribution,

Ž .however, lies in the implementation of Eq. 7 . Toisolate pure knowledge spillovers, he uses the degreeof overlap in the distribution of firms’ patents toconstruct the proximity weights, since patents areclassified according to technological criteria. Fur-thermore, he uses the constructed spillover pool asan explanatory variable in a knowledge productionfunction, utilizing count data on patents as a proxyfor output. The coefficient on the spillover pool istherefore quite likely to represent pure knowledge

43 Note that this way of getting around the lack of quality-ad-justed price indexes will miss out on ‘spillovers’ in final-productmarkets, i.e., the increases in consumer surplus represented by the

Ž .fourth term on the right-hand side of Eq. 6 .44 Ž . Ž .Cf., e.g. Jaffe 1989 , Jaffe et al. 1993 , and Adams and Jaffe

Ž .1996 for the relevance of the geographical dimension.


spillovers. By studying the effect of the samespillover pool on profits and market value, he alsosheds light on the effect of negative rent spilloversthrough increased competition, as the estimated coef-ficient then is likely to be a mixture of this effect andknowledge spillovers. Without underplaying themethodological difficulties associated with his work,Jaffe argues that the sum of ‘circumstantial evi-dence’ brought out is enough to make a good casefor the existence of spillovers.

Positive rent spillovers from research embodied inintermediate inputs may best be investigated using aspillover pool whose weights are based on intermedi-ate input flows. There are several weaknesses associ-ated with this approach, however. First, as technicalinformation may be exchanged between suppliersand customers, such a measure may pick up somepure knowledge spillovers as well. Second, data onfirm-level input-output flows are extremely rare andwe do not know any microeconometric studies ofthis type. 45 Finally, rent spillovers to final con-sumers, i.e., increased consumer surplus associatedwith new goods or production techniques, cannot bemeasured using this framework.

In theory, rent spillovers should be measured asthe area under the final good’s demand curve. As

Ž .noted by Bresnahan 1986 , this may be done eitherby econometric techniques or by index-number tech-niques. 46 Bresnahan uses index numbers to measurethe rent spillovers from the computer industry toconsumers through the effect of computers as inputsin the financial sector. The computer industry ischosen because quality adjusted input prices areneeded, and these have been estimated for this indus-try using hedonic techniques. However, hedonictechniques are not suited to handle large productchanges, such as the introduction of qualitativelynew product characteristics. Partly for this reason,Bresnahan only covers the period up to 1972 whentraditional mainframes were challenged by software

45 Note, however, that there exist a large number of studiesbased on industry level data which use input–output matrices toconstruct the weights.

46 Ž .Cf. Mansfield et al. 1977 for an early study utilizing thisidea.

advances and large mini computers. 47 The study ofŽcomputer tomography scanners by Trajtenberg 1983;

.1989 , on the other hand, deals explicitly with theproblem of measuring the welfare gain from theintroduction of qualitatively new goods. The chal-lenges involved in correctly measuring the welfaregains from new goods are also discussed in a recent

Ž .book edited by Bresnahan and Gordon 1997 . Theevidence gathered there indicates that the increase inconsumer surplus associated with the introduction ofnew goods may be substantial, and that ordinaryprice indexes are likely to underestimate the welfaregains.

4.4. SurÕeying surÕeys and adding a grain of scepti-cism

Ž .Griliches 1997, ch. 5 summarizes availableeconometric studies of social rates of returns toR&D, and he concludes that these social rates ofreturns tend to be several times larger than the

Ž .private ones. Mohnen 1996 , listing more than 50studies, concludes that ‘‘spillovers exist and have tobe taken into account when evaluating the returns ofgovernment-financed R&D’’. Other surveys, such as

Ž . Ž .Griliches 1992 , Nadiri 1993 , the Australian In-Ž . Ž .dustry Commission 1995 , Hall 1996 , and Jaffe

Ž .1996 , agree, and their conclusions are not contro-versial.

There is no reason to doubt the existence ofpositive spillovers, but considering first the difficul-ties involved simply in constructing a measure of thestock of knowledge and next the uncertainty overwhat is an appropriate lag length, it is somewhatremarkable that almost all studies trying to estimatesomething as intangible as knowledge spillovers ac-tually report significant results. 48 There are at least

47 Another reason was that the regulation regime in the financialsector changed about that time.

48 Ž . Ž .Cf. Geroski 1991 and Geroski et al. 1998 for studies thatdo not find significant spillovers. These studies differ from othersin that they base the spillover pool on innovation count data ratherthan on R&D investments. The Bureau of Industry EconomicsŽ .1994 also finds rather modest spillovers in its 16 case studies.


three possible pitfalls that justify some concern. First,Ž .the results may be subject to what Griliches 1992

calls a publication filter, self-imposed by researchersworking in the field or imposed by editors andreferees considering non-significant coefficients tobe of little interest. Second, some of the effectsinterpreted as spillovers may actually be knowledgetransfers that are internalized in the market, e.g.,through cooperative agreements. Third, the reportedsignificant coefficients could to some extent be spu-rious, reflecting correlated unobservables across

Ž .technologically related firms. Griliches 1998, p. 281mentions in particular common technological oppor-tunities, but correlated productivity shocks or mea-surement errors would have the same effect. Thispotential bias is closely related to the problem ofestimating the counter factual outcome in the pres-ence of spillovers, discussed in Section 3.

