82086 AUTHOR ACCEPTED MANUSCRIPT PRELIMINARY INFORMATION Expanding Microfinance in Brazil : Credit Utilization and Performance of Small Firms Accepted for publication in Journal of Development Studies To be published by Taylor and Francis THE FINAL PUBLISHED VERSION OF THIS ARTICLE WILL BE AVAILABLE ON THE PUBLISHER’S PLATFORM This Author Accepted Manuscript is copyrighted by the World Bank and published by Taylor and Francis. It is posted here by agreement between them. Changes resulting from the publishing process—such as editing, corrections, structural formatting, and other quality control mechanisms—may not be reflected in this version of the text. This Author Accepted Manuscript is under embargo for external use and is made available for internal World Bank use only. It is not for distribution outside the World Bank. © 2013 The World Bank Expanding Microfinance in Brazil: Credit Utilization and Performance of Small Firms Emmanuel Skoufias, Phillippe Leite, and Renata Narita Poverty Reduction and Equity Unit THE WORLD BANK January 2013 Abstract We take advantage of the natural experiment generated by the exogenous change in government policy towards microcredit to evaluate the impact of the increased supply of microcredit on the utilization of credit by micro-entrepreneurs. Based on micro-entrepreneurs survey and administrative data from a microcredit program in Brazil, we show that (i) the increased supply of microcredit raised formal credit utilization and this does not crowd-out the use of informal credit sources, (ii) formal credit taking improves business performance, and (iii) returns are larger for women than for men-owned firms, but males employ significantly more workers after taking formal credit than females. JEL classification: O16, G21, L25 Keywords: Average treatment effect on the treated, intent to treat effect, microfinance, firm performance, CrediAmigo. Corresponding author: Emmanuel Skoufias, The World Bank (Mail Stop: MC4-415), 1818 H Street NW, Washington DC 20433-USA. tel: (202) 458-7539. fax: (202) 522-3134. e-mail: eskoufias@worldbank.org. I. Introduction Although Brazil is now one of the world’s ten largest economies, high levels of poverty and inequality continue to pose major challenges for the country. The spatially uneven development of the Brazilian economy among the five major regions of the county, with the north and northeast regions lagging behind the south and southeast regions, accompanied by the persistently high level of inequality have raised many economic and social concerns. Primary among these is the belief that high levels of inequality may compromise economic efficiency and growth. Credit and insurance market failures, for example, may prevent poorer households from investing in and contributing to the economy at an optimal level, thereby undermining efficiency and growth. As a consequence, many of the efforts of the Brazilian government to date have focused on fostering the availability of financial services and expanding access to microfinance. Directly or indirectly, micro- and small enterprises in Brazil are estimated to provide the primary source of income for almost 60 million people and generate nearly 20 per cent of GDP (Kumar, 2005). Lack of access to finance for these micro-entrepreneurs can be a major constraint to their ability to effectively perform their activities, or to increase their scale of operations. This suggests that microfinance has the potential to induce significant favorable impacts in the country. A leading example of such efforts is the Banco do Nordeste do Brasil (BNB) - a state- owned development bank created in 1952 to promote the development of the northeast region of Brazil - aimed at expanding access to credit in this relatively poor region of Brazil. 1 The CrediAmigo program is a flagship program of BNB that initiated its operations in April 1998 and offers loans to established micro-entrepreneurs for the financing of their working capital and fixed asset needs. 1 In this study, we take advantage of the natural experiment generated by the exogenous change in government policy towards microcredit to evaluate the impact of the increased supply of microcredit on the utilization of credit by micro-entrepreneurs in the northeast region of Brazil. The survey data available do not identify BNB or the CrediAmigo program explicitly as a source of credit. However, in consideration of the fact that CrediAmigo was by far the largest microcredit program operating in the northeast region of Brazil between 1998 and 2003, our analysis presumes that the credit from public or private banks reported by micro- entrepreneurs in the national survey, at least in the northeast region of Brazil, originates from the CrediAmigo program. The first question we address is whether the increased supply of credit in the northeast region resulted in higher utilization of formal (or bank) credit in the region. Based on administrative data from the CrediAmigo program, we identify all the municipalities in the northeast region of Brazil where the program operated and combine this information with the 1997 and 2003 detailed micro-entrepreneur surveys on the informal urban economy (ECINF) collected by the Brazilian Institute of Geography and Statistics (IBGE) for firms with five or fewer employees. We employ the Difference-in-Differences estimator to estimate the “intent to treat effect” of the increased availability of credit between 1997 and 2003 on the utilization of credit by micro-entrepreneurs in the municipalities where the CrediAmigo program operated. The second question of interest explored is whether the utilization of formal credit “crowds-out” credit from other informal sources, such as credit from friends and relatives, or product suppliers. The interest rate on loans offered by CrediAmigo is substantially lower than that charged by informal moneylenders and trade credit providers. Thus, it is quite possible that the availability of relatively cheaper credit through the formal banks may simply replace the credit obtained through other sources with little or no increase in the overall access to credit 2 by micro-entrepreneurs who never had credit before (for example, Jain (1999)). This implies that it is also important to investigate whether the utilization of credit increased, irrespective of whether the source is formal or informal. By 2003, CrediAmigo had been operating in the northeast region for more than four years, which is a sufficiently long period of time in which to expect changes in the behavior and performance of credit-taking micro-entrepreneurs. For this reason, we also examine whether the use of credit from formal banks is associated with better performance, such as a higher profitability and size in terms of number of employees. For this purpose, we use data from the 2003 ECINF survey and compare the performance of firms utilizing credit from public or private banks with the performance of a comparison group of firms using credit from informal sources or no credit at all. In this analysis we employ the Propensity Score Matching (PSM) method to match one or more entrepreneurs to a micro-entrepreneur who used credit from a private or public bank and estimate the “Average Treatment on the Treated” effect of credit on firm performance. Our paper relates to a broad literature on microfinance and business development in developing countries; see, for example, Paulson and Townsend (2004), McKenzie et al (2008a), Banerjee et al (2010), Karlan and Zinman (2010), Dupas and Robinson (2011). For Brazil, there is little evidence on this topic. The closest research to ours is Neri (2008) who uses ECINF data for 1997 and 2003 to compare average credit utilization for the northeast, where most CrediAmigo loans were applied, vis-à-vis that of the rest of the country. He finds that access to credit increased by 1.6 percent. In terms of evidence for the impact of credit on firm performance, Neri uses data from beneficiaries of CrediAmigo and matches it with a control group constructed from the longitudinal employment survey of Brazil (PME). Using a one-year window from 2005-2006, he finds that the average impact of CrediAmigo on profits vis-à-vis 3 that on similar households sampled in PME is 7.7 percent, and statistically significant at the 5 per cent level of significance. However, one major shortcoming of this approach is that the profit measure across treatment and control groups is different thus raising concerns about the reliability of the estimated effects of credit on performance. 2 The paper is organised as follows. Section II describes the data. Section III contains the description of the empirical methods employed to address the three questions posed above, and a discussion of the findings. Section IV concludes. II. Data The Brazilian survey of small firms Our analysis is based on data from the 1997 and 2003 Economia Informal Urbana (ECINF) 3 survey, which gathers socio-economic information from micro-entrepreneurs throughout Brazil. In particular, the ECINF is a repeated cross-section survey that collects data on owners of informal enterprises, identified as self-employed or employers with less than five employees, in urban areas of Brazil. The survey does not include: (i) non-agricultural units in the rural areas, (ii) homeless population in urban areas; (iii) people connected to illegal activities, and (iv) domestic workers. The dataset is rich in information on firm and entrepreneur characteristics, but also on the entrepreneur’s personal and household characteristics. The sampling in each year (1997 and 2003) of the ECINF survey is carried out following a two- stage methodology aimed at determining the role and dimension of the informal activities in the Brazilian economy. First, IBGE carries out a listing of all households in different enumeration areas randomly selected 4 from a sample frame setting each one of 27 Brazilian 4 states as a stratum 5. This sort of census in selected enumeration areas intends to identify target micro-economic units. Then, on the basis of the listing of micro-economic units, IBGE selects the final sample of units applying a very detailed questionnaire. The sample was designed to cover all different types of