Public R&D Support and Firms Performance

Transcription

1 June 2018 Publc R&D Support and Frms Performance A Panel Data Study Øvnd A. Nlsen, Arvd Raknerud, Dana-Crstna Iancu

2 Impressum: CESfo Workng Papers ISSN (electronc verson) Publsher and dstrbutor: Munch Socety for the Promoton of Economc Research CESfo GmbH The nternatonal platform of Ludwgs Maxmlans Unversty s Center for Economc Studes and the fo Insttute Poschngerstr. 5, Munch, Germany Telephone +49 (0) , Telefax +49 (0) , emal offce@cesfo.de Edtors: Clemens Fuest, Olver Falck, Jasmn Gröschl group.org/wp An electronc verson of the paper may be downloaded from the SSRN webste: from the RePEc webste: from the CESfo webste: group.org/wp

3 CESfo Workng Paper No Category 1: Publc Fnance Publc R&D Support and Frms Performance A Panel Data Study Abstract We analyse all the major sources of drect and ndrect R&D subsdes n Norway n the perod and compare ther effects on ndvdual frms performance. Frms that receved support are matched wth a control group of frms that dd not receve support usng a combnaton of stratfcaton and propensty score matchng. Changes n performance ndcators before and after support n the treatment group are compared wth contemporaneous changes n the control group. We fnd that the average effects of R&D support among those who obtaned grants and/or subsdes are postve and sgnfcant n terms of performance ndcators related to economc growth: value added, sales revenue and number of employees. The estmated effects are larger for start-up frms than ncumbent frms when the effects are measured as relatve effects (n percentage ponts), but smaller when these effects are translated nto level effects. Fnally, we do not fnd postve effects on return to total assets or productvty for frms who receved support compared wth the control group. JEL-Codes: C330, C520, D240, O380. Keywords: publc polcy, frm performance, treatment effects, stratfcaton, propensty score matchng, productvty. Øvnd A. Nlsen Norwegan School of Economcs Department of Economcs Norway Bergen ovnd.nlsen@nhh.no Arvd Raknerud Statstcs Norway Research Department Oslo / Norway rak@ssb.no Dana-Crstna Iancu Statstcs Norway Dvson of Statstcal Methods Oslo / Norway cr@ssb.no We are grateful for valuable comments durng presentatons at UNU-MERIT n Maastrcht and at workshops arranged by Oslo Insttute for Research on the Impact of Scence (OSIRIS) n Manchester and Oslo. We also thank Jørgen Modalsl, Terje Skjerpen and Perre Mohnen for frutful comments and dscussons. Fnancal support from the Research Councl of Norway through the grant s acknowledged.

4 1. Introducton There s a mutual understandng among economsts that technologcal progress s closely lnked to economc growth and that t s spurred by nvestments n Research and Development (R&D) (see e.g. Romer, 1990). The mechansms connectng nnovaton and productvty are related to both exstng and new frms, and to destructon of less effcent ones. Among exstng frms, R&D mght show up as creaton of new goods and servces, leadng to ncreased demand for the products of the frm, or to changes n the way the frm operates through process- and organzatonal nnovatons. R&D may also lead to entry of more effcent frms and new frms on the technology fronter. If frms themselves would get the whole economc benefts of ther R&D nvestments, there would be no need for publc support of R&D n the busness sector. Thus, when polcy-makers emphasse the mportance of publc R&D support, t s based on the understandng that there are market falures and spllovers related to nvestments n R&D (see e. g. Grlches, 1992). A common source of market falure s knowledge externaltes. Such externaltes may occur f t s dffcult to establsh ownershp rghts of new producton methods or technologes, enablng compettors to take advantage of nvestments n R&D wthout bearng the costs. The government mght subsdze R&D nvestments to reduce the margnal cost of R&D and/or the perceved rsk of external nvestors or lenders. In ths way government support may lead to ncreased R&D and/or nnovaton actvtes (see e.g. Hall and van Reenen, 2000.) The average gross domestc spendng on R&D as percentage of GDP n the OECD countres has been qute stable over the perod , varyng from 2.1 to 2.4 (OECD 2018). At the same tme OECD (2016, Fgure 4.7) shows that the ntensty of publc support has ncreased as a percentage of GDP n most OECD countres over the last ten years. Wth ths as a background, and knowng that n many countres there are several co-exstent and potentally complementary support schemes, the goal of ths paper s to evaluate quanttatvely the economc benefts of R&D subsdes on frms performance. The results presented are based on ndrect methods where t s assumed that nvestments n R&D can lead to product and process nnovatons, whch n turn can be reflected n performance ndcators measured over tme. The outcome varables studed are (frm-level) sales revenue, number of employees, value added, labour productvty (valued added per employee), and return on assets. These outcome varables are hghly relevant from a polcy perspectve as the analysed subsdy nstruments are ntended to contrbute to ncreased value added and employment n R&D-ntensve frms. 1

