DRAFT Research and Development Composition and Labor Productivity Growth in 16 OECD Countries

Transcription

1 DRAFT Research and Developmen Composiion and Labor Produciviy Growh in 16 OECD Counries Ram C. Acharya Indusry Canada Serge Coulombe Universiy of Oawa and Indusry Canada December 2005 Absrac Using daa for 16 OECD counries from 1973 o 2000, we show ha growh in labor produciviy is highly responsive o business research and developmen (R&D) expendiures. Increasing business R&D by 0.1 percen of GDP increases per capia income beween 0.9 and 1.8 percen in he long run. In public R&D, only expendiures on higher educaion have a significan posiive effec on labor produciviy growh. In our decomposiion of he secoral R&D ino a pure R&D inensiy effec and a secoral size effec, resuls show ha elasiciy of labor produciviy wih respec o hese variables differs by secor. The posiive size effec dominaes he high-ech manufacurer, whereas i is he inensiy effec ha drives he posiive correlaion beween medium-low-ech manufacurer R&D and labor produciviy. (JEL O32, O40) Corresponding auhor: Ram Acharya 10 Eas Queen Sree Oawa, Onario Canada K1A 0H5 acharya.ram@ic.gc.ca The papers canno be circulaed and quoed wihou obaining wrien approval from he auhors. Views expressed in his paper are hose of he auhors and do no necessarily reflec hose of Indusry Canada. We would like o hank Renee S-Jacques, Richard Roy, Someshwar Rao, Jianmin Tang and Jean McCardle for heir commens on an earlier version of he paper.

2 1 Research and Developmen Composiion and Labor Produciviy Growh in 16 OECD Counries Absrac Using daa for 16 OECD counries from 1973 o 2000, we show ha growh in labor produciviy is highly responsive o business research and developmen (R&D) expendiures. Increasing business R&D by 0.1 percen of GDP increases per capia income beween 0.9 and 1.8 percen in he long run. In public R&D, only expendiures on higher educaion have a significan posiive effec on labor produciviy growh. In our decomposiion of he secoral R&D ino a pure R&D inensiy effec and a secoral size effec, resuls show ha elasiciy of labor produciviy wih respec o hese variables differs by secor. The posiive size effec dominaes he high-ech manufacurer, whereas i is he inensiy effec ha drives he posiive correlaion beween medium-low-ech manufacurer R&D and labor produciviy. (JEL O32, O40) I. Inroducion One of he imporan lessons learned from he wave of empirical growh research in he las 15 years is ha no single facor deermines a counry s growh rae. Using a broad se of 88 counries ha includes boh developed and less-developed counries, Xavier Sala-i-Marin e al. (2004) idenify 18 variables (ou of a se of 67 poenial candidaes) ha are significanly relaed o growh. However, mos of hese significan variables represen differences beween he less- and mos-developed counries. The lis of poenial variables o explain economic growh becomes much smaller when one aemps o explain sandard-of-living differences across developed counries. From boh a heoreical and an empirical poin of view, he various measures of R&D expendiures are generally considered he mos imporan of hese poenial variables. Since R&D is carried ou mosly in developed counries and he exen of he R&D varies hrough ime and across counries, is abiliy o explain he variaion in produciviy growh in developed counries can be poenially very high. The findings of Rachel Griffih e al.

3 2 (2004) indicae ha R&D simulaes produciviy growh hrough a echnological cach-up and an innovaion channels in a panel daa sudy of indusries and welve developed counries. Along his line, he objecive of his paper is o evaluae empirically he conribuion of various componens of R&D expendiure o promoing economic growh across a se of counries in he Organizaion for Economic Co-operaion and Developmen (OECD). More specifically, we will address he following quesions: How effecive are privae and public R&D expendiures in raising labor produciviy in developed counries? As for he performance of public R&D, do governmen inramural R&D and higher educaion R&D have similar effecs? Does a dollar spen on R&D in every secor of he economy has he same effec or does i differ by secor? Is i he R&D inensiy in he secor or he size of he secor or boh ha maers for overall labor produciviy growh? Wha is he role of foreign R&D in promoing domesic labor produciviy? To answer hese quesions, we use a ime-series and cross-secion (TSCS) empirical model of annual daa from 1973 o 2000 for 16 OECD counries. 1 Due o daa limiaions, he empirical analysis dealing wih public R&D covers only he period of 1981 o 2000, and he decomposiion of business R&D ino R&D inensiy and size of he secors is analyzed for he 1979 o 2000 period. The quesions posed are no enirely new; previous researchers have aemped o answer some of hem. Broadly speaking, wo hreads of lieraure deal wih cross-counry R&D sudies. One esimaes he impac of R&D expendiure on various measures of labor produciviy growh; he second focuses on R&D s impac on oal facor produciviy (TFP) growh. Thus he esimaed coefficiens on he R&D variables in he firs ype of model capure he oal effec 1 The counries included in he sudy are Ausralia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Ialy, Japan, Neherlands, Norway, Spain, Sweden, he Unied Kingdom, and he Unied Saes.

