R&D Intensity, ICT Investment and Industry Productivity Growth*

Transcription

1 R&D Inensy, ICT Invesmen and Indusry Producvy Growh* Theo S. Echer Deparmen of Economcs, Unversy of Washngon Ifo Insue for Economc Research a he Unversy of Munch Thomas Srobel Ifo Insue for Economc Research a he Unversy of Munch Deparmen of Economcs, Unversy of Munch Aprl, 2009 Absrac Ths paper akes a sep oward denfyng he ulmae drvers of producvy growh a he ndusry level. Based on Neoclasscal and New Growh Theory we develop an emprcal mehodology ha allows us o es for he fundamenal drvers of ndusry producvy growh. We presen wo ses of ndusry classfcaons, one based on ICT nensy and anoher based on nnovaon ypes. Usng OECD daa we show how smlar he conrbuons of ICT-Inensve or Scence-Based ndusres are o value-added. However, he dsrbuon of R&D n hese ndusres across OECD counres s shown o vary dramacally. Our panel daa resuls confrm ha for boh ypes of ndusry classfcaons he source of srong capal deepenng conrbuons o producvy growh are ndeed o be found n ndusry-level R&D. Bu no any ndusry s R&D. We show ha s eher ICT-Producng ndusres R&D or Scence-Based Innovang ndusres R&D ha are he crucal deermnan of producvy growh. JEL Classfcaon: O3, O4 Keywords: Growh accounng, ndusry producvy analyss, research and developmen, nformaon and communcaon echnology, panel daa. * Ths paper s based on a book chaper forhcomng n Echer and Srobel (2009). We hank Elke Kronjäger for preparng he Ifo ndusry-level asse-specfc nvesmen daa, and he German Scence Foundaon Gran # for fnancal suppor.

2 Echer and Srobel 2 1 Inroducon The hallmark of he 20 h cenury has been rapd nnovaon and echnologcal progress. The evens mached he mplcaons of he Solow (1956) model, whch predced ha susaned long-run growh of per capa GDP s mpossble n he absence of echnologcal progress. A new srand of he leraure, branded as The New Growh Theory (Romer, 1990, Aghon and How, 1992, Grossman and Helpman, 1991), hghlghed he forces ha generae echnologcal change whn he economy o produce susaned long-erm growh. The key nsgh was ha susaned growh requres ever more effcen use of avalable resources, and ha hs ncrease n effcency s ulmaely drven by ndusry research and developmen (R&D). The New Growh Theory oulnes exacly how echnologcal progress s drven by produc and process nnovaons ha are underaken by frms ha seek o manan her compeve edge n a marke economy. Frm R&D effors preven a declne of he margnal produc of capal, as ever newer echnologes are emboded no cung-edge capal socks. Snce Solow (1957) nroduced growh accounng, a debae has raged as o how much of economc growh can be arbued o facor accumulaon or echnologcal change. In he prevous chapers we have followed he growh accounng leraure closely by focusng on facor accumulaon. In hs chaper we drll one level deeper and aemp o solae he fundamenal facors ha may have been he orgns for specfc ndusry paerns of facor accumulaon. Based on New Growh Theory mplcaons, our hypohess s ha such ncreases n effcency and accumulaon orgnaed wh R&D. The purpose of hs chaper s hen o denfy he separae nfluences of R&D and facor accumulaon on producvy growh. A he aggregae (naonal) level, he lnk beween R&D and growh are well documened (Jones, 1995, Dnopoulos and Thompson, 2000, Zacharads, 2004, Agnger and Falk, 2005). However, a he ndusry level, daa analyss s usually hampered by he absence of qualy daa on nvesmen, R&D and producvy. In prevous chapers, we oulned how crucal s o undersand ndusry-level dynamcs o hghlgh he drvers of naonal producvy rends. Our focus n hs chaper s herefore on he se of key ndusres ha have prevously been denfed as he major sources of producvy growh. We focus especally on he R&D dynamcs n nformaon and communcaon echnology ndusres (ICT). There s a wde spread noon ha he emergence of he 2

3 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 3 New Economy n he md 1990s was crucally drven by nnovaons n ICT secors. These nnovaons generaed dramac prce decreases and acceleraed effcency gans a unprecedened raes. Bu neres n hese ndusres s also based on heorecal foundaons. New Growh Theory developed he erm general purpose echnology (GPT), whch denfes echnologes ha rgger complemenary nnovaons and effcency gans across a broad range of secors. Already n 1995, Bresnahan and Trajenberg (1995) suggesed ha ICT echnology s a parcularly fng echnology o be labeled as GPT. Ths s due o ICTs unque ably o rgger complemenary nnovaons and R&D expendures n oher ndusres (see also Helpman and Trajenberg, 1998a,b). We wll use he concep of general purpose echnology o dssec he crucal oal facor producvy conrbuons ha are he promnen feaures n a counry s producvy make up. The conens of he resdual ha s usually observed n aggregae (naonal) producon funcons can hen also be clarfed, snce we characerze he economc phenomenon (R&D) ha s hough o gve rse o much of hs resdual. We consruc ndusry-level R&D socks, usng OECD mehodology oulned n Guellec and Poelsberghe (2001) and R&D daa provded by he OECD STAN R&D Daabase (2006). Usng growh accounng mehodology, we hen dssec he sources of producvy o examne f R&D conrbuons from specfc ndusres produced general purpose echnologes ha drove producvy growh. A dealed ndusry sudy s made possble by he Ifo Indusry Growh Accounng Daabase (IIGAD), whch provdes suffcen deal o oban a clear pcure across a fnely dsaggregaed se of 22 ndusres. 1 We also provde descrpve sascs for a broader se of counres o compare R&D nensy by ndusry ype across a larger se of OECD counres. The only OECD counres wh relevan R&D daa a suffcenly hgh level of dsaggregaon nclude France, Ialy, Span, Fnland, Sweden, Neherlands, and Denmark. 2 Our resuls confrm exacly ha R&D nensy has been he crucal drver of producvy growh n ICT-Inensve ndusres. Ineresngly, however, he GPT effec s weak, afer we conrol for possble endogeney, and only R&D n ICT- Producng ndusres s shown o possess GPT characer. An alernave classfcaon 1 For a dealed descrpon of he daabase see Roehn e al. (2007) or chaper 1. 2 To mach OECD counres R&D daa wh EUKLEMS Growh and Producvy Accouns (for ICT daa) reduces he level of ndusry dsaggregaon excessvely o render ndusry-level analyss raher meanngless. GPT characerscs a he ndusry level are masked by hgh levels of aggregaon.