5 The man novelty of ths paper les n our ablty to study all the major sources of R&D subsdy programs n one country Norway smultaneously; both drect subsdes (grants) and tax credts. Ths s potentally of great mportance, as a large share of frms that receve drect support also receve tax benefts (but not vce versa). Our data have two key features that we explot: (1) we can merge frms wth nformaton about receved publc support over a long tme perod: , (2) the data have unversal coverage of ncorporated frms wth detaled accountng, employment and ownershp nformaton. The combnaton of three contemporaneous polcy nstruments, the long length of the analysed perod, and the possblty to lnk polcy nstruments wth frm level data make our study unque compared to the exstng lterature. Although there are other studes that address the ssue of a frm smultaneously usng multple sources of publc support (e.g. Czarntzk and Lopes-Bentoa, 2013, and Dumont, 2017), we are, to the best of our knowledge, the frst to analyse all major sources of support n one country and matchng these data to the whole populaton of (ncorporated) frms. Usng panel data, we can montor the outcome varables over tme before and after support and compare wth a control group of frms that do not receve support,.e. frms that are representatve of the counterfactual outcomes for those recevng support. The advantage of such an approach s that the outcomes are possble to measure both for the treatment group and for the control group. However, there can be a large element of randomness n such comparsons, whch necesstates large datasets to dstngush systematc dfferences from spurous correlatons. Thus, one needs both a suffcently large treatment group and a large reference populaton from whch one can draw the control group. If the premses of the matchng are met, one can nterpret the estmated effect as a causal effect of the polcy nstruments among the frms that actually receve support: the average treatment effect on the treated (ATT). Our emprcal fndngs seem to ndcate the followng: Frst, our estmates of the average effects of support are postve and sgnfcant n terms of performance ndcators related to economc growth but mostly non-sgnfcant regardng labour productvty or returns to assets. Second, there s a clear tendency that (1) the estmated effects are hgher for start-up frms than for ncumbent frms when the effects are measured as relatve effects n percentage ponts, but lower when the relatve effects are translated nto level effects, (2) hgher for the tax credt scheme and the Research Councl of Norway than for Innovaton Norway, and (3) that the effects ncrease wth the amount of support. In partcular, 2

6 support that amounts to under NOK ( Euro) 1 have lttle or no effect, whatever the polcy nstrument. The remander of the paper s structured as follows: Secton 2 surveys the exstng lterature. Secton 3 presents nformaton about the nsttutonal settng n Norway. Secton 4 presents the data. Secton 5 descrbes the matchng procedure and the general econometrc model used for the analyss. Secton 6 provdes the man results and several senstvty analyses, and dscusses the results n lght of exstng fndngs. Fnally, Secton 7 concludes. 2. Extng Lterature The ssue of whether publc R&D spendng and government subsdes are complementary and addtonal to prvate spendng or tend to crowd out prvate R&D has been dscussed n many pror studes. A crtcal survey of some early mcroeconometrc studes s gven by Klette et al. (2000), wth focus on the problem of establshng a vald control group n a non-expermental settng. Also Davd et al. (2000) are crtcal to the earler lterature and conclude that the many estmates of crowdng out effect and nput addtonalty found n the earler lterature are generally based due to selecton ssues. The more recent mcroeconometrc lterature on the effects of publc programs to spur prvate R&D, are generally more aware of and explctly address -- methodologcal problems. Examples nclude: Wallsten (2000) (U.S. frms), Duguet (2004) (French frms), Czarntzk and Lcht (2006) (German frms), Cappelen et al. (2012) (Norwegan frms), Lokshn and Mohnen (2013) (Dutch frms), Dumont (2017) (Belgan frms), and Dechezlepretre et al. (2016) (UK frms). The most commonly used emprcal framework for studyng the economc mpact of frms R&D and nnovaton actvtes s the so-called CDM model. In ther orgnal paper, Crepon et al. (1998) propose a three-stage model. Frst, they specfy a probt model of the decson to undertake an nnovaton actvty. Condtonal on a postve outcome of ths (bnary) choce, the frm chooses ts R&D ntensty and, fnally, the economc outcome varable of nterest (e.g., productvty) s analysed wthn a standard regresson framework. Later developments and applcatons of the CDM framework are revewed n Lööf et. al. (2017). Takalo et al. (2013) propose a smlarly structured approach, but where the focus s on the mpact of publc polcy: the dependent varable n the frst step s a dummy of whether the frm has a project wth publc support or not. 1 Based on the mean exchange rate of about 9 NOK per EURO durng the perod analysed. 3

7 Whle our model s consstent wth both the three-stage CDM framework and the approach of Takalo et al. (2013), we use a more reduced form framework that does not requre R&D or nnovaton data at the frm level to study the effect of R&D support on economc outcomes. In our approach, the frm decdes to apply for tax credts and/or a grant; f t s accepted, the frm undertakes the project and may thereby ncrease ts knowledge stock, whch agan may have postve effects on performance ndcators, such as economc output or productvty. The crtcal prerequste for our analyses s that our control group of frms s representatve of the counterfactual outcomes for the frms that receve support (.e., the outcomes f they had not receved support). The average dfference between the actual and the counterfactual (hypothetcal) outcomes s the treatment effect we want to estmate. 3. Insttutonal settng Snce 2002, the three man government nstruments to promote R&D and nnovaton n the busness sector n Norway have been Innovaton Norway (IN), the Research Councl of Norway (RCN), and the R&D tax credt scheme Skattefunn (SKF). IN as we know t today, was formed through a merger of the Norwegan Industral and Regonal Development Fund, the Norwegan Trade Councl, Government Consultatve Offce for Inventors and the Norwegan Tourst Board. IN's actvtes consst of dstrct programs and nnovaton applcatons, and are manly fnanced by local governments, the Mnstry of Regonal Development, and the Mnstry of Fsheres. Our evaluaton covers only the nnovaton programs, as the other IN-programs are not amed at promotng R&D and nnovaton, but gve drect support to sparsely populated areas or the agrcultural sector. The nnovaton programs nclude grants, nnovaton loans, captal loans and advce to companes to develop a new product or new technology, or promote organsatonal nnovatons. The tax credt scheme Skattefunn (SKF) was ntroduced n 2002 for small and medum frms (SMEs). The scheme was expanded n 2003 to nclude all frms. SKF s regulated by Norwegan tax law and s subject to supervson by the EFTA Survellance Authorty (ESA). Through the SKF tax credt scheme, frms get tax credts for R&D; ether tax deductons or cash refunds (see below). Only approved costs provde the bass for tax deductons. 2 From 2003, the SKF scheme granted large frms 18 percent of R&D expenses related to an approved project up to a lmt of 4 mllon NOK. From 2009 to 2013, the maxmum lmt was 5.5 mllon NOK. Hence, the maxmum tax relef for a large frm 3 2 Only projects approved by the Skattefunn dvson of the Research Councl of Norway provde a bass for tax deductons. Ths only apples to costs that the frms have ncurred n the ncome year the approval was granted. The tax authortes control the stated costs and aggregate publc support for the enterprse under the State Ad Code. 3 Snce then the lmt has been ncreased several tmes and s now NOK 25 mll. 4

