EVALUATING MICRO- AND MACRO-EFFECTS OF REGIONAL DEVELOPMENT POLICIES: THE CASE OF THE NORTHEAST REGIONAL FUND (FNE) IN BRAZIL, RESUMO

Transcription

1 EVALUATING MICRO- AND MACRO-EFFECTS OF REGIONAL DEVELOPMENT POLICIES: THE CASE OF THE NORTHEAST REGIONAL FUND (FNE) IN BRAZIL, Gulherme Mendes Resende (PhD canddate LSE) RESUMO Este trabalho busca em parte preencher uma lacuna na lteratura através da realzação de duas avalações de resultados do Fundo Consttuconal de Fnancamento do Nordeste (FNE). Com este ntuto, o artgo reúne dos tpos de avalação de resultados, geralmente mplementadas separadamente na lteratura de avalação. Os resultados das mcro e macro-avalações mostram que, embora haja um mpacto postvo e estatstcamente sgnfcatvo dos empréstmos do FNE ndustral na cração de emprego no nível da empresa (mcro), este mpacto anda é bastante lmtado para ter qualquer efeto sgnfcatvo sobre o crescmento do PIB per capta em nível muncpal (macro) a fm de reduzr as desgualdades regonas no Brasl. ABSTRACT The goal of ths paper s to contrbute to fllng the gap n the lterature by carryng out two outcome evaluatons of the northeast regonal fund (FNE) n Brazl. Wth ths am, the paper assembles two types of outcome evaluaton often mplemented separately n the evaluaton lterature. The results of the mcro- and macro-evaluatons show that although there s a postve and statstcally sgnfcant mpact of the FNE ndustral loans on job creaton at the (mcro) frm level, ths mpact s stll too lmted to have any sgnfcant effect on GDP per capta growth at the muncpal (macro) level and thus reduce the regonal nequaltes n Brazl.

2 1. Introducton It s well known that regonal nequaltes have perssted n Brazl for decades. For example, the Gross Domestc Product (GDP) per capta of the poorest regon, whch s the Northeast 1, was only 43% of the natonal average n 1989 and 47.5% n On the other hand, the per capta GDP n the Southeast regon, the wealthest regon, was 139% of the natonal average n 1989 and 133% n 2006 (see Table A.1 n Appendx A). The man Brazlan regonal development polcy has been enacted snce 1989 by the Consttutonal Fnancng Funds for the Northeast (FNE), the North (FNO), and the Central-West (FCO) (henceforth referred to as regonal funds), seekng the economc and socal development of laggng regons through loans at nterest rates below market rates, manly to agrcultural and ndustral small frms. However, the evaluaton of regonal polces n Brazl has not occurred over the years, although Brazlan governments have always ntervened n regonal economes n pursut of reductons n regonal nequaltes. Thus, the goal of ths paper s to contrbute to fllng at least partally ths gap n the lterature by carryng out two outcome evaluatons of the FNE. Gven the avalablty of data, the analyss ncludes only the FNE loans for frms n the ndustral, commerce and servces sectors, whch represents roughly 40% of the FNE loans granted durng 2000 and Wth ths am, the study nvolves two types of outcome evaluaton that are often mplemented separately n the evaluaton lterature. Frst, t measures the effectveness of the FNE (does the FNE create jobs and/or ncrease productvty at the frm level?). Second, t assesses the mpact of the FNE on regonal nequaltes (does the FNE reduce regonal GDP per capta gap?). Ths combned approach s useful because evaluatons can prove, for nstance, that the regonal funds create jobs and/or ncrease productvty (proxed by wage growth) at the mcro level. However, t s stll necessary to demonstrate that the program has solved, or at least reduced, the regonal nequaltes (at the macro level). These mcro and macro effects have been overlooked n the lterature dealng wth the mpacts of regonal funds. In ths sense, ths paper mproves the mcro effectveness estmates for employment and wage growth done by Slva et al. (2009) by usng the frst-dfferencng method. It also goes further than pror publcatons n assessng the mpact of FNE on the regonal mbalances n Brazl usng the nstrumental varables technque. The former evaluaton uses a frm-level (mcro) dataset, whereas the latter evaluaton employs aggregate (macro) data at the muncpal level to evaluate the mpact of FNE on regonal nequaltes. Wth regard to ths macro analyss, other levels of spatal aggregaton of the observatonal unts are employed because an aggregaton problem [Modfable Areal Unt Problem (MAUP) and Ecologcal Fallacy Problem] mght prevent us from dentfyng the actual spatal scale on whch processes operate. The paper s organsed as follows. In Secton 2, the background of the regonal funds s dscussed, wth a revew of the prevous studes evaluatng regonal polcy n Brazl. Secton 3 descrbes the mcro and macro data employed n ths paper. Secton 4 carres out the effectveness mcro estmates for employment and wage growth usng the frst-dfferences method. Secton 5 shows the macro emprcal model n whch the nstrumental varable approach s used to assess the macro mpact of the FNE on regonal nequaltes n Brazl. Secton 6 presents the conclusons. 2. Background of the Regonal Funds and Lterature Revew Ths secton ams to explan how the FNE works as long as there exsts emprcal evdence of regonal polcy evaluaton n Brazl Background of the Regonal Funds The regonal funds (FNE, FNO, and FCO) were created by Federal Law 7827 n 1989 based on artcles 159.I.c and 161.II of the Federal Consttuton of An equal porton (3%) of ncome taxes (from ndvduals and frms IR ) and of the tax on ndustralsed goods ( IPI ) represents the transfer of resources from the Natonal Treasury to the regonal funds. The total resources allocated 1 The Northeast Regon ( Nordeste ) has about 53 mllon nhabtants, whch represents 29% of the total n the country. 1

3 to these funds each year s dvded as follows: 60% goes to the FNE, 20% to the FNO, and 20% to the FCO. These resources are transferred from the Natonal Treasury to the operatng bank va the Mnstry for Natonal Integraton ( Mnstéro da Integração Naconal ). Beyond the 3% IR and IPI taxes, the revenues for these funds come from the repayment of the loans (prncpal + nterest). In ths way, Federal Law 7827 defnes the source of fundng and desgnates the regonal banks as beng the operators of the regonal funds. The operator bank of FNE s the Bank of the Northeast (Banco do Nortedeste/BNB). The regonal funds granted loans of 10 bllon 2 n laggng Brazlan regons between 2000 and Ths amount represents 1.2% of the natonal GDP n For detals, see Ferrera (2004) and Almeda Junor et al. (2007), who conduct comprehensve studes of the resource allocaton each year for these funds. Among other observatons, these authors show that the rate of non-performng FNE loans reached 31% n As noted by Ferrera (2004), before 2001, most bad loans were consdered under renegotaton ; although n fact, they were never pad back. Ths hgh default rate lmted the Bank of the Northeast n grantng new loans durng the perod. In 2002, a federal balout plan captalsed the Bank of the Northeast. For ths reason, durng the followng years, the bank was able to ncrease the number of loans granted. For the FNO, loan defaultng s also hgh, reachng 13.2% n On the other hand, the FCO presented the lowest default rate, approxmately 3% n Furthermore, these regonal development funds are not the only resources avalable from a publc bank for laggng regons n Brazl. The Brazlan Development Bank (BNDES) a federal publc bank establshed n 1952 also offers loans (nterest rates are below market rates but are hgher than those of the regonal funds) to companes of any sze and sector n all Brazlan regons. At the end of 2007, BNDES credt represented 17% of the total credt n Brazl and was the man source of long-term credt n the country. Whle the focus of the regonal funds s the producers n the agrcultural sector (60% of total loans), BNDES loans are drected toward large-scale ndustral and nfrastructure projects (75% of the total loans). Moreover, unlke the operator banks of the regonal funds that work only n the laggng regons, BNDES mnsters to the demand for fundng n all Brazlan regons and does not have an explct mandate regardng regonal polcy. In Appendx A, Table A.2 compares the regonal funds (FNE, FNO and FCO) loans and the BNDES loans by regon for the perod between 2000 and Specfcally, the operator banks of the regonal funds the BNB n the Northeast regon are the agents responsble for analysng and decdng whether to award the subsdsed loans to applcants. The nterest rates of the loans are fxed but vary dependng on the sze of the benefcary and the sector. In rural FNE operatons, the nterest rates are between 6.00% and 10.75% per annum, and for the other operatons, they are between 8.75% and 14.00% per annum (small busnesses have the lower rates). It s worth notng that the average nterest rates for the producton sector n other banks were around 35% n 2000 (Banco Central, 2000). Furthermore, good payers wn complance bonuses n the form of an nterest rate reducton of 25% for those located n the sem-árdo regon (the hotdry hnterland of the Northeast) and 15% for those located n other areas. Applcants can be ndvduals, small busnesses, enterprses or cooperatves/assocatons that want to fnance a new busness or an exstng one located n the Northeast regon. There are some general gudelnes that the bank follows when analysng applcatons: preference s gven to () productve actvtes of ndvdual and small farmers and () small frms n other sectors, () actvtes that ntensvely use raw materals and are labour-ntensve and produce basc food for the populaton, and (v) new centres, actvtes or clusters that can reduce the economc and socal dfferences between regons. Moreover, by law, 50% of the FNE loans must be drected toward the sem-árdo regon (Fgure A.1 n Appendx A shows the boundares of the sem-árdo regon). 2 In terms of R$ (Reas), the Brazlan currency, ths amount represents R$ 28 bllon (n current prces). 3 It s nterestng to note that between 2000 and 2006, the European Unon (EU15) whch has provded a paradgm for regonal polcy for the Brazlan government allocated 135 bllon to regons wth less than 75% of the average EU15 GDP per capta. Concdentally, ths expendture also represented 1.2% of the EU15 GDP n