4 3 (boh direc and spillover) of R&D in he economy. The second hread, sudying he relaionship beween TFP growh and R&D, esimaes only he spillover effecs of R&D. 2 The firs ype of lieraure, pioneered by Waler Nonneman and Parick Vanhoud (1996), uses he now-sandard convergence-growh regression framework, derived from he ransiional dynamics of he neo-classical growh model (N. Gregory Mankiw e al., 1992). 3 Nonneman and Vanhoud s (1996) conribuion exended Mankiw e al. (1992) by incorporaing he R&D capial as an addiional explanaory variable ino he exended Solow growh model. 4 In his framework, hey find ha he Solow model, augmened furher by he addiion of R&D, has subsanially greaer explanaory power compared wih ha of Mankiw e al. (1992), and ha he R&D variable is significan a he 10 percen level. The posiive parial correlaion beween aggregae R&D expendiure and GDP per capia or labor produciviy growh across OECD counries has been confirmed in Kaarina Keller and Panu Pouvaara (2003) and OECD (2003). In he second ype of lieraure, several sudies have esimaed he impac of R&D (boh domesic and foreign) on TFP growh. 5 Domininque Guellec and Bruno van Poelsberge de la Poerie (2001), in a sudy closely relaed o ours, invesigae he long-erm effec of various ypes of R&D using an error-correcion model (ECM) for 16 OECD counries over he period 2 As explained in Charles R. Hulen (2001), since he calculaion of TFP includes oal producion coss including R&D capial any posiive relaionship beween TFP and R&D capial is considered a spillover effec. This is paricularly so in business R&D. In his case, he privae resources devoed o R&D are included in he economy s sock of capial and pool of labor a he aggregae level. Thus, if here are no spillovers of R&D and he sandard assumpions underlying he calculaion of TFP hold (noably, perfec compeiion and consan reurns o scale a he aggregae level), hen here should be no relaionship beween hese wo variables. The lieraure showing a posiive relaionship beween hese wo variables is herefore he proof of he exisence of spillover effecs of R&D. 3 In a TSCS framework, a convergence-growh model migh be viewed as a special case of an error-correcion model (ECM) describing he adjusmen of labor produciviy oward is long-run equilibrium (seady-sae) level. See Serge Coulombe (2005) for a discussion of he relaionship beween an ECM framework, TSCS models and convergence-growh regressions. The empirical convergence-growh analyses as in Mankiw e al. (1992) and Nonneman and Vanhoud (1996) are based on pure cross-secion daa. 4 Mankiw e al. (1992) is an augmened Solow growh model, where hey include physical capial as in Solow and augmen i wih he inclusion of human capial as anoher deerminan of per worker GDP. 5 Some of hese are Griliches (1979; 1994), Jeffrey I. Bernsein and M. Ishaq Nadiri (1988), David T. Coe and Elhanan Helpman (1995), Wolfgang Keller (2002), and Kul B. Luinel and Mosahid Khan (2005).

5 o Our resuls regarding he relaive effec of public and privae R&D expendiures differ from heirs o some exen. As we will see laer, i also appears ha some of heir resuls lack robusness for alernaive economeric specificaions. Furhermore, in conras o heir reliance on jus aggregae-level daa, we decompose privae R&D expendiure by secor and esimae he secoral impac of R&D. Despie he usefulness of he second approach in esimaing R&D spillover, we believe ha he firs approach is preferable for our purpose for he following reasons. Firs, for crosscounry sudies, he convergence-growh model focusing on labor produciviy or per capia GDP adjusmen is based on a sound heoreical framework. Unlike labor produciviy convergence, o our knowledge here in no heoreical model predicing ha TFP levels should adjus in a similar paern across counries. Second, considering he daa complexiy relaed o measuremen errors when compuing even one counry s TFP, i is very challenging o use cross-counry TFP growh. Third and mos imporan as shown in a seminal work of Rober E. Hall (1988) if here are nonconsan reurns and imperfec compeiion, TFP will be a biased measure of acual unobserved produciviy growh. Following Paul M. Romer (1990), we know ha R&D is associaed wih increasing reurns o scale and imperfec compeiion. If ha happens o be he case, hen wha we consider spillover effecs of R&D on TFP would be a biased measure. Hence we follow he firs approach and use labor produciviy growh, a erm closely relaed wih per capia GDP growh, as a dependen variable. 6 The paper conribues o he exising lieraure in hree ways. Firs, we esimae he effec of he various componens of R&D by decomposing R&D ino privae and public caegories and by dividing public R&D ino higher educaion and governmen. Given he fac ha privae and

6 5 public R&D can complemen as well as subsiue, i is imporan o incorporae boh ypes of R&D separaely. Furhermore, in he exising Solow augmened models, here is no conrol for foreign R&D spillover. By inroducing foreign R&D as a conrol, his paper no only sharpen he precision of he esimae of reurns o domesic R&D bu will also allow us, as a by-produc, o measure he impac of foreign R&D on domesic produciviy. Second, we provide a secoral dimension o hese models by esimaing separaely he impac of business R&D on high-ech manufacurers, medium-high-ech manufacurers, medium-low-ech manufacurers, low-ech manufacurers, and services indusries. This exension is paricularly noeworhy as here is a general belief ha R&D in high-ech secors should generae a higher reurn o he economy compared wih R&D in oher secors. This paper, as far as we are aware, will be he firs o be able o es his percepion. We also decompose he impac of secoral R&D on labor produciviy ino a pure inensiy and secoral-size effecs which allows us o undersand he relaive imporance of R&D inensiy increase in a given secor and size increase of a secor wih given inensiy. Third, raher han sudying he impac of R&D on GDP per capia and TFP, we sudy he impac of R&D on labor produciviy, conrolled for invesmen inensiy. We hereby escape he problem relaed o esimaing TFP. Our resuls should no be biased even if he usual assumpions of consan reurns o scale and perfec compeiion, used in compuing TFP growh, do no hold. The main findings of he paper are as follows. The oal effec of business R&D is posiive and srongly significan in all economeric specificaions. The resul shows ha a counry wih business R&D inensiy 10 percen higher han he ypical OECD counry ends up 6 Noe ha labor produciviy growh is equal o he growh in oupu per working-age person plus he growh in he employmen o working-age-populaion raio. Since he laer raio does no change subsanially among OECD

7 6 on a seady-sae growh pah wih a labor produciviy level beween 2.4 and 5 percen higher han he ypical OECD counry. In public R&D, only he higher educaion componen produces a posiive effec on labor produciviy bu is effec is no as grea as ha of business R&D. The effec of governmen inramural R&D on labor produciviy growh is no significan. However, governmen has by and large number of R&D relaed objecives o fulfill. The resuls in secoral analysis show ha he high-and medium-ech secors R&D is he main source of labor produciviy growh, whereas he R&D aciviies in low-ech and services secors are no significanly correlaed wih labor produciviy growh. In he high-ech secor, he posiive effec of R&D on produciviy is driven only by he size of he secor, no by he R&D inensiy in his secor. Neiher size changes nor R&D inensiy changes in medium-highech manufacurer affecs labor produciviy. Ineresingly, he R&D inensiy and size of medium-low-ech manufacurer affecs labor produciviy in opposie direcions: he R&D inensiy effec is posiive whereas he size effec is negaive. The role of foreign R&D spillover in raising domesic labor produciviy growh is posiive bu no robus. This lack of robusness migh be aribued o he fac ha foreign R&D spillover may no be measured adequaely wih he rade ransmission channel. Furhermore, i appears ha foreign R&D is correlaed wih public R&D and some secoral business R&D. Overall, once hose correlaions are aken ino accoun, he impac of foreign R&D becomes prominen in raising domesic labor produciviy growh. The res of he paper is organized as follows. In Secion II, we develop he empirical mehodology. The daa are described briefly in Secion III. In Secion IV, we discuss resuls for he relaionship beween labor produciviy growh and public and privae R&D. In his secion, counries, growh in oupu per working age populaion mimics very closely he growh in labor produciviy.