activities in sectors such as: manufacturing, construction, trade, lodging and food services, transportation, rendering services, technical and auxiliary services, and others services. The first listing of households in randomly selected enumeration areas contained more than one million households, of which 30 per cent had at least one person working as own account or micro-entrepreneur, the micro-economic unit. Hence, 49,664 micro-economic units were selected over 27 Brazilian states (urban areas only). Table A1 in the online Appendix shows by year a summary of the micro enterprise data used in the paper. In 2003, the complete data set has 48,813 respondents, representing 10,526,074 entrepreneurs in Brazil. When restricting the attention to the northeast, we have 19,200 respondents representing 2,763,987 entrepreneurs. When we look at whether formal credit has impacts on the performance of small businesses, the “treatment” is identified by whether the entrepreneur has been using credit or a loan from a bank (public or private) in the previous three months for his/her activity. Only 6.2 per cent of all entrepreneurs in the northeast did use some type of credit, with 3.6 per cent using formal credit sources ( from a bank) and 2.6 per cent from informal sources (from friends or family, other enterprises, or a supplier). Unfortunately we are unable to determine whether an entrepreneur has been using both types of sources, since the manner in which the questions are posed allow only the main source of credit to be identified. The use of credit is slightly more common and statistically different for female entrepreneurs (6.97 per cent of them declare to have used credit in the previous three months) 5 than among male entrepreneurs (5.69 per cent). This is mainly due to the wider difference in utilization of informal credit (3.04 per cent of women entrepreneurs used informal credit, as opposed to 2.30 per cent of men, which are statistically different at 5%). Considering business performance, the average profit of small firms in Brazil is 821 Reais (US$ 355) while in the northeast this is smaller at 596 Reais (US$ 258). The average size of small firms is 1.48 and 1.42, respectively, in the entire country and in the northeast region. Although credit is more used among female entrepreneurs, businesses run by females seem to underperform at least when we look at profits. On average, men earn 947 Reais while women 592 Reais. In the northeast, the gap is even larger with men earning 706 Reais and women 416 Reais, both differences being significant at 1%. Men also run slightly larger businesses with 1.50 (1.46) workers on average and women with 1.43 (1.37), respectively, in Brazil and northeast region. These differences are also found significant at 1% level. In terms of the main characteristics of the entrepreneurs, 65 per cent are males, they are on average 41 years old and have started working at age 15; 57 per cent have less than secondary education, 31 per cent have secondary, and only 12 per cent beyond secondary. About 85 per cent are self employed and the remaining employers, 10 per cent have partners, 29 per cent have their domicile as main workplace, 33 per cent did not need capital to start up, and 50 per cent did not register the business. From 1997 to 2003, significantly, formal credit utilization increased by 2.4 percentage points and profits decreased 1.6 percent in the northeast region following a lower economy growth in 2003. CrediAmigo program CrediAmigo is a microcredit program which is operated by the state-owned Banco do Nordeste do Brasil (BNB). Its operations began in April 1998 aiming to improve the 6 development of the northeast through expanding access to microloans in this relatively poor region in Brazil. The program operates in municipalities from nine states in the poor northeast region and in some poorer municipalities in the state of Minas Gerais and Espirito Santo in the southeast region. Priority in the implementation of CrediAmigo was given to those cities where microfinance was nonexistent or very small. Unlike other microfinance programs focusing on informal micro-enterprises in rural areas, in Brazil most of the informal micro-enterprises are located in urban areas. The CrediAmigo program remained unrivaled in scale in Brazil at least until 2003 and constitutes a significant part of the microfinance expansion of the past two decades. The design of the CrediAmigo incorporates key features of successful microfinance interventions around the world. 6 Loans are collateral-free, but are extended using the solidarity group technique to small groups of 3-5 borrowers who cross-guarantee each other’s loan. The average size of CrediAmigo loan was 590 Reais (or US$256 at the prevailing exchange rate). Loans usually last for 3 months. First loans are limited to 300 to 1000 Reais (about US$130 – 433), but repeat loans can be up to 4,000 Reais (US$1,733). The program initiated with a five per cent flat monthly rate, but the interest rate has since decreased to two per cent. 7 As mentioned earlier, interest rates are substantially lower than those charged by informal moneylenders and trade credit providers. After a borrower has successfully paid back two loans under the solidarity program, the borrower becomes eligible for individual credit, with a loan maturity of up to 6 months. Fixed investment credit is also offered, with a maturity of up to 18 months. 8 CrediAmigo is the largest microfinance institution in Brazil and is among the top microfinance institutions in Latin America in terms of geographical penetration, numbers of clients and depth of outreach. In 2002, CrediAmigo served nearly 60 per cent of microfinance clients and held about 45 per cent of their outstanding loans (Kumar, 2005). According to the 7 administrative data of the program, as of end 2002 roughly 50 per cent of clients were women, 40 per cent were 35 years old or under, 57 per cent had reached the primary school level, and 41 per cent were poor, earning between US$1-$2 per day. III. Empirical Analysis This section is composed of two parts. First, we show the impacts of the availability of the CrediAmigo program on total credit utilization and on credit from formal sources. Second, we analyze the extent to which higher formal credit utilization led to impacts on the performance of small firms. Supply of formal credit and credit utilization The empirical literature provides evidence of high marginal productivity of capital from formal sources. This seems to be the case for small Sri Lankan (McKenzie et al , 2008b), small Mexican (McKenzie and Woodruff, 2008) and medium-sized Indian firms (Banerjee and Duflo, 2008). These studies find that returns to capital are higher than the interest rates charged by banks, so that entrepreneurs would be happier to borrow even more than they are offered. These suggest that credit matters at least for smaller entrepreneurs in developing countries. Here we examine whether in Brazil the utilization of credit has increased after the start of the CrediAmigo program in 1998. Increased access to credit does not necessarily imply increased utilization of credit. The utilization of credit is the end result of interaction between the supply of credit by formal entities and the demand for credit by micro entrepreneurs. To the extent that the conditions for loans are not attractive it is possible that a greater supply of credit may not be associated with an increased utilization of credit. 8 One potential constraint associated with the estimation of the impact of the increased access to microcredit through CrediAmigo on credit utilization by micro-entrepreneurs in Brazil is that the ECINF survey does not contain a specific question about the source of credit from formal banking institutions. Thus, while one can identify whether the source of credit was from a private or public bank, one cannot attribute the source of credit only to CrediAmigo with any certainty. However, as mentioned above, by 2003 this program was still the largest microfinance institution in the country and in the northeast in particular, having been implemented in cities with very little or no source of microcredit. To our knowledge, there were no other important sources of microcredit in the northeast region of Brazil other than CrediAmigo in 2003 or earlier. In September 2003, the Federal Law No. 10735 established that banks must earmark 2% of cash deposits to microcredit transactions (Soares and Sobrinho, Central Bank of Brazil, 2008). Given that the ECINF survey took place in October 2003, we believe that it is practically impossible for other potentially important credit sources from banks other than BNB to be reflected in the survey data. 9 The ECINF survey of micro-entrepreneurs collects information on whether the entrepreneur used any credit in the last three months 10 and conditional on the answer being positive, the primary source of credit is identified based on five alternative sources. 11 The sources of credit identified in the survey include friends and relatives, private or public banks, suppliers, and other persons/firms. 12 Bearing in mind these caveats about the survey instrument, it is important to identify as closely as possible the municipalities where CrediAmigo operated in 2003. Based on administrative data from the program, CrediAmigo as of December 2003 was active in 1172 municipalities (out of a total of about 5,500 municipalities in the country as of 2003) in the northeast region and in the north of the states of Minas Gerais and Espirito Santo). Of the 1172 9 municipalities covered by CrediAmigo, 850 municipalities had more than 20 active clients in 2003. Service is provided through a logistics structure comprised of 166 branches, 44 service centers and 857 loan officers. The publicly released versions of the ECINF surveys in 1997 and 2003 contain the state code but not the municipality code for the residence of the microenterprise. The state code, however, provides a rather crude approximation of the areas of operation of the CrediAmigo program that might in turn lead to a biased estimate on the utilization of credit from CrediAmigo. 