Micronance at the Margin: Experimental Evidence from Bosnia and Herzegovina

Transcription

1 Micronance at the Margin: Experimental Evidence from Bosnia and Herzegovina Britta Augsburg, Ralph De Haas, Heike Harmgart, Costas Meghir This version: March 2012 Abstract We use a randomized controlled trial to analyze the impact of microcredit on poverty reduction among households without prior access to nance. The participants consist of 1,200 individuals that EKI, a Bosnian micronance institution, would normally have just rejected on the basis of its regular screening procedures. As part of the trial a random selection of half of these potential borrowers nevertheless received a loan. We nd that access to credit allowed borrowers to start and expand smallscale businesses but that the impact on consumption and other outcome variables was heterogeneous. To develop a unied interpretation of these ndings we construct a model of investment decisions when a minimum amount of capital is required. In line with only partially relaxed liquidity constraints and other model predictions, we document that borrowers needed additional resources to meet minimum investment costs. Households that already had a business and those that were highly educated could run down their savings. In contrast, less-educated households had insucient savings and needed to reduce consumption. Moreover, in the latter case young adults aged started to work more and attend school less. Keywords: JEL Codes: Micronance; liquidity constraints; human capital; randomized controlled trial 016, G21, D21, I32 The authors thank Erik Berglöf, Maren Duvendack, Karolin Kirschenmann, David Roodman, Jeromin Zettelmeyer, and participants at the `Consumer Credit and Bankruptcy' conference in Cambridge, the 3ie Mexico Impact Evaluation Conference `Mind the Gap: From Evidence to Policy Impact', the 3ie-London International Development Centre seminar, the 5th CEPR/EBC Winter Conference on Financial Intermediation, and seminars at the EBRD, Paris School of Economics, and University of Toulouse for useful comments. This project was conceived and promoted with the invaluable help of Francesca Cassano and beneted from continuous support from Borislav Petric and Ryan Elenbaum. Carly Petracco provided excellent research assistance. Institute for Fiscal Studies (britta_a@ifs.org.uk). European Bank for Reconstruction and Development (dehaasr@ebrd.com). European Bank for Reconstruction and Development (harmgarh@ebrd.com). Institute for Fiscal Studies and Yale University (c.meghir@yale.edu).

2 1 Introduction An increasing body of empirical evidence shows that nancial development contributes to long-run economic growth (Rajan and Zingales, 1998; Demirgüç-Kunt and Levine, 2001). A well-functioning nancial system ensures that savings are allocated eciently to productive investments in the form of loans. Yet, a substantial part of the world's population has no or only limited access to formal nancial services. Informal funding channels have been shown to be incomplete as well. The poor in particular often face binding liquidity constraints and credit rationing may constrain potential entrepreneurs in executing protable investment projects. The inability of the poor to access credit led to the emergence of micronance institutions (MFIs) at the end of the 1970s. MFIs, as pioneered by the Bangladeshi Grameen Bank, started by lending small amounts of money to groups of low-income individuals on an uncollateralized basis but with joint liability. 1 The rapid growth of microcredit over the last three decades has, however, been accompanied by a move towards individual-liability loans. The empirical evidence on the impact of either type of microcredit on the economic lives of poor borrowers is still scarce. This paper presents some such evidence from a randomized controlled trial (RCT) in Bosnia and Herzegovina (BiH). The aim of our RCT was to analyze the eect on entrepreneurial activity and poverty reduction of a program that gave a random selection of poor Bosnian households access to individual-liability microcredit. The clients did previously not have access to credit from EKI, the micronance institution that funded the program. We nd that access to microcredit relaxed the liquidity constraints of households but only to a certain extent. As a result, borrower impact diers according to whether the client was self-employed at baseline and according to his or her education level. In line with the idea of only partially relaxed liquidity constraints and other predictions of the model we present, borrowers with higher education levels and with an existing business at baseline were able to supplement their loans with savings. In contrast, those without savings - mainly the less-educated - had to reduce consumption. At the time of the follow-up survey, we did not nd any eects on business prots or household income, independent of the characteristics of the client. We do nd, however, that in borrower households children aged 16 to 19 started to work signicantly 1 See Ghatak and Guinnane (1999) for an early summary of the theoretical literature and Giné, Jakiela, Karlan, and Morduch (2010) for recent experimental evidence on the mechanisms through which join liability aects loan repayment. 1

