SME Policy and Firms Productivity in Latin America

Transcription

1 DISCUSSION PAPER SERIES IZA DP No SME Polcy and Frms Productvty n Latn Amerca Pablo Ibarrarán Alessandro Maffol Rodolfo Stucch October 2009 Forschungsnsttut zur Zukunft der Arbet Insttute for the Study of Labor

2 SME Polcy and Frms Productvty n Latn Amerca Pablo Ibarrarán Inter-Amercan Development Bank and IZA Alessandro Maffol Inter-Amercan Development Bank Rodolfo Stucch Inter-Amercan Development Bank Dscusson Paper No October 2009 IZA P.O. Box Bonn Germany Phone: Fax: E-mal: za@za.org Any opnons expressed here are those of the author(s) and not those of IZA. Research publshed n ths seres may nclude vews on polcy, but the nsttute tself takes no nsttutonal polcy postons. The Insttute for the Study of Labor (IZA) n Bonn s a local and vrtual nternatonal research center and a place of communcaton between scence, poltcs and busness. IZA s an ndependent nonproft organzaton supported by Deutsche Post Foundaton. The center s assocated wth the Unversty of Bonn and offers a stmulatng research envronment through ts nternatonal network, workshops and conferences, data servce, project support, research vsts and doctoral program. IZA engages n () orgnal and nternatonally compettve research n all felds of labor economcs, () development of polcy concepts, and () dssemnaton of research results and concepts to the nterested publc. IZA Dscusson Papers often represent prelmnary work and are crculated to encourage dscusson. Ctaton of such a paper should account for ts provsonal character. A revsed verson may be avalable drectly from the author.

3 IZA Dscusson Paper No October 2009 ABSTRACT SME Polcy and Frms Productvty n Latn Amerca * Very lttle s known about the effectveness of SME polces, and a careful look at the structure, mechansms and ncentves provded by these polces suggest cauton n ther mplementaton and, most mportantly, the need to carefully and closely montor ther results. Ths paper reles on the mcroeconometrc analyss of a homogeneous dataset of sxteen Latn Amercan and Carbbean countres to analyze the magntude and determnants of the productvty gap between large and SME frms and to smulate of the mpact on productvty of varous polcy scenaros. JEL Classfcaton: D24, L53, L60, O38, O54 Keywords: SMEs, SME polcy, productvty, Latn Amerca Correspondng author: Pablo Ibarrarán Offce of Strategc Plannng and Development Effectveness Inter-Amercan Development Bank 1300 New York Avenue Washngton, DC USA E-mal: pbarraran@adb.org * The vews expressed n ths paper are those of the authors and do not necessarly represent those of the Inter-Amercan Development Bank.

4 1 Introducton In the last decades, governments have spent an ncreasng amount of resources n polces to support small and medum enterprses (SMEs). The extended use of these polces may be justfed based on some of the challenges that SMEs face and that could hnder ther productvty and growth. However, very lttle s known about the effectveness of these polces, and a careful look at the structure, mechansms and ncentves provded by SME polces suggest cauton n ther mplementaton and, most mportantly, the need to carefully and closely montor ther results. As n most publc polces, strong opnons exst n favor and aganst SME support programs. On one hand, t has been argued that SMEs employ a large proporton of the populaton and therefore by subsdzng SMEs more employment wll be created. It has also been argued that SME polces can promote growth and reduce poverty. Defenders of these polces argue that SMEs are more flexble, more dynamc, and more productve than large frms, even though they are usually constraned by specfc market falures, whch requres specfc polces. On the other hand, detractors pont out that these polces generate dstortons that affect the allocaton of resources and affect the sze dstrbuton of frms, and thus may have a negatve effect on productvty. Also, those aganst the wdespread use of SME polces pont to the lack of documented results on the effectveness of such nterventons. The relatonshp between sze and productvty s of partcular nterest n Latn Amerca and Carbbean (LAC) countres. There s a wde consensus that the dfference n the growth rate of the GDP per capta between LAC countres and the rest of western economes s due to low productvty growth rates (see Cole et al 2005). Ths fact, plus the fact that a large percentage of establshments are small but they account for a relatvely small share of producton, calls for a deeper understandng of the relatonshp between sze and productvty and of the role played by SME polcy n the regon. In ths paper we dscuss the relatonshp between frm sze and productvty n LAC and provde a framework to gude the analyss of the ratonale, structure and potental mpacts of the SME polces on the aggregate level of productvty. Ths analytcal framework assumes that SME polces mght affect the aggregate level of productvty through two channels: frst, SME polces can affect aggregate productvty by drectly nfluencng the productvty of SMEs; second, f SMEs and large frms dffer n productvty (and there s ample evdence that they do sgnfcantly), SME polces can nfluence aggregate productvty through the reallocaton of resources across SME and large frms. We perform three emprcal exercses usng comparable data for sxteen LAC countres. We frst compare the productvty of SMEs wth that of large frms. Second, we study how the varables that can be affected by SME polcy such as tranng, access to credt, product nnovaton, and ISO certfcaton affect the productvty gap. Fnally, the thrd exercse 2