8 was about one mllon NOK (about Euros). For SMEs the rate s 20 percent. The tax refund takes place the year after the actual R&D expenses have occurred (and the project was approved). If the frm s taxes are less than the refund, the remanng tax credt s gven as a drect grant (see Cappelen et al., 2010 for more detals). In fact, each year about three fourth of the total tax subsdes are gven as drect grants. 4 Whle SKF s a general nstrument, support from the Research Councl of Norway (RCN) s a selectve nstrument, where frms compete for funds. The man argument for a selectve support scheme s that the publc can drect support towards projects expected to have major postve external effects and consequently hgher socal than prvate economc return. A theoretcal bass for such project selecton s found n Jaffe (1998), who evaluates the potental for postve externaltes (spllovers), prvate fnancal returns and addtonalty. 4. Data The man data nclude nformaton about the tmng and amount of support from the three man R&D polcy nstruments: IN s nnovaton program, drect R&D subsdes from the RCN, and SKF the R&D tax credt scheme. Our data cover the perod , and are merged wth frm level regster data collected by Statstcs Norway. The data have the advantage that they are collected for publc regsterng and have unversal coverage for lmted lablty companes. Furthermore, they are scrutnsed by Statstcs Norway before beng made publc. Hence, the data are n general of a hgh qualty. We lmt the populaton of frms to lmted lablty companes (ncorporated frms), snce our performance ndcators are based on accountng nformaton. Incorporated frms contrbute to percent of value added n the market-based ndustres (the excluded ndustres are: prmary ndustres, health care and the government sector), and a roughly equal share of government R&D support. Furthermore, we exclude frms wth ther man actvty n the ndustry Research and Development (NACE 72). The reason s that frms n ths ndustry receve, drectly or ndrectly, R&D support on a regular bass. Thus, a proper control group cannot be establshed. 4 Ths share has been remarkably stable over tme. See the web-artcle (n Norwegan). 5

9 In all our analyses, we dstngush between (1) support to entrepreneural frms, defned as frms that are less than three years of age (counted from the date of ncorporaton) at project start and (2) support to ncumbent frms. Operatonalsatons One could argue that n an deal envronment one should observe both project dentfers and the outcome varables at the project level. However, n practce, all the outcome varables are avalable from regster data collected at the level of the frm who obtans support. Hence, some form of aggregaton from the project to frm level must be made. A further complcaton s that tax credts are avalable from tax accounts data and hence refer to a frm-year (a frm observed n a partcular year), not to a specfc research project. In the lack of unque project dentfers, we have to operatonalze the concept of a research project. Ths concept must be applcable to all forms of monetary support, and co-fundng of projects by multple nstruments. In partcular, we must take nto consderaton that most projects have a duraton beyond one year: the medan duraton of RCN-projects s three years, and the medan number of years wth consecutve tax refunds s two years. Moreover, the same project may get support from several sources. We proceed by makng the followng smplfyng assumptons: (1) A project s trggered by an award. (2) The project s beleved to start the year after the frst occurrng award (subsdy). (3) The project length s standardsed to three years (the normal length of projects n the RCN). (4) If a frm gets addtonal subsdes durng the project perod of three years (regardless of source), ths wll be regarded as support for the same project. (5) The overall project support ncludes the sum of all support to the frm from all fundng sources over the 3-year project perod. To classfy projects accordng to source of fundng, we dentfy the man polcy nstrument, defned as the man source of fundng durng the three-year project perod. Descrptve statstcs to be detaled below, show that the man polcy nstrument accounts for the man part of fundng at the project level. For projects wth ether RCN or IN as the man polcy nstrument, tax credts s the clear secondary source of support, whereas for projects wth SKF as man polcy nstrument, support from addtonal sources s almost neglgble. Gven our operatonalzaton of a research project, we estmate ATT at the project level for all three man polcy nstruments. Furthermore, we dstngush between whether the treatment was gven to a start-up frm (defned as beng at most 3 years old at project start-up) or an ncumbent frm, and the 6