3 more if the borrower had at most completed primary education. At the same time, school attendance decreased signicantly for these young adults. To summarize, we document that access to microcredit stimulated investments in both new and existing businesses but we cannot yet quantify the overall welfare eects nor can we fully assess the possible negative and unintended consequences of the lending program (such as a reduction in school participation). 2 Finally, we nd that the program was not commercially viable as defaults were signicantly higher among marginal compared to regular borrowers. This paper contributes to three main strands of the literature. First, we add to the still limited empirical evidence on the poverty impact of micronance. While the micronance evaluation literature was sparked by a non-experimental study, Pitt and Khandker (1998), more focus has recently been placed on RCTs to gather rigorous evidence on the eectiveness of micronance programs. 3 Our work falls within this group of studies, which generally conrm that many poor households face binding liquidity constraints and that these can be (partly) relieved by oering microcredit. 4 Banerjee, Duo, Glennerster, and Kinnan (2010) nd this to be the case across a large number of slums in the Indian city of Hyderabad where MFIs had not been oering credit before. The introduction of microcredit boosted business creation. Attanasio, Augsburg, De Haas, and Harmgart (2011) present evidence from Mongolia, where group and individual-lending programs were randomly introduced across villages. The probability of enterprise ownership increased by 10 per cent more in the group-lending (but not the individual-lending) villages compared to the control villages. While these and other studies 5 conrm for a number of settings that reduced liquidity constraints may stimulate business creation, the impact on borrowers and their households remains ambiguous. Attanasio et al. (2011) nd no clear benets for clients that were oered individual-lending contracts but document positive impacts, 2 The share of the Bosnian labor force younger than 25 that was unemployed was 48.7 per cent in 2009 (European Commission, 2010, p.63). 3 Morduch (1998), Morduch and Roodman (2009), and Roodman (2012) point to the scope for selection bias in non-experimental studies. They replicate Pitt and Khandker's (1998) study and fail to reproduce the positive impacts. Another prominent non-experimental study is Kaboski and Townsend (2005) who nd a positive impact of microcredit on consumption but not on investments in Thailand. 4 Other micronance RCTs analyze more specic issues, such as the impact of contract design on repayment rates. For example, Giné and Karlan (2010) analyze how repayment rates dier between individual and joint-liability loans while others look at the impact of the frequency of mandatory meetings on repayment (Field and Pande, 2008) and informal risk sharing (Feigenberg, Field, and Pande, 2010). 5 See for example Karlan and Zinman (2011), De Mel, McKenzie and Woodru (2009), and Fafchamps, McKenzie, Quinn, and Woodru (2011). 2