5 smulates the effect of several polcy scenaros on the productvty of SMEs and aggregate productvty. We use plant level data for sxteen LAC countres from the World Bank Enterprse Surveys (WBES) dataset. The prncpal advantage of WBES s that t provdes homogeneous nformaton for all the countres. The use of comparable mcro data for all the countres s an mportant advantage wth respect to prevous studes n the regon. Those studes used mcro level data for only a small set of countres or aggregate data constructed from dfferent sources that n many cases were not comparable between countres. In most of the LAC countres, WBES provdes us wth a cross-secton of plants for the perod However, for some countres the nformaton s for the perod 2003 and for a reduced group of countres 2003 and Because of the lack of panel data nformaton, we estmate plant s productvty usng non-parametrc technques that assume constant returns to scale (CRS). Ths assumpton can be problematc when comparng frms of dfferent scale. To overcome ths lmtaton we analyze the robustness of our results by consderng dfferent values of the parameter of returns to scale. As an addtonal robustness check, we estmate the producton functon wthout assumng CRS. Ths would be the preferred method but wth cross-secton data t s hard to fnd proper nstruments for the nputs. The fndngs reported here show that, as expected, large enterprses tend to be more productve than SMEs. On the determnants of ths productvty gap, some key assocated factors are the dfferent levels of use of credt, use of tranng, ntensty of nnovaton and qualty certfcaton, all varables related to frms acquston of mproved technologes Unfortunately, data constrants precludes dstngushng whether the postve correlaton between productvty and these factors s drven by selecton of more productve frms nto these actvtes or whether t reflects the effect of these actvtes on the productvty of SMEs. Fnally, a mcro-smulaton analyss shows that properly targeted SME polces mght have a sgnfcant postve effect on aggregate productvty, mostly because of reallocaton of resources toward more productve frms. However, the results of the smulaton also suggest that the same polces not specfcally targeted at SMEs, but appled to all frms, mght produce a hgher ncrease of aggregate productvty. The man contrbuton of ths study to the lterature on SMEs and productvty n LAC s to systematcally reexamne the topcs for a large number of LAC countres usng a consstent methodology and comparable data. Ths work s related to Van Besebroeck (2005) who studed growth and productvty growth n Afrcan manufacturng wth partcular focus on frms s sze. For LAC countres the evdence s scarce and manly focused n few countres (Peres and Stumpo, 2000; Ledholm, 2002; Alvarez and Cresp, 2003, OECD 2007, and Goldberg and Palladn, 2007). The rest of the paper s organzed as follows. Secton 2 dscusses the objectves, ratonale, and nstruments of the SME polcy. It also presents a revew some of the SME polces appled n LAC countres. Secton 3 presents the relatonshp between frm sze and productvty. Secton 4 presents the dataset. Secton 5 presents evdence on the productvty gap between SMEs and large frms. Secton 6 dscusses the determnants of the gap. Secton 7 smulates several polcy scenaros. Secton 8 dscusses the evdence on the effectveness of the SME polcy. Fnally, secton 9 concludes. 3

6 2 The SME Sector and SME Polces: Objectves, Ratonale, and Instruments 2.1 The SME sector n Latn Amerca and the Carbbean The relatve sze of the SME sector vares across countres and t s an endogenous characterstc of each country. The endowment of natural resources, technology, polces, and nsttutons help determne a naton s ndustral composton and optmal frm sze. For nstance, some countres may have endowments that gve the country a comparatve advantage n the producton of goods that are produced effcently n large frms, whle other countres may have a comparatve advantage n goods produced more effcently n small frms (You, 1995). Smlarly, countres that are open to nternatonal trade may have larger optmal frm sze than countres that are less ntegrated nternatonally (Caves et al. 1980). In addton, the frm sze dstrbuton can also be affected by economc polcy. For example, smplfed tax regmes for SMEs may affect ther ncentves to grow, as frms may not fnd t proftable to move to the standard tax regme. Internatonal comparsons of the sze of the SME sector are affected by the partcular defnton of SME. SMEs are those frms wth a quanttatve measure, e.g. number of employees, lower than a certan threshold. However, not all the countres use the same quanttatve measure. Moreover, those that use the same quanttatve measure do not necessarly use the same threshold. The quanttatve measures used LAC countres nclude number of employees, sales, assets, and a combnaton of these crtera. The most commonly crtera used to defne SMEs are number of employees and monthly or annual sales. In general, the upper lmt of the number of employees used n the SME defnton n LAC countres s lower than n EU countres and US. The exceptons are the manufacturng sectors n Brazl wth a lmt of 500 equal to the one n US- and Mexco wth a lmt of 250 employees equal of the one n the EU. 4 [TABLE 1 ABOUT HERE] As shown n Table 1, the largest fracton of small frms s composed of very small frms, those wth less than fve or ten employees, and that consttute the mcro frms n the context of the MSME sector. Because data constrants our emprcal analyss focuses on the SME sector, namely those wth fve or more employees. 4 Accordng wth the EU defnton, a SME s a non-subsdary and ndependent frm that employs less than 250 employees and has an annual turnover lower than EUR 50 mllons or an annual balance sheet lower than EUR 43 mllons. 4