10 amount of support that was gven: small; less than 0.5 mll. NOK, medum; between 0.5 and 1.5 mll. NOK, or large; above 1.5 mll. NOK. Thus, there are three dmensons n our reportng of ATT estmates: (1) man polcy nstrument of the project, (2) frm-age of the frm at project start-up, and (3) amount of support gven to the project. These operatonalsatons mean that, condtonal on the major source and total amount of fundng, publc support to a gven project s assumed to have the same effect regardless of the presence of secondary sources of fundng. Our approach can be seen as crcumventng the endogenety ssues that would arse f we ncluded control varables ndcatng secondary sources of fundng. The correspondng coeffcent estmates (of these control varables) could reflect the qualty and nature of the project, rather than the causal effect of havng multple sources of fundng. 5 Our approach s also justfed by Czarntzk and Lopes-Bentoa (2013), who fnd that the estmated treatment effects of a regonal R&D program n Belgum do not depend on dummy varables ndcatng the presence of a subsdy mx. The estmaton of ATT s based on the standard approach n the treatment lterature by formng a treatment group and control group by statstcal matchng. The control group ncludes frms that never get nnovaton related support from RCN, IN, or SKF. Ths group of frms conssts of a selecton of the reference populaton wth observed characterstcs smlar to the companes that receved fundng. The matchng method s a combnaton of stratfed (exact) matchng and propensty score matchng (wthn each stratum). A more detaled descrpton of the matchng procedure s deferred to Secton 5. Descrptve statstcs Total R&D support from the three nstruments to our frm-populaton durng the perod amounts to 28 bllon NOK. Table 1 reports total support from each polcy nstrument, both before and after matchng. We see from the frst four columns n Table 1 that the total amount of R&D subsdes before match s 7.9 bllon NOK for IN, 9.2 bllon NOK from RCN, and 10.9 bllon NOK from SKF, of whch a lttle less than 80 percent s covered by the frms n the matched sample (see the last four columns n Table 1). The total support s dvded nto projects that may consst of fundng from several sources throughout the project duraton. As explaned above, we classfy each project accordng to the man polcy nstrument of the project supported. For example, whle total support from IN s 7.2 bllon (before matchng), total support to projects wth IN as man polcy nstrument s 7.9 bllon. There are two sources of the (postve) dscrepancy of 0.7 bllon: (1) some of the support from IN are 5 The stuaton s smlar to attemptng to estmate the returns to nvestments usng the project sze as a control varable. The fndng that, say, large nvestments have hgher average returns than small ones, does not mply that an ncrease n a gven (small) nvestment would yeld a hgher return. 7

11 gven to projects labelled RCN- or SKF (n total 1 bllon), and (2) some of the support to INprojects come from SKF (1.4 bllon) or RCN (0.4 bllon). The net effect s 0.7 bllon. Table 1. Support n mllon NOK, by man polcy nstrument. Before and after match Support from Man polcy nstrument Before match After match IN RCN SKF Total IN RCN SKF Total IN RCN SKF Total The number of projects by year and man polcy nstrument s shown n Table 2. We note that there has been a substantal decrease n the number of SKF-projects compared to the frst years after the extenson n Regardng IN, the number of projects was nearly doubled n 2010 compared to Ths was the result of the government s fnancal crss stmulaton package. The number of projects has remaned at a much hgher level also n the years after the crss compared to the pre-crss level, reflectng an ncreased mportance of ths nstrument. In contrast, the number of RCN projects has been qute stable over tme. Table 2. Number of projects by man polcy nstrument Before match After match IN RCN SKF IN RCN SKF Total

12 From the numbers n the three last columns of Table 2 after match and the correspondng numbers n Table 1, t follows that the average amount of fundng per project wth respectvely IN, RCN, and SKF as the man polcy nstrument s: 3.4 mllon, 5.3 mllon, and 1.0 mllon. 6 Table 3. Share of project support from each polcy nstrument, by the projects man nstrument. After match Support from Man polcy nstrument IN RCN SKF IN RCN SKF Total In the rest of ths secton, we focus on the after-match sample. Frst, n Table 3, we report the share of support comng from the man nstrument versus other sources. We see that projects wth IN as the man nstrument, obtan 84 percent of total project support from IN, projects wth RCN as the man nstrument obtan 82 percent from RCN and, fnally, SKF-projects receve 91% of the support from ths nstrument. For both IN- and RCN-projects, SKF s the largest secondary source of fundng, provdng between 15 and 20 percent of total publc support. The relatvely hgh share of SKF-fundng among the RCN-project s not surprsng, as RCN-approved R&D projects are legally enttled to tax credts (wth an upper lmt at the frm level due to EEA rules). On the other hand, projects wth SKF as the man nstrument obtan a very small share of fundng from RCN (4 percent). Table 4a. Number of projects by man polcy nstrument, support amount, and frm age category. After match Man polcy nstrument Start-up frms wth support Incumbent frms wth support Support amount (mll. NOK) Support amount (mll. NOK) Small (<0.5) Medum Large (>1.5) Small (<0.5) Medum Large (>1.5) IN RCN SKF Total The calculatons are: 7408/2178=3.4, 5597/1061=5.3 and 8750/8735=1.0. 9

13 Table 4b. Share of support by man polcy nstrument, support amount, and frm age. After match Man polcy nstrument Start-up frms wth support Support amount Incumbent frms wth support Support amount Small (<0.5) Medum Large (>1.5) Small (<0.5) Medum Large (>1.5) IN RCN SKF More nformaton about the support and the recpents s gven n Tables 4a and 4b, whch categorzes the project support along three dmensons: (1) man polcy nstrument, (2) support amount (small, medum or large) and (3) frm-age (start-up or ncumbent). About one fourth of the projects are gven to start-up frms. Approxmately one thrd of the projects belong to each of the support amount categores, but wth consderable heterogenety across the man polcy nstruments. In partcular, RCN-projects have a much hgher share of large projects than IN and (n partcular) SKF. Table 5. Share of project support by ndustry, after match NACE Share Manufacturng Producton of chemcals Producton of chemcals rubber and plastc products Producton of computers and el. and optcal nstr Producton of motor vehcles 0.09 Retal trade 0.04 Informaton and communcaton 0.16 Professonal and scentfc servces 0.12 Admnstratve servces 0.03 Total 1 Table 5 provdes nformaton about the dstrbuton of support across ndustres. We see that support s hghly concentrated n a few ndustres, wth 2/3 of total support gong to Manufacturng (wth Producton of chemcals as the largest 2-dgt NACE level ndustry). Then comes Informaton and communcaton (16 %) and Professonal and scentfc servces (12 %). An almost neglgble share of the support goes to other ndustres. 10