5. R&D spillovers and the case for governmentalsupport

As emphasized already, the concerns raised abovedo not imply that we have doubts about the existenceof spillovers, but there remain some questions con-cerning the existing estimates. In this section, weraise another set of questions concerning R&Dspillovers, now taking a closer look at what policyimplications can be drawn, given that spillovers ex-ist. If spillovers can be received costlessly, it is quiteobvious that the arguments in favor of subsidies arevalid. Firms performing R&D do not reap the wholebenefit, and as they equate marginal cost to marginalpriÕate benefit, their investments will be below thesocial optimum. There is, however, a number ofreasons why this argument is incomplete, and belowwe will discuss four issues that deserve further atten-tion in the evaluation of the net welfare gains associ-ated with R&D subsidies. In Section 5.1, we con-sider how private investment in R&D is affected byspillovers when a firm cannot receive such spilloverswithout incurring own R&D activity, and Section5.3 discusses some of the recent insights from stud-ies of R&D spillovers when such spillovers affectforeign as well as domestic firms and consumers. InSection 5.4 we give some remarks on coordination

through R&D joint ventures and similar marketarrangements, while Section 5.5 considers implica-tions of spillovers transmitted through the mobilityof research workers. Discussing these issues, wehope to make clear why and how evaluation studiesoften need to go beyond the topics reviewed inSections 2 and 4.

5.1. Costless spilloÕers Õs. complementary R&DactiÕities

Ž .Geroski 1995 points out that, even if one ac-cepts that involuntary diffusion of knowledge hap-pens and that this knowledge has commercial valueto some of the recipients, it is still one thing to arguethat spillovers exist and another to argue that theyundermine incentives to innovate. Geroski’s point isthat firms must typically invest in research them-selves in order to benefit from external knowledgepools. This argument is emphasized in Branstetter

Ž .and Sakakibara 1998 , cf. Section 2 above, and isperhaps most forcefully stated by Cohen and

Ž . 49Levinthal 1989 . Cohen and Levinthal discuss indetail how a firm’s own R&D activity tends toenhance the absorptive capacity of R&D resultsproduced in other firms. If such a complementaryrelationship exists, ‘‘the analogy between spilloversand manna from heaven’’ is misleading and it is‘‘not clear exactly what ‘bits’ of knowledge havebeen produced by one’s own learning efforts and

Ž .which have spilled over from rivals’’ Geroski, 1995 .In this situation, spillovers may actually stimulateR&D. 50 The returns to own R&D increase in thesize of the spillover pool, and this creates a positivefeedback mechanism between the R&D investmentsin technologically related firms. 51 A negative effectdue to imperfect appropriability still exists, but it iscounteracted by an ‘absorption’ incentive, and conse-

49 Important empirical evidence is presented by MansfieldŽ .1981 , finding that imitation costs on average are about 65% ofthe original innovation costs.

50 Ž .Cf. Cohen and Levinthal 1989 for a formal analysis.51 It has been argued that a similar complementarity exists

between the knowledge stock and new investments in R&D withinŽ . Ž .individual firms, cf. Klette 1996 , Klette and Johansen 1998 ,

and the references cited therein.


quently the net effect of spillovers on R&D invest-ments is ambiguous. 52

The empirical evidence regarding the relationshipbetween own and others’ R&D suggests that com-plementarities in R&D are important in many cases.In addition to the empirical results presented by

Ž .Cohen and Levinthal 1989 and Branstetter andŽ . Ž .Sakakibara 1998 , Jaffe 1986 and Geroski et al.

Ž .1993 find a complementary relationship betweenown and others’ R&D. 53 Despite the rapid growthin the theoretical literature on R&D investment,spillovers and welfare, however, little attention hasbeen paid to the role of such complementarities andno results seem to be available discussing to whatextent a market equilibrium will lead to too littleinvestment in R&D in this case. 54 A rather boldsuggestion is that technology policies may be toofocused on sectors such as aircraft, semi-conductors,computers, electronics components and communica-tion equipment, where innovations tend to be com-

Ž .plementary according to Levin 1988 . The apparentparadox, that one observes coinciding high spilloversand high R&D investments in industries like theseŽ .Spence, 1984 , may indicate that these are industrieswhere spillovers do not undermine the incentives toinnovate, and where rivalry and strategic interactionmay even lead to excessive R&D investments.

5.2. Is ‘a big push’ from goÕernment R&D subsidiesneeded?

Complementarity in R&D activities, as discussedabove, is related to the discussion of governmentalsupport to emerging industries. A significant portionof the support to commercial R&D is targeted to-wards new, high-tech businesses and emerging tech-

52 Note, however, that if R&D investments can be divided intoinnovative research on one hand and imitation costs on the other,it may be that only imitation costs are complementary to thespillover pool. If this is the case, there will be no positivefeedback, i.e., it might be that firms with a deliberate imitationstrategy contribute little or nothing to the spillover pool.

53 Ž .Bernstein 1988 finds a complementary relationship in R&Dintensive sectors while firms in sectors performing little R&Dtend to substitute spillovers for own R&D.

54 Ž .A study that comes close is Kamien and Zang 1998 , whichcontains a model emphasizing the complementarities in R&Dactivities across firms.

nologies, and it seems to be based on infant industryarguments. That is, support to targeted high-techsectors is often rooted on the view that governmentsupport is needed to get emerging industrial activi-ties to ‘take off’ and reach ‘a critical mass’.