7 2.2 SME Polces and Instruments SME polces are those targeted at frms below a certan sze. Wth ths broad defnton, a wde range of polces falls n ths category. In general, almost all LAC countres have a smplfed tax regme or dfferental labor regulatons for SMEs, as well as programs to facltate access to credt and a set of subsdes and servces amed at supportng SMEs n dfferent aspects of ther actvtes. Ths chapter centers on those polces amed at ncreasng frms productvty, most often through the promoton of tranng, nnovaton, and qualty certfcaton. Although access to credt has been analyzed n another chapter of the report, t wll also be dscussed here as t s an nstrument wdely adopted n SME polces. Most of SME polces menton as ther goal the achevement of hgher rates of economc growth and the reducton of poverty. 5 However, they often have other objectves, such as employment, n addton to the two ones mentoned. In general those objectves are not clearly specfed. Addtonally, when desgnng and mplementng SME polces, polcy makers are often faced wth restrctons and ncentves that go beyond concerns about market falures, and that are more related to poltcal cycles, equalty, and other concepts n the realm of poltcal economy. These features should be acknowledged when examnng the coherence of SME nterventons. From an economc pont of vew, there are two broad sets of justfcatons for SME polces. The man economc justfcaton s the exstence of market falures that mght stunt growth of SMEs, whch n turn can lead to lower aggregate growth. In ths context, publc nterventons targeted at SMEs are justfed n terms of addressng market falures that affect SME, such as asymmetrc and/or ncomplete nformaton, non-convextes, and local externaltes. For nstance n the fnancal market, banks often fal to accurately assess the rsk of lendng to SME and therefore reject some proftable projects and compromse the survval and growth of many promsng frms. Therefore, by removng nformaton asymmetres that preclude access to fnancng, economc polcy may enhance the growth of the most effcent frms and a net gan n terms of productvty. In some cases owners of small frms do not have nformaton on the prvate benefts of takng certan courses of acton, e.g., offerng tranng to ther employees or obtanng external advce from specalsts or consultant, llustratng nstances of ncomplete nformaton. In other cases, those actons are not taken because of scale problems, suggestng falures related to ndvsbltes and non-convextes. In ths context, SME effcency may be constraned by the ncapacty or unwllngness of crucal supplers to scale down ther servces to meet the demand of smaller frms. The most common example s perhaps the provson of small loans by credt nsttutons, but the concept could be expanded to several servces that mply an ntal assessment of the costumer needs and characterstcs (such, for nstance, techncal assstance and tranng). In these cases, the fxed costs of provdng the servces do not sgnfcantly decrease wth the sze of the clents, whle the revenues decrease sgnfcantly, makng the provson of servces to smaller frms hghly unproftable. Ths 5 See Ayyagar et al. (2007). 5

8 argument has frequently been used to justfy publc polces that am at supportng the coordnaton efforts that the SMEs need to undertake n order to organze a jont demand for servces. Assocatvty, networkng and clusterng polces, for nstance, have often been justfed for ths reason, among others. 6 The other set of justfcatons for SME polces relate to the (presumed or real) characterstcs of SMEs that are deemed partcularly desrable. Frst, t s argued that SMEs enhance competton and entrepreneurshp and hence have external benefts on economywde effcency, nnovaton, and aggregate productvty growth. These externaltes would justfy publc support to SMEs. Second, SMEs are portrayed as labor-ntensve, thus an expanson of the SME sector would boost employment more than an expanson of large frms. Fnally, proponents clam that SMEs may be effcent frms n an early stage of development, whose growth s constraned by nsttutonal falures that favor large frms. There are also skeptcal vews questonng the effcacy of SME support polces. Frst, some authors stress the advantages of large frms, for example, argung that large frms beneft from economes of scale and create more stable and hgher qualty jobs. Second, some research fnds that SMEs are nether more labor ntensve, nor better at job creaton than large frms. Thrd, other authors queston the valdty of consderng frm sze as an exogenous determnant of economc growth. Accordng wth ths vew, pro-sme subsdes could actually dstort frm sze and potentally hurt economc effcency. A fourth crtc vew stresses the mportance of mprovng the busness envronment for all sze of frms. Low entry and ext barrers, well defned property rghts, and effectve contract enforcement characterze a busness envronment that s conducve to competton and prvate commercal transactons. Whle these factors may encourage SMEs, the focus of the busness envronment vew s not on SMEs per se; t s on the envronment facng all busnesses. Fnally, a ffth skeptcal vew of SME polcy (Levne, 2005) argues that these polces are probably most needed where they are less lkely to succeed: f SME face nsttutonal obstacles due to some sort of regulatory capture, t s very lkely that SME programs wll also be captured. Although ths debate s far from reachng any consensus, as a matter of fact many LAC Governments have over tme ntroduced an ncreasngly complex set of SME polces. Fgure 1 summarzes some SME polces appled n Latn Amerca aggregated n terms of the varable they am to affect. The nstruments commonly used by the SME polcy are credt, matchng grants, techncal assstance, and fscal ncentves. One mportant characterstc of the SME polcy s that there are many programs and t s not possble to 6 The emergence of hghly compettve clusters and ndustral dstrcts often composed by SMEs or by combnatons of large frms and small provders has ntroduced new justfcatons for polces targeted at SMEs. In ths case, the local concentraton of SMEs specalzed n a specfc producton has been seen as a potentally effcent combnaton of producton scale and flexblty, whch, however, requres agan mportant coordnaton efforts by the frms nvolved n rather complex producton systems. In ths case, specfc polces have been advocated to support the process through whch SMEs dentfy and fnance jont actvtes amed at mprovng the cluster systemc effcency and at dealng wth local externaltes (for examples on LAC, see Petrobell and Rabellott 2004). 6

9 nclude all of them n one fgure and therefore some countres and polces were selected. [FIGURE 1 ABOUT HERE] 3 The relatonshp between sze and productvty SME Polcy affects the aggregate level of productvty through two channels. Frst, SME polces can affect aggregate productvty by drectly nfluencng the productvty of SMEs and, thus, the productvty gap between SMEs and large frms. Second, f SME and large frms dffer n productvty, SME polces can nfluence aggregate productvty through the reallocaton of resources across SME and large frms. The emprcal framework below decomposes aggregate productvty n a country nto these two channels. The framework takes nto account that SMEs mght be more or less productve than large frms and that aggregate productvty also depends on how resources are allocated between these two groups of frms. Therefore t s thus well suted to examne the mportance of SMEs for aggregate productvty through these two channels. The aggregate level of productvty n perod t, P t, s gven by P t = N ω tpt, (1) = 1 where P t and ω t are the market share and the productvty level of frm n perod t. The market share s gven by the rato between frm s sales Yt and the total sales Y t defned as the sum of frms s sales n perod t. For smplcty we are currently performng productvty decomposton for the economy as a whole wthout focusng on frms' ndustry afflaton. 7 We are nterested n evaluatng the mpact of the SME Polcy, so we can decompose P t nto productvty contrbuton attrbuted to SME frms and large frms as follows Ths expresson can be wrtten as. (2) P = ω P + ω P t t t t t SME Large 7 When dfferent ndustres are consdered ths should be done by ndustry and after that the data needs to be weghted by the ndustry shares. Ths pont s mportant when workng wth sample data. In a stratfed random sample, all populaton unts are grouped wthn homogeneous groups and smple random samples are selected wthn each group. Ths method allows computng estmates for each of the strata wth a specfed level of precson whle populaton estmates can also be estmated by properly weghtng ndvdual observatons. 7