4 Echer and Srobel 4 of ndusres, accordng o he ype of research underaken, ndcaes ha R&D n Scence-Based ndusres also conans srong GPT characer ha generaes enormous producvy gans. 2 Theorecal Underpnnng The emprcal model ress on he basc enans of R&D-based growh models ha were nroduced by Romer (1990) and exended o he open economy by Rvera- Baz and Romer (1991). The qunessenal R&D-based growh model bulds on he neoclasscal model (Solow 1956) where echnology s exogenous and oupu n ndusry s gven by [ K, L A ] Y = F,. (1) Oupu s denoed by Y, capal s K, Labor s L and echnology s A. Wrng he equaon n erms of labor producvy, y=y/l, he per capa capal sock, k=k/l, and assumng ha he producon process s homogenous o degree one, we fnd [ k A ] y = F,. (2) We can ake logs and dfferenae (2) o fnd he growh rae of ndusry producvy, yˆ = α kˆ + βaˆ (3) s a funcon of ndusry capal deepenng, k ˆ, ndusry echncal change, Â, and her assocaed srucural elasces. Capal accumulaon s assumed o be sandard, and gven by he excess of ncome over consumpon. The novel n he feaure of R&D-based growh models s he endogenous rae of echnologcal change. Specfcally, he growh rae of echnology s deermned explcly va dedcaed R&D ha s fnanced by prof maxmzng frms ~ ~ [, L A] A ˆ = G K,, (4) R&D expendures are hen gven by he cos funcon assocaed wh G[.] and deermned by he capal and labor socks allocaed o R&D, K ~ and L ~, respecvely. The dynamcs and he soluon o he general R&D-based growh seup descrbed by equaons (1) o (4) s provded n Echer and Turnovsky (1999). 4

5 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 5 Indusry R&D effors affec s own knowledge sock and ncreases he effcency of he ndusry s producon process. In addon, each nnovaon conrbues o he aggregae knowledge sock, A, ha evenually splls over o all ndusres. As specfed n (4), he aggregae sock of knowledge eners each ndvdual frms R&D producon funcon. Ths s he mechansm by whch GPT echnology n secor may conrbue o he effcency n oher ndusres. 3 Indusry Classfcaons Prevous chapers oulned ha dfferen ypes of ndusres generae dfferen raes of producvy growh snce 1990, especally when ndusres are grouped by he ICT ype. Below we accoun for such ndusry dfferences and nroduce wo ndusry classfcaons. Frs we presen resuls usng he convenonal classfcaon accordng o ICT nensy. Then we adop an ndusry classfcaon ha focuses on he nnovaon poenal of an ndusry. 3.1 ICT Taxonomy Accordng o he sandard ICT classfcaon, ICT-Inensve ndusres exhb ICT capal servces shares ha exceed he medan n he oal economy, see e.g. Sroh (2002, 2006) and Echer and Roehn (2007). The classfcaon of ICT-Inensve ndusres s hen furher dsaggregaed no ICT-Producng and ICT-Usng ndusres. ICT-Producng ndusres are defned by he German Federal Sascal Offce (GFSO, 2006). ICT-Usng ndusres are deermned ex pos as he resdual. A full ls of he ICT classfcaons by ndusry s provded n Table 1. Classfyng ICT-Inensve ndusres accordng o hs approach s crczed by Daver (2004) due o s cu-off level dependency of ICT-Inensve and Non-ICT- Inensve secors. The argumen becomes relevan for large marke servces, e.g. Wholesale and Real Trade, as oulned by Van Ark and Inklaar (2005). For nsance, wheher Real Trade s classfed ICT-Usng or Non-ICT-Inensve depends heavly on usng he mean or medan of oal economy ICT capal servces as cu-off level. An alernave approach o classfy ndusres s o ulze a daa-deermned cu-off level. Ths can be accomplshed by cluser analyss, whch aemps o deermne naural groupngs of ndusres. Cluser analyss uses dfferen algorhms and

6 Echer and Srobel 6 mehods for groupng ndusres of smlar knd no respecve axonomes based on meanngful srucures n he observed daa. Those approaches are separaed no herarchcal and k-means cluserng mehods, where herarchcal mehods comprse varous submehods, lke e.g. average lnkage, sngle lnkage, or complee lnkage mehods. Neverheless, cluser resuls are remarkably sensve o small changes n he daa and o he dverse cluserng echnques (see e.g. Inklaar, 2005). Evenually, here exss no clearly defned way of how o group ndusres accordng o ICT ulzaon. Snce we wan o keep n lne wh prevous sudes on producvy growh for Germany and oher OECD counres, whch employ ICT- Inensve-ype classfcaons and snce he selecon of approprae cluser mehods s no less bu even more arbrary han Sroh s hreshold measure, we absan from an ICT-Inensy classfcaon derved by cluser mehods. 3.2 Innovaon Taxonomy Our second ndusry classfcaon sysem s based on nnovave acvy and R&D ype. I follows Pav (1984) and Van Ark and O Mahony (2003), who separae goods-producng ndusres no four caegores ha reflec he dversy of R&D ypes (produc /process R&D), marke srucures (prce and performance sensvy), and means of appropraon (paens, lcensng, ec). 3 Those four caegores comprse: Scale-Inensve Indusres (SII), Suppler-Domnaed Goods-Producng Indusres (SDG), Specalzed Goods Supplers (SGS), and Scence-Based Innovaors (SBI). Dealed goods-producng ndusres as classfed accordng o he nnovaon axonomy are lsed n Table 1. The frs group n hs classfcaon refers o SII secors ha use large-scale assembly producon processes. R&D n hese ndusres focuses on process nnovaons ha faclae capal-labor subsuon and reduce producon coss. The naure of SII ndusres s n process nnovaons locaed n-house R&D deparmens (e.g. process engneerng specals groups) or R&D ha s delvered by upsream supplers. Typcal SII ndusres are Vehcles or Meal Producon. 3 Pav s (1984) research s based on he Scence Polcy Research Un daabase a Sussex Unversy, whch provdes daa on approxmaely 2000 nnovaons produced n UK frms ha belonged o he goods-producng secors. The daa covers he me perod from 1945 o Usng hs nformaon on nnovaons and a choce of oher frm characerscs enabled hm o se up four ndusry groupngs. For a more dealed descrpon of he daa se and he classfcaon see Pav (1984). 6