4 It s mportant to note that the goal of the FNE s to reduce regonal nequaltes through the fnancng of productve sectors n the Northeast. Ths mprecsely defned objectve (or broad objectve) s the major obstacle to outcome evaluatons. As ponted out by Jann & Wegrch (2007, p.54) gven the strong ncentve of blame-avodance, governments are encouraged to avod the precse defnton of goals because otherwse poltcans would rsk takng the blame for obvous falure. In the Brazlan case, the combned approach of mcro and macro evaluaton of the FNE s relevant because the FNE goal s broadly defned at the macro level (reducng regonal nequaltes) along wth the polcy tool operatng at the mcro level by means of subsdsed loans to producers wthn the Northeast regon. Whle the objectves at the frm level are not defned by federal law, some offcal documents 4 have hghlghted that the FNE loans seek more effcent resource allocaton to ncrease the productvty of frms and generate new jobs. In ths sense, gven the avalablty of the data, ths study defned the reducton of the GDP per capta gap as the polcy objectve at the macro scale level and job creaton and ncreases n productvty (proxed by wages growth) as the objectves at the frm (mcro) level. Another ssue that requres explanaton s that the mcro-evaluaton focuses only on the frms that can be traced n RAIS 5 durng the perod under analyss. Most of the FNE loans (approxmately 60%) are granted to ndvduals who have small farmng busnesses n the nformal sector.e., they do not have a CNPJ dentfer 6 and for ths reason, they are not covered by RAIS, whch s the source of nformaton for the mcro-evaluaton. The formal rural frms found n RAIS are few and are not statstcally representatve of the FNE rural populaton. For ths reason, the agrcultural sector was excluded from the mcro-evaluaton n ths paper. The government stll needs to formulate a specfc survey to cover the ndvduals and small rural busnesses n the Northeast to evaluate ths mportant, targeted FNE populaton. For the sake of comparablty between the mcro- and macroevaluatons, the macro-evaluaton at the muncpal level was restrcted to assessng only the amount of FNE loans granted to frms n the ndustral and commerce/servces sectors Lterature Revew In ths paper, the man queston s related to outcome evaluaton, whch s the fnal stage of the polcy process, wheren polcy s assessed aganst ntended objectves and mpacts 7. Ths s clearly an ex-post evaluaton. However, evaluaton can be defned n several ways: n terms of tme (e.g., exante, ex-post), levels of complexty (e.g., montorng daly tasks or assessng mpact on the problem) and as an nternal or external evaluaton. Specfcally n terms of tme, ex-ante evaluaton ams to verfy targets; at the md-pont, t establshes correctve actons; and fnally, ex-post evaluaton assesses outcomes. Although ths can be a useful way of defnng of types of evaluaton, I prefer to descrbe the types of evaluaton based on ther levels of complexty because the qualty and the objectves of evaluaton studes may be relatvely uneven and dverse. Thus, I follow the defnton of Bartk & Bngham (1995), whch looks at evaluaton as a contnuum movng from the smplest form of evaluaton, montorng daly tasks, to the more complex assessng mpact on the problem, as llustrated n Dagram 1. 4 For example, see the webpage of the Mnstry for Natonal Integraton < consttuconas/dretrzes.asp?d=dretrzes> and Banco do Nordeste (2001, 2009). 5 Annual Socal Informaton Report (Relação Anual de Informações Socas) of the Mnstry of Labor. The RAIS data were used under a cooperaton agreement between the Labor Mnstry and the Insttuto de Pesqusa Econômca Aplcada (IPEA). More detals n secton Cadastro Naconal da Pessoa Jurídca lterally, natonal jurdcal person regstraton as opposed to the CPF number for persons. CNPJ s an dentfcaton number for Brazlan companes assgned by the Brazlan Mnstry of Revenue ( Mnstéro da Fazenda ). The CNPJ number s comprsed of a base of 8 dgts, a 4-dgt radcal and 2 check dgts, such as / Regonal polcy follows the general publc polcy cycle, usually dvded nto fve stages: () analyss (agenda-settng), () formulaton (desgn), () choce (decson-makng), (v) mplementaton, and (v) outcome evaluaton. In other words, frst, problems are defned and put on the agenda, next polces are developed, adopted and mplemented; and, fnally these polces wll be assessed aganst ther effectveness and effcency and ether termnated or restarted (Jann & Wegrch, 2007, p. 44). 3

5 Dagram 1 Types of Evaluaton by Levels of Complexty Source: Bartk & Bngham (1995). As ponted out by Bartk & Bngham (1995), these sx levels of evaluaton provde a framework for assessng the qualty of evaluatons. To demonstrate that a program (or polcy) accomplshes ts targets, the evaluaton must be at the hghest level: measurng effectveness (for nstance, t actually does create jobs) or assessng mpact (there has been an mprovement n the stuaton). Wth regard to the Brazlan regonal polcy, an evaluaton could prove, for nstance, that the FNE creates jobs. However, t s stll necessary to demonstrate that the program has elmnated, or at least reduced, the regonal nequaltes 8. In the nternatonal context for example, n the context of the European Unon (EU) regonal polcy most of the studes focus on the macro mpacts of EU-regonal funds on regonal nequaltes. These studes nclude Rodríguez-Pose & Frates (2004), Dall erba (2005), Leonard (2006), Espost & Bussolett (2008) and Dall erba & Le Gallo (2008). Other studes are focused on the mcro mpact of specfc polces: e.g., Romero & Noble (2008), whch evaluates England s New Deal for Communtes programme. The contrbuton of ths paper s that t brngs together these two types of outcome evaluaton, whch are often mplemented separately n the evaluaton lterature, showng a more complete pcture of FNE performance. There are few papers related to the FNE evaluaton at the frm level (namely, Slva et al., 2009; Soares et al., 2009) and none at the macro level. Slva et al. (2009) measure the effectveness of regonal fund (FNE, FNO, and FCO) loans usng propensty score estmates of frms that receved loans (treatment group) and those of others that dd not receve them (control group) over The results show that FNE has a postve mpact on the growth rate for employment and no mpact on the growth rate for wages. Wth regard to FNO and FCO, there s no observed mpact of the regonal funds on the two varables under study. Ths orgnal research was sponsored by the Mnstry for Natonal Integraton (MI/Government of Brazl) and generated polcy reports usng other tme perods that bascally showed the same results. Specfcally, Slva et al. (2007) show a postve mpact of FNE on employment growth between 1995 and 1998, but ths effect wanes and dsappears over a large perod, from 1995 to Soares et al. (2009) employ the same propensty score method and expand the evaluaton of FNE by Slva et al. (2009) by ncreasng the number of granted frms by enlargng the tme horzon under analyss. The results show sgnfcant mpacts of FNE on employment growth for all perods between 1999 and 2005 and no mpact on the growth rate for wages was found. However, t s worth pontng out that nether paper examnes the loans granted to ndvduals, and t s these loans that are most common n the agrcultural sector; ndeed, they represent roughly 60% of the total loans funded. For ths reason, these results can be vewed as a partal evaluaton of the regonal funds, and much more effort needs to be made n ths feld. There are a few other papers that have tred to evaluate Brazlan regonal funds: for nstance, Olvera & Domngues (2005) and Carvalho et al. (2006). Olvera & Domngues (2005) employ a muncpal data set to examne f the Brazlan regonal funds (only FNO and FCO are analysed) have a postve mpact on regonal nequalty. The results show that regonal growth n Brazl between 8 Also, a cost-beneft analyss needs to be carred out to prove that the program benefts outwegh the costs. However, because the data necessary to carry out ths cost-beneft analyss are not avalable, ths type of evaluaton s left for future research. 4