4 including increased food consumption, for those oered group loans. Banerjee et al. (2010) document heterogeneous eects on non-durable consumption. Those that start an enterprise reduce consumption in order to pay for the xed start-up cost, which typically exceeds the loan amount, whereas non-entrepreneurs increase their consumption. Crépon, Devoto, Duo and Parienté (2011) study the eect of making micronance available in rural areas of Morocco. They nd that self-employment increases and that households who already had a business at baseline decrease their consumption to expand the activities. Karlan and Zinman (2010), in a study on consumer loans oered by a South African lender, nd net positive benets for borrowers along a broad range of outcomes. In a later paper on microcredit, the same authors nd that access to loans led to a small decline in subjective well-being (Karlan and Zinman, 2011). They interpret this nding as indicating that microcredit mainly helps borrowers to manage risk and smooth consumption, rather than it leading to protable investments. This latter study is similar in experimental set-up to ours. Nevertheless, theirs is an intention-to-treat analysis given a non-negligible amount of non-compliance, whereas we are able to present actual treatment eects. Second, our paper contributes to the sparse empirical literature that deals specifically with the impact of micronance in post-conict Bosnia and Herzegovina. Hartarska and Nadolnyak (2007), using a non-experimental approach, nd that access to microcredit has alleviated Bosnian rms' nancing constraints. Demirgüc-Kunt, Klapper, and Panos (2011) nd similar results for nancing constraints at the household level. Their ndings suggest that in particular households that received microcredit were able to make a successful switch from informal to viable, formal entrepreneurs over the period Our contribution lies not only in the experimental nature of our approach, which prevents the selection bias that is dicult to rule out in the aforementioned nonexperimental papers, but also in the time period that we study. During , Bosnia was strongly aected by the global nancial crisis and various Bosnian MFIs experienced, after years of rapid credit expansion, an increase in non- and late repayment (Maurer and Pytkowska, 2011). Our paper is one of the rst to study the impact of microcredit on borrowers during an economic downturn and amidst widespread concerns about overindebtedness. Third, our ndings also relate to the literature on the relationship between access to nance and human capital investment. Endogenous growth theory suggests that such investment is crucial for growth and that credit-market imperfections that de- 3

5 ter investment in human capital may consequently constrain long-term development. Empirical evidence indicates that liquidity constraints can indeed lower investments in the education of children. Conversely, access to credit may allow households to smooth consumption through other channels than the use of child labor. 6 While correcting credit market imperfections may reduce the need for child labor, Wydick (1999) shows theoretically that the eect may also go in the other direction. For instance, if access to credit allows businesses to expand and if child labor cannot be perfectly substituted by hired labor, then access to credit may actually reduce school attendance. Kring (2004) and Menon (2005) provide empirical evidence to this eect in the eld of micronance. Our ndings suggest that similar mechanisms may play a role in Bosnia and Herzegovina. 2 The experiment 2.1 Experimental design We conducted our eld experiment together with EKI, a Bosnian MFI. EKI was interested in nding out whether it could protably expand lending to poorer segments of the Bosnian population that it had hitherto not served and, second, whether access to credit would have a benecial impact on this new client group. 7 These marginal clients are of interest because they had less or no access to other sources of formal credit and were poorer than the average MFI client at the time. Lending to a random selection of these clients therefore allows us to measure the causal impact of microcredit on entrepreneurial activity and various poverty indicators. At the start of the experiment, loan ocers across all EKI branches were asked to identify potential marginal clients over a period of several months. During training sessions loan ocers were instructed to nd clients that they would normally just reject, but to whom they would consider lending if they were to accept slightly more risk, for instance because a client possessed insucient collateral, was less-educated or poorer than average, or was perceived as somewhat more risky for other reasons. 8 6 See Jacoby (1994), Wydick (1999), and Karlan and Zinman (2010). Jacoby and Skouas (1997) show that seasonal uctuations in school attendance act as a form of self-insurance in rural India. Likewise, Beegle, Deheija, and Gatti (2003) study household enterprises in rural Tanzania and nd that credit-constrained households use child labor to smooth income. 7 EKI - Ekonomska Kreditna Institucija - was created by World Vision International in 1996 and has currently about 36,000 clients across both the Federation BiH and the Republika Srpska. 8 The loans oered as part of the experiment were very similar to EKIs regular loan product in terms of interest rate (22 per cent per annum in both cases) and average maturity (around 11 4