7 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 7 Table 1 Indusry Classfcaon ICT Inensy Innovaon Acvy 1 Food and Tobacco Ohers SII 2 Texles Ohers SDG 3 Apparel Ohers SDG 4 Leaher Ohers SDG 5 Wood Producs Ohers SDG 6 Paper, Pulp Ohers SDG 7 Publshng, Prnng ICT-Usng SDG 8 Coke, Peroleum, Nuclear Fuels Ohers SII 9 Chemcals and Chemcal Producs Ohers SBI 10 Basc Meals Ohers SII 11 Fabrcaed Meal Producs Ohers SII 12 Machnery ICT-Usng SGS 13 Offce Machnery and Compuers ICT-Producng SBI 14 Elecrcal Apparaus n.e.c. ICT- Usng SBI 15 Rado, TV and Comm. Equpmen ICT-Producng SGS 16 Insrumens, Opcs and Waches ICT-Producng SGS 17 Moor Vehcles ICT- Usng SII 18 Oher Transpor Equpmen ICT- Usng SII 19 Furnure and Manufacurng n.e.c. Ohers SDG 20 Recyclng ICT- Usng SDG 21 Elecrcy, Gas, and Waer Supply Ohers SII 22 Consrucon ICT- Usng SDG Noes: SII = Scale-Inensve Indusres. SDG = Suppler-Domnaed Goods-Producers. SGS = Specalzed Goods-Supplers. SBI = Scence-Based Innovaors. Insead, SGS are producon-nensve secors wh a sronger focus on produc nnovaons, whch are ofen smaller, echnologcally specalzed and funcon as complemenary secors for SII ndusres. Those ndusres usually hold close relaonshps wh her larger cusomers and provde hem wh her specalzed knowledge and experse. SGS secors are ofen locaed n producon processes of vercal dsnegraon n whch varous frms/ndusres of dfferen economes of scale/scope spl a producon process no separae producon unes, each performng a lmed subse of acves requred o creae a fnal produc. The concep of vercal dsnegraon s que mporan n very volale markes, when frms/ndusres need o adap o sudden marke changes quckly. Machnery or Rado, TV and Communcaon Equpmen are characersc SGS secors. SDG secors are radonal manufacurng secors whose prmary focus s coscung and less on echnologcal nnovaons. Cos-cung nnovaons are usually ransmed va upsream supplers, why R&D expendures on process and produc

8 Echer and Srobel 8 nnovaon are less sgnfcan n hose secors. These ndusres rely grealy on ndusry specfc professonal sklls or rademarks. SDG secors are Texle and Prnng, bu also Furnure and Manufacurng n.e.c. Fnally, R&D n SBI secors focuses on produc and process nnovaons based on bes avalable scence. SBI ndusres upsream R&D supplers/parners are unverses and oher organzaons of hgh academc sandards. SBI ndusres approprae her nnovaons hrough paenng, specalzed sklls, and dynamc learnng economes. In mos cases, hose ndusres are characerzed by rapd growh poenals and srong marke posons. Secors ha mach SBI characerscs are Chemcals and Elecrcal Apparaus n.e.c., as well as Offce Machnery and Compuers. 4 R&D, ICT Inensy and Producvy Growh We commence our analyss wh nernaonal comparsons of secoral conrbuons o aggregae value-added and R&D socks o gan a general overvew of he daa. 4 Fgure 1a summarzes he conrbuons for 8 OECD counres ha feaure suffcen daa for an ndusry level comparson. Fgure 1a shows ha he conrbuons o aggregae value-added are relavely smlar across counres. ICT-Usng ndusres generae abou 40 50% of aggregae value-added, whle Oher (Non-ICT-Inensve) ndusres generae abou, 45 56%. Ths leaves abou 3 6% for ICT-Producng ndusres, whch seems small a frs, bu snce hs segmen ncludes only hree relavely small secors of he economy, s neverheless mpressve. Fnland s he sole ouler n erms of ICT-Producng ndusres; whch added over 10% o aggregae value-added. Whle he conrbuons o aggregae value-added across ICT ndusry ypes are roughly smlar for hese OECD counres, Fgure 1b shows ha he R&D conrbuons of ICT ndusres dffer dramacally across OECD counres. For Germany, France and Ialy, he vas majory of R&D s generaed by ICT-Usng ndusres (57 72%), whle Sweden, Neherlands, Span, and Denmark fnd hemselves beween 26% and 41% of R&D generaed by ICT-Usng ndusres. 4 Conrbuons reflec ndusry shares n value-added and R&D socks, respecvely. 8

9 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 9 Fgure 1 Conrbuons o Aggregae R&D and Value-Added By ICT Indusry Classfcaon a) Conrbuon o Aggregae Value-Added b) Conrbuon o Aggregae R&D Socks 100% 100% 80% 80% 60% 60% 40% 40% 20% 20% 0% GER FR ITA ESP FIN SWE NLD DEN 0% GER FR ITA ESP FIN SWE NLD DEN ICT-Producng ICT-Usng Ohers ICT-Producng ICT-Usng Ohers By Innovave Acvy Indusry Classfcaon c) Conrbuon o Aggregae Value-Added d) Conrbuon o Aggregae R&D Socks 100% 100% 80% 80% 60% 60% 40% 40% 20% 20% 0% GER FR ITA ESP FIN SWE NLD DEN SII SDG SGS SBI 0% GER FR ITA ESP FIN SWE NLD DEN SII SDG SGS SBI Sources: EUKLEMS (2008), OECD STAN R&D (2006) and IIGAD (2008). Fnland excels dramacally n ICT-Producng ndusres; s 28% R&D conrbuon o he aggregae R&D sock s wce as hgh as hese ndusres conrbuons n Denmark. In general, he ICT-Producng conrbuons range whn 10 20%. Fnally, he R&D conrbuons of Oher ndusres also vary subsanally across counres. In Fnland, Neherlands and Denmark hese ndusres conrbue he majory of R&D wh shares of over 50%, whle n Germany, France and Ialy less han 26% of R&D s generaed n Oher ndusres. I s surprsng ha he R&D nenses of

10 Echer and Srobel 10 ndusres dffer so dramacally across counres alhough her conrbuons o value-added are so smlar. Ths suggess srong dffuson of echnology, eher va basc research or va mulnaonal horzonal or vercal producon chans. We underake he same comparson of value-added and ndusry conrbuons across OCED counres usng our n secon 3.2 oulned nnovaon-acvy-based classfcaon of ndusres. Fgure 1c shows ha he conrbuons o aggregae value-added by ndusry ype are agan exceedngly smlar across counres. SII and SDG ndusres conrbue roughly equal shares o aggregae value-added across OECD counres, rangng from 30 o 40%. SGS and SBI shares are also smlar across counres; hey conrbue abou 10 20% (he only ouler beng Span wh an unusually low SGS conrbuon of 6%). Overall hese conrbuons reflec he smlar homogeney n value-added conrbuons ha we observed n Fgure 1a. The dfference beween R&D conrbuons and value-added conrbuons s even more pronounced when we examne ndusry classfcaons based on nnovave acvy. Fgure 1d hghlghs he dramac dversy of he R&D conrbuons across nnovave ndusry ypes across counres. SII ndusres n Fgure 1d are shown o conrbue over 70% o R&D n France, whle hey conrbued less han 16% o R&D n Neherlands. The Neherlands nsead receved massve conrbuons o R&D from SBI ndusres (over 65%), whle mos oher counres SBI conrbuons range beween 15% and 30%. Mos smlar are SDG conrbuons across counres, bu hey also consue he smalles conrbuon o overall R&D n all counres, rangng from 1% (n Ialy) o 10% (n Denmark). SGS conrbuons are hghes n Germany, Fnland, Sweden and Denmark (30 40%), whle oher counres show consderably lower conrbuons (10 23%). Ths dramac dversy of R&D conrbuons by ndusry ype hghlghs he nernaonal dversfcaon of producon and R&D across OECD counres. Whle value-added shares are largely smlar across counres, he ndusry conrbuons o R&D vary dramacally when ndusres are grouped by her ype of R&D acvy. 10