6 1991 and 2000 was not affected by these funds (FNO and FCO). Carvalho et al. (2006) addresses a relatvely lmted queston dd the regonal funds successfully nduce entry by frms nto Brazl s laggng regons over the years ? Usng the RAIS dataset, whch contans nformaton on frms n the formal sector, one fnds that the regonal funds have stmulated entry nto the laggng regons of Brazl partcularly n the case of machnery, transportaton equpment, wood and furnture, textles, and food and beverages. The contrbuton of the current paper based on those prevous studes s twofold. Frst, as regards mcro-evaluaton, the paper employs the frst-dfference estmaton technque to elmnate the unobservable factors that are constant over tme and may be basng the prevous results. Ths method has not been used n prevous studes. Moreover, ths paper assesses the mpact of FNE on employment and wage growth untl Secondly, concernng macro-evaluaton, ths s the frst study to evaluate the macro effects of FNE loans awarded to the frms n the ndustral and commerce/servces sectors on GDP per capta growth at the muncpal level. In ths sense, ths paper seeks to provde a more complete pcture of FNE performance durng It s far to stress that regonal polcy evaluaton s not an easy or trval task. In an deal world, the evaluator could exactly determne what would have happened f the polcy had not exsted. However, outcome evaluatons mght suffer from at least three threats to valdty (Bartk & Bngham, 1995): () hstorcal events occurrng durng the program; () the maturaton of a frm or an ndustry whose growth pattern for a new frm or ndustry approxmates an s shaped curve; and () selecton bas, gven that frms that partcpate to the program are not chosen randomly. These problems may cause dffculty n determnng how much of the polcy outcome s actually due to the polcy, how much s due to natonal economc performance, and how much s due to the frm s performance n the absence of such a polcy. Thus, statstcal technques have been desgned to attempt to overcome these problems. Some methods for evaluatng the Northeast Regonal Fund (FNE) are dscussed and employed n the next sectons. Frst, however, the datasets are descrbed. 3. Data descrpton Ths secton descrbes the mcro and macro datasets. The mcro-analyss reles on a frm-level dataset and on polcy evaluaton models n whch treatment and control groups are used to nvestgate the effects of FNE on employment and wage growth. The complete descrpton of ths dataset s n secton 3.1. At the macro level, the objectve of the outcome evaluaton s to nvestgate the effects of FNE on economc growth at the muncpal level n the Northeast regon, so I use econometrc growth models that nclude the FNE loans varable along some other condtonng varables that are descrbed n detal n secton Mcro Data Frst, t s mportant to hghlght that the mcro-evaluaton approach reles only on nformaton about frms that are found n the annual censuses for frms n the formal sector (those that have the CNPJ dentfer 9 ), namely RAIS 10. The annual RAIS s supposed to cover all Brazlan formal frms; however, as explaned earler, some of the FNE loans (approxmately 60%) are granted to ndvduals who have small farmng busnesses.e., they do not have a CNPJ dentfer and are not covered by RAIS. For ths reason, the agrculture sector was excluded from ths mcro-evaluaton. Thus, the entre mcro-analyss s only carred out for those frms n the ndustral and the commerce/servces sectors. Based on the nformaton provded by the Bank of Northeast (BNB) t was possble to dentfy those frms ncluded n the RAIS dataset from 2000 that had receved the FNE loans 11 (the treatment group) and those that had not receved the FNE loans (the control group). Ths matchng between the 9 See footnote See footnote The amount of money receved at the frm level was not provded by the BNB. Ths nformaton was only gven at the muncpal level and s used n the macro evaluaton. 5

7 BNB nformaton and the RAIS dataset was made possble by usng the CNPJ dentfers 12. The dataset used here s bascally the same as that employed by Slva et al. (2009). These authors constructed a dataset contanng several frms characterstcs n both groups (treatment and control) usng the RAIS dataset and evaluated the varaton n ther employment and wages at the frm level between 2000 and 2003 usng the propensty score technque dscussed earler. However, three mprovements have been made to the method used n the current evaluaton. Frst, I set up two other control groups, one based on a sample matched to the treated group and another contanng all frms n RAIS, to check the robustness of the results. Second, I extended the perod of analyss by collectng nformaton about the poston of those two groups of frms (the treatment and control groups) n 2006 based on RAIS2006. In ths way, t s possble to capture the short-term ( ) and medum-term ( ) effects of the FNE loans granted n I have also collected nformaton at the frm level for 1998, whch represents a prevous perod n the lendng year (2000). Wth ths nformaton from 1998, t s possble to control for unobservable characterstcs of the frms n the treatment and control groups by means of the frst-dfferences estmaton technque. These procedures enhance prevous evaluatons of FNE, as dscussed n secton 4.1. Wth regard to the varables of nterest employment and wage growth t s worth notng that the former corresponds to an mplct objectve of FNE (.e., job creaton), whle the latter corresponds to the ncrease n ncome or value added and may represent a good proxy for measurng mpacts on productvty, another mplct 13 goal of FNE. Moreover, because t s possble that productvty gans nvolve job losses, at least n the short term, the jont consderaton of these two varables allows a more consstent evaluaton of FNE. All nformaton s found n the RAIS datasets from 1998, 2000, 2003 and The varables are the followng at the frm level: number of employees, average wage 14, average age of employees, average years of schoolng of employees, dummy for the commerce/servces sectors defned by the CNAE/IBGE 15 (the ndustral dummy was excluded) and dummes for states where the frm s located (the dummy for the Ceará state was excluded). As explaned earler, the data are exclusvely for frms n the ndustral and commerce/servces sectors. The loans granted to ndvduals, 16 most workng n the agrcultural sector, cannot be evaluated n ths mcro-analyss. The dataset s comprsed of two dfferent samples. The frst sample, or the treatment group, corresponds to the 85 frms that receved FNE loans only n 2000 and could be traced to RAIS 2000 and the datasets for the years 1998, 2003 and Ths treatment group s the same as that used n the study of Slva et al. (2009). However, the current study uses fewer frms because I have excluded 17 frms n the agrcultural sector and also used only the frms that could be traced n the years 1998, 2000, 2003 and 2006; n contrast, Slva et al. (2009) consdered the perod alone. The second sample s the control group, whch s comprsed of frms that dd not receve FNE loans n any year analysed and was selected from ndustral and commercal/servces frms n the Northeast regon that were n the RAIS n 2000 and whose progress could be traced through tme. I developed estmates usng three dfferent control samples: () one usng all frms dentfed n RAIS durng (97,452 frms); () a sample matched to the treatment group (9,338 frms); and () the control sample used n the Slva et al. (2009) paper, whch employed 727 frms drawn from a random 12 It s worth notng that the constructon of the dataset has followed procedures necessary to guarantee the confdentalty of nformaton. 13 As dscussed earler, Federal Law 7827/1989 has not precsely defned the objectve of the regonal funds. It only states that the regonal polcy ams to reduce regonal nequaltes through the fnancng of productve sectors n the Northeast, North and Central-West regons. For ths reason, I need to suppose that ths reducton n regonal nequaltes s va job creaton and/or ncreasng wages/productvty at the mcro level. As explaned n footnote number 6, offcal reports also state that the FNE loans am to ncrease the productvty of frms and generate new jobs. 14 The average wages n all years were converted to constant prces n 2000 usng a CPI ndex, namely IGP-M from Fundação Getúlo Vargas (FGV). 15 Natonal Economc Actvty Classfcaton (Classfcação Naconal de Atvdades Econômcas CNAE) from the Brazlan Insttute of Geography and Statstcs (IBGE). 16 It s worth notng that only persons wth a CNPJ the dentfer for a jurdcal person can be traced n the RAIS. 6

8 and representatve sample of the populaton of the Northeast wth records n RAIS. The control sample matched to the treated group ams to ensure that the treatment and control groups have smlar characterstcs, makng the two groups more comparable. Table 3.1 shows some summary statstcs of the RAIS dataset n 2000 for (a) the FNE sample and (b) the three control samples: (b.) all frms (excludng the FNE sample); (b.) the sample matched to the treatment group; and (b.) a random sample used n Slva et al. (2009). Table 3.1 Summary statstcs for the dataset at frm level n the RAIS 2000 (a) FNE sample* (treatment group) (b.) Control sample: All frms n RAIS2000 (excludng the FNE sample)* (b.) Control sample: Matched sample to the treatment group* (b.) Control sample: Random sample* Varables (85 frms) (97,318 frms) (9,250 frms) (727 frms) Average age of employees Average schoolng Average Wage (R$ n 2000) Sector of actvty (%) Agrculture 0.0% 0.0% 0.0% 0.0% Industry 64.7% 13.5% 65.0% 16.8% Commerce/Servces 35.3% 86.5% 35.0% 83.2% Sze (%) Small frms (1-49 employee) 77.6% 92.9% 77.5% 82.7% Medum frms (50-99 employees) 7.1% 3.1% 11.6% 7.4% Large frms (>99 employees) 15.3% 4.0% 11.0% 9.9% Northeast states (%) Maranhão 9.4% 5.7% 5.5% 5.4% Pauí 7.1% 5.1% 5.6% 4.7% Ceará 16.5% 16.0% 19.3% 19.4% Ro Grande do Norte 5.9% 7.1% 7.7% 7.0% Paraíba 21.2% 7.9% 8.2% 6.2% Pernambuco 15.3% 19.6% 19.9% 20.8% Alagoas 3.5% 4.8% 4.8% 3.9% Sergpe 7.1% 4.8% 4.3% 4.4% Baha 14.1% 29.0% 24.7% 28.3% Note:* Only frms that can be lnked through tme n the RAIS n 1998, 2000, 2003 and 2006 are selected. Own elaboraton based on RAIS data n Macro Data To evaluate the mpact of FNE on economc growth n the Northeast regon n Brazl, I employ a muncpal data set from the Brazlan Natonal Accounts (IBGE), the Bank of the Northeast (BNB), the Brazlan Census of 2000 (IBGE), and IPEADATA. It s worth notng that ths macro-evaluaton s also restrcted to frms n the ndustral and commerce/servces sectors (as n the mcro-analyss). The macro data set provdes the total amount of FNE resources at the muncpal level granted to the frms n the ndustral and commerce/servces sectors, whch represents roughly 40% of the total amount of FNE loans. The dependent varable s the average annual growth of the per capta gross domestc product (GDP) at the muncpal level over the perod. The GDP per capta of the muncpaltes has been computed annually by the IBGE snce However, due to changes n the IBGE methodology for calculatng the muncpal GDP n 2002, only the GDP from 2002 onwards s comparable and used here. In the results secton (5.2), some robustness checks are also carred out usng other sub-perods such as , and Due to the permanent creaton of new muncpaltes durng , there are few muncpaltes that dd not exst n To construct a consstent dataset, I use the boundares of the 1,787 Northeastern muncpaltes n 2000 nstead of those of the exstng 1,793 muncpaltes n The soluton was to assgn the new muncpaltes to the exstng muncpaltes n Fnally, all GDP per capta varables were obtaned drectly at constant 2000 prces n R$. 7