13 In order to generate a more precise identification of the areas where access to credit was expanded as a result of the CrediAmigo program, we utilised the complete list of municipalities in the northeast and southeast regions covered by the CrediAmigo program, made available to us by the program administration. We then managed to obtain permission from the Brazilian Institute of Geography and Statistics (IBGE) to access the confidential municipality codes for the location of the micro-enterprises by working with the ECINF surveys within the premises of IBGE in Rio de Janeiro. 14 In our analysis we employ the difference-in-differences (DD) estimator. This estimator compares differences in the utilization of credit between the treatment group (micro- entrepreneurs in the municipalities of Brazil where CrediAmigo operated) and the control group (micro-entrepreneurs in areas where CrediAmigo did not operate) before and after the start of the CrediAmigo program (in 1998). The DD estimator offers the advantage that any time invariant pre-program unobserved heterogeneity between the treatment and control groups is eliminated in the estimation of impacts. The untested maintained assumption behind the application of the DD estimator is that the time or trend effect is identical across the treatment and control groups. We also include a number of control variables that may be useful for reducing any remaining statistical bias. 15 10 The following regression equation defines a model that can nest various “difference” estimators controlling for individual, household and other observed characteristics: K Y (i, t ) = β 0 + β T T + β R D 2003 + β TR (T * D 2003) + ∑ θ k X k (i, t ) + η (i, t ) (1) k =1 where Y(i,t) denotes the outcome of interest equals 1 if the household i in period/round t reports using credit in the last three months prior to the month of the survey and equals zero otherwise, β and θ are fixed parameters to be estimated, T is a binary variable taking the value of 1 if the household resides in a municipality where CrediAmigo operates (the treatment region), and 0 otherwise. The binary variable D2003, is equal to 1 for the 2003 round of the ECINF survey, and equal to zero for the baseline (1997 round) observations. The vector X summarises observed individual and microenterprise characteristics. The last term in equation (1), η denotes the influence of unobserved factors (- is a pure random error term with the usual properties). The coefficient βT allows the conditional mean of the outcome indicator to differ between households in treatment and control municipalities before the initiation of the CrediAmigo program. A statistically significant value of βT suggests that in 1997 there are pre-existing differences in the use of credit between the treatment municipalities (mostly in the northeast) and the control municipalities. Along similar lines the coefficient βR allows the conditional mean of the outcome indicator to differ between households across the 1997 and the 2003 rounds. The coefficient of primary interest in our analysis is βTR, the difference-in-differences estimate which summarises the difference in the outcome means over time between the treatment and control municipalities, that is β TR = [E (Y | T = 1, D 2003 = 1, X = X ) − E (Y | T = 1, D 2003 = 0, X = X )] − [E (Y | T = 0, D2003 = 1, X = X ) − E (Y | T = 0, D2003 = 0, X = X )] 11 Put differently, the parameter βTR provides an estimate of the “intent or offer to treat effect” (ITE) which summarises the effect of the opportunity to access credit in municipalities where CrediAmigo operates on the use of credit from formal banks (public or private) (see Heckman, La Londe, and Smith, 1999). 16 The ITE is an estimate of the impact of making the credit available in specific municipalities regardless of whether micro-enterprises actually take advantage of the opportunity (by choice or not) to get the credit. The ITE is of particular interest to policy makers since it captures the operational efficiency in the implementation of the CrediAmigo program, as well as any potential general equilibrium effects of the credit availability on non-borrowing micro-enterprises in the municipalities where CrediAmigo operated. With this background in mind, regression equation (1) is estimated for each of the four sources of credit for two different samples from the 1997 and 2003 ECINF. Sample A consists of the full sample of micro-entrepreneurs in the ECINF surveys. In this sample, the treatment variable T equals 1 if the microenterprise is located in a municipality in the northeast region of Brazil or in a municipality in the states of Minas Gerais and Espirito Santo, where CrediAmigo operates/has clients, and equals 0 otherwise. In this case the comparison group consists of all other municipalities in the ECINF survey not covered by CrediAmigo in the north, northeast as well as in the centre, south, and southeast regions. 17 Given potential concerns that the trend in the comparison group may not be an adequate representation of the trend that would have prevailed in the municipalities covered by CrediAmigo in the northeast region, we also employ an alternative comparison group. Sample B consists of micro-entrepreneurs from municipalities in the northeast region of Brazil - covered by CrediAmigo. In this sample, the treatment variable T equals 1 if the microenterprise is located in a municipality of the northeast region where CrediAmigo operates 12 -, and equals 0 otherwise. In this case, the comparison group consists of all other municipalities in the northeast not covered by CrediAmigo. Table 1 below provides a detailed description of the sample of municipalities contained in the ECINF surveys in 1997 and in 2003 as well as of the subsample of municipalities where the CrediAmigo program operates. 18 Table 1 We begin with a discussion of the results on the impact of CrediAmigo on the utilization of credit (by source). In this case the dependent variable Y(i,t) equals 1 if the household i in round t reports using credit in the last three months prior to the month of the survey and equals zero otherwise. Equation (1) is estimated using a linear probability model (LPM). 19 The set of covariates used in place of the vector X to account for socio-economic characteristics of the enterprise owner and characteristics of the enterprise itself. These include: age categories (15-24, 25-34, 45-54, 55-64, 65+), gender (male=1), completed high school (incomplete high school is omitted category), more than high school, born in the municipality of current residence, employer (as opposed to self-employed), micro-enterprise is at home, if business uses own equipment, age at which started working, if initial capital was borrowed, if start-up capital was not needed, if entrepreneur has business partner, if business is open throughout the year (all months), has fixed clients, and if the business is registered. Table 2 shows the DD estimates of the impact of the increased access to credit through CrediAmigo on credit utilization overall and by specific source summarised by the coefficient βTR in regression equation (1). 20 Significance level tests are based on robust standard errors that control for the sampling design of the survey in each year. 21 Table 2 13 The DD estimates reported in table 2 reveal that overall the utilization of credit (independently of source) increased by 1.2 percentage point in the full sample or 48 per cent of growth in credit utilization between 1997 and 2003. Moreover, the utilization of formal credit from private or public banks as a primary source of credit increased by 1.5 percentage point without any indication that the utilization of credit from other sources declined in a statistically significant sense. Thus, the increased access to credit offered through CrediAmigo not only increases the utilization of credit from the private or public sector, but it also appears to be associated with an overall increase in the utilization of credit, without replacing pre-existing sources of credit such as own suppliers or friends and relatives. These findings are consistent with the presence of binding constraints in accessing credit prior to the initiation of loans by CrediAmigo. The results obtained for the full sample are mostly driven by the results found using only the northeast region (sample B). The DD analysis in this sample shows that the utilization of formal credit from private or public banks increased by 1.8 percentage point, that is 72 per cent of the observed increment in formal credit taking between 1997 and 2003. While there is a significant increase in the use of credit from any source in the full sample, there is no apparent (statistically significant) increase on the overall utilization of credit in the northeast region. To a large extent, this rather peculiar result can be attributed to the fact that the comparison group is different in each of the samples. Is credit take-up higher among women? Unlike most commercial banks, which tend to offer credit to men mainly because they run larger businesses, microfinance is a different business (Armendáriz and Morduch, 2010). Pioneer programs such as the Grameen Bank were built around serving women. Essentially, 14 women are mostly involved in small businesses, contributing to the large informal sector in developing economies. Women also tend to be more credit-constrained than men. They are reported to be more conservative in their investment strategies, thus are better than men about repaying loans. Finally, women are overrepresented among the poorest. For these many reasons, given that overall formal credit utilization increased as a result of more credit supply in municipalities, one wonders if the formal credit take-up is higher among women than men in Brazil. 