11 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 11 5 Economerc Modellng 5.1 Tme Seres Properes To assess he me seres properes of our ndusry panel we conduc un-roo ess for boh balanced and unbalanced panel daa ses (see Table A1 n he Appendx). One mporan ssue wh panel saonary s spurous correlaon (Granger and Newbold, 1974). Non-saonary daa, whch s characerzed by a non-mean reverng generang daa process, ofen exhbs some knd of rend componen. Hence, he economerc analyss of non-saonary rended varables augmens he rsk of falsely denfyng sgnfcan relaonshps beween hose varables. Tesng panels for her saonary properes abaes false sascal nferences and assures ha fxed-effecs or frs-dfference mplemenaons fully remove dosyncrac me-nvaran componens. Saonary properes become especally mporan when regresson echnques are appled o longer me dmensons. We apply wo ess for balanced and unbalanced daa ses. For our four regresson varables average labor producvy (ALP) growh, growh n ICT and Non-ICT capal deepenng, and TFP growh we employ panel daa un-roo ess proposed by Levn, Ln and Chu (2002) and Im, Pesaran and Shn (2003), henceforh LLC and IPS. Boh ess only apply n case of balanced daa ses and are based on he es approach nroduced by Dckey and Fuller (1979). Whle LLC assumes homogenous auoregressve coeffcens for all ndvdual seres under H 0 ha equal zero (ndcang a un roo) and under he alernave H 1 all auoregressve coeffcens o be smaller han zero, IPS allows some of he ndvdual seres o conan a un roo (.e. some of he auoregressve coeffcens may be equal o zero) under he alernave H 1. Hence, he IPS s less resrcve under H 1 and perms saonary as well as (few) non-saonary seres n he panel. Those ess pronounce he saonary behavor of he enre panel nsead of sngle seres. In case of unbalanced daa ses oher ypes of panel un-roo ess become necessary. In hs sudy we employ he class of non-paramerc Fsher-ype ess ha are based on combnaons of ess conduced for ndvdual me seres. Smlar o he IPS es, whch combnes es sascs of sngle me seres, Fsher-ype ess combned sngle seres sgnfcance levels and relax he assumpon of LLC havng homogeneous auoregressve coeffcens under he H 1. Those Fsher-ype ess com-

12 Echer and Srobel 12 bne sgnfcance levels from N ndependen un-roo ess (Maddala and Wu, 1999) and assume ha all seres are non-saonary under he H 0. Under H 1, a leas some of he ndvdual seres n he panel may be non-saonary. Those ess have he advanage (especally over IPS) of allowng each me seres n he panel o dspose of dfferen sample sze. Ths s parcularly mporan for our R&D sock varable. A second advanage s ha hose ess can also be conduced for dfferen un-roo ess. In our case we conduc Fsher-ype ess for augmened Dckey-Fuller (1979) and Phllps-Perron (1988), henceforh ADF and PP, respecvely. Furher advanages of he Fsher-ype ess are her prevalence over LLC/IPS under cross-seconal correlaon and when saonary and non-saonary seres are ncluded under H 1. Laer s due o Fsher-ype ess hgher power o dsngush beween he null and he alernave hypohess. 5 Snce our prmer neres s growh regresson we es all varables n growh raes. The resuls n Table A1 for LLC and IPS rejec he non-saonary of varables n he panel for all lag srucures. Takng he unbalanced naure of R&D sock growh no accoun and applyng Fsher-ype ess, boh ADF and PP rejeced he nonsaonary of varables. Bascally, as Fsher-ype ess are characerzed by hgher power under saonary and non-saonary seres n he alernave hypohess we regard her rejecon of panel non-saonary as srong addonal suppor. Accordng o hose es resuls we decde o no furher dfference our varables. 5.2 Economerc Esmaon To model our economerc specfcaon we ake he operaons of he R&D secor n (4) as gven and focus on esmang (3) usng Δln y, = α + β Δlnk, + γ ΔlnTFP, + ε,, (5) where labor producvy, y,, s denfed as value-added per hours worked n ndusry a me ( ), k s capal deepenng measured by capal servces per hours worked, and TFP s oal facor producvy. To accoun for he vas producvy dfferenals ha have been documened for ICT and Non-ICT nvesmens, we dsaggregae capal socks o allow for separae effecs of ICT and Non-ICT capal ICT NICT servces ( and ). Ths renders (5) no k, k, 5 For a more deal descrpon on he dfferences n panel-un roo es procedures see e.g. Maddala and Wu (1999). 12

13 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 13 Δln ICT NICT y, = α + β1δln k, + β 2Δln k, + γ ΔlnTFP, + ε,. (6) Our key neres s, however, o solae he effecs of R&D nensy on labor producvy growh and wheher such effecs dffer across ndusry ypes. Therefore we replace he generc TFP resdual wh acual daa on R&D socks,. These socks are calculaed usng he perpeual nvenory mehod suggesed by Guellec and Poelsberghe (2001). 6 In a frs sep we furher allow for ndusry parameer heerogeney n R&D effecs and growh rends, and fxed me effecs by nroducng me dummes denoed byδ. Thus (6) can be rewren no R, Δ ln y α R I + e. (7) ICT NICT, = + α j I j + δ + β1δ ln k, + β 2Δ ln k, + γ 1Δ ln R, + γ 2 Δ ln, j, A dummy varable I ndcaes dfferen caegores of ndusres j ha we have dscussed n secon 3 (lsed n Table 1). They relae o alernave ndusry classfcaons accordng o her R&D ype and ICT nensy. Under he assumpon ha he error erm of equaon (7) s..d and exhbs no correlaon wh one of he regressors can be esmaed wh OLS. In case of endogeney (semmng from smulaney, omed varable bas or measuremen error) ha s capured by ndusry heerogeney and nduces a correlaon beween e, and one of he regressors, he applcaon of sandard fxed-effecs echnques may be approprae. To accoun for ndusry heerogeney we apply he LSDV (Leas Squares Dummy Varable) approach and nroduce ndusry effecs, α, measurng ndusry-specfc growh rends. The applcaon of LSDV urns (7) no (8) Δ ln y α R I + e. ICT NICT, = + δ + β1δ ln k, + β 2Δ ln k, + γ 1Δ ln R, + γ 2 Δ ln Oher endogeney problems leavng one (or all) of he regressors correlaed wh he error erm despe havng conrolled for ndusry heerogeney can be specfed n he followng assumpon: E [ x e ] s s 0, (9) where x represens he regressors on he rgh hand sde of equaons (8) and e he prevously..d. assumed error erms. Assumpon (9) allows for conemporaneous correlaon beween a curren shock n e and x as well as feedback effecs from pas, j, 6 Dealed dervaons of R&D socks are provded n he Appendx.