Perhaps surprisingly, this view might be entirelyconsistent with the discussion of complementary R&D activities above, where it was argued that suchcomplementary spillovers may encourage invest-ments in R&D. The point is, as emphasized by

Ž .Matsuyama 1995 , that complementarities tend tocreate multiple equilibria where, e.g., one equilib-rium corresponds to little or no R&D activity ineach of the firms, while another equilibrium corre-sponds to high R&D activities in several or allfirms. 55 That is, with an emerging industry or newtechnology, the firms might get trapped in a low-levelequilibrium where the lack of complementaryspillovers renders R&D unprofitable in all firmswith the result that the emerging industry neverreaches ‘the critical mass’ and ‘takes off’. Thissuggests that the government might play a coordinat-ing role by triggering higher activity, e.g., through anR&D program, until the firms and the industry havereached the high R&D-activity equilibrium.

Ž .Klette and Møen 1999 argue that the rationalefor government funding of IT-related research pro-grams in Norway can be well understood in theseterms. As pointed out in Section 2, their findingssuggest, however, that the Norwegian IT programswere not very successful in initiating new manufac-turing activities related to IT, and their case studyelaborates on the informational difficulties involved.

Ž .Inspired by Matsuyama 1997 , they conclude:

In contrast to the situation with illustrative andsimplistic game theoretic models, in real coordi-nation problems, information is a serious obstacle;what is the nature of the game, which players areinvolved, what does the pay-off structure looklike and how rapidly is it likely to change? Or inless formal terms; exactly which firms and whatactivities should be coordinated and in what way?

55 Ž .Klette and Møen 1999 elaborate on this point and givefurther references.


These serious questions are very hard to answer ina rapidly developing field such as informationtechnology and might be particularly hard to solvein a small open economy where a large majorityof the innovations take place abroad. We believethat industrial innovation is an activity wherecoordination problems and ‘market failure’ oftenare pervasive, but it is probably also an activitywhere policy makers and bureaucrats often lackthe information needed to improve on the marketsolution.

Hence, even though complementarities make itpossible, in theory, to improve on the market solu-tion, it is necessary to analyze whether the govern-ment, in practice, has the necessary capabilities to doso before initiating coordination programs.

5.3. International spilloÕers and high-tech policy

Complementarity between firms in R&D andother activities is a central idea in the ‘new tradetheory’ and ‘new economic geography’ literature ofthe last two decades. Much of the policy debate oversupport for R&D and innovation is concerned withinternational competition and ‘dynamic comparativeadvantage’, and those in favor of public technologyprograms are clearly inspired by the infant industryarguments discussed above. The work of Grossman

Ž .and Helpman 1991 is of particular relevance toŽtechnology policy. Grossman and Helpman 1991,

.ch. 8 show that if spillovers are geographicallybounded, then history matters, and countries with ahead start in accumulation of knowledge can widentheir lead over time. Moreover, they show that gov-ernments of lagging countries can improve theirgrowth prospects by offering a temporary R&Dsubsidy. This may eliminate these countries’ disad-vantages in high-tech industries. Similar results are

Ž .obtained by Krugman 1987 in the context of learn-ing-by-doing spillovers. However, as demonstrated

Ž .by Grossman and Helpman 1991, ch. 7 , the scopefor national policies disappears if knowledgespillovers are perfectly international, i.e., if ideasflow as easily between nations as they do withinnations. The extent to which spillovers are ‘intrana-tional’ or ‘international’ is therefore an important

empirical question. Inspired by these findings,Ž .Branstetter 1996 presents a microeconometric in-

vestigation using panel data for US and Japanesefirms, and he finds evidence that spillovers arestronger within each of the two countries than be-tween them. These results are supported by Narin et

Ž .al. 1997 , finding substantial ‘excessive’ self-cita-tion when comparing citations across countries, and

Ž .by Eaton and Kortum 1994 , finding that technologydiffusion is considerably faster within than betweencountries. 56 Branstetter concludes that ‘‘the ideathat promotion of R&D can have an impact oncomparative advantage is one that trade economistsshould take more seriously’’.

Trade economists working on growth and devel-opment are, naturally, focused on export orientedand import competing sectors. However, it is notobvious that these are industries where the case forgovernment support is particularly strong. The totalgain from national R&D investments includes notonly knowledge spillovers, but rent-spillovers to cus-tomers and buyers of intermediate goods as well. Asargued in Section 3, these may be considerable, andin the extreme case of monopolistic competition,often assumed in theoretical models, all profits arecompeted away such that only rent spillovers arerelevant for policy. If a substantial part of thespillovers created through R&D subsidy programs isto the rest of the world, e.g., because the targetedR&D intensive industries are highly export oriented,one may question why the government of the sourcecountry should bear the financial burden. This reser-vation seems particularly relevant to small openeconomies, but it has also been emphasized by sev-eral commentators in the debate over the funding ofthe ATP-program in the US. 57

At a general level, it is not difficult to outline theimplications of international R&D spillovers. Gov-ernments should only subsidize R&D up to the pointwhere the marginal cost equals the marginal socialbenefit accruing to its own nationals. When evaluat-ing the marginal social benefit, potential negativerepercussions from increased competition due to un-

56 Ž .Cf. Mohnen 1998 for a recent review of the literature oninternational R&D spillovers.