9 The FNE varables come from the Bank of the Northeast (BNB), whch provded nformaton about the amount of loans for ndvduals and frms n the aggregate by programme at the muncpal level over the /October perod 17. Usng ths nformaton, I constructed the FNE ndustral rato, whch s the amount of FNE loans to the ndustral sector between 2000 and 2001 as a proporton of GDP n the ndustral sector n For the sake of comparablty wth the mcro-evaluaton, the FNE ndustral rato also ncludes the loans to commerce/servce programmes 18, whch totalled 1% over the perod (the ndustral programmes alone represented 36% of the total loans). The results secton focuses on the analyss of the FNE ndustral loans between 2000 and 2001; however, t also provdes a robustness check usng the FNE ndustral rato between 2000 and 2003, whch totalled 43% of the total loans (the ndustral programmes alone represented 37% of the total loans, and the programmes for the frms n the commerce/servces sectors represented 6%). See Table 3.2 for the summary statstcs. The FNE loans were denomnated n nomnal prces, and I converted them to R$ at the 2000 level usng a CPI ndex (namely, the IGP-M ndex from Fundação Getúlo Vargas-FGV). IPEADATA provdes the varables at the muncpal level: (log of) average years of schoolng of the populaton, (log of) the nfant mortalty rate n 2000, (log of) the Gn ndex n 2000, (log of) populaton densty n 2000, and (log of) the transportaton cost to São Paulo n 1995, whch s the result of a lnear program procedure for calculatng the mnmum transportaton cost between the muncpaltes majors headquarter to São Paulo. The local nfrastructure ndex s made usng the prncpal components analyss by Da Mata et al. (2007b) that employs the Census2000 nformaton. It takes nto account seven dmensons of publc housng servces and utltes: electrcty, sewage, water provson, garbage collecton, publc lghtng, pavement and addressng. Ths varable s supposed to capture the quantty of housng nfrastructure n Brazlan muncpaltes 19. Also, the econometrc models nclude dummes for the Northeast states (the dummy for the Ceará state was excluded). Table 3.2 provdes the summary statstcs for these varables. Table 3.2 Summary statstcs of the data set at muncpal level Varables Obs. Mean Mnmum Maxmum St. Dev. Amount of FNE loans to the ndustral sector between 2000 and 2001 as a proporton of ndustral GDP n 2002* Amount of FNE loans to the ndustral sector between 2000 and 2003 a proporton of ndustral GDP n 2002* Average years of schoolng of the workng force populaton n Infant mortalty rate n Housngl nfrastructure ndex n Gn ndex n Transportaton cost to São Paulo n Populaton densty (nhab./km2) n Note: * Ths varable also ncludes the loans to commerce/servces sectors and ndustral GDP ncludes commercal/servce GDP. Own elaboraton based on IBGE, IPEADATA and BNB datasets. 4. Measurng Effectveness: The Mcro Approach Frst, the emprcal strategy for evaluatng the mpact of the FNE loans on frms employment growth and wage growth s dscussed. As noted earler, the analyss of both varables s mportant because, at least n the short-term, productvty gans may ental job losses. The estmates are based on polcy evaluaton regresson models that control for observable and unobservable varables. More precsely, 17 Prevously, ths dataset was used n Almeda et al. (2007) to analyse the spatal dstrbuton of FNE across muncpaltes. 18 Moreover, to calculate the FNE ndustral ratos, the ndustral GDP also ncludes the GDP n the commerce and servces sectors. 19 I do not take the log of ths varable because t has postve and negatve values. 8

10 the technques try to solve the problem of evaluaton under nsuffcent nformaton about the treated. Next, the results regardng the effectveness of the FNE loans at frm level are shown and dscussed Method: Frst-Dfferencng (FD) The man challenge of any polcy evaluaton s to deal wth the selecton bas ntroduced when random assgnments are not possble. To descrbe ths problem more precsely, we can thnk about the FNE loans granted to frms n the Brazlan Northeast by means of the bnary varable that represents partcpaton, D = {0,1 } whch s 1 f the frm s treated (.e., receved the loan) and 0 otherwse. The observed outcome of varable Y (for nstance, employment growth between 2000 and 2003) for frm would be as follows: Y DY1 + ( 1 D) Y0 =, (4.1) where Y 0 s the employment growth of a frm had t not receved the FNE loan and rrespectve of whether t was actually receved, whereas Y 1 s the frm s employment growth f t receves the loan. The result of nterest s the dfference between Y 1 and Y 0, whch s the causal effect of the FNE loan for frm. Ths analyss would be possble f we could to go back n tme and change a frm s treatment status. However, Angrst & Pschke (2009) pont out that because we never see both potental outcomes for any one frm, we must learn about the effect of ths polcy by comparng the average employment growth of those who were and were not granted the FNE loans. The comparson of the averages of those frms granted and not granted FNE loans s formalsed n the followng equaton: E[ Y D = 1] E[ Y D = 0] = E[ Y1 D = 1] E[ Y0 D = 1] + E[ Y0 D = 1] E[ Y0 D = 0] (4.2) Observed _ dfferences _ n _ average _ employment _ growth Average _ treatment _ effect _ on _ the _ treated Selecton _ bas The term average treatment effect on the treated (ATT) represents the averages dfferences between the employment growth of frms that receved the FNE loan, E [ Y 1 D = 1] and what would have happened to them had not they been granted, E [ Y 0 D = 1]. Nevertheless, we must also consder selecton bas n addton to ths causal effect. Ths last term s the dfference n the average for Y 0 between those frms that were ( D = 1) and those that were not ( D = 0) granted the FNE loans. The pont here s that we do not know what would happen f the granted frms had not receved the loans, E [ Y 0 D = 1]. The selecton bas may be negatve or postve, causng us to underestmate or overestmate the treatment effect. For nstance, one may argue that those frms seekng the loans are more motvated and entrepreneural and that even wthout the FNE loans, they would perform better than others. The random assgnment of FNE loans ( D ) solves the selecton problem. Formally, n the absence of the selecton bas, we would have the followng: E Y D 1] E[ Y D = 0] 0, (4.3) [ 0 = 0 = that s, on average, there would be no dfferences between the potental outcomes for the untreated and treated frms f they had not been granted. Ths does not mean that the FNE loans should be randomly granted, but t demonstrates that the goal of the emprcal economc research s to overcome ths selecton bas usng the approprate polcy evaluaton technques. The benchmark estmaton carred out n the next secton s the dfference between the means of employment growth wthout controls for those who are treated and those who are not. The regresson of Y on D can be used to test the sgnfcance of the treatment effect: Y = α + ρ + η, (4.4) D 9

11 where Y s the employment growth of frm, α s the constant term, ρ s the coeffcent of the treatment effect and η s the random term. In ths case, the treatment effect estmaton reles on the weaker assumpton that D s ndependent of Y 0 wthout placng any restrcton on the relatonshp between D and Y 1 (Wooldrdge, 2002, p.606). It s worth notng that under ths assumpton of there beng no selecton bas, t s unlkely that these estmates provde a relable value for the mpact of the polcy because the frms were not selected randomly. However, these estmates are shown n the results secton for the sake of comparablty wth other emprcal strateges dscussed next. One way to mprove the estmaton of Equaton (4.4) s to add some controls ( X ) that tend to affect both drectly and ndrectly motvatng partcpaton, D the varables of nterest ( Y ): Y = α + ρd + X ' β + υ, (4.5) In ths case, the assumpton s that there are dfferences among frms n the treatment and control groups n terms of ther observable characterstcs. For ths reason, t s necessary to explctly nclude all covarates ( X ) that are mportant for the determnaton of Y and partcpaton, D. In ths stuaton, as explaned by Angrst & Pschke (2009, p.59), the resdual υ s uncorrelated wth the regressors D and X, and the regresson coeffcent ρ s the causal effect of nterest. Ths s the selecton-on-observables assumpton for regresson models (Barnow et al., 1981), whch assumes that the observable characterstcs X are the only reason why η and D are correlated n Equaton (4.4) 20. It s worth notng that the estmates mght be based f the Equaton (4.5) does not consder all varables mportant n determnng partcpaton that also affect the varable of nterest, Y. An alternatve to the econometrc specfcaton (4.5) strategy s to use matchng or propensty score technques. Matchng based on observable characterstcs assumes that the selecton of the FNE loans occurs only based on observable characterstcs, so that frms wth such characterstcs have the same probablty of partcpaton. Then, the average effect of FNE can be obtaned by averagng weghted effects for subgroups of frms wth smlar characterstcs. The drawback of ths estmator s ts mplementaton when there are large number of varables or when these varables are contnuous. The propensty score estmates overcome ths problem, summarsng the smlar characterstcs of frms va the estmaton of a logt or probt model that ndcates the probablty of recevng the FNE loan. Although these matchng approaches are appealng, they are accompaned by the same explct statement of the condtonal ndependence assumpton requred to provde a causal nterpretaton of regresson coeffcents, and for ths reason, we can say that matchng and regresson are both control strateges (Angrst & Pschke, 2009, p.69) 21. In the current work, I use only the regresson analyss approach because the propensty score strategy has already been used n other studes (Slva et al., 2009; Soares et al., 2009). More mportantly, an ssue that has not been dealt n the evaluaton lterature of FNE s the lkely bas due to unobservable characterstcs. For nstance, ths s the case for some dmensons of motvaton/ablty/entrepreneurshp as related to applyng for or recevng the FNE loan. Here, the frst-dfferences (FD) method s employed to elmnate the unobservable effects that are constant 20 In other words, Wooldrdge (2002, p.607) hghlghts that when D and ( Y 0, Y 1 ) are allowed to be correlated, we need an assumpton n order to dentfy treatment effects. Rosenbaum & Rubn (1983) ntroduced the followng assumpton, whch they call gnorablty of treatment (gven observed covaratesx ): Condtonal on X, D and ( Y 0, Y 1 ) are ndependent. 21 Angrst & Pschke (2009, p.69) argue that snce the core assumpton underlyng causal nference s the same for the two strateges, t s worth askng whether or to what extent matchng really dffers from regresson. Our vew s that regresson can be motvated as a partcular sort of weghted matchng estmator, and therefore the dfferences between regresson and matchng estmates are unlke to be of major emprcal mportance. 10