14 Echer and Srobel 14 shocks n e -s ono curren x. The regressors are no longer assumed o be srcly exogenous bu predeermned n he sense ha x may depend on pas shocks and no on expeced shocks n he fuure. Prncpally, hs can be assumed for all varables n our producon funcons, hence, all x are reaed as predeermned. Only me dummes are consdered o be srcly exogenous. Unless he economy has converged o s seady sae, ranson mples feedback effecs from ncome o nvesmen and R&D. Ths endogeney hen mples ha a leas one (or all) of he regressors may well be correlaed wh he error erm. Insrumens are dffcul o movae and he bes soluon s o rely on lagged varables as suggesed by he General Mehod of Momens approach (GMM). Dfference GMM employs frs dfferences of he varables, whch elmnae ndvdual growh rends. Ths renders lagged values of each varable as suable nsrumens. Snce huge ses of nsrumens end o overf endogenous varables and produce nvald nsrumens n fne samples, we examne our resuls for nsrumens ses of dfferen lag lenghs. 7 6 Capal and R&D Effecs on Producvy Growh Fgure 1 has clearly esablshed he dversy of R&D conrbuons across counres. Whle he relaonshp beween R&D and producvy a he aggregae level s nformave, we examne ndusry-level conrbuons o deermne he drvng ndusres n more deal. As we have shown n prevous chapers, ndusry nvesmen decsons vary grealy o generae profoundly dfferng raes of reurns and producvy conrbuons. To accoun for such dfferences, we apply he economerc mehodology oulned n secon 5, whch allows for parameer heerogeney for dfferen ndusry groups as well as for dfferen nvesmen ypes. The adopon of he esmaon approach derved n equaon (7) requres suffcenly dsaggregaed R&D daa. The mos dealed ndusry-level daa s provded by he Ifo Indusry Growh Accounng Daabase, whch feaures secoral daa for 52 ndusres (WZ2003 classfcaon, oal economy excludng prvae household servces). Inernaonal R&D daa (as feaured n he OECD STAN R&D Daabase) s avalable a a slghly hgher level of aggregaon ha reduces he number of appro- 7 Furher drawbacks n he applcaon of oo many nsrumens are horoughly addressed n Roodman (2008). 14

15 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 15 prae ndusres o he 22 goods-producng ndusres repored n Table 1. 8 To mplemen equaon (7), we requre addonal nformaon on ICT and Non-ICT nvesmens ha can be obaned from he EUKLEMS Growh and Producvy Accouns (Tmmer e al., 2007a, b). However, EUKLEMS ICT nvesmen daa s no avalable a he same level of dsaggregaon as R&D daa, whch would reduce he level of dsaggregaon o 12 ndusres. We vew such a hgh level of ndusry aggregaon s excessvely resrcve such ha dealed ndusry producvy analyses are rendered raher meanngless. We choose no o reduce our daa from 22 o 12 ndusres, and nsead focus on Germany, where he Ifo Indusry Growh Accounng Daabase provdes suffcen daa o denfy ndusres by ICT nensy and nnovaon acvy ype. 6.1 OLS Esmaes Table 2 repors he resuls of equaon (6) and (7) for OLS. Our sarng regresson uses a general producon funcon wh capal and oal facor producvy (column I) and connues wh an asse spl n ICT and Non-ICT capal (column II). Boh regressons show ha much of he producvy growh a he ndusry level s generaed by growh n TFP. A 1.0 percen ncrease n TFP growh generaes abou 3.0 percen n ALP growh accordng o column I. Growh n capal deepenng has n fac a posve and sgnfcan effec bu s noably lower. Tesng consan reurns o scale (CRS) shows ha he null hypohess of he sum of he esmaed npu coeffcens equals 1.0 can be rejeced on he 1% sgnfcance level. In column II we sae a posve and sgnfcan ALP growh effec from Non-ICT capal deepenng, bu TFP s sll que hgh. ICT capal deepenng conrbuons are no sgnfcanly dfferen from zero (and wll reman so hroughou all of he specfcaons n Table 2). Jon capal deepenng and srong TFP growh ndcae enormous ncreasng reurns o scale, whch s confrmed by he CRS es. To ge a more dealed pcure of he nfluence generaed by echnologcal change, we replace he TFP resdual by acual R&D socks (column III) and apply he ndusry dsaggregaon of equaon (7) hroughou columns IV VII. R&D ndcaes ha an ncrease n echnology generaes a much smaller effec as compared o he prevously esmaed TFP growh effecs, however, R&D s esmaed sascally 8 A hs level of dsaggregaon we have daa for seven major European counres (France, Ialy, Span, Fnland, Sweden, Neherlands, and Denmark).

16 Echer and Srobel 16 nsgnfcan. In hs specfcaon he CRS es canno rejec he null hypohess of he sum of npu facors beng equal o one. Also, n conras o prevous resuls here s a srong ndcaon for decreasng reurns o scale (around 0.8 percen). As he esmaed R&D effec capures an average effec across all ndusres and as we are neresed n he ndusry characerscs ha caalyze R&D o mprove secoral labor producvy, n he followng specfcaons we allow for parameer heerogeney by ndusry ypes. Msspecfcaons n he regresson model may be a reason why here s such srong ndcaon for decreasng reruns o scale. As formulaed n equaon (7) we addonally conrol for parameer heerogeney n R&D effecs for ICT-Inensve and SGS/SBI ndusres. Employng he ICT-Inensy spl (column IV, V) obvously here are que dfferen R&D growh effecs for ICT- Inensve and Non-ICT-Inensve ndusres. Whle he R&D effec on ALP growh of Non-ICT-Inensve ndusry s very low and sascally nsgnfcan, ICT- Inensve ndusres exhb a growh effec around 0.8 percen. Calculang he compose R&D effec (from he margnal R&D effec of he reference caegory and he dummy-neraced margnal R&D effec of ICT-Inensve ndusres ncludng he ICT-Inensve ndusres average rend growh) generaes a sgnfcan overall R&D effec for ICT-Inensve ndusres of approxmaely 0.9 percen. Takng he esmaes for ICT and Non-ICT capal deepenng no accoun oal scale effecs of he npu facors (around 1.3 percen) ndcae slghly ncreasng reurns o scale. Neverheless, he CRS es canno be rejeced, hus consan reurns o scale are sascally suppored. Splng ICT-Inensve furher no ICT-Producng ndusres reveals ha R&D growh effecs are prmarly locaed n he ICT producon secors (column V). 9 Especally hose ndusres are known for havng experenced srong surges n producvy performance durng he launchng of he New Economy n he md 1990s. The margnal growh effec for ICT-Producng ndusres s esmaed by nearly 1.0 percen. The compose R&D growh effec amouns o abou 1.1 percen and s hghly sgnfcan. Hence, akng all npu facors no accoun he CRS propery for German goods-producng ndusres s rejeced a he 5% sgnfcance level. The esmaed reurns o scale are abou 1.6 percen. 9 Regresson analyses esng separae R&D growh effecs for ICT-Usng ndusres revealed no sgnfcan R&D neracon erm. The resuls for hose regressons are no depced n hs paper. 16