22 Table 3 confirms this prediction showing that females take significantly more credit than males regardless of the sample used. The overall credit utilization (independently of source) increased by 2.1 percentage points among women but did not increase among males. The utilization of formal credit increased by 2.5 percentage points for women and less than 1 percentage point for men, in both cases statistically significant. Focusing exclusively in the northeast (sample B), this gap is even higher with utilization of credit from private or public banks increasing by 3.8 percentage points among women and showing no statistically significant increase among males. Thus, in principle, this analysis suggests that there are large differences in credit usage by gender. In view of these findings, the analysis of how credit usage reflects on firm performance allows for heterogeneous impacts across male and female micro-entrepreneurs. Table 3 Formal credit utilization and performance of small firms The preceding analysis provided evidence that the increased access to microcredit in the municipalities where CrediAmigo operated was associated with increased use of credit from formal sources, from either private or public banks. In this section we investigate whether the profits of microenterprises increase with the utilization of formal sources of credit. Given that 15 microenterprise profits are an important component of household welfare we can infer indirectly the impacts of the increased availability of credit on household welfare. Our analysis of the impact of credit on small business performance is carried out using data from the 2003 ECINF survey by comparing the performance of firms utilizing credit from public or private banks with the performance of a comparison group of firms using credit from informal sources or no credit at all. We use two measures of performance (i) business profit and (ii) number of employees. The profit variable is constructed by the difference between revenues and expenditures per month (in October). 23 We take into account substantial differences in living costs across areas in Brazil, by adjusting our measure of profits using spatial price indices (World Bank 2007a). In the ECINF 2003, there is also detailed information on entrepreneur’s characteristics which we use as control variables. Specifically, our estimate of the impact on credit on business performance is based on two different samples using the 2003 ECINF survey and focusing on cities with CrediAmigo program. Sample C consists of the sample of micro-entrepreneurs in the northeast region and in the states of Minas Gerais and Espirito Santo of the southeast region. Sample D consists of the sample of micro-entrepreneurs in the northeast region exclusively. In each of these samples some micro-entrepreneurs have credit from public or private banks and some (or the majority) do not have any credit at all. We use the Propensity Score Matching (PSM) method to match one or more entrepreneurs with no credit or credit from informal sources to a micro-entrepreneur who used credit from a private or public bank and thus estimate the “Average Treatment Effect on the Treated” (ATT): ATT ≡ E{Y1 − Y0 | T = 1} = E{Y1 − Y0 | T = 1, p( X )} = 16 = E{E [Y1 | T = 1, p( X )] − E [Y0 | T = 0, p( X )] | T = 1} where Y1 and Y0 are the outcomes of interest (profits or number of employees), under treatment and non-treatment; T is the indicator for exposure to treatment (=1 if the individual has credit form a public or private bank, and equal to 0 if not); p(X)≡Pr(T=1|X)=E(T|X) is the conditional probability of using credit, given pre-treatment characteristics X. The set of pre-treatment characteristics used in the logit specification of whether a micro-entrepreneur has credit from public or private banks includes : binary variables for age category (15-24, 25-34, 45-54, 55-64, 65+), gender (male=1), if male, completed high school (incomplete high school is omitted category), more than high school, born in the municipality of current residence, employer (as opposed to self-employed), micro-enterprise is at home, if business uses own equipment, age at which started working, if initial capital was borrowed, if start-up capital was not needed, if entrepreneur has business partner, if business is open throughout the year (all months), has fixed clients, if registered, and municipality dummies. 24 Appendix table A.4. shows the means of covariates by treatment status and relevant statistics, before and after the matching. The first column of Tables 4 and 5 show the impact on profits and the number of employees, respectively. The second column of these tables shows the impact on indicators of high (above average) profits and number of employees. This is useful because measured profits are very dispersed, including negative values. In addition, we use the results for binary outcomes to later perform some sensitivity analysis. In terms of profits, the results show that the use of formal sources of credit is associated with significantly higher profits. In sample C, results suggest that using formal credit versus having no credit or other sources such as family leads to an increase of 281 Reais in profits (46 per cent of average profit in 1997). In sample D, returns are smaller but not dissimilar 247 Reais, representing 41 per cent of the average profits in this region. 17 The results using binary outcomes also confirm a significant and sizable effect of formal credit taking on the probability of having high profits (that is, profits that are greater than the sample mean). This effect is 0.11 percentage points or a 63 per cent increase in the probability for the northeast sample and 0.12 or a 49 per cent increase for full sample. Table 4 In terms of firm size, the use of formal sources of credit does lead to a significant increase of 0.18 employees (about 42 per cent of sample average) in the northeast sample and a similar increase in the number of employees in the full sample (about 43 per cent of sample mean). The probability that a business is large also increases significantly in the northeast by 8.4 percentage points and in the whole sample by 8.6 percentage points, which stands for 35 and 37 per cent respectively of the proportion of large businesses in each sample. Table 5 Difference in returns by gender Returns to capital may also differ across males and females. The literature points out many reasons as to why. For example, men tend to re-invest a larger share of their profits generated into business, reflecting among other things their power within the household. Social conventions may constrain women to devote relatively more time to house work, thus limiting the sector and the type of businesses or job they work on; women may also have a more discontinuous participation in the labor market which may imply a smaller social and business- related network; finally, women and men may differ in entrepreneurial ability and risk aversion. Evidence from a randomised experiment in Sri Lanka which gave grants to small business owners shows that neither owner characteristics such as ability, risk aversion or entrepreneurial attitudes nor sector of work or social norms are important at explaining higher returns in businesses owned by men (De Mel, McKenzie and Woodruff, 2008). The authors do 18 find evidence that women invest differently than men: a smaller share of the grants remain in businesses owned by women and men are more likely to invest the grant in working capital while women on equipment. Our results in Table 6 show evidence of increased returns with formal credit taking for both male and female-run firms. The percentage increase in profits is smaller for males than females. When we restrict our attention to the northeast sample, the coefficient loses significance for males while it is significant and positive for females. Together, the results on profits indicate that business run by females experience larger returns to capital. There are also regional differences, males seem to succeed relatively more in the southeastern states of Minas Gerais and Espirito Santo than in the northeast region. Table 6 The last two columns in Table 6 show that firms run by men employ significantly more workers after taking credit. Access to formal sources of credit by male-run firms in the northeast increases employment by 0.232 (51 per cent of the sample average). On the other hand, women continue to run businesses with few workers or on their own. This is consistent with women owning relatively lower capital firms. With low capital, incentives to reinvest profits in capital rather than hiring labor are high at least in the short run (return to capital equals 56 per cent for females in the northeast). Alternatively, anecdotal evidence suggests that women tend to use profits to help the family by keeping the business small or informally employing their family members. We also carried out sensitivity tests for the propensity score estimates following Becker and Caliendo (2007) to test for several deviations from the conditional independence assumption (the procedure and main results are in the online Appendix). We find that our estimates are robust considering both performance outcomes and different subsamples. 19 IV. Concluding Remarks and Policy Considerations Our main findings suggest that an exogenous increase in the supply of credit through the CrediAmigo program raised formal credit utilization and that the use of formal credit did not crowd out credit from informal sources such as family and friends. Overall, the utilization of credit independently of source increased by 1.2 percentage points while formal credit utilization increased by 1.5 percentage points. Moreover, these effects differ across male and female owners. Females take more credit from banks or formal institutions than males. This is consistent with the focus of microfinance on women, reinforcing the argument that amongst other factors women are more credit-constrained than men, they are overrepresented among the poorest, and they are better in repaying their loans. We also show that the use of credit from formal sources is associated with better business performance, meaning larger business profits and number of employees. Individuals who use formal sources of credit seem to re-invest it towards higher returns and larger businesses. The results for males and females are twofold: on the one hand, returns are large and higher for female than male-owned firms; on the other hand, existing businesses run by males employ significantly more workers after taking formal credit while firms owned by women do not. This means that if microcredit policies were to focus on women, they would achieve relatively more credit taking and profits but less employment, at least in the short run, with high returns to capital. On the other hand, the availability of formal credit targeted to women is likely to be an effective way for encouraging women to establish and develop small businesses, increase their participation in the labor market and their economic empowerment. 