57 Ž .See, e.g., Yager and Schmidt 1997 .


intentional spillovers received by foreign firms shouldbe included, 58 but also potential positive effects ofeconomic growth abroad. 59 Such positive effectscould, e.g., be larger export markets and increasedpolitical stability in developing countries. Empiricalresults have obviously not been accumulated to alevel which makes it possible to determine whatamount of subsidies is optimal according to thistheoretical criteria.

Empirical results suggest, as noted above, thatspillovers to some extent are geographicallybounded. 60 This may justify national technologyprograms, but the point we want to emphasize is thata careful analysis of the likely distribution ofspillovers is necessary as the share of spilloversaccruing to non-nationals may be substantial in somesectors. Note also that the existence of internationalspillovers gives scope for increased global effi-ciency through R&D cooperation between countries.The fact that technology policy and R&D programswithin the European Union to some extent have beenmoved from individual member states to the unionlevel since the 1980s, can be interpreted as a re-sponse to this understanding. 61

58 This effect need not be negative, as increased competition inthe home market benefits consumers and other industries throughinput linkages.

59 There is a large theoretical literature on optimal R&D poli-cies, exploring what may happen under various assumptions re-garding degree of competition, degree of intra- and interindustryspillovers, degree of openness, degree of international spillovers,whether there is strategic behavior or not, whether there is R&Dcooperation or not and whether R&D of the firms in question are

Ž .strategic complements or substitutes. Cf. Leahy and Neary 1997Ž .and Neary 1998 for recent reviews of this literature. The not

surprising policy advice in this literature, as we read it, is that itall depends on the assumptions. An important challenge forempirically minded economists, therefore, is to sort out whatassumptions are the relevant ones. Alternatively, one can follow

Ž .Neary 1998 and many others and conclude that the detailedinformation required to improve on the market solution is unlikelyever to be available to the policy maker.

60 This is not only a finding of the literature on internationalspillovers. There is also a literature utilizing national data which

Ž .strongly supports the view, cf., e.g., Jaffe 1989 , Jaffe et al.Ž . Ž .1993 , and Adams and Jaffe 1996 .

61 Well known examples of such programs include ESPRIT,EUREKA, and TSER, among others.

5.4. R&D joint Õentures and the Coase theorem

Ž .Klette and Møen 1999 argue that firms seem tointernalize spillovers through various market ar-rangements largely ignored in many of the theoreti-cal models of R&D investment. 62 One aspect ofthis is that the empirical findings emerging from theliterature reviewed in Section 4 might be quite mis-leading, since some of the effects interpreted asspillovers may actually be knowledge transfers thatare internalized in the market, e.g., through coopera-tive agreements. It seems reasonable to believe thatfirms know who their customers, suppliers, and ri-vals are, and according to the ‘Coase theorem’, firmswould tend to sign contractual arrangements govern-ing the knowledge flows between them. 63 A largenumber of cooperative agreements observed in themarket indicate that this may be an aspect of theexternality issue which has been grossly underem-

Ž .phasized. Freeman 1991 reports that ‘‘almost all ofŽ .the top 20 information technology IT firms in US,

EU and Japan made more than 50 cooperative ar-rangements of various kinds in the 1980s and somemade more than a hundred’’. With respect to smaller

Ž .companies, Aakvaag et al. 1996 report that about60% of Norwegian electronic firms participate intechnological cooperation schemes. Partner firms of-ten have an interrelated ownership structure, and thisis obviously a simple and basic market mechanismfor internalizing externalities. 64

Related evidence is provided in the two studies byŽ . Ž .Zucker et al. 1998a and Zucker et al. 1998b . In

Ž .the 1998b study, Zucker et al. demonstrate that thelocation of academic experts at the leading edge ofbasic bioscience strongly influenced the location ofnew biotechnology enterprises in the US. Further

Ž .exploring this in the 1998a study, it was revealedthat firms and star scientists were not merely locatedin the same area, but that the scientists were deeply

62 Ž .See Leahy and Neary 1997 and references cited in that studyfor a review of recent theoretical studies of R&D joint ventures.

63 More precisely, firms will perfectly internalize externalities inthe absence of information and transaction costs. See UsherŽ .1998 for a critical view.

64 Ž .See Klette 1996 for a study of spillovers between firms withan interlocking ownership structure.


involved in the operations of the firms. Hence, whatmight have been interpreted as localized knowledgespillovers using standard methodologies and data

Ž .sets cf., e.g., Jaffe, 1989 , was to a large extent amatter of market exchange.

Our point is not that spillovers are fully taken careof by contracting, and we recognize that it is notori-ously hard to write complete contracts for uncertainand unpredictable activities such as R&D. What weargue is that both in theoretical and empirical analy-sis, more attention should be paid to the manycontractual arrangements utilized and invented bythe firms to overcome the potential spillover prob-lems generated in innovative activities.