8 7 we also describe he resuls by decomposing he aggregae R&D expendiure ino five secors. We conclude in Secion V. II. Theoreical model and empirical mehodology We invesigae he relaionship beween various R&D indicaors and economic growh across 16 OECD counries by using he convergence-growh framework ha is now he sandard empirical approach for analyzing cross-counry aggregae daa. This empirical approach, based on he heoreical work of Mankiw e al. (1992) and Rober J. Barro and Xavier Sala-i-Marin (1992), is derived from he convergence propery of he neo-classical growh model and describes he dynamic adjusmen oward he seady-sae. The convergence propery saes ha he growh rae of labor produciviy (log difference) measured in efficiency unis of labor! Y i,, for counry i during period, is a funcion of he gap beween is seady-sae (log) level value Y a ime and is iniial (log) level i, 1 * i, * Y! as given by Yi,! ( Yi, Yi, " 1) # = " ". For a model wih yearly daa, "! is he annual speed of convergence oward he seadysae. The convergence propery is usually esed empirically wih he growh rae of labor produciviy being regressed on is iniial level and a se of conrol variables, Z i,, used as proxies for he seady-sae level, Y. Following Serge Coulombe and Frank C. Lee (1995) and Nazrul * i, Islam (1995), he adjusmen model of labor produciviy is ofen esed by pooling ime-series and cross-secion (TSCS) daa cross-secion being eiher cross-counry or cross-region. Wih he TSCS echnique, he error erm can be modeled in a wo-way error-correcion model wih ime dummies and fixed effecs: (1) Yi,! Yi, % 1 " Zi, # i $ ui, & = ,

9 8 where ui, is he idiosyncraic disurbance ha capures he effec of counry-specific shocks emporarily affecing he economy i during period. The fixed effecs! i capure he unobserved ime-invarian heerogeneiy across counries, such as he iniial level of echnology. 7 The use of ime dummies! in a TSCS empirical model implies ha all variables are ransformed using a cross-secional demeaned procedure. In his case, all common shocks o counries are deleed from he analysis. The firs resuls repored in his sudy come from various esimaions of he following specific formulaion of equaion (1): (R1) ( Y =! Y + " CC + " I + " N + " OPEN + " BR + i i' 1 c i I i n i o i, br i + " PR + " FR + # ( BR + # ( PR + # ( FR + $ G + % + & + u pr i fr i br i pr i fr i g i i. Since we are working wih an annual daa framework, a measure of cyclical componen, CC i,, has been added o conrol for shor-run business cycles. The nex hree variables used as proxies in i, Z are he logarihm of he invesmen o GDP raio, I i, ; he annual growh rae of he populaion, N i, ; and he logarihm of he rade (expor plus impor) o GDP raio, OPEN i,. Among he conrol variables, he firs wo, I i, and N, i, are he deerminans of he seady-sae level in he closed economy version of he Solow growh model; and he rade share, OPEN i,, is he usual proxy for openness in he rade and growh lieraure (Jonahan Temple, 1999). 8 And G is he German 1991 dummy. 7 The drawback of using fixed effecs in a TSCS framework of his ype is ha poenial conrols, i Z, which are ime invarians or change slowly hrough ime such as a democracy or a rule of law indexes, canno be used in he regression. 8 In (non-repored) regressions, we used he log of average years of schooling, a measure of human capial, based on Angel de la Fuene and Rafael Doménech s (2002) schooling daa as addiional conrol. The poin esimaes of he human capial variable were no close o being significan. Furhermore, he inroducion of he human capial variable modifies only marginally he poin esimae and he saisics of he oher variables. These negaive resuls regarding he effec of he schooling indicaor on labor produciviy across OECD counries concur wih he findings

10 9 The main conribuion of his paper is he consideraion of a variey of R&D indicaors in he Z i,. In he regression se-up given by (R1), we focus on separaing he effecs of public and privae R&D expendiure o GDP raios on labor produciviy growh. To his end, we use BR i,, he log of business performed R&D expendiure o GDP raio, and hree alernaive measures of public R&D o GDP raios, PR i,, namely, public R&D in higher educaion, governmen R&D, and finally he sum of he wo. Furhermore, along he line of an open economy model, we also include an index of foreign R&D expendiure o foreign GDP raio, FR i,, o conrol for inernaional R&D spillovers. The Coe and Helpman (1995) sudy riggered a subsanial number of sudies which ry o explain he relaionship beween foreign R&D capial sock and domesic TFP growh. In his effor, some researchers use inernaional rade as a ransfer mechanism of foreign R&D spillover, as did Coe and Helpman (1995); ohers use foreign direc invesmen, while ye ohers use boh (Keller, 2004, provides a deailed survey of foreign echnology diffusion). In his paper, we use rade flows as a ransmission mechanism of foreign R&D. As shown in Appendix 1, we use he share of impors normalized by foreign counry GDP as weigh o aggregae he foreign R&D of 16 OECD counries. Finally, in he spiri of an ECM, we have also included he firs difference of all R&D variables in he se of conrols. 9 of Coulombe e al. (2004). They argue for using lieracy es scores raher han average years of schooling, sressing ha lieracy es scores are more comparable across counries han years of schooling. Unforunaely, he lieracy daa are no available for he se of 16 OECD counries used in his paper. As a resul, human capial is no properly conrolled in our model. To he exen ha R&D and human capial are complemens, he esimaed effecs of R&D on labor produciviy growh may be overesimaed. 9 Before proceeding furher, one echnical commen migh be useful here. Noe ha he poin esimaes and - saisics of BR, PR and FR variables will be he same wheher hey are used in lagged forms or conemporaneously as long as heir firs differences are also included in he lis of regressors. Consequenly, adding he firs differences in he lis of regressors is useful in ha we do no have o es for various lag srucures for he level of R&D variables.