6 The training stressed that marginal clients were not applicants with a poor credit history, that were over-indebted, or that were expected to be fraudulent. Note that EKI does not use an automated credit-scoring system. Because identifying and monitoring marginal clients could take additional eort, and to compensate for the possible loss of bonus payments as a result of holding a higher risk portfolio, loan ocers received a fee of 10 BAM (~USD 6) 9 for each marginal client to whom a loan was disbursed. While one may be concerned that loan ocers would divert regular clients to the marginal group, this concern is mitigated by the fact that they would not want to take the 50 per cent risk of losing a solid client due to the randomization process. Once a loan ocer identied a potential marginal client, she would submit the application to the loan committee, including all applicant information such as credit history, collateral, and a project description. Those potential clients that were approved by the loan committee, which used slightly more lax criteria than for regular clients, then had a face-to-face meeting with the loan ocer. She explained the aim of the study and stressed that the applicant would normally not qualify for an EKI loan but that, if agreeing to be interviewed now and in one year's time, he would have a 50 per cent chance of getting a loan. The loan ocer also explained that on the basis of the results of the study, EKI may decide to expand lending to this new client group on a permanent basis, meaning that the current marginal clients could eventually continue to borrow as regular clients. The pilot started in November 2008 in two EKI branches in Gradacac and Bijeljina and two months later the experiment was extended to all 14 branches (see Figure 1a in the Appendix). This process continued until a total of 1,241 `marginal applications' were submitted to the loan committee. In total 1,198 of these marginal loan applicants were then interviewed. 10 This baseline survey was conducted after the individual was judged to be eligible for participation in the program but before the randomization took place. This ensured that responses were not inuenced by the outcome of the randomization process. We also made sure that respondents were aware that their answers would in no way inuence the probability of receiving a loan. At the end of each week, the research team in London would allocate these newly interviewed applicants randomly with a 50 per cent probability to either the treatment months). 9 The exchange rate at baseline was USD 1 to BAM The interview lasted up to 60 minutes and was conducted by a professional survey company using computer-assisted telephone interviews (CATI). 5

7 (receiving a loan) or the control group (no loan). The results of the randomization were communicated to EKI the same day and the EKI branches then made sure that loans were disbursed to the treatment group during the next week. Applicants that were allocated to the control group were notied about their status and did not receive a loan from EKI for the duration of the study. The last interview and loan disbursal took place in May During February-July 2010, 14 months after the baseline survey, all RCT participants - both those who received a loan and those who did not - were called back and invited to be re-interviewed. 3 Data 3.1 Sample description We collected detailed data during the baseline and follow-up interview rounds on the applicant's household structure, entrepreneurial activities and other sources of income, income expectations, household consumption and savings, asset ownership, outstanding debt, exposure to shocks, and stress levels. Table 1 below and Table 20 in the Appendix present summary statistics for the main characteristics of the marginal clients and their households. In each case we rst present the variable mean for the control group and then the value for the average dierence between the control and the treatment groups (with the standard error reported below this dierence). In both tables, columns 1 and 2 provide statistics for the full baseline sample, while columns 3 and 4 provide statistics for the sub-population of households that we re-interviewed at follow-up. Table 1 shows that almost 60 per cent of the (potential) marginal clients are male and that their average age is 37 years. Just over 60 per cent of the potential clients are married and slightly more than half of them were employed at the time of the baseline survey. The average respondent worked 49 hours a week, of which 34 hours were spent in a small-scale business. A third of the marginal clients only attended primary school while ve per cent of the sample went to university. We also show information on household income of the marginal clients. The average income was BAM 18,175 (USD 11,123) in the year previous to the baseline survey, of which on average 7,128 (USD 4,362) was earned through self-employment and BAM 267 (USD 163) as wages from agricultural activities. Columns 5 and 6 in Table 1 allow us to compare the average marginal client to the population of Bosnia and Herzegovina as a whole and to regular (i.e. non-marginal) 6