17 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 17 To examne he R&D growh effecs beyond he role of ICT Inensy we apply our alernave Innovaon axonomy, where ndusry dummes now represen SGS & SBI ndusres. Accordng o column VI we sae srong posve and sgnfcan producvy effecs from R&D growh n SGS & SBI ndusres. Those secors managed 0.9 percen hgher ALP growh from ncreased R&D han oher ndusres. Calculang he compose R&D effec derves a sascally sgnfcan 1.0 percen ncrease for SGS & SBI ndusres. The CRS propery s sll suppored, even hough he p- value for he CRS es barely fals o rejec he null hypohess of reurns o scale resemblng uny. Breakng SGS & SBI no SBI and SGS separaely dscloses a hgher margnal ALP growh from R&D generaed n SBI secors (column VII). More precsely, an ncrease n R&D growh by 1.0 percen ransfers no a 1.0 percen ncrease n SBI secors ALP growh. 10 As he compose R&D growh effec already amouns o a sgnfcan 1.2 percen ncrease for SBI ndusres accounng for all oher npu facor esmaes he CRS propery s no longer suppored. The CRS es now rejecs he null hypohess of reurns o scale beng equal o one on he 1% sgnfcance level wh an esmaed reurns-o-scale coeffcen of abou 1.7 percen. The absence of a sgnfcan ICT capal deepenng coeffcen may nally be profoundly surprsng. However, he reason for s nsgnfcance s drecly relaed o he relaon beween ICT ndusres R&D nensy and ICT nvesmen suggesed by he heory. Neoclasscal heory clearly oulnes ha long-erm growh n oupu per worker s exclusvely due o echnologcal progress, no maer how much of ha growh can be accouned for by capal accumulaon. Table 2 confrms ha he ulmae drver of producvy growh s ndeed R&D as proposed by he heory. I s no any R&D however; s R&D ha orgnaes n he hghly dynamc ndusres ha produce ICT or generae scence-based nnovaons. Ths fndng sheds an enrely new lgh on prevous fndngs ha ndcae a dramac mpac of ICT capal deepenng on producvy growh. The mplcaon beng ha he source of he capal deepenng producvy conrbuons was n fac R&D growh ha orgnaed n ICT-Inensve ndusres. 10 Regresson analyses esng SII, SDG and SGS effecs separaely showed no sgnfcan R&D neracon erms. The resuls for hose regressons are no depced n hs paper.

18 Echer and Srobel 18 Table 2 ALP Growh Regressons by ICT Inensy and Innovaon Taxonomy, OLS Esmaes, Dep. Var.: ALP Growh I II III IV V VI VII Consan * [0.0048] [0.0053] [0.0273] [0.0254] [0.0270] [0.0271] [0.0264] ICT-Inensve [0.0226] ICT-Producng [0.0329] SGS & SBI * [0.0193] --- SBI *** [0.0186] Capal Deepenng *** [0.0217] ICT Capal Deepenng [0.0279] [0.1261] [0.1013] [0.0941] [0.1017] [0.0865] Non-ICT Capal Deepenng *** * *** *** *** *** --- [0.0374] [0.2322] [0.1345] [0.1560] [0.1506] [0.1441] TFP *** *** [0.1339] [0.1342] R&D Sock [0.2203] [0.0627] [0.0610] [0.0632] [0.0652] R&D Sock ICT-Inensve ** [0.3046] ICT-Producng *** [0.1995] SGS & SBI *** [0.2941] --- SBI *** [0.0872] R&D Compose ICT-Inensve ** [0.3190] R&D Compose ICT-Producng *** [0.2170] R&D Compose SGS & SBI *** [0.3057] --- R&D Compose SBI *** [0.0520] Reurns o Scale CRS Tes Observaons R Adj. R No. of Indusres Tme Dummes yes yes yes yes yes yes yes Noes: All varables are n exponenal growh raes. An exreme ouler n all specfcaons s dropped: he Peroleum and Coke ndusry. Excluded s also Consrucons due o s very low R&D expendures shares. Compose effecs are calculaed va he Dela mehod. Reurns o scale provdes he sum of he esmaed npu coeffcens. CRS es repors he p-value assocaed wh a es of he null hypohess ha he sum of he esmaed npu coeffcens equals 1.0. Robus sandard errors n brackes allowng for correlaon whn ndusres over me. Sgnfcance levels: * sgnfcan a 10; ** sgnfcan a 5; *** sgnfcan a 1. Sources: OECD STAN R&D Daabase (2006) and IIGAD (2008). 18

19 R&D Inensy, ICT Invesmen and Indusry Producvy Growh LSDV Esmaes In Table 3 we presen regresson resuls for equaon (8) where we nroduce ndusry dummes o conrol for ndvdual secor heerogeney. Includng ndusry dummes subsanally lowers he neracon effecs for all ndusry caegores and generaes more reasonable compose esmaes and CRS properes. The F-es for jon sgnfcance of ndusry dummes rejecs he null hypohess of all dosyncrac ndusry effecs equal zero hroughou all specfcaons; hence we regard he ncluson of ndusry dummes as jusfed. Compared o Table 2 he capal deepenng and TFP growh effecs are only slghly reduced by accounng for ndusry heerogeney (column I). Column II reveals a sgnfcan and posve growh effec from ICT capal deepenng by almos 0.1 percen. In fac, hs s n range of prevous OLS esmaes bu hs me wh sgnfcance a he 10% level. The Non-ICT capal deepenng effec nsead s subsanally reduced by almos half n sze and sascally sgnfcan. TFP growh sll exhbs he sronges effec. Boh specfcaons (column I, II) rejec he CRS properes as n case of he OLS esmaes. Regardng he R&D effec (column III), whch represens an average across all ndusres, s now esmaed sascally sgnfcan on he 10% level bu wh srong reducon. Boh capal deepenng growh effecs, ICT and Non-ICT, urn ou o be sascally nsgnfcan n hs specfcaon. The CRS propery s rejeced a he 5% sgnfcance level and reurns o scale are srongly reduced. They dsplay decreasng reurns o scale of abou 0.5 percen. The sample spl no ICT-Inensve and Non-ICT-Inensve ndusres reveals mos subsanal changes compared o he OLS esmaes (column IV). Non-ICT capal deepenng growh s sll sascally sgnfcan bu reduced n magnude, whle ICT-/Non-ICT-Inensve ndusres show a reduced margnal R&D growh effec. 11 Regardng he margnal R&D growh effec for ICT-Producng ndusres we also sae a decreased magnude (column V). The reducons n margnal R&D effecs reduce he compose R&D effecs and generae much more reasonable CRS properes. Boh columns IV and V show reurns o scale locaed around uny. Those CPS properes are also srongly suppored by he CRS ess. 11 Regresson analyses esng separae R&D growh effecs for ICT-Usng ndusres revealed no sgnfcan R&D neracon erm. The resuls for hose regressons are no depced n hs paper.