20 Acknowledgements The authors are grateful to the CrediAmigo administration (Stelio, Iracema and Marcelo) that made possible this analysis and to Susana Sanchez, Pedro Olinto, Edinaldo Tebaldi, Alinne Veiga, Francisco Haimovich, and two anonymous referees for valuable comments and inputs throughout the long gestation of this project. The findings, interpretations, and conclusions in this paper are entirely those of the authors and they do not necessarily reflect the view of the World Bank. ENDNOTES 1 In 1998, the per capita GDP in 2000 prices for the northeastern states was from 1.51 to 3.65 thousand dollars, whereas for the relatively richer southeastern states the figure was about three times as much or more: 5.89 to 10.35 thousand dollars. Source: IPEADATA, IPEA, Brazil. 2 In fact, the PME contains information on the labor incomes of the self employed and employers. Neri - (2008) uses these data as a measure of profit. 3 The reference month is October. It is also important to note that the ECINF surveys took place during the same calendar months in months in 1997 and 2003, thus minimizing problems due to seasonal variation in the need for credit. 4 Enumeration areas (EA) were selected using a probabilistic sample proportional to size on the basis of the urban census enumeration areas. Each EA contains on average 300 households. 5 Actually each state was segmented into 3 strata: i) the municipalities of the capital city; ii) urban areas of the municipalities of the metropolitan area; and iii) urban areas of the remaining municipalities of the state. 6 These key design features included solidarity group lending, targeting the informal sector, charging interest rates high enough to provide a return on assets sufficient, and to permit financial sustainability with gradually increased loans, amortizing loans regularly, offering incentives for regular repayment through discounts on the last installment, and penalizing borrowers if repayment falls behind schedule. 7 This rate only applies to loans under 1000 Reais. Clients are also required to pay a few other fees when obtaining a loan. 8 Loan overdue rates were initially 4.2 percent in the first year of the program, but by writing off bad loans and modifying the performance-based incentive scheme for staff, by 2003, only 1.8 percent of program loans were overdue more than one day (Banco do Nordeste, 2003). 9 The fraction of credit that is not from CrediAmigo can be interpreted as measurement error in the dependent variable. Therefore, coefficients remain unbiased. This only decreases efficiency. Yet, our coefficients are significant and most at 1 or 5% level, as we will show later. 10 We use variable V4331 which equals 3 for occasional or 5 for frequent access to credit. 11 The emphasis on the primary source of credit, implies that we cannot distinguish whether micro- entrepreneurs used more than one source of credit. 12 We use variable V4332 which equals 1 if credit is from friends or relatives, 2 if from private or public banks, 3 if from suppliers, and 4 or 5 if from other firms, people or other sources. 13 For example, Neri (2008), reports difference-in-differences estimates of the increased access to credit through CrediAmigo using the same surveys by simply comparing the utilization of credit in the 9 states comprising the northeast region of Brazil with all other states outside the northeast region. 14 The ECINF surveys can be obtained directly from IBGE (http://loja.ibge.gov.br). Please note that the municipality codes cannot be made available due to a confidentiality agreement. 21 15 Table A.1. in the appendix shows means by treatment status. As the simple t-test suggest, there are significant differences between treatment and control groups for most control variables in year 1997. Thus these factors need to be accounted for. In any case, later we also present estimates of the treatment effects with and without the control variables. 16 Treatment in this case is “exposure to credit” from CrediAmigo. Because exposure is involuntary, β TR provides the intent-to-treat rather than the treatment-on-treated effect. 17 As a means of testing the sensitivity of the results, we have also estimated equation (1) using as the treatment variable T=1 if the microenterprise is located in any of the nine states (instead of the specific municipality) in the NE region and in the Minas Gerais and Espirito Santo states in the SE region, =0 otherwise. In this case the comparison group consists of all other states not covered by CrediAmigo. The DD estimates in this case were slightly lower overall (for example, the utilization of formal credit as primary source increased by 1.3 percent instead of 1.5 percent). 18 Given the sampling design of the ECINF surveys, not every municipality sampled in the 1997 ECINF necessarily appears in the 2003 survey. However when we use the DD estimator we do not have to limit our analysis on the set of municipalities that appears in both survey rounds. We run the DD regressions using the sampling weights in each year to account for the stratification procedure explained in endnote 5. 19 Given that very similar estimates were obtained using a probit and logit model we chose to report the LPM estimates since the coefficients of the LPM provide direct estimates of the marginal effect on the probability of utilizing credit. 20 The full set of estimates of equation (1) based on samples A and B are presented in Appendix tables A.2 and A.3. In particular, these show the bias correction by including the covariates. In the full sample it accounts for the treatment group being worse than the control group in terms of human capital and business capacity in 1997, thus with more need of capital. In the northeast sample, it accounts for the treatment group being relatively better in 1997. 21 Specifically, we cluster our standard errors by enumeration area in each municipality and each year of the survey. As pointed out by Bertrand, Duflo, and Mullainathan (2004) the DD estimator has potentially serious limitations especially if there is serial correlation in the error term of the regression. We have also investigated the sensitivity of our results taking into account the potential of serial correlation within municipalities (and states). We found that the correction for serial correlation did not have any substantial change in the results reported here. 22 The CrediAmigo program focused on established entrepreneurs who are found among likely vulnerable groups, such as the self employed and small businesses, in which women tend to be overrepresented. 23 Revenues are defined as the sum of the value of sales of own product, resale of merchandize, provision of services and other receipts. Expenditures are the sum of expenditures on inventories (primary material and items for resale), labor costs, contributions to social security (INSS), severance pay (FGTS), electricity, water, telephone, rental, machines, equipment, vehicles, gasoline, repair and maintenance, taxes, financial and other expenditures. 24 The inclusion of a set of binary variables for municipality implies that micro-entrepreneurs with credit in any given municipality are matched with micro-entrepreneurs without credit in the same municipality. 22 REFERENCES - Armendáriz, Beatriz and Jonathan Morduch. 2010. The Economics of Microfinance, 2nd edition, Cambridge, MA: MIT Press. Banco do Nordeste. 2003. “CrediAmigo: The Experience of the Bank of Northeast of Brazil with Microcredit.” In Fostering the Sustainable Development of Microcredit in Brazil, ed. Stelio Gama Lyra Junior, Fortaleza, Brazil. Banerjee, Abhijit and Esther Duflo. 2008. “Do Firms Want to Borrow More: Testing Credit Constraints Using a Targeted Lending Program.” BREAD Working Paper No. 005, 2004, revised 2008, Duke University-. Banerjee, Abhijit, Esther Duflo, Rachel Glennerster and Cynthia Kinnan. 2010. “The Miracle of Microfinance? Evidence from a Randomized Evaluation.” BREAD Working Paper No. 278, Duke University. Becker, Sacha O. and Marco Caliendo. 2007. “Sensitivity Analysis for Average Treatment Effects.” The Stata Journal, 7(1): 71-83. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. 2004. “How Much Should we Trust Differences-in-Differences Estimates?.” The Quarterly Journal of Economics, 119(1): 249-275. Blundell, Richard, and Monica Costa Dias. 2000. “Evaluation Methods for Non-Experimental Data.” Fiscal Studies, 21(4): 427-468. Dupas, Pascaline and Jonathan Robinson. 2011. “Savings Constraints and Microenterprise Development: Evidence from a Field Experiment in Kenya.” Cambridge, MA: NBER Working Paper # 14693. Heckman, James, Robert LaLonde, and Jeffrey Smith. 1999. “The economics and econometrics of active labor market programs.” In Orley Ashenfelter and David Card, Eds., Handbook of Labor Economics Vol. III. Amsterdam, North-Holland. Jain, Sanjay. 1999. “Symbiosis vs. crowding-out: the interaction of formal and informal credit markets in developing countries.” Journal of Development Economics, 59(2): 419-444. Karlan, Dean and Jonathan Zinman. 2010. “Expanding Credit Access: Using Randomized Supply Decisions To Estimate the Impacts.” Review of Financial Studies, 23(1): 433-446. Kumar, Anjali. 2005. Access to Financial Services in Brazil, Washington, D.C.: World Bank.- LaLonde, Robert. 1986. “Evaluating the Econometric Evaluations of Training Programs with Experimental Data.” The American Economic Review, 76(4): 604-620. Leuven, E and Barbara Sianesi. 2003. “PSMATCH2: Stata module to perform full Mahalanobis and propensity score matching, common support graphing, and covariate imbalance testing.” (http://ideas.repec.org/c/boc/bocode/s432001.html) McKenzie, David, and Christopher Woodruff. 2008. "Experimental Evidence on Returns to Capital and Access to Finance in Mexico," The World Bank Economic Review, 22(3):457-482. McKenzie, David, Suresh de Mel, and Christopher Woodruff. 