11 10 In his modeling, he shor-run dynamics of R&D spending are capured by he ˆ! parameers, and he long-run level effecs of he various R&D measures and he oher seadysae deerminans on he level of labor produciviy are capured by #! ˆ " ˆ. I is imporan o noe a his poin ha, as in any oher dynamic regression se-up of his ype, as long as " 1<! < 0, he model converges o a seady-sae regime where he growh rae of labor produciviy in all counries is equal o he common growh rae of echnological progress. 10 The second empirical model used in his paper is a varian of (R1) and is designed o j differeniae he effec of various secoral business R&D expendiures o GDP raio, BR i, : (R2) 5 j j i! i' 1 " c i " I i " n i " o i, )" br i, j= 1 ( Y = Y + CC + I + N + OPEN + BR 5 j j frfri ) br BRi, fr FRi ' 1 gg i ui j= 1 + " + # ( + # ( + $ + % + & +, where j indexes he five secors: high-ech manufacurers, medium-high-ech manufacurers, medium-low-ech manufacurers, low-ech manufacurers, and services. In his se-up, he public R&D variables are wihdrawn from he conrol lis, allowing us o expand he sudy period, from 1973 o The esimaion of he above equaion allows us o undersand which secors are more imporan in labor produciviy growh. However, since he conrol variable is he secoral R&D o aggregae GDP, i does no ell us wheher he secoral effec is driven by secoral R&D inensiy or he size of he secor. To undersand his decomposiion, we esimae a slighly differen equaion given in (R3): 10 Bu his does no mean ha he long-run level of labor produciviy will be he same across counries: each counry will converge o is own seady-sae level. The relaive long-run levels of labor produciviy across counries could be differen as hey are deermined by heir respecive Z i,. Convergence implies ha, saring from an iniial seady-sae siuaion, a shock o one of hese conrol variables exers a emporary effec on he growh rae of labor produciviy during he ransiion o he new seady-sae, and a long-run effec on he relaive level of labor produciviy.

12 11 (R3) 5 5 j j j j i! i' 1 " c i " I i " n i " o i, )" ri i, )" s i, j= 1 j= 1 ( Y = Y + CC + I + N + OPEN + RI + S 5 5 j j j j frfri ) ri RIi, ) s Si, fr FRi ' 1 gg i ui j= 1 j= 1 + " + " ( + " ( + # ( + $ + % + & +, The only difference beween (R2) and (R3) is ha in he laer, he variable BR he log of indusry level business R&D o GDP raio is decomposed ino log of indusry business R&D o indusry value added raio (RI) and he log of indusry value added o oal GDP raio (S). More, specifically, ln( ) ln( ) i, BR = r Y! r y y Y! RI S, where r i denoes business j j j j j j j j j i, i, i, i, i, i, R&D in secor i, y i denoes value added in secor i, and Y is oal GDP. In his specificaion, RI measures he parial effec of R&D inensiy, for a given size in erms of value added for he secor, and (S) capures a pure secoral-size effec, for a given secoral R&D inensiy. Regression models (R1) and (R2) are esimaed using various TSCS echniques. To ackle he heeroscedasiciy problem an imporan issue in TSCS analysis we presen resuls from ieraed feasible generalized leas-square (FGLS) esimaions wih cross-secional weighs. Wih his echnique, we also repor consisen sandard errors (HCCME) ha are robus o he remaining ime-series heeroscedasiciy. We also repor resuls from seemingly unrelaed regression (SUR) esimaions. SUR, which is a Parks esimaor, is he leas resriced TSCS esimaion echnique as he residuals are assumed o be boh cross-secional heeroscedasic and conemporaneously correlaed. Parks esimaions, however, are also known o poenially produce sandard errors ha lead o exreme confidence, paricularly when he number of ime series is no much larger han he number of cross-secions (Nahaniel Beck and Jonahan N. Kaz 1995). For example, in all regression specificaions dealing wih model (R1) for which he sample period was resriced o due do daa availabiliy on public R&D, SUR esimaions generaed high -saisics for mos poin esimaes compared o FGLS.

13 12 We also repor resuls of (R1) and (R2) using sysem esimaion wih insrumenal variables (IV). Firs, in all IV esimaions, he log level of labor produciviy, lagged by wo years (Y i,-2 ), has been included in he lis of insrumens. This procedure is known o decrease he propensiy o overesimae he convergence speed due o measuremen error in his ype of model (Barro and Sala-i-Marin, 2004) and o reduce Sephen Nickell s (1981) bias associaed wih fixed-effec esimaions of dynamic TSCS models. We also use IV echniques o surmoun he problems of endogenous explanaory variables. To his end, we have included one-year lag of populaion growh, invesmen o GDP raio, and openness variables in he lis of insrumens. We repor he resuls of wo differen sysem esimaions wih IV. The firs se of resuls comes from ieraed weighed wo-sage leas-squares (2SLS) esimaion, which is he IV sysem equivalen o FGLS esimaions. The second se is from hree-sage leas-squares (3SLS) esimaion, which is he IV esimaor analogue o SUR. I hus has he same problem as SUR regarding overconfidence since i is also a Parks esimaor. III. Daa Mos of he daa used in his sudy come from he OECD (a more deailed accoun of he daa is given in Appendix 1). The aggregae R&D daa are from he Main Science and Technology Indicaors daabase of he OECD. The daa series we used are in 2000 consan prices and in purchasing power pariies (PPPs). The indusry level R&D daa are from he Analyical Business Enerprise R&D (ANBERD) daabase of he OECD. These curren-price PPP daa for each counry have been convered o consan-price PPPs by dividing hem by heir own counry s GDP deflaor (calculaed as a raio of curren and consan-price GDP for each counry from he OECD daabase).