8 Table 1: Marginal clients: treatment-control balance, attrition, representativeness Variable Baseline sample Population Regular Full Re-interviewed (LITS '10) EKI client Mean Di. C-T Mean Di. C-T C (std.dev) C (std.dev) (1) (2) (3) (4) (5) (6) Age (0.708) (0.767) (17.04) Female (0.028) (0.031) (0.50) Marital status Married (0.028) (0.031) (0.49) Economic activity Empl (0.029) (0.032) (0.49) Highest education Prim (0.027) (0.030) (0.390) (0.306) Sec (0.028) (0.031) (0.48) (0.355) Univ (0.012) (0.014) (0.29) (0.199) Working hrs (week) Total (1.572) (1.765) Business (1.706) (1.913) HH income Total 18, , (834.3) (880.4) Self-empl. 7, , (551.4) (557.1) Agriculture (62.90) (70.10) No of observations T: 637 T: 551 C: 569 C: 443 This table provides summary statistics on both the potential marginal clients that received credit (T) and those in the control group (C). Column 1 provides the sample mean for the entire control group at baseline. Column 2 shows the mean dierence between the control and the treatment group with the corresponding standard error. Columns 3 and 4 show the same statistics for those households that were re-interviewed at follow-up. These rst four columns are based on data from the baseline survey. Columns 5 contains statistics for the entire Bosnian population based on the EBRD Life in Transition Survey (LiTS) Column 6 shows comparative statistics on regular rst-time EKI borrowers based on data from EKI's management information system. For variable denitions see Table 22. rst-time EKI borrowers, respectively. In column 5, we use 2010 data from the Life in Transition Survey (LiTS) in which 1,000 households were interviewed in Bosnia and Herzegovina, a sample representative at the national level. LiTS sampled two types of respondents. The rst one is the household head or another household member with sucient knowledge about the household. The second sampled person (if dierent from the rst) is the person aged 18 years and over who last had a birthday in the household. We compare our marginal clients to these latter, randomly sampled persons and constrain the sample to the same age range we observe for our marginal clients. We nd that, compared to this population, the average marginal client is younger and more likely to be male and married. We also nd that on average the marginal client is less educated as relatively many of them completed at most primary 7

9 education. Comparing the marginal client to rst-time EKI borrowers shows that they are younger, less likely to be married, and have less education. Marginal clients are also less likely to be full-time employed. We present further information on the households of the marginal clients in the Appendix in Table 20. The average household consists of 3.2 persons. On average 1.1 household members are employed, 0.7 are unemployed and a further 0.7 attend school. At the time of the baseline survey, a typical marginal household spent on average BAM 110 (USD 67) on food per week, BAM 236 (USD 144) on other nondurable items (such as rent, combustibles, clothes, and recreation) per month, and BAM 2,434 (USD 1,490) on durable items (such as education expenses, repairs, and furniture) per year. This adds up to consumption expenditures of approximately BAM 11,000 (USD 6,732) a year. 3.2 Randomization and treatment-control balance As the allocation of marginal applicants into the treatment and the control group was random, we expect no systematic dierences between both groups. To check whether this is the case, column 2 in Table 1 and 20 show the dierence in means between the treatment and the control group and the corresponding standard error. For almost all variables we observe no statistically signicant dierence between the means of the two groups. The only exception is the number of children aged 11 to 15 (Table 20). Although the dierence is statistically signicant, the number of young children is only 0.07 higher in the treatment group and the economic relevance of this dierence is negligible. When we conduct a joint signicance test for all variables together we nd no systematic overall dierence between the two groups. The Chi 2 statistic for this test is with a p-value of We conclude that the randomization process was successful in that it ensured the absence of selection bias. This means we can attribute any post-treatment dierences in outcomes to the microcredit program. 3.3 Attrition A total of 1,206 individuals were interviewed before the program and 995 of these were re-interviewed as part of the follow-up survey. 12 The attrition rate was thus relatively 11 The full regression results for this test are displayed in Table 21 in the Appendix. 12 Eight baseline respondents were eliminated as they were not approved by the loan committee or decided not to borrow after all (thus reducing the original baseline sample to 1,198). Thirteen of the 995 interviews were not fully completed. Table 19 in the Appendix provides more details on the 8