14 5. Matchng methodology The classcal matchng estmator pars the treated frms wth a control group that s assumed to represent the counterfactual (non-treated) outcomes of the treated frms. The control group s selected based on a vector of matchng varables, S, where subscrpt denotes frm. Under certan condtons, a treated frm and the matchng frms to whch the treated frm s pared are dentcal n all respects, except a random term,. The most mportant condton s that the untreated outcome (the counterfactual outcome of a treated unt) s ndependent of treatment assgnment condtonal on S. Ths s called the Condtonal Independence Assumpton (CIA), often referred to as unconfoundedness. In our context, ths means that f a frm obtans (s assgned to) R&D support n perod T, ths assgnment s per se unnformatve about the counterfactual outcome of the dependent varable (gven S T ) n the post-treatment perod, T +1. We wll now frst consder the estmaton of a smplfed verson of our model, wth a bnary treatment ndcator, D {0,1}, assgned at a fxed pont n tme, T (whch may dffer across frms). Specfcally, we wll consder usng a combnaton of dfferencng and matchng, advocated by Blundell and Costa Das (2009). Let y t (1) and y (0) denote the dependent varable, the outcome of R&D support (treatment), when t the (same) frm,, obtans treatment and non-treatment, respectvely. We assume that y (1) f m ( S ) 1( t T ) t t t t1 y (0) f m ( S ) t t t t 0 where f s a fxed frm effect, m( S ) s a non-parametrc (unknown) common trend functon, s t t the frm-specfc treatment effect and 1( A ) s the ndcator functon whch s one f the statement A s true and zero otherwse. The vector S t conssts of a mnmal set of observable varables that makes both error terms E( t 0 St ) 0 and E( t1 St ) 0. The realzed (observed) value of yt s then yt ( D ). Thus, f D 0, nether yt (1) nor the assgnment year, T, are observed. The ncluson of the common trend functon m( S ) n the model of y ( D) s mportant as the t treatment group and the control group (the non-treated outcomes) must have the same trend. By 11 t t

15 ncludng m( S ), we mtgate the potental problem that the observed (pre-treatment) characterstcs, T t t S, whch determne the treatment assgnment may also nfluence the outcome varable, y ( ), 1 D. T Based on the above model, we can formally defne the average treatment effect on the treated (ATT): ATT E( y (1) D 1) E( y (0) D 1), T1, T1 Here E( y, T 1(1) D 1) s the expected (post-treatment) outcome for frms n the treated group,.e. those who were assgned to R&D support at tme T. Ths means that the post-treatment outcome, y (1) T,, s observable for all frms n ths group. On the other hand, 1 y (0) T, s not observed f D 1 = 1. Usng the mean outcome of the frms that do not get R&D support: E( y, T 1(0) D 0) may not be approprate for estmatng E( y, T 1(0) D 1). Ths non-nterchangeablty of, T 1 E( y (0) D 0) and E( y, T 1(0) D 1) s due to the fact that characterstcs that determne whether a frm gets R&D support are also lkely to determne the future outcome of ths frm. To deal wth ths potental effect, often referred to as the selecton effect, we combne stratfcaton and propensty score matchng. 7 The specfc motvaton for stratfcaton n our case s that cell characterstcs are key determnants of both the probablty of obtanng support, e.g. through regonal programs and programs targetng young frms, and of the dependent varables, e.g. through ndustry-specfc market condtons and local labour market condtons. 8 More specfcally, we do as follows: Frst, n any gven year t, we stratfy frms nto ndustry regon age cells ( j, r, s ) consstng of frms that belong to 2-dgt NACE ndustry j, regon r and age group s (1-3, 4-6, 7-9, or >9 years old). Next, wthn these ndustry-regon-age specfc cells, we construct a contnuous matchng varable, S t, whch n our applcaton s a measure of the frms sze (total assets). Now, wthn each stratum we use propensty score matchng to match treated and non-treated frms usng the matchng varable, S t. The probablty of treatment gven S : P ( S ) Pr( D 1 S ) s T T T 7 See the semnal contrbuton by Rosenbaum and Rubn (1983) who effectvely reduced the mult-dmensonal matchng problem to a unvarate one, by matchng on the probablty of treatment gven S : P ( S ) Pr( D 1 S ). T T T 8 A general dscusson of the pros and cons of matchng wth stratfcaton are dscussed n Calendo and Kopeng (2008). 12

16 the so-called propensty score. Moreover, the log-odds P( S ) / (1 P( S )) s a non-lnear functon of S t, specfed as a pecewse lnear splne. The knk ponts of the splne are located at the quartles of the (cumulatve) emprcal dstrbuton of the sze varable (specfc to each strata). T T The propensty score we estmate s a transton probablty: the probablty of transton at T from prevously havng had no support, to obtanng frst-tme support from ether IN, RCN, or SKF. The estmated propensty score of each frm n the group of treated frms s then matched wth one or more of the nearest neghbours. See Appendx A for techncal detals. The matchng procedure used s the STATA routne psmatch2 wth 1 to 5 nearest neghbour matchng wth trmmng. 9 The combnaton of stratfcaton and propensty score matchng yeld a sample of comparable matched frms wth an approxmately balanced dstrbuton of the observed characterstcs,.e. when we compare ths dstrbuton for the group of frms recevng R&D support and the group of matched frms not recevng such support. By further combnng ths matchng procedure wth a dfference-ndfference approach (DID), we are able to control for unobserved frm specfc effects, f. 10 The classcal DID estmator when appled to a matched sample, can be expressed as: 1 1 DID T y y # N T N # ( ), T1 j, T1 C jc () where T N s the set of treated frms, C () s the control group of frms matched to frm T N and #A denotes the number of elements n (any set) A. The estmaton strategy s to contrast each posttreatment outcome, y,, n the treatment group, wth the average outcome n the control group 1 T C () (the frms matched to the treated frm ): 9 We use the command: psmatch2 common trm(10). Ths mposes a common support by droppng treatment observatons whose propensty score s hgher than the maxmum or less than the mnmum propensty score of the controls. It also drops 10 percent of the treatment observatons at whch the propensty score densty of the control observatons s lowest. See Leuven and Sanes (2003) for documentaton. 10 It s an ongong debate n the lterature whether one would beneft from combnng the two approaches; see for nstance Blundell and Costa Das (2009), Imbens and Wooldrdge (2009), Lechner (2010), and Chabé-Ferret (2015). We emphasse the argument used by Blundell and Costa Das (2009, p. 604); the combnaton of matchng wth DID as proposed n Heckman et al. (1997) can accommodate unobserved determnants of the nontreated outcome affectng partcpaton for as long as these are constant over tme. 13