14 13 All indusry-level daa have been compued from he Inernaional Sandard Indusrial Classificaion, Revision 3 (ISIC Rev. 3) and Revision 2 codes. Revision 3 has more indusry disaggregaion compared wih he earlier version (Revision 2). Since he indusry codes in Revision 2 and Revision 3 differ, we use a concordance developed by OECD o conver he Revision 2 ino Revision 3 daa. Finally, he manufacuring indusry daa are aggregaed ino four ypes of manufacurers (high-ech, medium-high-ech, medium-low-ech and low-ech) using he echnology classificaion (given in Appendix 2) adoped by OECD. We also sudy aggregae services as a separae secor. The labor produciviy, populaion growh and invesmen inensiy (invesmen o GDP raio) daa are aken from Penn World Table (Alan Heson e al., 2002). For labor produciviy, we use he series Real GDP chain per worker a 1996 prices. Our measure of he business cycle correcion, GDP gap in percenage, is obained from quarerly real GDP daa using Hodrick and Presco s filer wih a smoohing parameer of The raw daa are GDP volume a 2000 consan PPP from OECD (Economic Oulook Quarerly daabase) for 32 years (1963 o 2004). The quarerly cyclical componen was hen annualized o fi ino our sample framework. The daa on inernaional impors are from he Bilaeral Trade daabase of OECD. For each counry, we used impor daa o calculae he impor shares of all individual counries in he sample. These shares were hen used as weighs o aggregae foreign business R&D expendiure (as shown in Appendix 1). Since impor daa are available only from 1980, for he years 1973 o 1979, we use he average impor shares of 1981 o 1985 as weighs. The impor shares are compued using daa in curren-price U.S. dollars.

15 14 IV. Resuls Resuls for he regression se-up (R1) are repored in Tables 1 o 3 for hree alernaive measures of public R&D where he sample runs from 1981 o Equaion (R1) was also esimaed by dropping he variables on public R&D and exending he sample period from 1973 o 2000 and he resuls are repored in Table 4. Resuls in Table 5 come from he (R2) regression se-up in which he sample period is as in Table 4 and public R&D variable is also aken ou, however, he business R&D variable has been divided ino five secors. Finally, resuls for he regression seup (R3), in which he sample is limied from 1979 o 2000 and secoral R&D share in aggregae GDP is decomposed ino secoral R&D inensiy and size, are repored in Table 6. In he firs subsecion below, we discuss he resuls of Tables 1 o 4 followed by he resuls of Tables 5 and 6 in he nex subsecion. IV.A. Regressions wih privae and public R&D Table 1 refers o he (oal) public expendiure in R&D (PERD), which is he sum of he governmen expendiure in R&D (GOVERD) and higher educaion expendiure in R&D (HERD). Resuls in Table 2 (Table 3) refer o he regressions for which HERD (GOVERD) is used as he measure of public R&D. The resuls for he four non-r&d variables (lagged labor produciviy, he business cycle, invesmen and openness) are consisen and robus in all hree ables and four specificaions/esimaion echniques (FGLS, 2SLS, SUR, 3SLS). Furhermore, heir poin esimaes have he expeced sign: a negaive fracion for lagged produciviy level, indicaing condiional convergence; posiive for he business cycle, invesmen inensiy and openness variables. The various poin esimaes for he lagged produciviy variable imply ha he

16 15 esimaed convergence speed varies beween 6.7 and 11.7 percen. This implies ha he economy moves halfway o seady-sae in abou 6 o 12 years. 11 More imporanly, he posiive effec of he domesic business expendiure in R&D (BERD) variable is ulra robus in his sudy. In he 12 repored regression resuls dealing wih privae and public R&D variables (all specificaions in Tables 1 o 3), he poin esimaes of he BERD variable are posiive and significan a he 1 percen level. On quaniaive grounds, he long-erm elasiciies of hese poin esimaes for his variable vary beween 0.24 and Thus, according o hese resuls, a counry wih a business R&D inensiy 10 percen higher han he ypical OECD counry has, in he long run, a labor produciviy level beween 2.4 and 5 percen higher. This findings imply ha, on average, if a counry increases is business R&D inensiy (percen of R&D in GDP) by 0.1 percen, i will increase is per capia income by 0.9 o 1.8 percen. I works as follows: wih given average R&D inensiy of 1.6 percen in sample counries, a 10 percen increase in R&D inensiy resuls in an increase of R&D by 0.16 percen of GDP. Thus, an increase of R&D by 0.1 percen of GDP increases labor produciviy by 1.5 (= 2.4/0.16 * 0.1) o 3 percen. Considering average employmen o populaion raio of 0.6, he effec will be an increase in per capia income of approximaely 0.9 o 1.8 percen. This range of elasiciies falls wihin he range of some earlier esimaes. For example, summarizing several sudies, Nadiri (1993) noes ha he elasiciies of oupu wih respec o business R&D is in he range of 8 o 30 percen. The OECD (2003) sudy esimaes he elasiciy of real oupu per working-age populaion o business R&D o be 0.72, which is much higher han our resuls. Our esimaes of long-run labor produciviy growh wih respec o business 11 This speed of convergence seems raher fas compared wih cross-counry sudies based on pure cross-secion informaion, sudies such as Barro and Sala-i-Marin (2004). However, our esimaes of he convergence speed concur wih many ime-series cross-counry sudies ha have used fixed effecs. For example, Islam (1995) repors a convergence speed of 9 percen in he fixed-effec specificaion. The poin esimaes of populaion growh also have