10 Y Y Y Y P = P + P, SME, t t Larg e, t, t t t t Yt SMEYSME, t Yt LargeYLarg e, t Y P = ω P + ω P, wth Ps t = Y then t SME, t SME, t Larg e, t Larg e, t Gven that ω, + ω arg, = 1, SME t L e t P t, t, s s, t Y s, t s, t = Yt ω and s = SME, L. ( ) P = P + ω P P. (3) t Larg e, t SME, t SME, t Larg e, t Ths expresson clearly shows that the level of productvty depends on the level of aggregate productvty of each sze group and the contrbuton to aggregate producton of the SME sector. Therefore, changes n the aggregate level of productvty ( P t ) can result from changes n the productvty of each sze group (.e., changes n P SME, t and P L arg e, t ) or changes n the allocaton of resources between sze groups (.e., changes n ω SME, t ). Polces amng to support SME may affect both P SME, t and ω SME, t. The mpact on aggregate productvty nduced by changes n the productvty of each sze group s clear: aggregate productvty wll be hgher f there s an ncrease n the productvty of the small or large frms. However, an ncrease n the share of the SME sector wll only ncrease aggregate productvty f SME frms tend to be more productve than large frms. Most of the exstng evdence suggests that, on average, large frms have hgher total factor productvty. 8 If SME are less productve than larger frms, polces amng to ncrease the proporton of the SME sector (e.g. employment polces) wthout ncreasng ts productvty may have a negatve mpact on aggregate productvty. Ths framework thus ndcates that t s mportant to consder the reallocaton and drect productvty channel when dscusson polcy optons for SMEs. If SME are less productve than larger frms, polces amng to ncrease the proporton of the SME sector (e.g. employment polces) wthout ncreasng ts productvty may have a negatve long run mpact on aggregate productvty. In ths context, aggregate productvty could be at least n prncple be mproved by SME polces that ncrease the productvty of the SME sector, P and that do not smultaneously reduce productvty of large frms. SME, t 4 Data and Descrptve Statstcs We use plant level data from The World Bank Enterprse Surveys (WBES). The surveys 8 For example, Söderbom and Teal, 2004 fnd that SMEs are more productve n Afrca, whle Van Besebroeck, 2005, also for Afrca, reaches the opposte concluson; other relevant papers are Alvarez and Cresp, 2003 for Chle; and Daz and Sanchez, 2008, for Span. 8

11 are appled to a representatve sample of frms n the non-agrcultural economy and cover plants wth more than fve employees. The sample ncludes manufacturng, servces, transport, constructon sectors. However, only manufactures has been surveyed n all LAC countres and therefore we restrct our analyss to the manufacturng sector. The nformaton for the majorty of countres was surveyed n However, there are some countres wth nformaton for 2003 and others wth nformaton for 2003 and The prncpal advantage of ths dataset s that t provdes homogeneous nformaton for sxteen LAC countres. 9 Surveys are desgned n the same way, wth smlar questonnare structure and follow smlar samplng frames. 10 Ths s mportant advantage wth respect to prevous studes that present evdence only for a small set of countres. WBES provdes us wth the necessary nformaton to estmate Total Factor Productvty (TFP) at the plant level. It also provdes nformaton about the frm (ownershp, other plants, organzaton) and some characterstcs of the plant (age, etc) and a set of varables that can be modfed by SME polcy, such as access to credt, tranng, and process nnovaton. See Appendx A for the lst of varables we were able to match across all the surveys. The datasets have two lmtatons. Frst, all the plants n the dataset belong to formal frms wth more than 5 employees. As mentoned n secton 2, ths s mportant because n LAC economes a large proporton of employment takes place n mcro and nformal frms. Unfortunately, there s no dataset coverng nformal or mcro frms for all the consdered countres. The second lmtaton s that the surveys do not provde us wth panel data nformaton. The lack of panel data lmts our ablty to control for the endogenety of both nputs and productvty determnants. The log of Total Factor Productvty (TFP) of frm s gven by p p = y α l α m α k, (4) l m k where y, l, m, and k are the log value of sales, log number of hours worked per year by temporary and permanent workers, the log value of materals and the log value of captal, respectvely. In all the surveys monetary varables are n local currency unts (LCUs) and were converted nto US dollars. Because of the lack of panel data, we estmate the nput-output elastctes ( αl, αm, α k) by the cost share of each nput. 11 An mportant assumpton behnd ths estmaton procedure s 9 The countres are Argentna (2006), Bolva (2006), Brazl (2003), Chle (2004, 2006), Colomba (2006), Costa Rca (2005), Ecuador (2006), El Salvador (2003, 2006), Guatemala (2003, 2006), Honduras (2003, 2006), Mexco (2006), Ncaragua (2003, 2006), Panama (2006), Paraguay (2006), Peru (2006), and Uruguay (2006). 10 The samplng frame n 2006 s dfferent than the samplng frame n prevous years. In 2006 there was stratfed saplng. 11 Alternatvely, the nput-output elastctes can be obtaned by estmatng the producton functon. 9