10 low at 17 per cent. In order to limit attrition, the survey company sent all participants a reminder letter at the beginning of the follow-up survey. This letter also announced a rae in which all who completed the survey could participate. Interviewers were trained to encourage participation and people who initially declined to participate were called back later by a senior interviewer and asked once more to participate. Persistent refusers were also oered a EUR 10 phone card. 13 We need to check whether a respondent's willingness to participate in the followup survey is orthogonal to whether he or she received a loan. When we regress a dummy that indicates whether the household was interviewed at follow-up or not on the treatment indicator, we nd that attrition is signicantly higher for respondents that did not receive a loan. These respondents were almost ten per cent more likely not to participate in the follow-up survey. This remains the case when we account for covariates. Importantly, however, when we analyze the observed baseline characteristics of only those who were surveyed at follow-up, we nd that these characteristics are still balanced between the treatment and control group (see column 4 in Tables 1 and 20). Thus, while less control individuals responded in the follow-up interview round than treated individuals, this non-response is not correlated with any of the observable characteristics we consider. From this we conclude that attrition did not undermine the balanced nature of the treatment and control samples and did not introduce bias in the reported results. 14 targeted and actual number of interviews at baseline and follow-up. 13 The average yearly income of potential marginal clients was BAM 13,381 at baseline. EUR 10 (BAM 19) therefore corresponds to 54 per cent of average daily earnings. 14 We also checked that pre-treatment characteristics are balanced across treatment and control groups in the following sub-samples: business ownership at baseline or not, high versus low education level, and gender of the marginal client. 9

11 4 A simple model To structure a unied interpretation of our empirical ndings, this section develops a simple model of investment decisions when production requires a minimum amount of capital in order to make lumpy investments. 4.1 Relaxing liquidity constraints when investments are lumpy Suppose there are two periods. In the rst period the household can invest in a business that will produce output in the second period. The production technology is Q = 1(K > Γ)δK, where Γ is the minimum amount of capital needed for production. Individuals have a period utility function u(c) = c1+ρ. The household has a constant 1+ρ income stream Y in each period. There is no access to capital markets except for a loan B subject to a ceiling B. The household thus faces the constraint c 1 +K Y + B, i.e. investment can be nanced by household income and the loan. The interest rate is r, the price of output is 1, and the discount factor for the household as well as the market is β. The household optimization problem can now be expressed by the Lagrangian L = c1+ρ ρ + β c1+ρ ρ + λ([y (1+β) + β(δ r)1(k > Γ)K c 1 βc 2 ] + µ(b + Y c 1 K) (1) If it is optimal to invest less than Γ then the solution is standard: the household does not produce and consumes c = Y in each period. Optimal utility is L = (1+β) Y 1+ρ 1+ρ Now suppose it is optimal (and feasible) for investment to exceed the minimum amount of capital needed for production K > Γ. In the absence of any borrowing ( B) we clearly have that c 1 < Y, since some of the current income must be diverted to investment. The values of consumption are: c 1 = Y (1 + β) + β[δ r]k 1 + β(1 + β(δ r)) 1/ρ (2) and c 2 = (1 + β(δ r)) 1/ρ c 1 (3) Now consider the eect of oering a loan to those households that are already investing in a business. The interesting case here is when the liquidity constraint is binding. The household then borrows the entire amount and c 1 + K = Y + B. Using the expression for c 1 that we derived above, we obtain an expression for optimal 10

12 investment: K = (B + Y )G Y (1 + β) G + β(δ r) where G = 1 + β(1 + β(δ r)) 1/ρ. The marginal propensity to invest out of a relaxation of the liquidity constraint is MP I = (4) G G + β(δ r) 1 (5) We nd that both consumption and investment increase when the borrowing limit is relaxed for households that were already investing. Consumption increases for two reasons. First because the liquidity constraint is less binding and second because of a wealth eect as the household benets from the business opportunity. Note, however, that consumption may fall for those who are starting up a business. To see this, note that at B = 0, c 1 = Y K. The optimal capital level is positive at that point so long as the marginal product of capital (δ) is larger than the interest rate r. If the optimal level of K is above the minimum threshold then the household will cut consumption and invest. As the loan ceiling increases current consumption will increase as well. We can thus distinguish between two cases. In the rst one, households who could not invest before are now given the opportunity to do so. They may crowd in additional resources by reducing consumption and then invest. In the second case, households who already had a business but were liquidity constrained increase both investment and consumption, depending on the wedge between the marginal product of capital and its cost. If the loan size is large enough we may observe an increase in consumption for everybody. However, in general consumption will increase for the liquidity-constrained households that are already running a business and decrease for those that are just starting one. 4.2 Labor supply of young adults Now suppose the household includes a young adult who can either go to college, and earn a lifetime return of τ/r, or work. We model the return to education as increasing the eciency units of labor. An untrained person has one unit of labor. With school attendance maximum eciency units become ˆl 2 = 1 + τ r (1 l 1) where l 1 is the labor supply in period one. Since leisure does not yield utility, the individual will work ˆl 2 in the second period. Let the wage rate per eciency unit be w so that 11