5.5. SpilloÕers and the mobility of research workers

We will end our review of spillover issues relatedto R&D policy by turning to the labor market. Anumber of authors have pointed to mobility of laboras an important mechanism for knowledgediffusion, 65 and it is most often thought of as a

Ž .spillover mechanism. Jaffe 1996 , making a cleardistinction between rent spillovers and knowledgespillovers, considers mobility of researchers to be of

w xthe second type, writing that ‘‘ k nowledge spilloversalso occur when researchers leave a firm and take ajob at another firm’’. Defining knowledge spilloversas ‘‘benefit leakages that occur in absence of amarket interaction between the innovator and thespillover beneficiary’’, this seems a bit inconsistentsince mobility of researchers takes place in the labormarket. Jaffe implicitly acknowledges this point,writing that ‘‘important innovative successes arelikely to increase the incentive for researchers tocapitalize on their tacit knowledge by moving toanother firm or starting their own’’. We will arguebelow that, from a theoretical point of view, it is notentirely clear to what extent labor mobility really is aspillover, but if it is, we believe it is most correct to

Ž .analyze it as a market i.e., rent spillover.

65 Ž . Ž .Cf., e.g., Geroski 1995 , Jaffe 1996 , Almeida and KogutŽ . Ž .1996 and Zucker et al. 1997 for some recent statements on the

Ž .importance of labor mobility. Almeida and Kogut 1996 , study-ing patent holders, are particularly interesting, showing empiri-cally that ideas are spread through the mobility of key scientists.

Our point of departure is that R&D investmentnot only increases the firms’ stock of innovations, italso increases the human capital of the researchworkers. After all, research is a learning process.This perspective introduces two interesting ques-tions. First, who captures the value of the humancapital from R&D activities, and second, how is thefirms’ investment incentives affected by the possibil-ity that research workers may quit? With perfectlabor and credit markets, the answer to the latterquestion is that the investment incentives are notaffected. To the extent that research work has a‘general training’ element and increases the re-searchers’ future marginal product, they can look

Žforward to corresponding future wage increases cf..Becker, 1964 . This gives the research workers in-

centives to bear the cost of the training throughlower wages in the beginning of their career, andconsequently, a research worker who quits does notimpose a cost on his or her employer. 66 If the fairlysteep wage profile thus associated with a researchcareer does not suit the researchers’ consumptionpreferences, they can borrow for current consump-tion towards future wage increases. With respect tothe question about who captures the value of thehuman capital from R&D, this analysis implies thatit is the research workers, but they also pay theinvestment costs. The flip side of this conclusion isthat labor mobility is not a mechanism that causesunderinvestments in R&D, and should not be con-sidered a spillover channel either.

A first objection to the analysis above is thatcredit markets are not likely to deliver all the neces-sary services given the moral hazard problems in-volved in borrowing on future income. This marketfailure will, evaluated in isolation, cause underin-vestment in R&D. 67 If there is larger uncertaintyover the future gains from research work than thereis over future income from alternative career paths,

66 Ž .See Pakes and Nitzan 1983 for a related, formal analysis.67 The utility loss associated with low consumption in the

beginning of the career will shift the supply-of-research-laborcurve downwards, increase the equilibrium wage of researchworkers and thereby the price on R&D investments. This willresult in R&D investments below the level associated with aperfect credit market.


risk aversion at the individual level will magnify theunderinvestment problem. 68 Imperfections in the la-bor market may, on the other hand, increase firms’incentives to invest in research work by reducing themobility of researchers across firms. Such labor mar-ket imperfections include search costs and asymmet-ric information about the human capital of the em-ployees. These effects will result in wages beingbelow marginal product, and hence give firms an

Žincentive to invest in workers’ general i.e., non-firm. 69specific human capital.

Determining the total effect of the ‘training as-pect’ of R&D on investments and wages is in theend an empirical task, and little can at this momentbe said. In order to investigate the issue, a frame-work explicitly linking R&D investments of firmswith human capital accumulation in research work-ers, must be developed. Given the increasing numberof matched employer-employee data sets now be-coming available, we think future research in thisdirection will prove fruitful. It might be essential,however, that these data sets are able to trace themobility of researchers across establishments, as suchmobility and entrepreneurship can be a major com-ponent of the pay-off for successful researchers.

6. Conclusions

We have not succeeded in answering all ourproblems. The answers we have found only serveto raise a whole set of new questions. In someways we feel we are as confused as ever, but webelieve we are confused on a higher level and

Ž .about more important things Øksendal, 1985 .

Estimates of the economic returns to R&D pro-jects have gone a long way since this line of research

68 It might be the individual’s aversion towards high skewnessrather than high variance which is the more important issue here.Notice that risk-neutrality at the firm level is irrelevant for theargument.

69 Ž .Cf. Acemoglu and Pischke 1999 for a review of somerelevant literature.

Ž .started more than 40 years ago cf. Griliches, 1958 .We have in this paper focused on a relatively smallnumber of recent studies that try directly to evaluatethe social returns from subsidies to commercial R&Dactivities. Four of the five studies suggest that thesubsidy schemes have had a positive effect on per-formance in the targeted firms. We have, however,pointed out some of the shortcomings in the avail-able studies and raised some questionmarks aboutthe conclusion that these subsidy schemes have re-duced market failures. Discussing similar shortcom-ings related to causal inference from observational

Ž .studies, Cochran 1965 notes that a reader, ‘‘if laterasked for a concise summary of the paper, may quiteproperly report: ‘He said it’s all very difficult’.’’ Werecognize that our paper may leave the same impres-sion on our readers, but we also believe, as Cochran

w xemphasizes, that ‘‘ a listing of common difficultiesis . . . helpful in giving an overall view of the prob-lems that must be overcome if this type of researchis to be informative.’’ Furthermore, we have tried toemphasize that many of the unresolved questions areready for further research with tools and data setswithin our reach.