12 the presence of constant returns to scale (CRS) at the ndustry level. Ths assumpton can be problematc n our case because our objectve s to compare the productvty of frms of dfferent sze and therefore the scale s the relevant dmenson. If the CRS assumpton does not hold, the nput-output elastctes have to be multpled by the returns to scale parameter to obtan a measure of productvty. Under non-crs, the log of productvty of frm s gven by [ ] p = y γ α l + α m + α k, (5) l m k whereγ s the parameter of the returns to scale. The comparson of productvty between SMEs and large frms s affected by the returns to scale parameter. Suppose for smplcty that = S, L (.e. we only have two frms; one small and one large). The dfference n productvty between these two frms s gven by ( ) γ α ( ) α ( ) α ( ) pl ps = yl ys l ll ls + m ml ms + k kl ks.(6) When there s ncreasng returns to scale (IRS) (.e. γ > 1) the assumpton of CRS (.e. γ = 1) ntroduces a bas n favor of large frms. On the other hand, when there s decreasng returns to scale (DRS) (.e. γ < 1) the assumpton of CRS ntroduces a bas n favor of small frms The productvty gap between SMEs and large frms The economc theory does not provde an answer to the queston regardng the relatve productvty of SMEs wth respect to large frms. On the one hand, large frms may take advantage of economes of scale and scope and more easly undertake some fxed costs. On the other hand, small frms may have hgher flexblty to face changes n ther envronment, they may use cooperaton to acheve economes of scale and scope smlar to those of larger frms; or they may smply focus on small and hghly specalzed markets. In ths secton we compare the productvty dstrbuton of SMEs wth that of large frms. To compare the productvty of frms n dfferent ndustres n ths secton we use a productvty ndex that measures the proportonal dfference of TFP for frm relatve to a reference frm that vares across ndustres (Caves, Chrstensen and Dewert, 1982). For a gven ndustry j, the frm of reference s defned as the frm that has: () ts output s equal Unfortunately, wthout panel data nformaton t s dffcult to control for the endogenety of the nputs (See Olley and Pakes, 1996; Levnshon and Petrn, 2003; and Ackerberg et al 2007). 12 For robustness n the emprcal analyss that follows we allowed the scale parameter to vary from 0.8 to 1.2, and the results dd not change sgnfcantly. We also estmated the producton functon gnorng the endogenety of the nputs. The estmaton rejects the CRS assumpton n favor of IRS. However, the value of the returns to scale parameter s not far from 1; t les between 1.03 and 1.05 dependng on the ndustry 10

13 to the geometrc mean of frm s output quanttes n ndustry j over the entre perod; () the quanttes of nputs are equal to the geometrc means of frms nput quanttes n ndustry j over the entre perod; and () the cost shares of nputs are equal to the arthmetc mean of frms cost shares n ndustry j. Therefore, when observatons of dfferent ndustres are pooled, productvty dfferences among ndustres are removed. If frm belongs to the ndustry j, then ts productvty ndex (n logs) s gven by αx, + αx, j ω = ( y y ) ( x x ), (7) j j x= { l, m, k} 2 where x= l (labor), m (materals), k (captal); j=1,2,, J (number of ndustres); and for a generc varable z whch can be y (sales), or x ; z j s the average value of z for the frms n ndustry j. Averages are computed wthn each country, therefore each country has ts own reference frm for each ndustry. Wth the TFP ndex defned n (7) t s possble to pool all the observatons wthn each country and to compare the productvty dstrbuton of SMEs and the productvty dstrbuton of large frms. Fgure 2 shows the cumulatve dstrbuton functon (cdf) of the productvty ndex n (7) of SMEs and large frms n selected countres. In Brazl, Mexco, Costa Rca and Panama, the cdf of large frms s at the rght of the cdf of small frms ndcatng stochastc domnance. On the other hand, n Argentna and Bolva the dfference s not clear. Hence, these results ndcate that large frms tend to be more productve than SMEs, although the evdence s less clear cut n Argentna and Bolva. [FIGURE 2 ABOUT HERE] To summarze the nformaton for all countres we use the relatve dstrbuton functon, R. 13 Ths dstrbuton allows us to compare a dstrbuton G wth a reference dstrbuton F. 1 The relatve dstrbuton functon s defned as R ( r) = G( F ( r) ) where 0 r 1. If both dstrbutons are dentcal then R(r) s the unform dstrbuton on [0, 1]. Fgure 3 shows the relatve dstrbuton functon for all the countres. [FIGURE 3 ABOUT HERE] The dagonal represents the unform dstrbuton,.e., the relatve dstrbuton f both dstrbutons were dentcal. The poston of the relatve dstrbuton below the dagonal suggests that the dstrbuton represented n the vertcal axs stochastcally domnates the dstrbuton n the horzontal axs. Wth the excepton of Argentna and Bolva, all the dstrbutons are below the dagonal ndcatng that the cdf of productvty n large frms stochastcally domnates the cdf of 13 Ths functon has been used by Delgado et al (2002) to compare the productvty cdf of exporters and nonexporters. 11