17 1 y # C ( ) jc () jt, 1. The DID estmator s gven as the average over these contrasts (dfferences) over all the treated frms T N. Above we have assumed that any effect of a treatment assgned n T s realzed mmedately afterwards (n T 1 ). However, n our applcaton we have annual data (t s a calendar year), whle treatment effects are naturally defned over longer tme ntervals, from project start ( T ) to project end or later ( T k ). If k 3(the project length), the average treatment effects of the treated s k, 3 modfed as follows: 3 k 1 ATT E( y N ) E E( y j C( )) N T, Tk j, Tk T (cf. the expresson for DID above). To estmate ATT we apply a regresson formulaton of the DID estmator (see Lechner, 2010), rather than the classcal DID estmator stated above. The regresson formulaton s: y t y T m 1( T t T 3) / 3, for N, C( ), t T t1 m, for jc( ). jt C( ), t jt0 where the error term s assumed to have a movng average (MA(q)) dstrbuton. The dvson by 3 n the expresson above means that T can be nterpreted as the 3-year change n yt nduced by the treatment. The treatment effects at the frm-level ( ) are allowed to be project-specfc, mplyng T ATT E N. Note that the expressons for yt nclude nteracton terms between tme T dummes and cell membershp, where the (cell-specfc) common trend s: mc ( ), t E( mt ( S jt )) j C( )). T 14

18 As ponted out by Lechner (2010), there are several advantages of the regresson formulaton of the DID dentfcaton and estmaton problem and, n fact, no dsadvantages when control varables are not ncluded (bd p. 195), as n our case. The frst advantage s the easness of obtanng the fnal estmates and ther standard errors. Second, the regresson formulaton naturally extends to an arbtrary choce of treatment nterval. For example, we wll examne long-term effects from project end to three years after project end,.e. from T 3 to T 6. Thrd, the regresson formulaton easly accommodates more treatments. Ths s mportant to us, as many frms do obtan repeated support. To accommodate repeated treatments, we replace T wth the year of the k th treatment assgnment n the regresson equaton. Note that, by defnton, a new project cannot overlap wth the (1) (2) precedng project: 1 T 3 T ( ) k T 6. Emprcal results Assessng the matchng qualty The comparsons n Table 6 show that the medan values of the outcome varables for the treatment and the control groups are smlar at the tme of matchng (for the matched non-treated frms we do not separate between amount of support gven to the correspondng treated frms). Ths ensures that the balancng propertes of the matchng hold. Fgure 1 depcts the dstrbutons of the estmated propensty scores n the treated and matched group of frms after the matchng. The extreme values (mnmum and maxmum) of the propensty scores are trmmed to ensure that the common support condton s met. The results ndcate that the dstrbutons of the treated and the control groups are very smlar. Estmates of relatve ATT In our applcaton, the dependent varable, y t,denotes ether: (1) log of sales revenue, (2) log of number of employees, (3) log of value added, (4) log-labour productvty (value added per employee), or (5) return on total assets. Thus, yt s approxmately equal to the annual relatve change n y. t 15

19 Table 6. Medan values of outcome varables at the year of matchng, by amount of project support and man polcy nstrument Treated group Control group Before matchng After matchng Matched frms Support amount (mll. NOK) Support amount (mll. NOK) Man polcy nst. Outcome varable Age Category IN No. of employees IN Value added per employee IN Return on total assets RCN No. of employees RCN Value added per employee RCN Return on total assets SKF No. of employees Small (<0.5) Medum ( ) Large (>1.5) Small (<0.5) Medum ( ) Large (>1.5) Start-up Incumbent Start-up Incumbent Start-up 3 % 0 % -2 % 2 % -1 % -6 % 0 % Incumbent 6 % 6 % 5 % 6 % 3 % 5 % 6 % Start-up Incumbent Start-up Incumbent Start-up -9 % 4 % 2 % -6 % 7 % -1 % 1 % Incumbent 7 % 11 % 6 % 9 % 14 % 3 % 8 % Start-up Incumbent SKF RCN Value added per employee Start-up SKF Return on total assets Incumbent Start-up 8 % 7 % 4 % 11 % 7 % 4 % 5 % Incumbent 11% 11 % 9 % 11 % 12 % 9 % 9 % 16

20 Fgure 1: Dstrbuton of propensty scores among treated and matched frms Tables 7-10 show the estmated treatment effects n percentage ponts over the three-year perod from project start, T, to T 3,.e. the average three-year growth dfference between the treated and matched frms. When presentng the results below, we dstngush between dfferent treatment groups accordng to the followng characterstcs of the supported project: (1) man polcy nstrument of the project (IN, SKF or RCN); (2) start-up frm (maxmum three years old at project start) or ncumbent frm; and (3) amount of project support (small, medum or large). Ths yelds n total possble combnatons. Each of these combnatons consttute a specfc treatment group, T N. To smplfy notaton, ATT wll denote the average treatment effect regardless of what treatment group s consdered (ths wll be clear from the context). Table 7 presents average treatment effects (ATT) for each combnaton of man polcy nstrument and age group and does not dstngush between amounts of support. Frst, let us clarfy the nterpretaton of the fgures n the table by focusng on an example: growth n the number of employees for start-up frms recevng support from IN. Accordng to Table 7, such frms wll have 21.4 percentage ponts addtonal ncrease n headcount three years after project start compared wth not recevng any support. Ths s a statstcally sgnfcant estmate at the 0.1 percent level (as ndcated by ***). For 17