On the methodological side, a more careful infer-ence of the magnitude of the impact parameters ofinterest can be made, drawing inspiration from therecent advances in the evaluation literature in labormarket econometrics. A more ambitious approachwould be to go beyond these largely non-parametrictechniques and try to merge the model of perfor-mance and subsidy impact with a structural model ofhow the government allocates the R&D subsidies. Astructural model of the allocation of R&D subsidiesshould address the question of how the governmentcan construct operational procedures to identify R&Dprojects with high social returns, 70 and the empiricalanalysis based on such a structural model can help usto identify to what extent the government agenciessucceed in implementing these procedures. These areclearly interesting and worthwhile research tasks in

70 The practical difficulties in selecting R&D projects with highsocial returns are discussed in some detail in Yager and SchmidtŽ .1997 . The ongoing ATPrNBER project is particularly notewor-thy in its attempt to draw on the insights from the econometricliterature to resolve some of these difficulties.


themselves, and if completed successfully they willgive us an alternative handle to eliminate the poten-tial selection biases that we discussed at some lengthin Section 3.

A large number of research papers on R&D andspillovers have emerged over the last decade, but wehave argued that several theoretical and empiricalaspects of spillovers deserve more attention beforeconclusions about R&D policy can be drawn. Manyof the issues we have raised seem to require a moredetailed investigation of the nature of the spilloversand also a more detailed investigation into the vari-ous contractual arrangements that prevail in the mar-ket between firms and between the researchers andtheir firms.

Finally, evaluation of the economic returns toR&D subsidy programs seems to require a combina-tion of empirical investigations at different levels ofobservation. We have argued that the microecono-metric approach that has been the focus of this papershould be supplemented with detailed case studies toget a more precise estimate of the economic returnsfrom the few, outstanding innovations that mighttypically generate a very large share of the economicbenefits emerging from risk-oriented R&D subsidyprograms. On the other hand, in order to estimate theimpact of an R&D subsidy program in the presenceof knowledge spillovers, we need to look beyond thedirect impact of the subsidies on the performance oftargeted firms and consider changes in performanceof the industries or ‘technological clusters’ to whichthe supported firms belong. This may lead us to amore aggregated, industry-level analysis. It is en-couraging to observe that economic researchers al-ready have many of these elements in their tool box,but they have not yet been fully tied together.

References

Aakvaag, T., et al., 1996. Utfordringen - forskning og innovasjonfor ny vekst, Report from a commission appointed by theNorwegian Ministry of Industry and Energy.

Abadie, A., Angrist, J., Imbens, G., 1998. Instrumental variableestimation of quantile treatment effects. NBER Technicalworking papers 229.

Acemoglu, D., Pischke, J.-S., 1999. Beyond Becker: Training inimperfect labor markets. Economic Journal 109, F112–F142.

Adams, J.D., Jaffe, A.B., 1996. Bounding the effects of R&D: an

investigation using matched establishment-firm data. RANDJournal of Economics 27, 700–721.

Almeida, P., Kogut, B., 1996. The economic sociology of thegeographic localization of ideas and the mobility of patent

Žholders. Mimeo Georgetown University and University of.Pennsylvania .

Angrist, J., Krueger, A., 1998. Empirical strategies in labor eco-Ž .nomics. In: Ashenfelter, O., Card, D. Eds. , Handbook of

Ž .Labor Economics forthcoming volume . North Holland, Ams-terdam.

Australian Industry Commission, 1995. Research and Develop-Žment, Report No. 44, Vol. 3, Appendicies Australian Govern-

.ment Publishing Service, Canberra .Averch, H., 1997. Review of Evaluating public sector research

and development. Journal of Economic Literature 35, 1424–1426.

Becker, G.S., 1964. Human Capital. Columbia University PressŽ3rd edn. by the University of Chicago Press, New York,

.1993 .Bernstein, J.I., 1988. Cost of production, intra- and interindustry

R&D spillovers: Canadian evidence. Canadian Journal of Eco-nomics 21, 324–347.

Blundell, R., 1998. Evaluation methods for non-experimental data.Ž .Mimeo Institute for Fiscal Studies, London .

Branstetter, L., 1996. Are knowledge spillovers intranational orinternational in scope? Microeconometric evidence from USand Japan. NBER Working Paper No. 5800.

Branstetter, L., Sakakibara, M., 1998. Japanese research consortia:a microeconometric analysis of industrial policy. Journal of

Ž .Industrial Economics 46 2 , 207–233.Bresnahan, T., 1986. Measuring the spillovers from technical

advance: mainframe computers in financal services. AmericanEconomic Review 76, 742–755.

Ž .Bresnahan, T., Gordon, R. Eds. , 1997. The economics of newgoods, The University of Chicago Press and NBER, Chicago.

Brown, M., Conrad, A., 1967. The influence of research on CESŽ .production relations. In: Brown, M. Ed. , The Theory and

Empirical Analysis of Production, Columbia University Pressfor NBER, New York, pp. 275–340.

Brown, M., Randall Curlee, T., Elliott, S., 1995. Evaluatingtechnology innovation programs: the use of comparison groupsto identify impacts. Research Policy 24, 669–684.