20 Echer and Srobel 20 Table 3 ALP Growh Regressons by ICT Inensy and Innovaon Taxonomy, LSDV Esmaes, Dep. Var.: ALP Growh I II III IV V VI VII Capal Deepenng *** [0.0290] ICT Capal Deepenng * [0.0307] [0.1320] [0.1305] [0.1267] [0.1280] [0.1252] Non-ICT Capal Deepenng ** * * * ** --- [0.0367] [0.2240] [0.1960] [0.1848] [0.1892] [0.1954] TFP *** *** [0.2091] [0.2097] R&D Sock * [0.1084] [0.0747] [0.0831] [0.0839] [0.0908] R&D Sock ICT-Inensve ** [0.2962] ICT-Producng ** [0.2463] SGS & SBI ** [0.2295] --- SBI *** [0.1498] R&D Compose ICT-Inensve ** [0.3475] R&D Compose ICT-Producng *** [0.1966] R&D Compose SGS & SBI *** [0.2439] --- R&D Compose SBI *** [0.0671] Reurns o Scale CRS Tes Observaons R Adj. R No. of Indusres Indusry Dummes yes yes yes yes yes yes yes Tme Dummes yes yes yes yes yes yes yes Indusry-Dummes Tes Noes: See Table 2. The ndusry-dummes es dsplays he p-value for jon sgnfcance of ndusry fxed effecs employng an F-es. Sgnfcance levels: * sgnfcan a 10; ** sgnfcan a 5; *** sgnfcan a 1. Sources: OECD STAN R&D Daabase (2006) and IIGAD (2008). Also, he margnal R&D growh effecs n SGS & SBI ndusres (column VI) dsplay a reduced coeffcen as well as n SBI ndusres (column VII). 12 Those reducons n esmaes generae lower compose R&D effecs for SGS & SBI and SBI ndusres and as n case for ICT-Inensve and ICT-Producng ndusres provde much more realsc esmaes for overall reurns o scale. CRS properes are closely esmaed o uny for boh secor ypes and canno be rejeced by he CRS ess. As 12 Regresson analyses esng SII, SDG and SGS effecs separaely showed no sgnfcan R&D neracon erm. The resuls for hose regressons are no depced n hs paper. 20

21 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 21 for ICT-Inensve and ICT-Producng ndusres he CRS properes for SGS & SBI and SBI secors are srongly suppored sascally. 6.3 Frs-Dfference GMM Esmaes The specer of endogeney looms large n hs shor panel and we use GMM echnques (Blundell and Bond, 2000, Roodman, 2005) o verfy our LSDV resuls. Our applcaon of frs-dfference GMM employs frs dfferences of he varables x (n growh raes), whch elmnae he ndvdual growh rends and leave lagged values of x (n growh raes) as suable nsrumens. Snce he me seres analyses n secon 5 have rejeced he assumpon of panel non-saonary frs-dfferencng urns an approprae procedure o elmnang ndvdual secoral growh rends and ensures us ha no furher dfferencng of varables s requred. Also, due o x beng gven n erms of growh raes lags -3 and deeper become vald nsrumens. Usng nsrumen ses of dfferen lag lengh we seek o mgae he conngen problem of weak nsrumens. Therefore we specfy our base lne esmaon and employ lags -3 and deeper as approprae nsrumens (Table 4). For comparson purpose we also employ shorer lag srucures n nsrumens: -3 o -4 (nsrumen se 1, Table 5) and -3 o - 5 (nsrumen se 2, Table 5). Sarng wh Table 4 we sae ha he auoregressve srucure suppors he valdy of he nsrumens. The auocorrelaon srucure of he error erms does no ndcae an auoregressve srucure of second order and hence no requremen for deeper lags as nsrumens as he employed -3 lag specfcaons. Evaluang he valdy of he employed nsrumens furher he Hansen ess canno rejec he null hypohess of overdenfcaon hroughou all specfcaons. Checkng he valdy of he srcly exogenous specfed nsrumens (.e. me dummes) he Dfference-n- Hansen es canno rejec he null hypohess of overdenfcaon eher. Despe of reducng he sample sze by usng lags as nsrumens n GMM esmaons and hus esmaon precson, frs-dfference GMM broadly confrms our LSDV resuls alhough does sugges smaller effecs. Once we conrol for endogeney, he coeffcen for he broadly specfed ICT-Inensve ndusres R&D growh effec becomes nsgnfcan, however. Only Non-ICT capal deepenng remans as he sole drver of labor producvy growh. If we, however, refne our ICT ndusry focus and examne R&D expendures underaken by ICT-Producng ndusres, we agan fnd srong and sgnfcan R&D conrbuons o ALP growh (columns IV, V).

22 Echer and Srobel 22 Ths furher refnes our prevous resuls. I was no R&D growh n ICT-Inensve ndusres (ncludng ICT-Usng ndusres) bu he source of he producvy growh acceleraons can ulmaely be raced o nnovaons ha orgnaed n he ICT- Producng ndusres. Columns VI and VII replcae he exercse usng our alernave ndusry classfcaon. Insead of examnng ICT caegores we group ndusres agan by her ype of R&D acvy. Qualavely, he resuls are roughly dencal o hose produced by he ICT ndusry classfcaons. ICT capal deepenng, whch was he crucal conrbuor o labor producvy growh n prevous analyses s no sgnfcan hroughou f we accoun for R&D nensy. Non-ICT capal deepenng s sll he only sgnfcan drver of ndusry ALP growh. Among hose SGS & SBI ndusres ha are crucal producvy drvers we fnd no suppor of hgher labor producvy growh for ncreased R&D nensy when GMM employs all avalable nsrumens. The same s rue for SBI ndusres. As already menoned he employmen of oo many nsrumens runs he rsk of overdenfyng endogenous varables and may produce nvald nsrumens n fne samples. Ths may be due o he nably of elmnang he endogenous componens n he seres and bases GMM resuls oward hose wh no nsrumens appled. Zlak (1997), for nsance, shows ha he GMM bas rses as deeper lags of varables are employed. Regardng he numbers of observaons and he hgh numbers of nsrumens esmae bases from nvald nsrumens may be an approprae crque. Therefore we es smaller ses of nsrumens for he R&D neracon specfcaons (.e. for columns IV VII, Table 4). In se 1 we employ lags from -3 o -4 (Table 5, column I IV) reducng he avalable nsrumens from around 166/164 o 88. Wh respec o he auocorrelaon and overdenfcaon ess sascal nference s smlar o hose of he regressons n Table 4. However, he sascal sgnfcance for margnal R&D growh and compose effecs dffers subsanally. In hs exercse merely he margnal R&D growh effec for SBI ndusres and s compose are esmaed sascally sgnfcan. Overall he drasc reducon n nsrumens produces mmense degeneraons n coeffcen esmaes. 22