21 ncumbents, the correspondng estmated addtonal ncrease (ATT) s much smaller 2.4 percentage ponts whch s not sgnfcant even at the 5 percent level. Table 7. Estmated ATT, by man polcy nstrument and frm age category. Three-year dfferences n percentage ponts 1 Outcome ndcator Age category Man polcy nstrument IN RCN SKF Effect z Effect z Effect z Sales Start-up *** * *** 7.56 Incumbent 8.56 * * *** No. of employees Start-up *** ** *** 4.92 Incumbent *** 6.15 Value added Start-up ** *** 6.39 Incumbent ** *** 9.09 Labor productvty Start-up * ** 2.91 Incumbent *** 4.34 Return on assets Start-up Incumbent Addtonal growth n percentage ponts durng the three-year perod from project start (year T ) to project end ( T +3). *, ** and *** denote sgnfcant estmates at 5, 1 and 0.1 percent level, respectvely There are three man takngs from the numbers n Table 7. Frst, sales growth s the only ndcator wth sgnfcant estmates across all the polcy nstruments and age groups. Second, all polcy nstruments lead to sgnfcant ncreases n employment among start-up frms. Thrd, none of the nstruments mproves returns on assets. Lookng at the results n Table 7 n more detal, we note the followng: (1) The estmates for start-up frms across all the man polcy nstruments ndcate sgnfcant estmates of percentage ponts ncrease n sales revenue (over three years), and percentage ponts ncrease n employment. (2) The estmated ATT for ncumbent frms s sgnfcant for all the nstruments regardng sales revenue (8-16 percentage ponts estmated ncrease), but only support from SKF has a sgnfcant postve effect on employment (5 percentage ponts estmated ncrease). (3) Comparng start-up frms vs. ncumbent frms, support from SKF leads to approxmately 25 vs. 10- percentage ponts ncrease n value added and 10 vs. 5-percentage ponts ncrease n labour productvty. (4) The correspondng results wth regard to support from RCN are of a smlar magntude as for SKF, but less sgnfcant. (5) Support from IN does not seem to have any sgnfcant effects on ncumbent frms, although the estmates for value added are close to beng sgnfcant at the 5 percent level. 18

22 The man polcy nstruments Tables 8-10 present results for each of the three man polcy nstruments along two dmensons: (1) amount of support (small less than 0.5 mll. NOK, medum between 0.5 and 1.5 mll. NOK, or large above 1.5 mll. NOK) and (2) age-group. Table 8. Estmated ATT for Innovaton Norway, by frm age category and amount of support. ) Three-year dfferences n percentage ponts 2 Outcome ndcator Age category Support amount (mll. NOK) Small (<0.5) Medum ( ) Large (>1.5) Effect z Effect z Effect z Sales Start-up *** 7.33 Incumbent *** 3.58 No. of employees Start-up *** 7.71 Incumbent Value added Start-up * 2 Incumbent Labor productvty Start-up * Incumbent Return on assets Start-up 5.80 * Incumbent Projects wth IN as man polcy nstrument. 2 Addtonal growth n percentage ponts durng the three-year perod from project start (year T ) to project end ( T +3 ). Note: *, ** and *** denote sgnfcant estmates at 5, 1 and 0.1 percent level, respectvely Results for IN-frms are reported n Table 8. We see that small or medum amounts of support have a margnal or even non-exstng effect, whle large amounts of ad gven to start-up frms ncrease sales revenue, employment and value added wth respectvely 88, 47 and 26 percentage ponts accordng to our estmates. The results for ncumbent frms are generally nsgnfcant also for large amounts of support, except wth respect to sales revenue (the estmated 18 percentage ponts of addtonal growth s sgnfcant at the 0.1 percent level). Table 9 reports the effects on the RCN-supported projects. Agan, we fnd that small or medum amounts of support have only a margnal or non-exstng effect. For large amounts gven to start-up frms, the results show strong and sgnfcant effects for the outcome ndcators sales revenue, number of employees, value added and labour productvty (more than 40 percentage ponts addtonal growth on all these ndcators). On the other hand, for ncumbent frms the estmates are generally nsgnfcant also for large amounts of support. The only excepton s wth respect to value added, where the estmated ATT of 12 percentage ponts s sgnfcant at the 5 percent level. 19

23 Table 9. Estmated ATT for Research Councl of Norway, by frm age category and amount of support. 1 Three-year dfferences n percentage ponts 2 Outcome ndcator Type of frm Support amount (mll. NOK) Small (<0.5) Medum ( ) Large (>1.5) Effect z Effect z Effect z Sales Start-up * 2.31 Incumbent * No. of employees Start-up *** 4.29 Incumbent Value added Start-up *** 3.88 Incumbent * 2.38 Labor productvty Start-up ** 3.11 Incumbent Return on assets Start-up Incumbent Projects wth RCN as man polcy nstrument. 2 Addtonal growth n percentage ponts durng the three-year perod from project start (year T ) to project end ( T +3 ). Note: *, ** and *** denote sgnfcant estmates at 5, 1 and 0.1 percent level, respectvely Table 10 reports effects of the tax credt scheme SKF. Here we fnd statstcally hghly sgnfcant effects of large subsdes wth respect to all outcome ndcators, except return on total assets, where there s no effect n the case of ncumbent frms. There also seem to be some statstcally sgnfcant benefts of small (<0.5 mllon) and medum amounts of support ( mll.) of ths fundng alternatve, although there s a clear tendency that the effects ncrease wth the amount of support gven. Agan, support gven to start-up frms yelds hgher addtonal growth n percentage ponts than support gven to ncumbent frms. To examne the robustness of our results, Table 11 presents correspondng results as n Table 7 over the three-year perod from project end to 3 years later (that s, from year T 3 to year T 6 ). These long-term effects are generally close to zero and nsgnfcant, although there are some estmated effects that are sgnfcant at the 1 or 5 percent level. To summarze, there s no clear tendency for the effects to appear after the three-year project nterval. Equally mportant s that we fnd no tendency of mean reverson: that gans acheved durng the frst 3-year nterval are reversed durng the next three years. 20