Bureau of Industry Economics, 1994. Beyond the innovator:Spillovers from Australian R&D. Bureau of Industry Eco-

Ž .nomics, Occasional Paper no. 16 Canberra .Cochran, W., 1965. The planning of observational studies of

human populations, The Journal of the Royal Statistical Soci-ety, 234–65.

Cohen, W., Levinthal, D., 1989. Innovation and learning: the twofaces of R&D. Economic Journal 99, 569–596.

David, P., 1997. From market magic to calypso science policy: areview of Terence Kealey’s ‘The economic laws of scientificresearch’. Research Policy 26, 229–255.

David, P.A., Hall, B.H., Toole, A.A., 1999. Is public R&D acomplement or substitute for private R&D? A review of theeconomic evidence. Research Policy, 499–532.

Eaton, J., Kortum, S., 1994. International patenting and technol-ogy diffusion. NBER Working Paper No. 4931.


Friedman, M., 1994. National Science Foundation grants foreconomics: correspondence. Journal of Economic Perspectives8, 199–200.

Geroski, P.A., 1991. Innovation and the sectoral sources of UKproductivity growth. Economic Journal 101, 1438–1451.

Geroski, P.A., 1995. Do spillovers undermine the incentive toŽ .innovate? In: Dowrick, S. Ed. , Economic Approaches to

Innovation, Edwar Elgar, Aldershot, UK, pp. 76–97.Geroski, P.A., Machin, S., Van Reenen, J., 1993. Innovation and

firm profitability. RAND Journal of Economics 24, 198–211.Geroski, P.A., Small, I., Walters, C.F., 1998. Agglomeration

economies, technology spillovers and company productivityŽ .growth. CEPR Discussion Paper No. 1867 London .

Gompers, P.A., Lerner, J., 1997. Money chasing deals?: Theimpact of fund inflows on private equity valuations. MimeoŽ .Harvard University .

Griliches, Z., 1958. Research cost and social returns: hybrid cornand related innovations. Journal of Political Economy 66,419–431.

Griliches, Z., 1979. Issues in assessing the contribution of re-search and development to productivity growth. Bell Journalof Economics 10, 92–116.

Griliches, Z., 1986. Productivity, R&D, and basic research at thefirm level in the 1970’s. American Economic Review 76,141–154.

Griliches, Z., 1992. The search for R&D spillovers. ScandinavianJournal of Economics 94, S29–S47.

Griliches, Z., 1994. National Science Foundation grants for eco-nomics: response. Journal of Economic Perspectives 8, 203–205.

Griliches, Z., 1995. R&D and productivity: econometric resultsŽ .and measurement issues. In: Stoneman, P. Ed. , Handbook of

the Economics of Innovation and Technical Change. Black-well, Oxford.

Griliches, Z., 1997. The Simon Kutznets Memorial Lectures.Ž .Mimeo Harvard University .

Griliches, Z., 1998. R&D and Productivity, the EconometricEvidence. University of Chicago Press, Chicago.

Griliches, Z., Regev, H., 1998. An econometric evaluation ofhigh-tech policy in Israel. Paper presented at ATP-conferencein Washington, DC, June 1998.

Grossman, G., Helpman, E., 1991. Innovation and Growth in theGlobal Economy, MIT Press, Cambridge.

Hall, B.H., 1996. The private and social returns to research andŽ .development. In: Smith, B., Barfield, C. Eds. , Technology,

R&D, and the Economy, Brookings Institution and AEI,Washington DC, pp. 140–162.

Hall, B.H., van Reenen, J., 1999. How effective are fiscal incen-tives for R&D? A review of the evidence. Research Policy,449–471.

Heckman, J., 1979. Sample selection bias as a specification error.Econometrica 46, 931–961.

Heckman, J., Smith, J., Clements, N., 1997. Making the most outof social experiments: accounting for heterogeneity in pro-gramme impacts. Review of Economic Studies 64, 487–535.

Heckman, J., Ichimura, H., Smith, J., Todd, P., 1998. Characteriz-ing selection bias using experimental data. Econometrica 66,1017–1098.

Irwin, D., Klenow, P., 1996. High-tech R&D subsidies — esti-mating the effects of SEMATECH. Journal of InternationalEconomics 40, 323–344.

Jaffe, A., 1986. Technological opportunity and spillovers fromR&D. American Economic Review 76, 984–1001.

Jaffe, A., 1989. Real effects of academic research. AmericanEconomic Review 79, 957–970.

Jaffe, A., 1996. Economic analysis of research spillovers: Implica-tions for the advanced technology program. National Instituteof Standards and Technology, Report GCR 97-708.

Jaffe, A.B., Trajtenberg, M., Henderson, R., 1993. Geographiclocalization of knowledge spillovers as evidenced by patentcitations. Quarterly Journal of Economics 108, 577–598.

Kamien, M., Zang, I., 1998. Meet me halfway: Research jointventures and absorptive capacity. Mimeo.

Kauko, K., 1996. Effectiveness of R&D subidies — a scepticalnote on the empirical literature. Research Policy 25, 321–323.

Kealey, T., 1997. The Economic Laws of Scientific Research.MacMillan Press, London.

Klette, T., 1996. R&D, scope economies and plant performance.RAND Journal of Economics 27, 502–522.

Klette, T., Johansen, F., 1998. Accumulation of R&D-capital anddynamic firm performance: a not-so-fixed effect model. An-nales d’Economie et de Statistique 49r50, 389–419.