2008a. “Who are the microenterprise owners?: Evidence from Sri Lanka on Tokman v. de Soto.” World Bank Policy Research Working Paper 4635. McKenzie, David, Suresh de Mel, and Christopher Woodruff. 2008b. “Returns to Capital in Microenterprises: Evidence from a Field Experiment.” The Quarterly Journal of Economics, 123(4): 1329-1372. Neri, Marcelo. 2008. Microcredito: O Misterio Nordestino e o Grameen Brasileiro, Rio de Janeiro: Editoria FGV. 23 - Paulson, Anna L. and Robert Townsend. 2004. “Entrepreneurship and financial constraints in Thailand.” Journal of Corporate Finance, 10(2): 229–262. Rosenbaum, P.R. 2002. Observational Studies, 2nd ed. New York: Springer Sianesi, Barbara. 2001. “Implementing Propensity Score Matching Estimators with STATA.” UK Stata Users Group, VII Meeting London. ( http://www.stata.com/meeting/7uk/sianesi.pdf)- Soares, Marden M. and Abelardo D.M. Sobrinho. 2008. “Microfinancas: O Papel do Banco Central do Brasil e a Importancia do Cooperativismo de Credito.” Brasilia: Banco Central do Brasil. World Bank. 2007a. “Brazil: Measuring Poverty Using Household Consumption.” Washington, D.C.: The World Bank (Report # 36358-BR). World Bank. 2007b. “Impact Evaluation for Microfinance.” Doing Impact Evaluation # 7. 24 Table 1: Number of municipalities contained in the 1997 and 2003 ECINF surveys 1997 2003 Total Sample A: Full sample Total no. of 752 916 1,668 municipalities No. of municipalities with CrediAmigo (T=1) 314 345 659 Sample B: NE region Total no. of 351 378 729 municipalities No. of municipalities 308 335 643 CrediAmigo (T=1) Notes: IBGE 2003 and 1997 ECINF surveys and administrative data from the CrediAmigo program 25 Table 2: Difference in difference estimates of the utilization of credit from different sources Private or Others Any one Friends & Public Own (persons source relatives Banks Suppliers or firms) Sample A: Full sample β ˆ TR 0.012* 0.002 0.015*** -0.004 0.000 (0.007) (0.003) (0.004) (0.003) (0.002) 90,699 90,665 90,665 90,665 90,665 Sample B: Northeast region β ˆ TR 0.014 -0.005 0.018** 0.002 -0.001 (0.013) (0.006) (0.008) (0.006) (0.002) 36,438 36,436 36,436 36,436 36,436 Notes: Cluster-robust standard errors are in parentheses. Sample size is in italic. *significant at 10percent; ** significant at 5percent; *** significant at 1percent Additional covariates (X) included in the regressions: Binary variables for year (=1 in 2003), age category (15-24, 25-34, 45-54, 55-64, 65+), gender (male=1), completed high school (incomplete high school is omitted category), more than high school, born in the municipality of current residence, employer (as opposed to self-employed), micro- enterprise is at home, if business uses own equipment, age at which started working, if initial capital was borrowed, if start-up capital was not needed, if entrepreneur has business partner, if business is open throughout the year (all months), has fixed clients, and if business is registered. 26 Table 3: Difference in difference estimates of the utilization of credit from formal or any sources – Males and Females Any one source Private or Public Banks Males Females Males Females Sample A: Full sample β ˆ TR 0.007 0.021* 0.009* 0.025*** (0.008) (0.011) (0.005) (0.007) 59,137 31,562 59,109 31,556 Sample B: Northeast region β ˆ TR -0.005 0.051** 0.007 0.038*** (0.015) (0.022) (0.011) (0.013) 22,993 13,445 22,991 13,445 Notes: Cluster-robust standard errors are in parentheses. Sample size is in italic. *significant at 10 percent; ** significant at 5 percent; *** significant at 1 percent Additional covariates (X) included in the regressions: Binary variables for year (=1 in 2003), age category (15-24, 25-34, 45-54, 55-64, 65+), completed high school (incomplete high school is omitted category), more than high school, born in the municipality of current residence, employer (as opposed to self-employed), micro-enterprise is at home, if business uses own equipment, age at which started working, if initial capital was borrowed, if start-up capital was not needed, if entrepreneur has business partner, if business is open throughout the year (all months), has fixed clients, and if business is registered. 27 Table 4: The Impacts of Formal Source of Credit on the profits of micro- enterprises Profits in 1 (if Reais Profits>mean) Sample C: Northeast region + Minas Gerais and Espirito Santo states 281.492*** 0.115*** (108.471) (0.020) 14,221 14,221 46 % 49 % Sample D: Northeast region 246.808*** 0.112*** (107.731) (0.020) 14,084 14,084 41 % 63 % Notes: Method used: PSM (kernel matching) Standard errors are in parentheses. Sample size is in italic. Percentages of sample mean are indicated in third line of each set of results. *significant at 10 percent; ** significant at 5 percent; *** significant at 1 percent 28 Table 5: The Impacts of Credit on the number of employees in the micro- enterprise # of 1 (if # Employees Employees >mean) Sample C: Northeast region + Minas Gerais and Espirito Santo states 0.183*** 0.086*** (0.055) (0.020) 14,221 14,221 43 % 37 % Sample D: Northeast region 0.179*** 0.084*** (0.055) (0.020) 14,084 14,084 42 % 35 % Notes: Method used: PSM (kernel matching) Standard errors are in parentheses. Sample size is in italic. Percentages of sample mean are indicated in third line of each set of results. *significant at 10percent; ** significant at 5percent; *** significant at 1percent 29 Table 6: The impacts of formal credit on small business performance - Males and Females Profits in Reais # of Employees Males Females Males Females Sample C: Northeast region + Minas Gerais and 331.215** 225.128* 0.236*** 0.006 Espirito Santo states (164.797) (126.341) (0.076) (0.078) 8,129 4,842 8,129 4,842 45 % 53 % 52 % 2% Sample D: Northeast region 254.493 234.072** 0.232*** 0.001 (163.246) (127.450) (0.077) (0.078) 8,059 4,802 8,059 4,802 36 % 56 % 51 % 0% Notes: Method used: PSM (kernel matching) Standard errors are in parentheses. Sample size is in italic. Percentages of sample mean are indicated in third line of each set of results. *significant at 10percent; ** significant at 5percent; *** significant at 1percent 30 APPENDIX Expanding Microfinance in Brazil: Credit Utilization and Performance of Small Firms (ONLINE APPENDIX) Table A1: Summary of Micro Enterprises Data Full sample (means) Northeast (means) Variables Year all treatment control p-value all treatment control p-value Sources of credit: Family and friends 1997 0.016 0.014 0.018 0.001 0.014 0.014 0.004 0.012 2003 0.010 0.011 0.010 0.559 0.010 0.010 0.009 0.757 Public or private banks 1997 0.020 0.012 0.025 0.000 0.012 0.012 0.014 0.718 2003 0.039 0.037 0.041 0.048 0.036 0.037 0.020 0.010 Suppliers 1997 0.005 0.006 0.004 0.079 0.006 0.006 0.018 0.000 2003 0.007 0.009 0.006 0.001 0.009 0.009 0.020 0.001 Other people or firms 1997 0.007 0.006 0.007 0.226 0.006 0.006 0.001 0.077 2003 0.006 0.006 0.005 0.143 0.006 0.007 0.001 0.053 All sources 1997 0.048 0.038 0.054 0.000 0.038 0.038 0.037 0.881 2003 0.062 0.063 0.062 0.932 0.062 0.062 0.050 0.156 Profits (in Reais, 1997) 1997 947.86 617.62 1168.49 0.000 605.43 616.49 368.62 0.000 2003 821.15 610.90 949.52 0.000 595.92 606.96 360.89 0.001 # Employees 1997 1.46 1.39 1.50 0.000 1.39 1.39 1.26 0.000 (includes owner) 2003 1.48 1.43 1.51 0.000 1.42 1.43 1.37 0.083 Entrepreneur/ship characteristics: Age 1997 39.8 39.5 40.0 0.000 39.6 39.5 40.4 0.053 2003 41.0 40.5 41.2 0.000 40.5 40.6 39.9 0.139 Males 1997 0.658 0.640 0.671 0.000 0.642 0.639 0.687 0.004 2003 0.645 0.618 0.662 0.000 0.621 0.617 0.693 0.000 High school completed 1997 0.242 0.225 0.254 0.000 0.220 0.225 0.117 0.000 2003 0.306 0.303 0.307 0.344 0.299 0.303 0.207 0.000 More than High school 1997 0.094 0.068 0.112 0.000 0.065 0.067 0.014 0.000 2003 0.124 0.100 0.139 0.000 0.097 0.100 0.039 0.000 Born in this municipality 1997 0.412 0.480 0.366 0.000 0.485 0.481 0.565 0.000 2003 0.415 0.491 0.369 0.000 0.493 0.490 0.561 0.000 Self employment 1997 0.852 0.870 0.839 0.000 0.874 0.872 0.923 0.000 2003 0.854 0.873 0.843 0.000 0.875 0.874 0.897 0.049 Age at which started 1997 14.7 15.0 14.5 0.000 15.0 15.0 13.7 0.000 1 APPENDIX working 2003 14.7 14.9 14.6 0.000 14.8 14.9 13.5 0.000 If business is located in domicile 1997 0.302 0.316 0.293 0.000 0.317 0.316 0.323 0.667 2003 0.288 0.305 0.278 0.000 0.307 0.306 0.329 0.155 If business uses own equipments 1997 0.740 0.733 0.745 0.003 0.733 0.732 0.766 0.030 2003 0.721 0.708 0.730 0.000 0.709 0.707 0.760 0.001 If initial capital was borrowed 1997 0.098 0.104 0.094 0.001 0.104 0.104 0.098 0.569 2003 0.115 0.127 0.108 0.000 0.127 0.127 0.128 0.968 If no start-up capital was needed 1997 0.338 0.323 0.347 0.000 0.321 0.323 0.286 0.023 2003 0.331 0.319 0.339 0.000 0.317 0.318 0.293 0.116 If she has partners 1997 0.104 0.062 0.131 0.000 0.060 0.061 0.033 0.001 2003 0.104 0.080 0.118 0.000 0.078 0.080 0.040 0.000 If business works all months 1997 0.905 0.913 0.900 0.000 0.911 0.912 0.876 0.000 2003 0.893 0.900 0.888 0.000 0.899 0.900 0.889 0.302 If it has fixed clients 1997 0.102 0.096 0.106 0.001 0.094 0.095 0.061 0.001 2003 0.126 0.121 0.129 0.014 0.119 0.121 0.074 0.000 If unregistered 1997 0.481 0.579 0.416 0.000 0.584 0.580 0.672 0.000 2003 0.504 0.571 0.462 0.000 0.572 0.572 0.580 0.642 # Observations 1997 44,711 18,621 2003 48,813 19,200 # Microentrepreneurs 1997 9,580,925 2,501,571 2003 10,526,074 2,763,987 Notes: Own calculations using ECINF, IBGE. Treatment indicates residence in a city with CrediAmigo program. 