17 16 R&D inensiy are also in line wih he elasiciies obained in augmened Solow-ype models. For example, in Nonneman and Vanhoud (1996), he implied esimaed elasiciy of real GDP per working-age populaion wih respec o R&D inensiy is I is no a sraighforward maer o compare he R&D elasiciies found in his paper wih hose esimaed using TFP growh (raher han labor produciviy growh) as he dependan variable. In a simple Cobb-Douglas producion funcion wih consan reurns o scale, we can show ha he average labor produciviy growh will be higher han TFP growh if he capial deepening is increasing and he share of labor force wih higher marginal produc is also increasing. 13 Since we expec his o be he case for our sample of OECD counries, we expec he average labor produciviy growh o be higher han TFP growh. However, in such a simple model, if we could conrol properly for he effecs of capial deepening and labor qualiy in he esimaion, hen he regression coefficiens in (R1) hrough (R3) can be inerpreed as an effec on TFP growh, despie he use of labor produciviy growh on he lef-hand side. However, he real world is cerainly differen from he simple exbook-ype model ha we have described: he relaionship beween labor produciviy and TFP growh is far from linear. Furhermore, he crucial assumpion of consan reurns o scale needed for his expression may no hold, furher complicaing he relaionship beween labor produciviy and TFP growh. he negaive sign, as expeced from neo-classical growh models. This resul, however, is no very robus since i is no significan a he 10 percen level wih FGLS and 2SLS in Table This number is compued from Table IV, upper window, las column, of Nonneman and Vanhoud (1996) who do no repor long-run elasiciies. 13! " Le he producion funcion be given by Cobb-Douglas specificaion such ha Y = AK L, where Y is oupu; K is physical capial; and L is number of workers. Wih he assumpion of consan reurns o scale, we can conver his funcion ino labor produciviy per hour by dividing boh sides of he equaion by hours of works, H. Then aking log difference, he per hour oupu growh funcion can be wrien as # ln Y H =!# ln K H + "# ln L $ H + # ln A H. This shows ha he per hour labor produciviy growh ( ) ( ) ( ) ( ) depends on hree facors. The firs erm on he righ-hand side is capial deepening, he capial services per hour. The second erm is he improvemen in labor qualiy, defined as he difference beween growh raes of labor inpu and hour worked. If he hours of work by workers wih higher marginal produc rises, i raises he overall labor

18 17 In his case, even a proper conrol of capial deepening and labor qualiy may no make TFP growh equivalen o labor produciviy growh. Besides, because of daa limiaion, we canno conrol all oher facors ha poenially affec labor produciviy growh. Hence he resuls derived using TFP and labor produciviy growh are no direcly comparable. More imporan, in he models ha use TFP as a dependen variable as he direc impac of business R&D on oupu is already a leas parly accouned for in TFP he elasiciies mus capure mainly spillovers and possibly exra reurns (coming in addiion o normal remuneraion of capial and labor) arising from R&D. However, elasiciies in our model more likely capure he oal effec (boh direc and spillover) of R&D. Thus i is quie reasonable ha our elasiciies are higher han hose esimaed in models wih TFP growh as a dependen variable. The OECD sudy by Guellec and van Poelsberge de la Poerie (2001) finds he longerm elasiciy of TFP growh o be 0.13 wih respec o business R&D. Using his number as a norm for spillovers, our long-erm elasiciies range of 0.24 o 0.50 means ha he direc effec of R&D is in he range of 0.11 (slighly smaller han spillover effec) o 0.27 (more han double spillover effec). For R&D carried ou by he public secor, here is evidence of a posiive effec on labor produciviy only in he case of HERD. The effec is significan a he 5 percen level in hree esimaion echniques and is significan a he 10 percen level only in 3SLS. 14 The long-run quaniaive effec of HERD is no esimaed wih grea accuracy since i is abou four imes produciviy. The las erm is TFP growh, which increases labor produciviy on a poin-for-poin basis (see Dale W. Jorgenson and Kevin J. Siroh [2001] for similar expression). 14 As for he R&D carried ou in he public secor, i is imporan o noe a he sar ha he HERD variable is highly posiively correlaed wih BERD. In a TSCS framework, his is bes illusraed by he resuls of a simple pooled leas-squares bivariae regression where BERD is regressed on HERD (wihou a consan since boh variables are expressed as a deviaion from he cross-secional sample mean). In his es, he slope parameer of he higher 2 educaion variable was 1.03 wih a -saisic of 18.8 and R of The correlaion beween GOVERD and BERD was much weaker. In he same ype of bivariae regression (wih BERD regressed on GOVERD), he poin esimae 2 of he slope parameer was 0.11; he -saisic was 2.23, and R was close o zero.

19 18 smaller (0.06 versus 0.24) when i is esimaed wih SUR and 3SLS compared wih FGLS and 2SLS. Neverheless, he long-erm elasiciy of governmen- and universiy-performed research on TFP of around 0.17 esimaed by Guellec and van Poelsberge de la Poerie (2001) falls wihin our wide range of esimaes. The poin esimaes and he sandard errors of he higher educaion variable change only marginally when he business R&D variable is dropped from he regression. This poin is worhy of aenion since as menioned earlier, higher educaion and business R&D are highly correlaed. The hypoheses ha he PERD (Table 1) and GOVERD (Table 3) have a posiive effec on labor produciviy are rejeced by he resuls. Tables 1 and 3 show ha he poin esimaes for hese wo public R&D variables are even negaive and significan a he 1 percen level wih he SUR and 3SLS esimaions. These resuls concur wih he recen OECD sudy (2003) ha repors a negaive and significan esimae for he coefficien of public R&D in per capia oupu growh regression wih he same ype of esimaion echniques (Table 2.6, Column 2). However, wih FGLS and 2SLS specificaions, he poin esimaes for PERD and GOVERD alhough insignifican are posiive. In our view, we should no give much weigh o SUR and 3SLS resuls when hey conflic wih hose in FGLS and 2SLS. Parks esimaions ha correc for he conemporaneously correlaed errors in TSCS models (such as SUR and 3SLS) are known, as discussed earlier, o someimes underesimae he sandard errors dramaically and lead o overconfidence. Accordingly, we believe ha i is more appropriae o conclude ha he hypohesis of a posiive effec of public R&D variables, oher han universiy educaion, is rejeced by he daa raher han o pu he emphasis on he nonrobus negaive effec of public R&D We also esimaed (R1) using GOVERD and HERD as wo separae variables (insead of using he sum, PERD, as done in Table 1). In his specificaion only HERD was posiively significan.