24 Table 10. Estmated ATT for SKF, by frm age category and amount of support. 1 Three-year dfferences n percentage ponts 2 Effectndkator Type foretak Support amount (mll. NOK) Small (<0.5) Medum ( ) Large (>1.5) Effect z Effect z Effect z Sales Start-up *** *** 7.06 Incumbent 7.81 ** *** *** 9.94 No. of employees Start-up ** *** 5.06 Incumbent 3.32 * * *** 6.68 Value added Start-up ** *** *** 5.66 Incumbent 8.20 *** *** *** 7.16 Labor productvty Start-up * ** 2.65 Incumbent 4.52 ** * * 2.34 Return on assets Start-up * 2.4 Incumbent Projects wth SKF as man polcy nstrument. 2 Addtonal growth n percentage ponts durng the three-year perod from project start (year T ) to project end ( T +3 ). Note: *, ** and *** denote sgnfcant estmates at 5, 1 and 0.1 percent level, respectvely Table 11. Estmated ATT measured from 3 to 6 years after project start, by man polcy nstrument and frm age category. 1 Three-year dfferences n percentage ponts Man polcy nstrument Outcome ndcator Age IN RCN SKF category Effect z Effect z Effect z Sales Start-up Incumbent No. of employees Start-up * Incumbent * Value added Start-up Incumbent ** Labor productvty Start-up Incumbent Return on assets Start-up * Incumbent Addtonal growth n percentage ponts durng the perod from project end ( T 3 ) to three years later ( T 6 ). Note: *, ** and *** denote sgnfcant estmates at 5, 1 and 0.1 percent level, respectvely From percentage ponts to level effects To say somethng about the estmated effects when converted nto level effects, we attempt to estmate the mpact per mllon NOK n project support for a representatve frm for each of the 12 treatment groups. How to defne such a frm s, however, far from obvous. One possblty s as the medan frm 21

25 Table 12. Characterstcs of the representatve frms n each treatment group 1) Man polcy nst. Outcome ndcator IN No. of employees Age category IN Value added per employee IN Return on total assets RCN No. of employees RCN Value added per employee RCN Return on total assets SKF No. of employees SKF RCN Value added per employee SKF Return on total assets Before matchng Support amount (mll. NOK) Small (<0.5) Medum ( ) Large (>1.5) After matchng Support amount (mll. NOK) Small (<0.5) Medum ( ) Large (>1.5) Start-up Incumbent Start-up Incumbent Start-up Incumbent Start-up Incumbent Start-up Incumbent Start-up Incumbent Start-up Incumbent Start-up Incumbent Start-up Incumbent ) Weghted average over frms wthn each treatment group at project start, wth weghts proportonal to amount of support n each treatment group, as defned by the medan values descrbed n Table 6. The weakness of ths approach s that equal weght s gven to all the frms n a gven treatment group (e.g. n the group of start-up frms, wth small amount of support and SKF as man polcy nstrument), regardless of how much support each frm n that group receved. Therefore, we have chosen to construct a representatve frm wthn each treatment group as a weghted average frm (at project start-up) where 22

26 the weght s proportonal to the amount of support gven to the project. The characterstcs of the representatve frms are shown n Table 12. If we compare the representatve frms reported n Table 12, wth the medan frms reported n Table 6, we see that the former s much larger as measured by number of employees (large frms get more support). Ths apples to all polcy nstruments, especally for frms wthn the treatment groups wth large amounts of support. We further see that frms wth large IN-funded projects score lower on the outcome ndcators value added per employee and return on total assets than do frms wth RCNand SKF-projects. To estmate the level effect (or return ) to a gven support scheme, we ntally estmate level effects per mllon n project support to the representatve frm wthn each treatment group. Ths s done by combnng the percentage ponts effect estmates from Tables 8-10 wth the ntal characterstcs of the representatve frm n each treatment group (see Table 12). Fnally, gven these level-estmates of treatment effects, we calculate the weghted-average level-effect for each man polcy nstrument as follows: the estmated level-effect n each category s weghted wth ts share of total amount of support, as reported n Table 4b. All these calculatons were done separately for ncumbent and startup frms. We can then nterpret the result as an expresson of the return on a representatve project portfolo consstng of ether ncumbent or start-up frms for the gven man polcy nstrument (6 portfolos n total). Each of the portfolos (e.g. support to start-up frms by IN) then conssts of a mllon NOK beng allocated to small, medum and large projects n accordance wth the sx portfolospecfc dstrbutons n Table 4b. The fnal results are shown n Table 13. Table 13. Estmated effects n levels (numbers or NOK) per mllon NOK n project support. Three years after project start, by man polcy nstrument and frm age category Outputndcator No. of employees Age category IN RCN SKF Effect Conf.ntervall 1) Effect Conf.ntervall 1) Effect Conf.ntervall 1) Start-up Incumbent Value added (n mll. NOK) Start-up Incumbent Value added per employee (n 1000 NOK) Start-up Incumbent Lower and upper boundary n 95 % confdence nterval 23