12 over tme (fxed effect). At least two tme perods are needed to carry out ths strategy. Now, I nclude the t subscrpt to ndcate the tme perod as shown below: Y t = α + ρd + X ' β + A ' γ + υ, (4.6) t t t More precsely, Y t s the employment growth of frm between 2000 and , α s the constant term, D t s the dummy varable ndcatng f the frm receved a loan n 2000, X t s a vector of observed tme-varyng covarates n 2000, and υ t s the random term. More mportant s the vector of unobserved but fxed covarates, A, whch wll be ruled out wth the frst-dfference strategy. As descrbed n the mcro dataset secton, I collected nformaton at the frm level for a prevous perod represented by the subscrpt t-1: Y t 1 = + ρdt 1 + X t 1 ' β + A ' γ + υt 1 α, (4.7) where, Y t 1 s the employment growth of frm between 1998 and 2000, D t s the dummy varable (whch s now zero for all frms n 1998), and X t 1 s a vector of observed tme-varyng covarates n Subtractng (4.7) from (4.6) yelds the followng: Y t Y t 1 ( t t 1 t t 1 t t 1 = α + ρ D D ) + ( X X )' β + ( υ υ ) (4.8) Note that n the frst-dfference regresson model, the unobserved fxed effect, A, s elmnated by subtractng the observaton for the prevous tme perod from the observaton for the current tme perod. Then, equaton (4.8) can be estmated by OLS, and the coeffcent of the dummy varable, ρ, means the average mpact on the dfferences for the varable of nterest. Another approach that could be used to deal wth the unobservable factors s the nstrumental varable (IV) tool. To use ths strategy, I would need some wsdom about the mechansms/rules that determne partcpaton ( D ) but are not correlated wth the varable of nterest, Y. In other words, I would need an nstrument to run the frst-stage equaton, D = Z ' π + X ' β + f, where Z s the nstrument, X are the covarates and f s the random term. However, I do not have access to the rules that determne f the FNE loan was granted or not for a specfc frm. Ths nformaton was not provded by the Bank of the Northeast; t s restrcted to the staff workng wth the loans. For ths reason, I have not used ths strategy n the mcro effectveness evaluaton. On the other hand, for the macro level evaluaton dscussed n secton 5, I follow the IV approach because I have an dea at the aggregate level as to why some regons have receved more FNE loans than others. The frstdfference approach s not pursued at the macro level because the nformaton on the covarates s only avalable for Results: The Mcro Approach In ths secton, the results of the mcro-evaluaton are dscussed. It s worth notng once agan that ths approach measures the effectveness of FNE loans only to frms n the ndustral and commerce/servces sectors, seekng to answer the queston of whether the regonal fund creates jobs and/or ncreases productvty (proxed by wage growth). As noted earler, the jont consderaton of these two varables employment and wages growth n the analyss s mportant because they correspond to mplct objectves of FNE and because, at least n the short term, productvty gans may nvolve job losses. Table 4.1 shows the results of the FNE mpact usng three dfferent control samples and coverng two perods ( and ). Also, heteroskedastcty-robust standard errors are 22 In the results secton, I also show the results usng employment growth between 2000 and

13 ndcated n parenthess for all estmates because the dagnostcs for ths problem are statstcally sgnfcant 23. The most mportant results of ths table are shown n the second part, whch employs a sample matched to the treated group. Ths s a better sample than the others for carryng out the mcroevaluaton estmates because t ensures that the treatment and control groups have smlar characterstcs, makng the two groups more comparable as shown earler n the summary statstcs n Table 3.1. For ths sample, the frst results (. OLS wthout covarates) are obtaned usng Eq. (4.4), whch does not control for observable and unobservable characterstcs. Wth ths assumpton n mnd, the results show statstcally nsgnfcant mpacts of FNE on employment growth and the growth of wages between and If we add controls for observable characterstcs whch means estmatng the FNE mpact usng Eq. (4.5) the results n Table 4.1 (. OLS wth covarates) show that there s only a negatve (although small) mpact of FNE on productvty (proxed by wages growth) over the perod. More precsely, wage growth s 3.18 percentage ponts smaller for the fnanced frms than for the non-fnanced frms between 2000 and On the other hand, t s not possble to verfy any mpact of FNE on employment growth n the two perods. Table 4.1 Mcro approach of FNE mpact on employment and wage growth usng three control samples. All frms Employment growth between Wage growth between Employment growth between Wage growth between OLS wthout covarates * * (0.2345) (0.0168) (0.2787) (0.0320). OLS wth covarates *** ** (0.2352) (0.0152) (0.2811) (0.0334).Frst Dfference wth covarates (FD) (0.2784) (0.0216) (0.3166) (0.0364) Observatons 97,403 97,403 97,403 97,403. Matched sample to the treatment group. OLS wthout covarates (0.2347) (0.0170) (0.2794) (0.0300). OLS wth covarates ** (0.2368) (0.0144) (0.2838) (0.0260).Frst Dfference wth covarates (FD) * (0.3055) (0.0230) (0.3375) (0.0342) Observatons 9,335 9,335 9,335 9,335. Random sample. OLS wthout covarates (0.2368) (0.0194) (0.2842) (0.0328). OLS wth covarates (0.2549) (0.0221) (0.3184) (0.0341).Frst Dfference wth covarates (FD) (0.2885) (0.0335) (0.3328) (0.0428) Observatons Note: Robust standard errors n parentheses; * sgnfcant at 10%; ** sgnfcant at 5%; *** sgnfcant at 1%. OLS Dependent varable: Employment growth (or wages growth)= (Et, E0,)/ E0,, where Et, and E0,, are, respectvely, the fnal perod and the ntal perod of employment stock (average wage) for frm. OLS estmatons wth covarates nclude: average age of the employees n 2000, average years of schoolng of the employees n 2000, dummy for the commerce/servces sector, average wage n 2000, number of employees n 2000, dummes for Northeast states. FD Dependent varable= [(Employment growth between ) - (Employment growth between )]. FD estmatons wth covarates nclude: dff. average age of the employees ( ), dff. average years of schoolng of the employees ( ), dff. average wage ( ), dff. number of employees ( ). 23 To save space, the results of dagnostcs for heteroskedastcty are not reported; they wll be provded upon request. 12