20 19 Our resuls regarding he effec of public R&D, however, are in conras o hose of Guellec and van Poelsberge de la Poerie (2001) where hey repor a posiive and significan effec for public R&D on TFP growh using SUR and 3SLS esimaes. While duplicaing heir empirical sudies using heir daa bank, we found ha heir repored resuls regarding public R&D were no robus wih FGLS and 2SLS. Therefore, he significance of heir esimaes regarding public R&D migh well be aribued o he endency of Parks esimaors o provide overconfiden resuls. Finally, he poin esimae of he foreign R&D indicaor is posiive hroughou all specificaions in Tables 1 o 3. I is significan, however, only wih SUR and 3SLS esimaions in Tables 1 and 2 only. The elasiciy of labor produciviy wih respec o foreign R&D varies significanly across various esimaions; whenever he foreign R&D variable is significan, is elasiciy is in he range of 0.32 o This is in line wih Guellec and van Poelsberge de la Poerie (2001) who esimae he long-erm elasiciy of TFP growh wih respec o foreign R&D o be in he range of 0.45 o In heir models, unlike ours, he foreign R&D variable is significan in all specificaions. Again, in duplicaing heir resuls, we found ha he significance of heir variable did no survive o FGLS and 2SLS esimaions. Neverheless, he insignificance of foreign R&D variable in 8 of he 12 cases and in all of our preferred esimaion echniques of FGLS and 2SLS is unexpeced. Having said ha, we would like o sress ha he insignificance of he foreign R&D variable may no necessarily mean ha here are no spillover benefis from foreign R&D capial. Wha i means is ha he spillovers of foreign R&D may no be compleely ransmied hrough impors, which is no new 16 Since he foreign R&D is no accouned for in GDP, even if we have labor produciviy on he lef-hand side, we will be esimaing spillover as if we are using TFP as dependen variable. This is because he domesic opporuniy cos of foreign R&D is zero.

21 20 in he lieraure. 17 I may also be due o he fac ha, alhough foreign R&D has spillovers ha pass hrough impors o domesic economy, he es fails o show his resul as he foreign R&D variable is correlaed wih oher conrol variables used in he model. (This issue will be discussed more in he following subsecion.) By dropping he public R&D variable from he lis of conrols, we are able o exend he sample o he 1973 o 2000 period. The resuls from his exended sample, displayed in Table 4, illusrae he robusness of he impac of business R&D. The poin esimae of business R&D is always posiive and significan a he 1 percen level. Furhermore, he long-erm elasiciies of labor produciviy o business R&D falls beween 0.28 and 0.53, a range very similar o he one esimaed wih he shorer sample. The poin esimaes of he foreign R&D indicaor are posiive, bu he effec is no ha robus as he poin esimaes are no significan a he 5 percen level in hree ou of four cases. As for he non-r&d variables, hree variables (lagged labor produciviy, he business cycles and he openness) display he same robusness as in Tables 1 hrough 3. They all have he expeced sign and are significan a he 1 percen level wih all four specificaions. However, he effec of he remaining non-r&d variable, invesmen o GDP raio, is significan a he 1 percen level only wih FGLS and SUR. In IV regression, i is significan only a he 10 percen level wih 2SLS esimaions, and he -raio is only 0.21 wih 3SLS. These resuls illusrae a well-known fac in growh empirics: he correlaion beween invesmen and growh does no 17 In heir sudy, Coe and Helpman (1995) find a posiive and qualiaively large effec from impor-weighed foreign R&D, where 1 percen increase in R&D capial sock in he Unied Saes raises he average produciviy of 22 OECD counries by abou 0.12 percen. However, Keller (1998) shows ha, o generae Coe and Helpman s resul, using impor share o aggregae foreign R&D variable (implying ha impor is he spillover channel) as was done in Coe and Helpman (1995) is no essenial. Specifically, Keller uses randomly creaed shares in place of acual bilaeral impor shares o creae he counerfacual foreign knowledge sock. Using his alernaive foreign R&D variable yields similarly high coefficiens. Given ha impor shares are no essenial o obain Coe and Helpman s (1995) resuls, heir analysis does no allow us o draw srong conclusions regarding he imporance of

22 21 always survive he use of insrumenal variables (Temple 1999, p. 137). Given he poenial endogeneiy problem, when invesmen raio is correlaed wih he error erm, he resuls of IV esimaions should be preferred. The nonsignifican or weakly significan effec of his variable wih he IV mehod somehow empers he robusness of a posiive effec of invesmen inensiy in labor produciviy growh. IV.B. Decomposing he effec of business R&D We now urn o he secoral analysis wih resuls presened in Table 5 using he same sample period of 1973 o 2000 as in Table 4 and decomposing he business R&D variable ino five secors: high-ech-manufacurers (HTM), medium-high-ech-manufacurers (MHTM), medium-low-ech-manufacurers (MLTM), low-ech-manufacurers (LTM) and services. The decomposiion appears very useful since i yields hree differen ses of resuls for echnologically differen secors. Firs, for he business R&D in high-ech and medium-low-ech, he poin esimaes are all posiive and significan a he 1 percen level. Furhermore, he long-erm elasiciies appear o be esimaed relaively precisely across he various esimaion echniques since hey vary beween 0.13 and 0.16 for he high-ech manufacuring and beween 0.10 and 0.19 for he medium-lowech. Second, he poin esimaes of he medium-high-ech manufacurers are all posiive, bu he resul is no robus. The poin esimaes are no significan a he 5 percen level wih FGLS and 2SLS bu he p values in his case are in he neighborhood of 10 percen. The esimaed longerm elasiciies for his secor s business R&D are, however, in he same range as hose for he impors as a vehicle for diffusion. Similarly, Griffih e al. (2005) also find only a small effec of rade in TFP growh (saisically significan only a he 10% level).