14 SMEs. Agan, ths shows that large frms have hgher productvty than small frms, except n Argentna and Bolva were the results are not that clear. In lght of ths productvty gap showed n ths secton, polces that shft resources toward SME frms (wthout affectng frms' productvty), mght have a negatve effect on aggregate productvty because less productve frms would account for a larger share of economc actvty. 6 Understandng the productvty gap Is there any room for publc polces n ncreasng the productvty of SME? The best way of answerng ths crucal queston 14 would be to produce a sgnfcant number of studes on the mpact of SME polces and ther cost effectveness, whch means ther effectveness compared to other polcy optons. Unfortunately, such a set of studes has not been developed yet, and the rare evaluatons that rgorously address ths ssue do not provde enough crtcal mass to draw sgnfcant conclusons. Whle polcy makers, scholars and nternatonal agences should drastcally ncrease the resources devoted to mpact evaluatons of SME polces, other approaches should also be explored to have at least a glmpse on the adequacy of the SME polces. Ths secton and the followng are an attempt of dong that. We decompose the productvty gap n ts major determnants, wth partcular attenton to the varables usually targeted by SME polces, for nstance access to credt, nnovaton, tranng and qualty certfcaton. Once the determnants of the productvty gap have been ndentfed and the relevance of the factors targeted by SME-polces has been tested, we smulate what would be the mpact of expandng these polces on the productvty level. To study how the varables that can be affected by SME Polcy affect productvty, we consder two procedures: () a lnear regresson model, and () quas-expermental technques. In the frst procedure we drectly address the queston of how much of the productvty gap can be explaned by varables that can be affected by SME Polcy. Let p be the productvty of frm defned n (4), S a set of sze dummes, X a set of varables that can be affected by SME polcy, Z a set of exogenous control varables, and C Y and I a set of country-year and ndustry dummes. We run two regressons; one restrctng the coeffcent of the varables n X to zero and other wthout any restrcton. In these models we compare the coeffcents of the sze dummes and the dfference n productvty by sze category that can explan each model. The regressons are the followng 14 Crucal because of the ncreasng amount of resources devoted to SME polces. 12

15 p = c + α S + α Z + α C Y + α I + ε (6.1) ' ' ' ' r s, r z, r c, r I, r r p = c + α S + α X + α Z + α C Y + α I + ε (6.2) ' ' ' ' ' u s, u x, u z, u c, r I, r u The varables that we nclude as polcy varables (varables that can be affected by SME polcy) are tranng, credt, product nnovaton, ISO certfcaton. The defnton of these varables s as follows: () Tranng s a dummy varable that takes value one f the frm offers tranng to ther employees, () Credt s a dummy varable that takes value one f the frm reports that t has a credt lne or overdraft facltes, () Product nnovaton s a dummy varable that takes value one f the frm reported a product nnovaton n the last three years, (v) ISO certfcaton s a dummy varable that takes value one f the frm has ISO certfcaton. These varables may result endogenous. Tranng s endogenous because t s lkely that more productve frms offer tranng to ther employees because they may be n a better fnancal poston to do that. In the same way, credt s endogenous because t s possble to thnk that more productve frms are those that receve credt from fnancal nsttutons. To deal wth the endogenety of these varables we apply nstrumental varables. We frst consder as nstrument the proporton of frms reportng tranng, credt, product nnovaton, and ISO certfcaton n two clusters of frms: () frms n the same country and ndustry and, () frms n the same ndustry and sze category. That s,.e., for each endogenous varable we consder two nstruments. After testng the valdty of those nstruments, we consder addng as addtonal nstruments the proporton of frms reportng tranng, credt, product nnovaton, and ISO certfcaton n a cluster of frms n the same country, ndustry and sze category. We test the valdty of ths extra set of nstruments usng the ncremental Sargan test. The use of IV methods s warranted gven the cross-secton nature of the data; however the dffculty of fndng approprate nstruments n ths data calls for cauton n the nterpretaton of results. Wth the set of varables n Z we control for: () the age of the frm, () the organzatonal form of the frm we nclude dummes for ncorporated companes, frms wth only one establshment, and frms wth only one propretary, () the level of unonzaton n the ndustry and sze, (v) the busness cycle we use the average at the country-ndustry level of the capacty utlzaton, and (v) the level of formalty of the frm we nclude an average at the country-ndustry-sze level of the manager s tme dealng wth bureaucracy. Wth respect to the busness cycle we expect a postve relatonshp because the lterature has found that productvty s pro-cyclcal. We average across country-ndustry to capture the busness cycle and avod consderng neffcences by frm. We defne the level of formalty as the manager s tme dealng wth bureaucracy because t s sensble to expect that the more formal the frm s the more tme the manager needs to allocate n dealng wth bureaucracy. We take the average across country-ndustry-sze to avod confusons wth effcency n dealng wth bureaucracy at the frm level. 13

16 Table 2 shows the result of the estmaton of equatons (6.1) and (6.2). We dd two dfferent exercses wth dfferent defntons of sze. Frst, we consder three sze categores wth large frms beng the excluded group. Table 2 shows the results as Dscrete Sze. Second, we consder the number of employees Contnuous Sze n Table In both cases, we estmated (6.2) by OLS and 2SLS. [TABLE 2 ABOUT HERE] Table 2 shows a large gap n productvty between SMEs and large frms. On average, small frms are 22% less productve than large frms and medum-szed frms are 15% less productve than large frms. When the regresson ncludes the varables n X the estmated coeffcent on sze both n the dscrete and contnuous case- s lower. In the OLS columns the estmated coeffcents on the sze varables are lower but contnue to be sgnfcant. The magntudes of the coeffcents on sze n the 2SLS regressons are substantally lower and no longer statstcally sgnfcant. Wth respect to the varables that can be affected by SME Polcy, access to credt, tranng and ISO certfcaton show postve and sgnfcant coeffcents. On the other hand, product nnovaton shows no sgnfcant coeffcents when estmatng by OLS and postve and sgnfcant coeffcent only n the 2SLS estmaton. Ths non robust result of product nnovaton can be reflectng the fact that the nnovaton needs tme to produce an ncrease n productvty. Table 2 also shows other nterestng results. Frst, more formal frms are more productve. Second, frms wth a larger the proporton of unonzed workers have lower productvty. Thrd, the sgn of the rest of varables s as expected; productvty s pro-cyclcal, older frms are more productve, ncorporated companes are more productve and one establshment and one propretary frm are less productve. What s nterestng to note s that the dfference n productvty s present even when the coeffcent of sze s smaller. Gven that the coeffcent of the polcy varables s the same for small, medum-szed and large frms, the dfference n productvty across sze s explaned by the dfference n the varables that can be affected by SME polcy. Fgure 4 shows the average productvty by sze category obtaned by the predcted values of equatons (6.1) and (6.2) and the average by sze of the varables that can be affected by SME polcy. [FIGURE 4 ABOUT HERE] As mentoned above, both OLS and 2SLS results go n the same drecton n the sense that when ncludng polcy varables there s a reducton n the sze coeffcents. However, the dfference between the OLS and 2SLS coeffcents s large. When estmatng by 2SLS, the 15 We also consdered the number of employees the plant had three years before. By consderng the sze the plant had three years before, we avod the possble endogenety that can appear f most productve frms are those that grow and ncrease ther sze. Results were robust. 14