14 However, the most mportant estmaton s the one (. frst dfference wth covarates, FD) that controls for observable and unobservable (constant over tme) characterstcs usng Eq. (4.8). When the frst dfferences are estmated, a postve mpact of FNE on employment growth s observed for the perod between 2000 and 2003 (t s statstcally sgnfcant at 10% level). Ths result suggests that employment growth among those frms that receved the FNE loans was about 53.4 percentage ponts hgher than employment growth for those frms that were not granted them. Furthermore, the latter result s robust to alteratons n the condtonng set of the varables, and the sgnfcance level ranges from 5% to 10%, dependng on the set of varables ncluded n the estmaton (n the Appendx B, Table B.1 shows the complete results for these FD estmatons). The frst () and the thrd () part of table 4.1 show, respectvely, the results usng the control sample wth all frms dentfed n RAIS durng the perod under analyss and the control sample used n the paper of Slva et al. (2009), whch employed 727 frms drawn from a random and representatve sample of the populaton of the Northeast wth records n RAIS. Actually, ths last data set s dfferent from that of the one used n Slva et al. (2009); I excluded the few frms n the agrcultural sector and those not present n the RAIS dataset durng the years 1998 and 2006 (see secton 3.1 for detals). The results n parts () and () of Table 4.1 demonstrate that the evaluaton of FNE mpact depends on the choce of control sample. As can be observed, there are no statstcally sgnfcant effects when the FD estmates are carred out usng samples () and (). The results n Table 4.1 suggest that the control sample should be analysed wth cauton. The results n () usng the frst-dfferences approach are n lne wth those of prevous studes. For nstance, Slva et al. (2009) assessed the effectveness of the FNE usng propensty score estmates and found that employment growth s approxmately 60 percentage ponts hgher for those frms that receved loans than for those that dd not receve them over the perod Wth regard to the results n Soares et al. (2009), the mpact of FNE on employment growth over the three-year perod s 33 percentage ponts hgher for fnanced frms. Furthermore, n nether study was any mpact on wage growth found. Altogether, the results that are presented n ths secton suggest usng a method that controls for observable characterstcs and unobserved fxed effects that the FNE loans granted n 2000 played a role n attractng and stmulatng employment growth durng the followng three years. However, even the FNE s beng effectve.e., creatng jobs does not mean that the FNE loans have been able to elmnate or even to reduce Brazlan regonal nequaltes. Indeed, the observed employment growth mght stll be too lmted to have any sgnfcant mpact on GDP per capta or on other socoeconomc varables. The macro-evaluaton presented n the next secton ams to nvestgate ths ssue. 5. Assessng the Impact of FNE on Regonal Inequaltes: The Macro Approach In ths secton, the evaluaton of the mpact of FNE on regonal nequaltes n Brazl s detaled; here, I tested whether FNE loans foster GDP per capta growth at a muncpal level. Frst, ths secton dscusses the emprcal model where the nstrumental varable approach s used to assess the macro mpact of FNE on GDP per capta growth for the targeted muncpaltes. Then, the results are analysed usng some robustness checks wth dfferent perods, controllng varables, and spatal scales Method: Instrumental Varables (IV) Many papers that examne the mpact of regonal funds on regonal economc growth for nstance, the studes about the EU structural funds, such as Rodrguez-Pose & Frates (2004), Dall erba & Le Gallo (2008), and Espost & Bussolett (2008) are based on the neoclasscal growth model descrbed n Barro & Sala--Martn (1991, 1992). Because the prmary purpose of ths outcome evaluaton (at the macro level) s to nvestgate the effects of the regonal fund on economc growth n the Northeast regon, I use the so-called Barro-regresson, whch ncludes the FNE loans varable as showed n Eq. (5.1) whch s estmated by means the Ordnary Least Squares (OLS) method: 13

15 g α + β y + β FNE + β + ε (5.1) = X 3 where g s the GDP per capta growth rate for each muncpalty 24, y 0 s the ntal GDP per capta, FNE represents the FNE ndustral loans as a proporton of ndustral GDP n 2002, and the X matrx ( N K ) represents other muncpal characterstcs that are mportant to economc growth dynamcs; the omsson of relevant varables s avoded. Fnally, α s the ( N 1) constant vector, andε s the N 1 vector of errors. A negatve correlaton between the growth rate and the ntal GDP per capta ( β 1<0) suggests ether mean reverson or condtonal β-convergence. Most mportantly, f β 2 s postve and statstcally sgnfcant, then FNE has a postve mpact on economc growth, ncreasng the transtonal growth rate of each muncpalty towards ts own steady state. However, the problem of endogenety remans n the estmaton of the FNE mpact because the regonal fund loans are not allocated randomly but are nstead concentrated n the wealthest areas of the Northeast regon. Olvera & Domngues (2005) suggest that the regonal funds are drven by the demand sde that s, they are requested by producers that fulfl the fund s requrements. Thus, t s lkely that only the most developed actvtes, located n muncpaltes wth good access to nformaton and bankng nfrastructure, have access to these funds. In ths sense, the FNE varable potentally ntroduces a very serous problem of endogenety nto the model and thus ntroduces bas nto the OLS estmates. Specfcally, the problem of endogenety stems from the smultanety of FNE loans and economc growth n an SEM (Smultaneous Equaton Model) framework. Alternatvely, we can thnk n the language of omtted varable bas (OVB), notng that frms or persons located n some muncpaltes have more motvaton or ablty to obtan a FNE loan and that ths characterstc s unobservable. Ths makes the FNE varable and the error term correlated. The trck for solvng the endogenety problem s to fnd a varable (the nstrument) that s correlated wth the causal varable of nterest (n ths case, FNE) but uncorrelated wth any other determnants of the dependent varable; or, equvalently, the nstrumental varable s uncorrelated wth error termε (Angrst & Pschke, 2009). More precsely, the relevant queston that the IV technque helps to answer s whether the Northeast muncpaltes have grown faster than they would have n the absence of the FNE loans. The two-stage least squares (2SLS) estmaton s the workhorse of the IV approach. To mplement ths technque, the followng frst-stage equaton s estmated by means of the OLS method: FNE = α + X π 1 + π 2 z + υ, (5.2) where X s the matrx of the covarates (ncludng y 0 ) that appear n Eq. (5.1), z s the nstrumental varable, α s the constant vector andυ s the error term. Then, Eq. (5.1) s estmated usng the ftted values of the FNE varable, FN ˆ E, whch s a lnear functon of the nstrument and therefore by assumpton uncorrelated wth errorε. The dea s to use only the good varaton n the FNE varable (or at least some of t) that s, some part of the FNE varable that s not correlated wth the unobserved effects n the error term, ε. As ponted out by Temple (1999) there s a lack of good nstruments because so many varables can be used to explan growth that t s dffcult to fnd varables that are hghly correlated wth the endogenous varables that can be excluded from the growth regresson. In the next secton, I show the results usng the number of loans granted to frms n the ndustral sector n 1999 as the nstrument because ths varable s correlated wth the FNE ndustral rato n the perod and can be excluded from the growth regresson. The ntuton s that the number of loans granted n 1999 may represent the lkelhood of success of applcatons for the loans n subsequent years by the frms located n a specfc muncpalty. 24 g = (1/T)*ln(y T, /y 0, ), where y T, and y 0,, are, respectvely, the fnal perod and the ntal perod of per capta GDP and where T s the tme perod n years. 14

16 5.2. Results: The Macro Approach As dscussed above, the goal of the macro-evaluaton s to test f the total amount of the FNE loans to frms n the ndustral and commerce/servces sectors fosters GDP per capta growth at the muncpal level n the Northeast regon to reduce the regonal nequaltes n Brazl. It s worth notng that to make the mcro- and macro-evaluatons more comparable, the FNE ndustral ratos also nclude the allocaton of funds to the commerce/servces programmes, whch totalled 1% durng the perod and 6% durng the perod (the ndustral programmes alone represented 36% and 37% of the total loans). Table 5.1 shows the results regardng FNE mpact on GDP per capta growth durng dfferent tme perods. Heteroskedastcty-robust standard errors are provded n parenthess for all estmates because the dagnostcs for ths problem are statstcally sgnfcant. The results are shown for four dfferent tme perods for the dependent varable: GDP per capta growth between 2002 and 2006, between 2003 and 2006, between 2002 and 2004 and between 2004 and As explaned earler, the muncpal GDP s only comparable from 2002 onwards due to an alteraton n the IBGE methodology n Moreover, the results usng OLS and 2SLS, both wth and wthout covarates, are shown. The frst step s to estmate Eq. (5.1) va OLS, payng attenton to the FNE ndustral coeffcents. Wth regard to the mpact of FNE ndustral rato n on the GDP per capta growth between 2002 and 2006, the OLS estmates (a.1) and (a.2) n Table 5.1 show a negatve and statstcally sgnfcant mpact. Muncpaltes wth hgh FNE ndustral ratos n the perod experence slower growth between 2002 and These results hold wth or wthout condtonng varables (a.2 or a.1, respectvely). However, t s mportant to note that the endogenety problem dscussed earler may be basng these FNE macro mpact estmates. An nstrumental varable s used n the 2SLS estmatons (a.3 and a.4) to overcome the endogenety problem. Dall erba & Le Gallo (2008) also deal wth ths problem n the context of EU structural funds usng the IV technque and hghlght that the endogenety of the allocaton of regonal funds has been overlooked n the lterature. In ths paper, only one nstrument s used. Ths s done avod the over-dentfcaton problem; the ncluson of many weak nstruments ncreases bas 25. The nstrument used s the total number of loans at the muncpal level granted to frms n the ndustral sector n Ths varable s used n the frst-stage equaton [Eq. (5.2)] and s supposed to have a clear effect on the FNE ndustral rato, but t should not have an effect on the GDP per capta growth other than through the frst-stage channel. The results of the frst-stage equaton n (a.3) and (a.4) show that the nstrument s able to explan the FNE ndustral rato because the nstrument coeffcent s statstcally sgnfcant, showng that muncpaltes wth a hgh number of FNE loans granted to ndustral frms n 1999 wll have hgh FNE ndustral ratos n the perod. For ths reason, the IV estmator may be used. The 2SLS estmate of FNE ndustral coeffcents (wthout covarates) remans negatve but now s statstcally nsgnfcant (a.3). Moreover, when condtonng varables are added n the model (a.4), the coeffcent becomes postve but s stll statstcally nsgnfcant. Ths result s robust to alteratons n the condtonng set of the controllng varables 26 and s reported n Table B.2 n Appendx B. The most mportant element of ths result s that the FNE ndustral rato does not have any negatve effect on the GDP per capta growth as suggested by the OLS estmates. Ths result shows a more plausble concluson regardng the mcro- and macro-effects of the FNE loans to the ndustral sector. In fact, the postve and statstcally sgnfcant mpact of the FNE ndustral on job creaton verfed n the mcro evaluaton mght stll be too lmted to have any sgnfcant mpact on the GDP per capta growth at the muncpal level and thus reduce the regonal mbalances n Brazl. 25 As dscussed by Angrst & Pschke (2009, p.209), just-dentfed 2SLS s medan-unbased. However, the authors note that ths s not to say you can happly use weak nstruments n just-dentfed models. Wth a weak nstrument, justdentfed estmates tend to be too mprecse to be useful. 26 See Brock et al. (2003), who revew the model uncertanty lterature. 15