2 APPENDIX Table A2: Difference-in-difference estimates of the utilization of credit from different sources based on Sample A (full sample) Others Explanatory variables Any source Any source Friends & Private or Own (persons or (no covariates) relatives Public Banks Suppliers firms) (yr 2003) * (Treated municipality) 0.016** 0.012* 0.002 0.015*** -0.004 0 [0.007] [0.007] [0.003] [0.004] [0.003] [0.002] Treated municipality -0.012** -0.001 0 -0.006*** 0.004* 0.001 [0.005] [0.004] [0.002] [0.002] [0.002] [0.001] yr 2003 0.007 0.009* -0.007*** 0.011*** 0.005* 0 [0.005] [0.005] [0.002] [0.003] [0.003] [0.001] Age 15-24 -0.004 0.006 -0.015*** 0.007** -0 [0.006] [0.004] [0.003] [0.004] [0.001] Age 25-34 -0.002 -0.001 -0.004 0.002 0.001 [0.004] [0.002] [0.003] [0.001] [0.001] Age 45-54 0.003 -0.003** 0.003 0.002 0.001 [0.004] [0.002] [0.003] [0.002] [0.001] Age 55-64 0.001 -0.007*** 0.008 0 0 [0.006] [0.002] [0.006] [0.002] [0.001] Age 65+ -0.009 -0.009*** -0.002 0.001 0.001 [0.006] [0.002] [0.005] [0.003] [0.002] Males -0.003 -0.001 0 -0.002 0 [0.004] [0.001] [0.002] [0.002] [0.001] High school completed 0.012*** 0 0.012*** -0.001 0.001 [0.003] [0.001] [0.003] [0.001] [0.001] More than High school 0.025*** -0.002 0.026*** -0.005*** 0.004* [0.006] [0.002] [0.005] [0.002] [0.002] Born in this municipality -0.009*** -0.005*** -0.001 -0.002 -0.001* [0.003] [0.001] [0.002] [0.001] [0.001] Employer 0.058*** 0.007*** 0.042*** 0.003 0.005** [0.006] [0.002] [0.005] [0.002] [0.002] Age at which started working -0.001*** -0.000** 0 -0.000** 0 [0.000] [0.000] [0.000] [0.000] [0.000] If business is located in domicile -0.008** -0.001 -0.007*** 0.001 -0.002** [0.004] [0.001] [0.003] [0.002] [0.001] If business uses own equipments 0.006** -0.002 0.006*** 0.001 0.001 [0.003] [0.001] [0.002] [0.001] [0.001] If initial capital was borrowed 0.076*** 0.028*** 0.031*** 0.006** 0.011*** [0.007] [0.003] [0.005] [0.003] [0.003] If no start-up capital was needed -0.020*** -0.005*** -0.011*** -0.002* -0 [0.003] [0.001] [0.002] [0.001] [0.001] If she has partners 0.029*** 0.002 0.023*** 0.002 0.003 [0.008] [0.003] [0.006] [0.003] [0.003] If business works all months 0.008** 0.002 0.010*** -0.004 0 [0.004] [0.001] [0.002] [0.003] [0.001] If it has fixed clients 0.005 0 0.002 0.001 0.002 [0.005] [0.002] [0.003] [0.002] [0.002] If unregistered -0.035*** -0.005*** -0.021*** -0.007*** -0.003*** [0.003] [0.001] [0.002] [0.002] [0.001] Constant 0.059*** 0.025*** 0.015*** 0.016*** 0.004 [0.008] [0.004] [0.005] [0.004] [0.002] Observations 92839 90699 90665 90665 90665 90665 R-squared 0.001 0.04 0.01 0.03 0.00 0.00 Standard errors in brackets * significant at 10%; ** significant at 5%; *** significant at 1% 3 APPENDIX Table A3: Difference-in-difference estimates of the utilization of credit from different sources based on Sample B (NE region) Others Explanatory variables Any source Any source Friends & Private or Own (persons or (no covariates) relatives Public Banks Suppliers firms) (yr 2003) * (Treated municipality) 0.004 0.014 -0.005 0.018** 0.002 -0.001 [0.014] [0.013] [0.006] [0.008] [0.006] [0.002] Treated municipality 0.017** 0.011 0.010** -0.005 0 0.006*** [0.009] [0.008] [0.005] [0.004] [0.005] [0.001] yr 2003 0.020 0.008 0 0.008 -0.001 0.001 [0.013] [0.012] [0.006] [0.008] [0.006] [0.001] Age 15-24 -0.01 0.004 -0.016*** 0.005 -0.003* [0.008] [0.005] [0.003] [0.004] [0.002] Age 25-34 -0.008* -0.002 -0.006* 0.001 -0.001 [0.004] [0.002] [0.004] [0.002] [0.001] Age 45-54 -0.004 -0.004* -0.001 0.001 0 [0.005] [0.003] [0.004] [0.002] [0.002] Age 55-64 -0.014** -0.010*** -0.006 0.001 0 [0.006] [0.003] [0.005] [0.002] [0.002] Age 65+ -0.015 -0.013*** -0.008 0.004 0.003 [0.009] [0.002] [0.006] [0.006] [0.004] Males -0.007 -0.005 0.002 -0.003 -0.001 [0.005] [0.003] [0.003] [0.002] [0.001] High school completed 0.002 -0.005** 0.011*** -0.003* -0.001 [0.005] [0.002] [0.003] [0.002] [0.001] More than High school 0.014 -0.003 0.019*** -0.007** 0.004 [0.009] [0.004] [0.007] [0.003] [0.004] Born in this municipality -0.008** -0.003 -0.005** 0.002 -0.001 [0.004] [0.002] [0.002] [0.002] [0.001] Employer 0.050*** 0.014*** 0.034*** 0.002 0 [0.008] [0.005] [0.007] [0.002] [0.002] Age at which started working -0.001* -0.000*** 0 -0.000* 0 [0.000] [0.000] [0.000] [0.000] [0.000] If business is located in domicile -0.004 -0.002 0.003 -0.003 -0.002 [0.005] [0.002] [0.004] [0.002] [0.001] If business uses own equipments 0.001 0 0.001 0.002 -0.001 [0.004] [0.002] [0.003] [0.002] [0.002] If initial capital was borrowed 0.067*** 0.030*** 0.029*** 0.003 0.006** [0.009] [0.005] [0.006] [0.003] [0.002] If no start-up capital was needed -0.019*** -0.004** -0.010*** -0.003** -0.002** [0.003] [0.001] [0.002] [0.002] [0.001] If she has partners 0.022* -0.002 0.012 0.008 0.004 [0.012] [0.004] [0.007] [0.009] [0.003] If business works all months 0.013** -0.002 0.013*** 0 0.001 [0.005] [0.003] [0.003] [0.002] [0.002] If it has fixed clients 0.007 0.006 0.005 -0.004** 0 [0.006] [0.003] [0.004] [0.002] [0.002] If unregistered -0.034*** -0.005** -0.019*** -0.009*** -0.001 [0.004] [0.002] [0.003] [0.002] [0.001] Constant 0.052*** 0.022** 0.01 0.018*** 0.002 [0.013] [0.009] [0.007] [0.007] [0.003] Observations 37441 36438 36436 36436 36436 36436 R-squared 0.003 0.04 0.01 0.03 0.00 0.00 Standard errors in brackets * significant at 10%; ** significant at 5%; *** significant at 1% 4 APPENDIX Table A4: Means of covariates, balancing statistics before and after matching Means (before matching) Means (after matching) Variables Treatment Control % bias Treatment Control % bias p-value Sample C: Full sample Age 15-24 0.033 0.090 -23.8 0.033 0.049 -6.5 0.159 Age 25-34 0.233 0.253 -4.7 0.233 0.240 -1.7 0.756 Age 45-54 0.242 0.212 7.1 0.242 0.235 1.5 0.786 Age 55-64 0.091 0.103 -4.0 0.091 0.093 -0.5 0.922 Age 65+ 0.033 0.040 -3.4 0.033 0.035 -0.6 0.911 Males 0.579 0.625 -9.5 0.579 0.591 -2.4 0.666 High school 0.383 0.297 18.2 0.383 0.366 3.6 0.526 completed More than High 0.176 0.095 24.0 0.176 0.159 5.0 0.401 school Born in this 0.438 0.493 -11.1 0.438 0.444 -1.2 0.826 municipality Self employment 0.707 0.881 44.2 0.707 0.739 8.2 0.187 Age at which started 15.3 14.9 8.0 15.3 15.3 0.3 0.961 working If business is located 0.299 0.305 -1.2 0.299 0.303 -0.8 0.883 in domicile If business uses own 0.813 0.706 25.3 0.813 0.785 6.5 0.207 equipments If initial capital was 0.267 0.116 39.1 0.267 0.237 7.8 0.206 borrowed If no start-up capital 0.132 0.329 -48.1 0.132 0.185 -13.0 0.008 was needed If she has partners 0.134 0.077 18.7 0.134 0.122 3.8 0.518 If business works all 0.965 0.899 26.4 0.965 0.946 7.7 0.089 months If it has fixed clients 0.141 0.121 6.1 0.141 0.137 1.4 0.810 If unregistered 0.299 0.587 -60.5 0.299 0.367 -14.3 0.009 Municipality (0.392, - - (-10.2,12.4) - - (-5.2,5.7) dummies (range) 1.000) Sample D: Northeast region Age 15-24 0.034 0.091 -23.5 0.034 0.050 -6.6 0.162 Age 25-34 0.235 0.252 -4.0 0.235 0.241 -1.5 0.791 Age 45-54 0.240 0.212 6.5 0.240 0.233 1.6 0.783 Age 55-64 0.093 0.103 -3.2 0.093 0.094 -0.4 0.946 Age 65+ 0.034 0.040 -3.1 0.034 0.035 -0.6 0.915 Males 0.580 0.624 -9.0 0.580 0.591 -2.2 0.692 High school 0.384 0.297 18.4 0.384 0.367 3.6 0.527 5 APPENDIX completed More than High 0.176 0.095 23.8 0.176 0.159 4.8 0.423 school Born in this 0.437 0.492 -11.0 0.437 0.443 -1.2 0.827 municipality Self employment 0.711 0.882 43.7 0.711 0.744 8.4 0.181 Age at which started 15.3 14.9 7.5 15.3 15.3 0.3 0.957 working If business is located 0.303 0.307 -0.7 0.303 0.304 -0.2 0.965 in domicile If business uses own 0.812 0.705 25.1 0.812 0.783 6.7 0.202 equipments If initial capital was 0.269 0.117 39.3 0.269 0.237 8.3 0.183 borrowed If no start-up capital 0.134 0.329 -47.5 0.134 0.187 -13 0.009 was needed If she has partners 0.131 0.076 17.9 0.131 0.121 3.2 0.598 If business works all 0.964 0.899 26.2 0.964 0.945 7.8 0.093 months If it has fixed clients 0.142 0.121 6.2 0.142 0.137 1.3 0.820 If unregistered 0.300 0.587 -60.3 0.300 0.370 -14.6 0.008 Municipality (-10.1, (0.370, - - - - (-3.9, 6.0) dummies (range) 12.5) 0.999) Notes: Own calculations using ECINF 2003, IBGE (only entrepreneurs in cities with CrediAmigo program). % bias is the standardized percentage difference of means from Rosenbaum and Rubin (1985). P-value tests for the difference of means between treatment and control groups after matching is carried out. Treatment refers to "use of formal sources of credit" and control to "use of other credit sources or no credit at all". 6 APPENDIX Sensitivity Analysis for the Propensity Score Matching estimates The propensity score matching estimates discussed in the performance analysis are derived based on the Conditional Independence Assumption (CIA) which presumes selection on observables. If there are unobserved variables such as ability or motivation that simultaneously affect the decision the use formal credit and the outcome variable such as profits then matching estimators may not be robust. We follow Becker and Caliendo (2007) and perform sensitivity analysis for the estimated ATT effects by testing for several deviations from the conditional independence assumption. The sensitivity tests are based on the Rosenbaum (2002) bounds that provide evidence on the degree to which any significance results depend on the untestable assumption that unobserved variables do not influence the selection into treatment (i.e. getting formal credit). Specifically, take Ti as a binary treatment, i.e. identifying entrepreneurs with formal credit and zero otherwise. Define the participation probability for individual i by Pi=P(Ti=1|xi,ui) where xi are observed and ui are unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no unobserved selection bias which might cause different odds of receiving credit between observationally similar individuals. Deviations of gamma from one thus indicate different degrees of bias. For each such deviation, the statistics reported are the Mantel and Haenzel (1959) test statistics which are obtained along with their respective p-values for testing the null-hypothesis of no treatment effect. MH-stat+ and MH-stat- are the statistics, respectively, under the assumption of over and under estimation of the treatment effect. P- 7 APPENDIX values less than some tolerance level (0.01, 0.05 or 0.10) show robustness of the treatment effect, under the conditional independence assumption, to specific unobserved biases which cause odds to deviate from one. As tables A5.a-A6.c for the Northeast sample show, the interpretation of such tests is not straightforward. The first column shows the odds ratio of credit-taking (gamma), as explained, which is one if there is no hidden bias or if the unobserved covariates do not imply that matched individuals have different probability of taking credit. Thus, we consider scenarios in which biases would increase the odds ratio of credit taking from 5 to 100percent. In the estimation of effects on profits, there could be positive (unobserved) selection when those most likely to take credit tend to have higher profits even when they do not take credit and given that they are similar to those in the comparison group. In this case, the default MH-stat is too high and we should look at the MH-stat+ which attempts to adjusts it down. If p- value+ is below 1, 5 or 10percent if means that the ATT effect is robust to biases leading to those specific gammas. In general, across all outcomes and samples, we found that the effects we presented earlier are robust to unobserved biases. 