Klette, T., Møen, J., 1998. R&D investment responses to R&Dsubsidies: a theoretical analysis and a microeconometric study.Mimeo, Presented at NBER Summer Institute 1998.

Klette, T., Møen, J., 1999. From growth theory to technologypolicy — coordination problems in theory and practice. NordicJournal of Political Economy 25, 53–74.

Krugman, P., 1987. The narrow moving band, the Dutch disease,and the competitive consequences of Mrs. Thatcher. Journal ofDevelopment Economics 27, 41–55.

Leahy, D., Neary, J., 1997. Public policy towards R&D inoligopolistic industries. American Economic Review 87, 642–662.

Lerner, J., 1998. The government as venture capitalist: the long-runŽ .impact of the SBIR program. Mimeo Harvard University .

Previously published as NBER WP 5753, 1996.Levin, R., 1988. Appropriability, R&D spending and technologi-

cal performance. American Economic Review, Papers andProceedings 78, 424–428.

Link, A., 1996. Evaluating public sector research and develop-Ž .ment Greenwood, Praeger, Westport, Conn. and London .

Link, A., Teece, D., Finan, W., 1996. Estimating the benefits fromcollaboration: the case of SEMATECH. Review of IndustrialOrganization 11, 737–751.

Luukkonen, T., 1998. The difficulties in assessing the impact ofEU framework programmes. Research Policy 27, 599–610.

Mansfield, E., 1981. Imitation costs and patents: an empiricalstudy. Economic Journal 91, 907–918.

Mansfield, E., 1996. Contributions of new technology to theŽ .economy. In: Smith, B.L.R., Barfield, C.E. Eds. , Technology

R&D and the Economy, Brookings Institution and AEI, Wash-ington, DC, pp. 114–39.

Mansfield, E., Rapoport, J., Schnee, J., Wagner, S., Hamburger,M., 1977. Social and private rates of return from industrialinnovations. Quarterly Journal of Economics 91, 221–240.


Manski, C., 1993. Identification of endogenous social effects: thereflection problem. Review of Economic Studies 60, 531–542.

Marshall, A., 1920. Principles of Economics. 8th edn., MacMil-lan, London.

Matsuyama, K., 1995. Complementarities and cumulative pro-cesses in models of monopolistic competition. Journal ofEconomic Literature 33, 701–729.

Matsuyama, K., 1997. Economic development as coordinationŽ .problems. In: Aoki, M., et al. Eds. , The Role of Government

in East Asian Economic Development, Clarendon Press, Ox-ford.

Mohnen, P., 1996. R&D externalities and productivity growth.STI Review 18, 39–66.

Mohnen, P., 1998. International R&D spillovers and economicŽgrowth. Mimeo Universite du Quebeck a Montreal and´ ´ ¨ ´

.CIRANO .Mowery, D., Rosenberg, N., 1998. Paths of innovation: technolog-

ical change in 20th century America, Cambridge UniversityPress, Cambridge.

Nadiri, M.I., 1993. Innovations and technological spillovers.NBER Working Paper No. 4423.

Narin, F., Hamilton, K.S., Olivastro, D., 1997. The increasinglinkage between US technology and public science. ResearchPolicy 26, 317–330.

Neary, J., 1998. Pitfalls in the theory of international trade policy:Concertina reforms of tariffs, and subsidies to high-technologyindustries. Scandinavian Journal of Economics 100, 187–206.

Nickell, S., 1981. Biases in dynamic models with fixed effects.Econometrica 49, 1417–1426.

Øksendal, B., 1985. Stochastic Differential Equations, Springer-Verlag, Berlin.

Pakes, A., 1986. Patents as options: some estimates of the value ofholding European patent stocks. Econometrica 54, 755–784.

Pakes, A., Nitzan, S., 1983. Optimum contracts for researchpersonnel, research employment, and the establishment of‘rival’ enterprises. Journal of Labor Economics 1, 345–365.

Scherer, F., Harhoff, D., 1999. Technology policy for a world ofskew-distributed outcomes. Research Policy, this volume.

Spence, A., 1984. Cost reduction, competition, and industry per-formance. Econometrica 52, 101–121.

Trajtenberg, M., 1983. Dynamics and welfare analysis of productinnovations, PhD thesis, Harvard University.

Trajtenberg, M., 1989. The welfare analysis of product innova-tions, with an application to computer tomography scanners.Journal of Political Economy 97, 444–479.

Usher, D., 1998. The Coase theorem is tautological, incoherent orwrong. Economics Letters 61, 3–11.

Yager, L., Schmidt, R., 1997. The Advanced Technology Pro-gram: A Case Study in Federal Technology Policy. AEI Press,Washington, DC.

Zucker, L.G., Darby, M.R., Torero, M., 1997. Labor mobilityfrom academe to commerce. NBER Working paper No. 6050.

Zucker, L.G., Darby, M.R., Armstrong, J., 1998a. Geographicallylocalized knowledge: spillovers or markets? Economic Inquiry36, 65–86.

Zucker, L.G., Darby, M.R., Brewer, M.B., 1998b. Intellectualhuman capital and the birth of US biotechnology enterprises.American Economic Review 88, 290–306.

Do subsidies to commercial R&D reduce market failures? …bhhall/others/... · 2012. 2. 3. ·...

Documents

Transcript of Do subsidies to commercial R&D reduce market failures? …bhhall/others/... · 2012. 2. 3. ·...