17 Table 5.1 Macro approach of FNE mpact on GDP per capta growth over dfferent tme perods a. FNE ndustral rato between 2000 and 2001 GDP pc growth between GDP pc growth between GDP pc growth between GDP pc growth between a.1. OLS wthout covarates ** *** (0.0671) (0.0649) (0.0820) (0.1047) a.2. OLS wth covarates *** *** (0.0510) (0.0543) (0.0691) (0.0948) a.3. 2SLS wthout covarates * (5.2132) (4.6772) (8.3922) (5.9426) Frst-stage equaton: ** ** ** ** Instrument (number of loans granted n 1999) (0.0001) (0.0001) (0.0001) (0.0001) a.4. 2SLS wth covarates (8.1137) ( ) ( ) ( ) Frst-stage equaton: * * * * Instrument (number of loans granted n 1999) (0.0001) (0.0001) (0.0001) (0.0001) Observatons b. FNE ndustral rato between 2000 and 2003 b.1. OLS wthout covarates ** * (0.0558) (0.0766) (0.1260) (0.1579) b.2. OLS wth covarates (0.0864) (0.0712) (0.1718) (0.1166) b.3. 2SLS wthout covarates ** (1.3803) (1.4035) (1.9730) (1.8022) Frst-stage equatons: *** *** *** *** Instrument (number of loans granted n 1999) (0.0002) (0.0002) (0.0002) (0.0002) b.4. 2SLS wth covarates (5.2372) (7.8534) (6.2065) (9.7993) Frst-stage equatons: Instrument (number of loans granted n 1999) (0.0003) (0.0003) (0.0003) (0.0003) Observatons Note: Robust standard errors n parentheses; * sgnfcant at 10%; ** sgnfcant at 5%; *** sgnfcant at 1%. Dependent varable = (1/T)*ln(yT,/y0,), where yt, and y0,, are, respectvely, the fnal perod and the ntal perod of per capta GDP and T s the tme perod n years. The estmatons wth covarates nclude: ln(gdp per capta n 2002), ln(average years of schoolng n 2000), ln(gn ndex n 2000), ln(nfant mortalty rate n 2000), housng nfra-structure ndex n 2000, ln(populaton densty n 2000), ln(transportaton cost to SP n 1995), and dummes for Northeast states. Another possble explanaton for ths lack of effect at the muncpal level may be the delayed effect of FNE ndustral loans on the GDP per capta growth, so that ther mpact does not appear n the tme perod analysed ( ). To overcome ths ssue, at least partally, I estmated the same models usng other tme perods, tryng to fnd any dfferent sgnfcant mpact. The GDP per capta growth of the tme perods of , and was used n those regressons as the dependent varable. In addton, estmatons of the FNE ndustral rato between 2000 and 2003 were carred out usng GDP per capta growth as the dependent varable for those four dfferent tme perods (part b n Table 5.1). Agan, the estmates obtaned wth these regressons do not ndcate any sgnfcant mpact when condtonng varables are added n the model (b.4). Fnally, from an emprcal pont of vew, we do not know f the results for a sngle scale hold for another scale choce. Ths fact s lnked to a measurement ssue that can cause varablty n the estmated coeffcents due to the use of dfferent levels of spatal aggregaton of the observatonal unts. Ths varablty could occur because of the exstence of the Modfable Areal Unt Problem (MAUP) (Gehlke & Behl, 1934; Robnson, 1950; Openshaw & Taylor, 1979) and the ecologcal fallacy problem (Stoker, 1993; Anseln, 2002). These two problems stem from the fact that there s an aggregaton problem that mght prevent us from dentfyng the real scale on whch processes 16

18 operate. For nstance, a cautous analyss of FNE mpact on economc growth on dfferent spatal scales may allow one to dentfy the approprate spatal scale for evaluatng ths regonal polcy. The macro dataset s avalable at the muncpal level and s merged to form other two spatal scales: 189 mcro-regons and 22 spatal clusters n the Northeast regon. The mcro-regons were defned by IBGE n 1990 as beng a group of contguous muncpaltes n the same state. They were grouped accordng to natural and producton characterstcs. The spatal cluster level proposed by Carvalho et al. (2007) employs a cluster methodology (algorthmc) that groups contguous muncpaltes that share smlar characterstcs usng 46 varables reported n the Brazlan Census of Wth regard to the results at both spatal levels (mcro-regons and spatal cluster), the conclusons are smlar to those at the muncpal level 28. Altogether, the results at varous spatal scale levels suggest that there are no statstcally sgnfcant postve effects of FNE ndustral loans on economc growth. 6. Conclusons Evaluatons are employed n several felds that am to answers questons such as when and how nterventons or treatments work. The task of the evaluaton carred out n ths paper was to ad n decsons about mprovements or modfcatons to and expansons of the regonal fund for the Northeast regon (FNE) n Brazl. The contrbuton of ths paper s that t brngs together two types of outcome evaluaton that are often mplemented separately n the evaluaton lterature, showng a more complete pcture of the FNE loans drected toward frms n the ndustral/commerce/servces sectors (coned as FNE ndustral). These mcro- and macro-effects have been overlooked n the lterature dealng wth the mpact of regonal development funds. Frst, t s worth notng that both evaluatons (mcro and macro) focus on the FNE loans granted to frms n the ndustral, commerce and servces sectors, whch represent roughly 40% of the total loans n the analysed perod. It was not possble to evaluate the rural programmes of the FNE n ths paper because there were no mcro data avalable so far. Ths s because a porton of the FNE loans (approxmately 60%) are granted to ndvduals that have small busnesses or who conduct smallscale farmng n the nformal sector.e., they do not have a CNPJ number and are not covered by RAIS, whch s the source of nformaton for the mcro-evaluaton. In the future, a specfc survey should be formulated to collect nformaton about the ndvduals and small rural busnesses n the Northeast for the treated and control groups to evaluate ths mportant targeted FNE populaton. The mcro-evaluaton seeks to answer the followng queston: Dd the subsdsed FNE loans cause an ncrease n employment (and wages) n the target frms? The results based on the control sample matched to the treated group usng the frst-dfferences method that controls for observable characterstcs and unobserved fxed effects suggest that the FNE ndustral loans play a role n attractng and stmulatng employment growth n the Northeast regon between 2000 and However, these effects on employment growth wane and dsappear over a long perod: Moreover, t s mpossble to verfy any mpact of FNE ndustral loans on frm productvty (proxed by wage growth) n the estmates that control for observable and unobservable characterstcs. As dscussed n the paper, the postve effects on job creaton at the frm level do not mean that the FNE loans have been able to elmnate or even to reduce Brazlan regonal nequaltes. For ths reason, a second outcome evaluaton was carred out to attempt to answer the followng queston: Have regonal nequaltes been reduced as a result of the FNE loans? The emprcal strategy was performed to test whether FNE loans foster GDP per capta growth at the muncpal level n the Northeast regon. The nstrumental varable approach was mplemented because of the potental endogenety of the FNE ndustral varable gven that FNE loans are not allocated randomly. The IV results do not ndcate any postve mpact of FNE ndustral ratos on economc growth between 2002 and 2006 or n other sub-perods. Ths concluson s robust to combnatons of the set of 27 The varable lst ncludes employment n 17 sectors of economc actvty, educaton, health, ncome, the urbansaton rate, the prevalence of volence, housng condtons and others. See Carvalho et al. (2007) for further detals. 28 To save space, these results are not reported; they wll be provded upon request. 17