8 APPENDIX Table A5.a: Males and Females in the Northeast (Outcome = Probability of High Profits) Gamma MH-stat+ MH-stat- p-value+ p-value- 1 11.712 11.712 0.000 0.000 1.05 11.062 12.369 0.000 0.000 1.1 10.448 13.001 0.000 0.000 1.15 9.866 13.611 0.000 0.000 1.2 9.312 14.201 0.000 0.000 1.25 8.785 14.772 0.000 0.000 1.3 8.281 15.326 1.10E-16 0.000 1.35 7.799 15.864 3.10E-15 0.000 1.4 7.337 16.388 1.10E-13 0.000 1.45 6.893 16.898 2.70E-12 0.000 1.5 6.466 17.396 5.00E-11 0.000 1.55 6.055 17.881 7.00E-10 0.000 1.6 5.657 18.356 7.70E-09 0.000 1.65 5.274 18.820 6.70E-08 0.000 1.7 4.902 19.274 4.70E-07 0.000 1.75 4.543 19.719 2.80E-06 0.000 1.8 4.194 20.155 0.000 0.000 1.85 3.856 20.583 0.000 0.000 1.9 3.527 21.003 0.000 0.000 1.95 3.207 21.415 0.001 0.000 2 2.896 21.820 0.002 0.000 Note: Define Pi=P(Ti=1|xi,ui) where Ti is the binary treatment, xi are observed and ui unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no bias which causes different odds of receiving credit between observationally similar individuals. MH-stat is the Mantel and Haenszel test statistic, it moves apart from its default value (when gamma is one) according to deviations from the conditional independence assumption. MH-stat+ and MH-stat- are the statistics under the assumption of over and under estimation of the treatment effect, respectively. P-values less than some tolerable value (e.g. <0.1) show robustness of the effects to specific unobserved biases which causes odds to deviate from one. Further details on this sensitivity test for propensity score estimates can be found in Becker and Caliendo (2007). 9 APPENDIX Table A5.b: Males in the Northeast (Outcome = Probability of High Profits) Gamma MH-stat+ MH-stat- p-value+ p-value- 1 9.327 9.327 0.000 0.000 1.05 8.829 9.832 0.000 0.000 1.1 8.357 10.318 0.000 0.000 1.15 7.910 10.786 1.30E-15 0.000 1.2 7.486 11.240 3.60E-14 0.000 1.25 7.081 11.679 7.10E-13 0.000 1.3 6.695 12.105 1.10E-11 0.000 1.35 6.326 12.519 1.30E-10 0.000 1.4 5.972 12.922 1.20E-09 0.000 1.45 5.632 13.315 8.90E-09 0.000 1.5 5.305 13.698 5.60E-08 0.000 1.55 4.990 14.072 3.00E-07 0.000 1.6 4.686 14.437 1.40E-06 0.000 1.65 4.392 14.794 5.60E-06 0.000 1.7 4.108 15.144 0.000 0.000 1.75 3.833 15.487 0.000 0.000 1.8 3.566 15.823 0.000 0.000 1.85 3.308 16.153 0.000 0.000 1.9 3.056 16.476 0.001 0.000 1.95 2.812 16.794 0.002 0.000 2 2.574 17.106 0.005 0.000 Note: Define Pi=P(Ti=1|xi,ui) where Ti is the binary treatment, xi are observed and ui unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no bias which causes different odds of receiving credit between observationally similar individuals. MH-stat is the Mantel and Haenszel test statistic, it moves apart from its default value (when gamma is one) according to deviations from the conditional independence assumption. MH-stat+ and MH-stat- are the statistics under the assumption of over and under estimation of the treatment effect, respectively. P-values less than some tolerable value (e.g. <0.1) show robustness of the effects to specific unobserved biases which causes odds to deviate from one. Further details on this sensitivity test for propensity score estimates can be found in Becker and Caliendo (2007). 10 APPENDIX Table A5.c: Females in the Northeast (Outcome = Probability of High Profits) Gamma MH-stat+ MH-stat- p-value+ p-value- 1 8.281 8.281 1.10E-16 1.10E-16 1.05 7.857 8.711 2.00E-15 0.000 1.1 7.456 9.124 4.50E-14 0.000 1.15 7.076 9.523 7.40E-13 0.000 1.2 6.715 9.909 9.40E-12 0.000 1.25 6.371 10.283 9.40E-11 0.000 1.3 6.043 10.646 7.60E-10 0.000 1.35 5.729 10.999 5.00E-09 0.000 1.4 5.428 11.342 2.80E-08 0.000 1.45 5.139 11.677 1.40E-07 0.000 1.5 4.861 12.003 5.80E-07 0.000 1.55 4.594 12.321 2.20E-06 0.000 1.6 4.336 12.633 7.30E-06 0.000 1.65 4.086 12.937 0.000 0.000 1.7 3.845 13.235 0.000 0.000 1.75 3.612 13.527 0.000 0.000 1.8 3.385 13.814 0.000 0.000 1.85 3.166 14.095 0.001 0.000 1.9 2.952 14.371 0.002 0.000 1.95 2.745 14.642 0.003 0.000 2 2.543 14.908 0.005 0.000 Note: Define Pi=P(Ti=1|xi,ui) where Ti is the binary treatment, xi are observed and ui unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no bias which causes different odds of receiving credit between observationally similar individuals. MH-stat is the Mantel and Haenszel test statistic, it moves apart from its default value (when gamma is one) according to deviations from the conditional independence assumption. MH-stat+ and MH-stat- are the statistics under the assumption of over and under estimation of the treatment effect, respectively. P-values less than some tolerable value (e.g. <0.1) show robustness of the effects to specific unobserved biases which causes odds to deviate from one. Further details on this sensitivity test for propensity score estimates can be found in Becker and Caliendo (2007). 11 APPENDIX Table A6.a: Males and Females in the Northeast (Outcome = Probability of High Number of Employees) Gamma MH-stat+ MH-stat- p-value+ p-value- 1 13.121 13.121 0.000 0.000 1.05 12.457 13.793 0.000 0.000 1.1 11.830 14.440 0.000 0.000 1.15 11.236 15.064 0.000 0.000 1.2 10.672 15.669 0.000 0.000 1.25 10.135 16.255 0.000 0.000 1.3 9.623 16.824 0.000 0.000 1.35 9.133 17.377 0.000 0.000 1.4 8.664 17.916 0.000 0.000 1.45 8.213 18.441 1.10E-16 0.000 1.5 7.780 18.953 3.70E-15 0.000 1.55 7.363 19.453 9.00E-14 0.000 1.6 6.961 19.942 1.70E-12 0.000 1.65 6.572 20.420 2.50E-11 0.000 1.7 6.197 20.889 2.90E-10 0.000 1.75 5.833 21.348 2.70E-09 0.000 1.8 5.481 21.798 2.10E-08 0.000 1.85 5.139 22.240 1.40E-07 0.000 1.9 4.807 22.674 7.60E-07 0.000 1.95 4.485 23.100 3.60E-06 0.000 2 4.171 23.519 0.000 0.000 Note: Define Pi=P(Ti=1|xi,ui) where Ti is the binary treatment, xi are observed and ui unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no bias which causes different odds of receiving credit between observationally similar individuals. MH-stat is the Mantel and Haenszel test statistic, it moves apart from its default value (when gamma is one) according to deviations from the conditional independence assumption. MH-stat+ and MH-stat- are the statistics under the assumption of over and under estimation of the treatment effect, respectively. P-values less than some tolerable value (e.g. <0.1) show robustness of the effects to specific unobserved biases which causes odds to deviate from one. Further details on this sensitivity test for propensity score estimates can be found in Becker and Caliendo (2007). 12 APPENDIX Table A6.b: Males in the Northeast (Outcome = Probability of High Number of Employees) Gamma MH-stat+ MH-stat- p-value+ p-value- 1 10.473 10.473 0.000 0.000 1.05 9.963 10.990 0.000 0.000 1.1 9.481 11.489 0.000 0.000 1.15 9.024 11.970 0.000 0.000 1.2 8.591 12.435 0.000 0.000 1.25 8.178 12.887 1.10E-16 0.000 1.3 7.785 13.326 3.40E-15 0.000 1.35 7.409 13.752 6.40E-14 0.000 1.4 7.048 14.167 9.10E-13 0.000 1.45 6.703 14.572 1.00E-11 0.000 1.5 6.370 14.967 9.40E-11 0.000 1.55 6.050 15.353 7.20E-10 0.000 1.6 5.742 15.730 4.70E-09 0.000 1.65 5.444 16.099 2.60E-08 0.000 1.7 5.157 16.461 1.30E-07 0.000 1.75 4.878 16.815 5.40E-07 0.000 1.8 4.608 17.163 2.00E-06 0.000 1.85 4.347 17.504 6.90E-06 0.000 1.9 4.093 17.839 0.000 0.000 1.95 3.846 18.169 0.000 0.000 2 3.606 18.493 0.000 0.000 Note: Define Pi=P(Ti=1|xi,ui) where Ti is the binary treatment, xi are observed and ui unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no bias which causes different odds of receiving credit between observationally similar individuals. MH-stat is the Mantel and Haenszel test statistic, it moves apart from its default value (when gamma is one) according to deviations from the conditional independence assumption. MH-stat+ and MH-stat- are the statistics under the assumption of over and under estimation of the treatment effect, respectively. P-values less than some tolerable value (e.g. <0.1) show robustness of the effects to specific unobserved biases which causes odds to deviate from one. Further details on this sensitivity test for propensity score estimates can be found in Becker and Caliendo (2007). 13 APPENDIX Table A6.c: Females in the Northeast (Outcome = Probability of High Number of Employees) Gamma MH-stat+ MH-stat- p-value+ p-value- 1 7.505 7.505 3.10E-14 3.10E-14 1.05 7.094 7.921 6.50E-13 1.20E-15 1.1 6.705 8.321 1.00E-11 0.000 1.15 6.336 8.708 1.20E-10 0.000 1.2 5.986 9.081 1.10E-09 0.000 1.25 5.652 9.443 7.90E-09 0.000 1.3 5.333 9.794 4.80E-08 0.000 1.35 5.027 10.135 2.50E-07 0.000 1.4 4.735 10.466 1.10E-06 0.000 1.45 4.453 10.789 4.20E-06 0.000 1.5 4.183 11.104 0.000 0.000 1.55 3.922 11.412 0.000 0.000 1.6 3.671 11.712 0.000 0.000 1.65 3.428 12.006 0.000 0.000 1.7 3.192 12.293 0.001 0.000 1.75 2.965 12.575 0.002 0.000 1.8 2.744 12.851 0.003 0.000 1.85 2.529 13.122 0.006 0.000 1.9 2.321 13.388 0.010 0.000 1.95 2.118 13.649 0.017 0.000 2 1.921 13.906 0.027 0.000 Note: Define Pi=P(Ti=1|xi,ui) where Ti is the binary treatment, xi are observed and ui unobserved characteristics. Assume that we have matched a pair of individuals i and j, using xi and xj. Gamma is the ratio of odds that individuals i and j take credit, i.e., Pi(1-Pj)/Pj(1-Pi). Gamma equals one when there is no hidden bias or no bias which causes different odds of receiving credit between observationally similar individuals. MH-stat is the Mantel and Haenszel test statistic, it moves apart from its default value (when gamma is one) according to deviations from the conditional independence assumption. MH-stat+ and MH-stat- are the statistics under the assumption of over and under estimation of the treatment effect, respectively. P-values less than some tolerable value (e.g. <0.1) show robustness of the effects to specific unobserved biases which causes odds to deviate from one. Further details on this sensitivity test for propensity score estimates can be found in Becker and Caliendo (2007). 14