17 coeffcents of the sze varables become statstcally nsgnfcant and the magntudes of the estmated coeffcents on polcy varables ncrease consderably. However, even when the nstruments pass the statstcal tests for beng consdered good nstruments, the way n whch they have been constructed may cast doubts on the 2SLS estmates and the endogenety of the polcy varables on the OLS results. Gven the nherent problems wth OLS and 2SLS, we propose another exercse to obtan the mpact of the varables that can be affected by SME polcy on plant s productvty. For each one of those varables we evaluate ther mpact on productvty applyng standard methods proposed n the treatment effect lterature. The startng pont s the estmaton of the treatment effect, whch for each partcpant s defned as α = Y Y (7) 1 0 In other words, the mpact s the dfference n the outcome (Y) dependng on whether the frm partcpated n the program ( Y 1 ) or dd not partcpate ( Y 0 ). In our case partcpaton means, dependng on the case, provdng tranng, havng access to credt, havng ISO certfcaton, or nnovatng. The fundamental problem of evaluaton s that, for any gven ndvdual, t s mpossble to observe both Y 1 andy 0 smultaneously. For partcpants, we are nterested n α ( 1 = Y1 Y0 T = ), but for them t s mpossble to observe ( Y 0 ). For the populaton of benefcares we are nterested n the average mpact of the program G = E Y Y T = 1 = E Y T = 1 E Y T = 1, and we need to estmate ( 1 0 ) ( 1 ) ( 0 ) ( ) E Y 1 0 T =, the counterfactual of what would the productvty be f the plant that, for example, offer tranng to ther employees would not has offered such a tranng. The general approach s to approxmate E( Y 1 0 T = ) wth E( Y 0 0 T = ) usng a comparson group, whch s formed by plants smlar to the one that offered tranng that dd not offered tranng and the same for the rest of varables. The mpact evaluaton must rest on an adequate and rgorous strategy for dentfyng a statstcally robust control group of non-benefcares. A central ssue s how to create a vald comparson group, n the sense that the dfferences n the productvty between partcpants and non-partcpants are explaned by the program and not by other characterstcs. The deal soluton s to have the treatment randomly assgned, n order to guarantee that on average the characterstcs of both groups are the same. In the absence of ths soluton, we am at dentfyng ndvduals that have the same observable relevant characterstcs as the partcpants through matchng methodologes. Ths assumes that the selecton of frms nto the treatment occurs only through observables, and there s no systematc sortng nto treatment based on unobservable frm characterstcs. The procedure s as follows: Frst, we estmate the probablty of partcpaton, 15

18 pˆ( S, Z, C Y, I ), by estmatng the propensty score 1 ' ' ' ' ( T 1 S, Z1, C Y, I) = Φ ( αss + αzz1 + αcc + αi ) P = Y I (8) the varables n vectors S, C Y,and I have been defned prevously defned. The vector Z 1 s a subset of Z. Summarzng, the observable characterstcs we consder are age, the organzatonal form of the frm (ncorporated companes, frms wth only one establshment, and frms wth only one propretary), ndustry, and country-year. The propensty score s computed for all ndvduals, those that partcpated and those that dd not partcpate. The next step nvolves usng matchng technques to compare the productvty of each partcpant wth the productvty of nonpartcpants that have smlar probablty of partcpaton. The general form of the matchng estmator s: M = 1 Y1 w (, jy ) 0 j { = (9) 1} { j = 0 } n1 T j T where n1 s the number of treated and w(, j) represents the weght gven to the jth observaton when constructng the counterfactual for the th treated observaton. Table 3 shows the results for each of the varables that can be affected by SME Polcy for all the frms and consderng separately SMEs and large frms. [TABLE 3 ABOUT HERE] The results not only show a consstent and robust relatonshp between access to credt, tranng and qualty certfcaton and productvty, but also that the relevance of the polcy varables s hgher n the case of SMEs than larger frms (except n access to credt). It s nterestng to note that the magntude of the mpact les between the one obtaned by OLS and 2SLS n Table 2. In sum, the regresson analyss above suggests that smaller and medum szed frms are on average less productve than larger frms. Ths confrms the fndng of the productvty gap between large and small frms n secton 5. Further analyss suggests that the productvty dfferences between SMEs and larger frms could be n part drven by dfferental partcpaton of these frms n worker tranng, credt avalablty, product nnovaton and ISO certfcaton; all varables that n prncple can be affected by SME Polcy. Gven the lack panel data, t s dffcult to rule out the possblty that the postve assocaton between tranng, credt avalablty, etc. and productvty does not smply reflect selecton of nherently better frms nto these programs or reverse causalty, whereby more productve frms have easer access to credt or can afford worker tranng. If self-selecton or reverse causalty drves the postve assocaton between tranng and productvty, then provdng SME frms wth access to credt or tranng mght not necessarly lead to productvty mprovements n SME frms. We attempt to address these concerns by usng 2SLS and 16