23 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 23 Table 4 ALP Growh Regressons by ICT Inensy and Innovaon Taxonomy, Frs-Dfference GMM Esmaes, Dep. Var.: ALP Growh I II III IV V VI VII Capal Deepenng *** [0.0427] ICT Capal Deepenng ** [0.0220] [0.1430] [0.1386] [0.1452] [0.1358] [0.1257] Non-ICT Capal Deepenng ** * ** ** ** ** --- [0.0492] [0.1539] [0.1776] [0.1482] [0.1565] [0.1675] TFP *** *** [0.2957] [0.2923] R&D Sock * [0.1308] [0.1328] [0.1344] [0.1315] [0.1425] R&D Sock ICT-Inensve [0.4270] ICT-Producng * [0.2162] SGS & SBI [0.3010] --- SBI [0.1794] R&D Compose ICT-Inensve * [0.3578] R&D Compose ICT-Producng *** [0.1875] R&D Compose SGS & SBI ** [0.2586] --- R&D Compose SBI *** [0.0764] Reurns o Scale CRS Tes No. of Insrumens AR1 Tes AR2 Tes Hansen Tes Dff-n-Hansen Tes (IV) Observaons No. of Indusres Tme Dummes yes yes yes yes yes yes yes Noes: See Table 2. AR ess dsplay he p-values for auocorrelaon up o he second order. Hansen es repors he p-value for he es for overdenfcaon resrcons, whle Dfference-n-Hansen es repors he p-value for he es for he subse of nsrumens ha are assume o be exogenous (.e. me dummes). Tmes lags from 3 onwards for all varables are used as nsrumens. Sgnfcance levels: * sgnfcan a 10; ** sgnfcan a 5; *** sgnfcan a 1. Sources: OECD STAN R&D Daabase (2006) and IIGAD (2008). We furher es he alernave se 2, whch employs a larger number of nsrumens. To ulze a parsmonous se of nsrumens and o employ as few addonal nsrumens as possble we add one lag furher, hus coverng lags -3 o -5 (Table 5, column V VIII). The auocorrelaon and overdenfcaon ess suppor he valdy of he nsrumens. Regardng he margnal R&D growh effec for ICT- Producng ndusres we sae qualavely smlar effecs as n Table 4 (column V).

24 Echer and Srobel 24 For SBI ndusres se 2 generaes comparable margnal R&D growh effecs as for se 1. The R&D growh composes for ICT-Producng and SBI secors are esmaed sascally sgnfcan of around 0.6 and 0.5 percen. CRS properes for boh ndusry ypes are sascally suppored once agan, however, he sgnfcance levels are less pronounced and esmaed reurns-o-scale coeffcens ndcae slghly decreasng reurns. The sensvy analyss for dfferen ses of nsrumens has shown ha here s a srong suppor for R&D growh effecs on ALP growh n ICT-Producng and SBI ndusres. We prefer se 2 of nsrumens over oher specfcaons as he longer he lags for nsrumens he weaker he power of es sascs. Shorer lag specfcaons obvously end o produce degeneraed esmaes. Asde from denfyng he roo cause of he srong conrbuons of capal deepenng o labor producvy growh, s also mporan o hghlgh ha wo dfferen ndusry classfcaons generae broadly smlar resuls. I s neresng n s own rgh ha he resuls are srkngly smlar despe he fundamenally dfferen ndusres ha comprse ICT-Producng or SBI ndusres. There s lle overlap beween SBI and ICT-Producng ndusres n boh defnon and classfcaon (see Table 1). Neverheless boh ndusry classfcaons are denfed as he crucal conrbuors o ndusry producvy growh n he goods-producng secors. 24

25 R&D Inensy, ICT Invesmen and Indusry Producvy Growh 25 Table 5 ALP Growh Regressons by ICT Inensy and Innovaon Taxonomy, Frs-Dfference GMM Esmaes, Reduced Se of Insrumens, Se 1 Se 2 Dep. Var.: ALP Growh I II III IV V VI VII VIII ICT Capal Deepenng [0.2187] [0.1691] [0.2228] [0.1920] [0.2016] [0.1870] [0.2104] [0.1857] Non-ICT Capal Deep * * *** * ** * [0.2291] [0.1910] [0.2502] [0.2323] [0.1780] [0.1884] [0.2290] [0.2425] R&D Sock [0.1334] [0.1068] [0.1241] [0.1080] [0.1342] [0.1267] [0.1314] [0.1347] R&D Sock ICT-Inensve [0.3554] [0.5190] ICT-Producng * [0.3047] [0.2108] SGS & SBI [0.3530] [0.4033] --- SBI * * [0.1944] [0.1890] R&D Compose ICT-Inen [0.2641] [0.4536] R&D Compose ICT-Prod *** [0.3108] [0.1954] R&D Compose SGS & SBI [0.3197] [0.3617] --- R&D Compose SBI ** *** [0.1494] [0.0956] Reurns o Scale CRS Tes No. of Insrumens AR1 Tes AR2 Tes Hansen Tes Dff-n-Hansen Tes (IV) Observaons No. of Indusres Tme Dummes yes Yes yes yes yes yes yes yes Noes: See Table 2. AR ess dsplay he p-values for auocorrelaon up o he second order. Hansen es repors he p-value for he es for overdenfcaon resrcons, whle Dfference-n-Hansen es repors he p-value for he es for he subse of nsrumens ha are assume o be exogenous (.e. me dummes). Insrumen se 1: Tmes lags from 3 ll 4 for all varables are used as nsrumens. Insrumen se 2: Tmes lags from 3 ll 5 for all varables are used as nsrumens. Sgnfcance levels: * sgnfcan a 10; ** sgnfcan a 5; *** sgnfcan a 1. Sources: OECD STAN R&D Daabase (2006) and IIGAD (2008). 7 Concluson Ths paper akes anoher sep oward denfyng he ulmae drvers of producvy growh a he ndusry level. Based on Neoclasscal and New Growh Theory we develop an emprcal mehodology ha allows us o es for he fundamenal drvers of ndusry producvy growh. We presen wo ses of ndusry classfcaons, one base on ICT nensy and anoher based on nnovaon ypes. Usng OECD daa we