19 controllng varables and to the use of dfferent levels of spatal aggregaton of the observatonal unts. Altogether, the mcro- and macro-evaluaton results suggest that the effect of FNE ndustral loans on employment growth at the frm (mcro) level mght stll be too lmted for t to have any sgnfcant mpact on GDP per capta growth at the muncpal (macro) level n the Northeast regon or thus to reduce the regonal nequaltes n Brazl. It s worth notng that the Brazlan regonal funds have the broad objectve of reducng regonal nequaltes as defned by federal law but that no varable or measure of nequalty was well defned when the polcy was mplemented. Ths study assumed that the reducton of the GDP per capta gap s the polcy objectve at the macro-scale level and that job creaton and ncreasng productvty (proxed by wage growth) are the objectves at the frm (mcro) level. The outcome evaluaton may be hampered by ths lack of a precsely defned objectve; t s always hard to defne a measure for polcy evaluaton f t does not actually exst 29. Fnally, despte some recent changes n Brazlan regonal polcy namely, the adopton of the Natonal Regonal Development Polcy (PNDR) mplemented by the Mnstry of Natonal Integraton (MI) through Decree n of t s stll necessary to defne a relevant system for apprasal, montorng and outcome evaluaton coverng all desgned nterventons at both the frm/ndvdual and macro levels. Furthermore, t s mportant to demonstrate to publc admnstrators and legslators the benefts and costs of more rgorous outcome evaluatons. As noted by Bartk & Bngham (1995), t s dffcult to convnce someone to do somethng (n ths case, outcome evaluaton) that has not been done before. In addton, those authors argue that once polcymakers have seen that a hgh-qualty evaluaton of the regonal development funds can help mprove polcy performance and poltcal vablty, the nterest n outcome evaluatons should ncrease. References ALMEIDA JUNIOR, M., RESENDE, G. M., SILVA, A. M. (2007) Dstrbução espacal dos fundos consttuconas de fnancamento do Nordeste, Norte e Centro-Oeste. Revsta de Economa, Edtora UFPR, v. 33, n. 2, p ANSELIN, L. (2002). Under the hood: ssues n the specfcaton and nterpretaton of spatal regresson models. Agrcultural Economcs, v. 27, p ANGRIST, J. & PISCHKE, J. (2009) Mostly harmless econometrcs: an emprcst's companon. Prnceton Unversty Press, London, UK. BANCO CENTRAL (2000). Juros e Spread Bancáro no Brasl. Departamento de Estudos e Pesqusas do Banco Central. Avalable at: BANCO DO NORDESTE (2001) Programação do Fundo Consttuconal de Fnancamento do Nordeste FNE para o ano de Fortaleza, p BANCO DO NORDESTE (2009) Programação do Fundo Consttuconal de Fnancamento do Nordeste FNE para o ano de Fortaleza, p BARNOW, B.; CAIN, G.; GOLDBERGE, A. (1981) Selecton on observables. Evaluaton Studes Revew Annual 5, BARRO, R. & SALA-I MARTIN,X. (1991) Convergence across states and regons. Brookng Papers on Economc Actvty, n.1, p BARRO,R. & SALA-I MARTIN,X. (1992) Convergence. Journal of Poltcal Economy, v.100, p BARTIK, T. & BINGHAM, R. (1995) Can Economc Development Programs Be Evaluated? Staff Workng Papers 95-29, W.E. Upjohn Insttute for Employment Research. 29 A clear example s the Central-West regon, whch has a regonal fund (FCO): GDP per capta n the Central-West regon was 75% of the natonal average n 1989 and 123% n Because the Brazlan regonal polcy does not defne macro- and mcro-level ndcators for polcy success that are verfable, measurable and drectly lnked to the nterventons fnanced by the regonal fund, we do not know f Central-West regon needs to be the focus of regonal polcy at ths tme. 30 Avalable at < 18

20 BRASIL. (2008) Consttução da Repúblca Federatva do Brasl. BROCK, W., DURLAUF, S., WEST, K. (2003) Polcy evaluaton n uncertan economc envronments. Brookngs Papers of Economc Actvty, p CARVALHO, A.; LALL, S. V.; TIMMINS, C. (2006). Regonal Subsdes and Industral Prospects of Laggng Regons. World Bank Polcy Research Workng Paper n CARVALHO, A. Y., DA MATA, D., RESENDE, G. M. (2007) Clusterzação dos muncípos brasleros, In: Dnâmca dos Muncípos, Brasíla: IPEA. CENSO DEMOGRÁFICO 2000: documentação dos mcrodados da amostra. Ro de Janero: IBGE. DALL ERBA, S. (1995) Dstrbuton of regonal ncome and regonal funds n Europe : An exploratory spatal data analyss. The Annals of Regonal Scence, v. 39, p DALL ERBA, S. & LE GALLO, J. (2008) Regonal convergence and the mpact of European structural funds over : A spatal econometrc analyss. Papers n Regonal Scence, v. 87, n. 2, p ESPOSTI, R. & BUSSOLETTI, S. (2008) Impact of Objectve 1 Funds on Regonal Growth Convergence n the European Unon: A Panel-data Approach. Regonal Studes, v. 42, n. 2, p FERREIRA, P. C. (2004) Regonal Polcy n Brazl: A Revew. Mmeo. GEHLKE,C. & BIEHL,K. (1934) Certan Effects of Groupng Upon the Sze of the Correlaton Coeffcent n Census Tract Materal. Journal of Amercan Statstcal Assocaton, v.29, n IPEADATA. Dados macroeconômcos e regonas. Avalable at: < peadata.gov.br> JANN, W. & WEGRICH, K. (2007). Theores of the Polcy Cycle. In Fscher, F; Mller, G.; Sdney, M. (orgs.) Handbook of Publc Polcy Analyss. LAW Nº 7827, (1989). < LE GALLO, J., ERTUR, C. (2003) Exploratory spatal data analyss of the dstrbuton of regonal per capta GDP n Europe, Papers n Regonal. Scence, v.82, p LEONARDI, R. (2006) Coheson n the European Unon. Regonal Studes, v. 40, n.2, p OLIVEIRA, H. & DOMINGUES, E. P. (2005). Consderações sobre o mpacto dos FNO e FCO na redução da desgualdade regonal no Brasl.In: XXXIII Encontro Naconal de Economa. OPENSHAW, S. & TAYLOR, P.J. (1979) A mllon or so correlaton coeffcents: three experments on the modfable areal unt problem. In: N. Wrgley, ed. Statstcal Applcatons n the Spatal Scences. London: Pon, p RAIS (1998, 2000, 2003, 2006) MINISTÉRIO DO TRABALHO. Relatóro Anual de Informações Socas. ROBINSON, W. S. (1950) Ecologcal correlatons and the behavour of ndvduals. Amercan Socologcal Revew, v.15 p RODRIGUEZ-POSE, A. & FRATESI, U. (2004) Between development and socal polces: The mpact of European structural funds n objectve 1 regons. Regonal Studes, v.38, p ROMERO, R. & NOBLE, M. (2008) Evaluatng England s New Deal for Communtes programme usng the dfference-n-dfference method. Journal of Economc Geography, v. 8, p SILVA, A. M., RESENDE, G. M., SILVEIRA NETO, R. (2007) Uma Avalação da Efcáca do FNE, no período Revsta Análse Econômca (UFRGS), v. 25, p SILVA, A. M.; RESENDE, G. M.; SILVEIRA NETO, R. (2009). Efcáca do gasto públco: uma avalação do FNE, FNO e FCO. Estudos Econômcos, v. 39, n. 1, p SOARES, R.; SOUSA, J.; PEREIRA NETO, A. (2009) Avalação de Impactos do FNE no Emprego, na Massa Salaral e no Saláro Médo em Empreendmentos Fnancados. Revsta Econômca do Nordeste, v. 40, n. 1, p STOKER, T. M. (1993) Emprcal approaches to the problem of aggregaton over ndvduals. Journal of Economc Lterature, v.33, p TEMPLE, J. (1999) The new growth evdence. Journal of Economc Lterature, v.37, n.1, p TEMPLE, J. (2000) Growth regressons and what the textbooks do not tell you. Bulletn of Economc Research, v. 52, p WOODBRIDGE, J. (2002) Econometrc Analyss of Cross Secton and Panel Data. Cambrdge, MA: MIT Press. 19

21 Appendx A Table A.1 Regonal GDP per capta as a proporton of Natonal GDP pc n Brazl between 1989 and 2006 Source: Own elaboraton based on IBGE data. Table A.2 Regonal Funds (FNE, FNO, FCO) and BNDES loans by regon ( ) (R$ mllon_current prces) Regon Source of loans Total Northeast BNDES 2,783 3,334 3,784 3,112 2,737 3,803 4,836 5,322 29,712 FNE ,019 3,209 4,174 4,588 4,247 18,362 BNDES/FNE North BNDES , ,954 1,616 1,626 3,461 13,039 FNO ,075 1, ,110 7,224 BNDES/FNO Centre- BNDES 2,064 1,703 2,589 2,831 5,161 3,271 3,659 5,755 27,032 West FCO , ,172 1,468 1,444 1,974 9,688 BNDES/FCO Southeast BNDES 13,008 14,494 23,074 20,036 21,299 28,740 31,415 37, ,646 South BNDES 4,261 4,826 6,092 6,842 8,683 9,551 9,783 12,773 62,809 Total BNDES all regons 23,046 25,217 37,419 33,534 39,834 46,980 51,318 64, ,239 BNDES NE+NO+CO regons (A) 5,777 5,897 8,254 6,656 9,852 8,689 10,121 14,538 69,784 FNE+FNO+FCO (B) 1,558 1,735 2,298 3,014 5,702 6,618 7,018 7,331 35,274 (A) / (B) Source: Own elaboraton based on BNDES and Mnstry for Natonal Integraton (MI) data. 20

22 Fgure A.1 Gross Domestc Product (GDP) per capta n 2002 n the Northeast (at muncpal level) Source: Own elaboraton based on IBGE data. GDP per capta n 2002 s at constant 2000 prces n R$. Fgure A.2 Spatal dstrbuton of FNE ndustral ratos at muncpal level Source: Own elaboraton based on IBGE and BNB data. The FNE ndustral ratos also nclude the loans to commerce/servces sectors and ndustral GDP n 2002 ncludes commercal/servce GDP. 21