19 propensty score matchng and contnue to fnd support for the above results. Nonetheless, gven the structure of our data and strong dentfcaton assumpton behnd these technques the concerns about the nterpretaton of results contnue to exst. Do these results mply that SME polces are justfed? Even takng the results at face value, one should not jump to the concluson that SME polces amed at mprovng access to credt, promotng tranng, nnovaton, and ncreasng qualty certfcaton would necessarly have a postve effect on the overall productvty level, as t s not evdent that polces would actually nduce the benefcary frms to engage n those actvtes that have a postve relatonshp wth productvty for nstance t may be that frms that partcpate n a tranng program would have traned ther workers anyway wth ther own resources. Also, the type of actvtes promoted wth publc polces s not necessarly the same type as those that mplemented by frms on ther own for nstance the tranng provded by SME programs may dffer substantally from the tranng provded by the most productve frms. 7 The effect of SME Polcy va smulatons In ths secton we combne the methods used n prevous secton to smulate the effect of several polces on frms s productvty. We use the reduced form estmates to evaluate the mpact of changes n access to credt, tranng, product nnovaton and ISO certfcaton on frms s productvty and the propensty score to select the frms that wll be recevng the polcy,.e., we use the propensty score as a targetng rule. Wth the targetng rule we am at promotng those actvtes among frms that have not yet mplemented them, but that have smlar observable characterstcs to the ones that have. We are nterested n two questons: () How much of the productvty gap can be closed by an SME polcy? and () Whch s the effect on the aggregate level of productvty? 7.1 The effect on the productvty gap We compare two polcy desgns: () () SME Polcy 1: SME Polcy targets those SMEs that are not tranng ther workers but have a propensty score hgher than 0.5. We do the same for credt, product nnovaton, and ISO certfcaton. SME Polcy 2: SME Polcy targets those SMEs that have the hghest propensty score of tranng and are not offerng tranng. In ths case, we select frms accordng ther propensty score up to the pont where the fracton of SMEs that engage n tranng s smlar to the fracton of large frms that tran. We do the same for credt, product nnovaton, and ISO certfcaton. BL SP To obtan the productvty level of frm n the baselne, p, and under polcy SP, p, (wth SP=tranng, credt, product nnovaton, and ISO certfcaton) we use equaton (6.2). 17

20 Then, the productvty of frm n the baselne s p = cˆ + ˆ α S + ˆ α X + ˆ α Z + ˆ α C Y + ˆ α I (10) BL ' ' ' ' ' u s, u x, u z, u c, r I, r and under polcy SP s p = cˆ + ˆ α S + ˆ α X + ˆ α Z + ˆ α C Y + ˆ α I (11) SP ' ' SP ' ' ' u s, u x, u z, u c, r I, r where SP X s the vector correspondng to the smulated polcy. In our case, all the varables that can be affected by SME Polcy are dummy varables and therefore the smulaton s computatonally smple; t just mples changng zeros by ones n the polcy varables for those frms recevng the polcy. As mentoned above, the rule for assgnng the polcy s mportant. Not all the frms the frms that do not offer tranng to ther employees (or are not nnovatng, etc) receve the polcy n the smulaton exercse; SP only those frms that satsfes the condtons descrbed above. Therefore, X s equal to X except for those frms that were selected to partcpate n the polcy. We do the exercse SP polcy by polcy and therefore X dffers from X only n the actvty consdered by the polcy. Fgure 5 shows the dstrbuton of productvty of SMEs and large frms under both scenaros. [FIGURE 5 ABOVE HERE] 7.2 The effect on the aggregate level of productvty In order to address the skeptc vew that productvty polces should not be targeted exclusvely or prmarly to SMEs we consder how the aggregate level of productvty would ncrease under two scenaros: one n whch polcy s targeted by sze (SME Polcy) and other one wth non-targeted by sze polcy. The smulated polcy desgns are as follows () () SME Polcy 1 of prevous subsecton. Non-Targeted by Sze: Polcy targets frms wth the hghest propensty score of tranng and not currently offerng tranng. We select frms up to the pont were the number of frms startng to tran ther workers s equal to the number of frms startng to tran ther workers n polcy desgn (). The same for credt, nnovaton, and ISO certfcaton. BL BL Let p and w be the productvty and the output-share of frm n the baselne. Then the baselne aggregate level of productvty s gven by 18

21 P = w p BL BL BL. (12) Let SP p and SP w be the productvty and output share of frm n the smulated polcy desgn SP, wth SP beng tranng, credt, product nnovaton and ISO certfcaton. Then the change n the aggregate level of productvty due to polcy SP can be decomposed followng Foster et al (1998) as follows: ( ), = + + SP BL BL BL P w p p P w w p where p = p p and SP BL w = w w. SP BL The frst term s the ncrease n aggregate productvty when frms ncrease ther productvty at baselne market share (the wthn-plant term). The second term s the ncrease n productvty resultng when plants wth above-average productvty expand ther market share relatve to plants wth below-average productvty (the between-plant term). The thrd term s the cross-plant term. The productvty level of frm n the baselne and under polcy SP are defned above. To obtan the change n market shares we use the producton functon. 16 Then the value of output and market shares after polcy SP are gven by 17 SP ( ) Y p L M K SP αl αm αk = exp, w SP = Y SP Y SP. Table 4 shows the results of ths smulaton. The results show that properly targeted SME polces mght have a sgnfcant postve effect on aggregate productvty n the sample adopted for the smulaton, they would nduce a 5.7 percent aggregate productvty ncrease (when all polces are performed smultaneously). The overall effect on aggregate productvty s mostly due to the between-frms term, whch accounts for 65 percent of the aggregate productvty ncrease, whle the wthn-frm term accounts for 45 percent aggregate productvty ncrease. 16 Smlar approach was proposed by Escrbano et al (2008) to obtan the market shares n a smulaton on the mpact of the nvestment clmate on the components of the Olley and Pakes decomposton. 17 We are consderng only the drect change n output due to changes n productvty. The nput demand functons are also functon of productvty and therefore the output s also affected by changes n nputs nduced by the productvty change. For smplcty, we are gnorng those changes. 19