23 22 high-ech and he medium-low-ech secors. In his secor oo, overall here is some evidence (bu no ha robus) in favor of a posiive effec of his R&D componen on labor produciviy. Third, he null hypohesis of a posiive and significan effec for he las wo componens, low-ech manufacurers and services, is clearly rejeced by he daa. Again, he negaive sign is significan even a 1 percen for he low-ech secor wih SUR and 3SLS. Bu we believe ha giving oo much imporance o hese resuls is misleading, given he endency of Parks esimaions o creae overconfidence in rejecing he null. Hence, he robus posiive impac of aggregae business R&D on labor produciviy in Table 4 should have been driven mainly by business R&D in high-ech and medium-low-ech manufacurers, and less so (if any) by medium-high-ech manufacurers. I seems somewha counerinuiive ha business R&D in medium-low-ech manufacurers is more effecive for labor produciviy han medium-high-ech manufacurers. However, i should be borne in mind ha his resul may no necessarily mean ha he rae of reurn o R&D in medium-low-ech manufacuring is higher han in medium-high-ech. Finally, he poin esimaes of foreign R&D have improved in he longer ime period model compared wih hose in he shorer period (resuls in Tables 4 o 5 versus resuls in Tables 1 o 3). Loosely speaking, hey improved afer we excluded he public R&D from he esimaion. I improves furher once we decompose he R&D ino five secors; he coefficiens of foreign R&D are now posiive and significan a leas a he 5 percen level in all four regressions. Nex, we would like o see wheher he resuls in Table 5 are driven by he R&D inensiies or by he pure size (indusrial srucure) effecs. For his purpose, we esimae he specificaion (R3) by limiing he sample from 1979 o 2000, as secoral value added daa prior

24 23 o 1979 were no consisenly available across counries. As he variables populaion growh raes, German dummy and he R&D in LTM secor were no significan, we dropped hem from he conrol and repored only he resuls on long-run elasiciies in Table 6. The resuls show ha he size effec in he HTM secor increases labor produciviy. A counry wih a HTM secor 10 percen larger han he ypical OECD counries ends up on a seady-sae growh pah wih a labor produciviy level of 2.6 o 3.4 percen higher. The impac of R&D inensiy in he HTM secor is no robus; i is posiively significan wih FGLS bu insignifican wih SUR. The relaive effeciveness of he size variable in he high-ech secor can resul from he fac ha he R&D inensiy (per uni of value added) is already very high in his secor compared wih ohers. Hence, mos poenially successful R&D opporuniies in his highly innovaive secor has previously been exploied in which case increasing he rae of R&D aciviy will no be ha profiable. However, increasing he number of firms or increasing firms size opens avenues for new producive R&D aciviies. Ineresingly, we ge he reverse resul in he case of he MLTM. In his secor, he overall posiive and significan effec of R&D spending (Table 5) is more han driven by he pure size effec as he parial effec of secoral inensiy is negaive (Table 6). These resuls sugges ha, in his less innovaive secor, some poenially producive R&D aciviies are no fully exploied. Increasing he size of he secor, leaving he R&D secoral inensiy consan, exers a negaive impac on produciviy growh. By conras, increasing R&D aciviy simulaes growh on impac and increases he produciviy level in he long run. The decomposiion of he oal effec ino a size and a secoral inensiy does no add more insigh o he case of MHTM, as he

25 24 poin esimaes are no significan. This resul concurs wih findings for he overall effec of his secor in he shorer sample. 18 Finally, for he services secor, boh increase in R&D inensiy and size lower labor produciviy growh. The resuling negaive size effec migh be inerpreed as evidence in favor of a Baumol s disease (William Baumol, 1967). However, noe ha in he longer sample in Table 4, services secor s R&D inensiy wih respec o overall GDP has no significan impac on labor produciviy. The conradicory resuls in Table 6 wih shorer sample, does no allow drawing any conclusions regarding he impac of R&D aciviies in he services secor. V. Conclusion The main lesson from sudies focusing on he deerminan of growh in a broad se of counries is ha rich counries, unlike poor counries, have mos of he fundamenal deerminans righ. Bu dispariies in growh performance among rich counries are also observed. The obvious quesion is: Wha can we learn from cross-counry sudies o help orien policy so rich counries, eager o close he gap wih he leaders, can do so? Our empirical invesigaion is an aemp o answer his imporan quesion. On heoreical grounds, R&D is considered a good candidae o accelerae growh as i could bring echnological progress and also increase he absorpive capaciy of he domesic economy o capure spillovers from foreign R&D. The conribuion of our paper, which conforms wih his view, is ha on empirical grounds, mos R&D componens, bu no all, are posiively and significanly correlaed o labor produciviy growh. The firs poin of ineres arising from his sudy is ha he growh effec of business R&D, unlike public R&D, is posiive and srongly significan in all specificaions. Resuls indicae ha a counry wih a business R&D inensiy 10 percen higher han he ypical 18 See noe o Table 5.

26 25 OECD counry ends up on a seady-sae growh pah wih a labor produciviy beween 2.4 and 5 percen higher. This esimaed elasiciy in he range of 0.24 o 0.50 is subsanially larger han he 0.02 share of business R&D in he GDP, an indicaion of a srong spillover effec of business R&D. On he oher hand, among public R&D, only expendiure in higher educaion affecs labor produciviy posiively and significanly; he effec of governmen R&D is a bes nil. Furhermore, higher educaion R&D is highly and posiively correlaed wih business R&D. A. key poin of ineres in our sudy is ha business R&D in differen secors does no enhance growh in he same way in every secor. We found ha high- and medium-ech secors R&D are he main source of labor produciviy growh. However, in he medium-ech secor, i is mainly he medium-low-ech area raher han he medium-high-ech ha generaes he sronger effec on labor produciviy. Finally, in he low-ech manufacuring and service secors, R&D aciviies are no significanly correlaed wih growh. This indicaes ha he reurn on R&D aciviies migh differ across secors because he direc reurn differs, he spillover effecs differ, or boh. I is enirely possible ha he knowledge generaed hrough R&D aciviies is more non-rival (and be useful for many oher secors of he economy) in he high-ech secor han in he service secor and low-ech manufacuring. I is also possible ha he high- and medium-ech manufacures are disproporionaely used more as inpus ino oher indusries han he low-ech manufacurers and services. If his is he case, R&D and he corresponding echnological improvemens in he high- and medium-ech secors reduce coss a a faser rae (hereby increasing produciviy) han hose in low-ech and services. Whaever he reasoning for he varied impac of R&D in he differen secors (some secors have negaive effecs on aggregae labor produciviy growh), his paper is able o measure more precisely he posiive impac of R&D by esimaing R&D s impac by secor.