13 labor income becomes w(1 + τ r ) + w(1 τ r )l 1. With a return that exceeds the interest rate (τ/r > 1), increases in l 1 will always reduce lifetime labor income, implying that young adults will never work in the rst period and will attend school instead. Of course, a young adult with low returns to education will work full time setting l 1 = 1. We are mainly interested in the former case, in particular in the question what happens to educational attendance when the household can obtain a loan and thus has the opportunity to start a business. We assume that labor market regulations imply a labor cost of (1 + ς)w, where ς represents taxes and other regulatory costs. To the extent that τ r < 1 + ς the household will therefore hire the internal labor of its young adults. Only when a capacity constraint is hit will the household also hire from the market. We discuss the case of an interior solution where only internal labor is hired and where the cost of internal labor is foregone future earnings. Take the production function to be Q = 1(K > Γ)K α H 1 α, where H is total labor employed. We dene exogenous income as M = Y (1 + β) + w(1 + τ ). Because r of liquidity constraints it is important to note that rst-period income is Y. The Lagrangian for the problem can be written as L = c1+ρ 1 1+ρ + β c1+ρ 2 1+ρ + + λ[m + w(1 τ r )l 1 + β(1(k > Γ)(K α (L + l 1 ) 1 α (1 + ς)wl w β l 1 rk)) c 1 βc 2 ]+ (6) +µ(y + B c 1 K (1 + ς)wl) So as to impose liquidity constraints in the above we assume that both capital and external labor costs have to be nanced before production by current income and the loan. The loan B is subject to a maximum B, which we assume is binding. Hence B = B. We only present the case where L = 0, that is internal labor is sucient for the optimal choice of inputs. With K < Γ no production takes place, l 1 = K = B = 0 and c = M/(1 + β). If the household does invest then consumption will grow according to the relationship c 2 = (1 + G) 1/ρ c 1 (7) where G = αβ [ w( τ ] α 1 r ) 1 α β(1 α) (8) 12

14 Now, because there is the opportunity to work in the family business, and as long as this work generates a high enough return to overcome the loss in future human capital, the labor supply of the household's young adults may be positive. Assuming an interior solution for both capital and labor, when borrowing is allowed we get a labor supply function of l 1 = (B + Y )(1 + β(1 + G) 1/ρ ) M G 1/(α 1) [1 + β[(1 + G) 1/ρ + (G α r)] w( τ r ) (9) while capital is given from the rst-order condition by K = G 1/(α 1) l 1. (10) The equilibrium labor input will be positive so long as the borrowing limit is high enough and the net wage low enough. Even if l 1 = 0 production may still take place by purchasing labor from outside as long as labor productivity is high enough to cover the regulatory costs ς. Following the same logic as in a model without internal labor supply, we can show that for business start-ups consumption may be lower than in the counterfactual world of no borrowing. However, consumption will increase as the borrowing limit increases for those households that are already in business. 4.3 Summary of model predictions Our model indicates that consumption will increase for all those who already have a business when the borrowing limit increases. In contrast, creating a new business may involve a period of lower consumption as households put resources together to nance the start-up. This will be the case if there are xed start-up costs or a minimum amount of required capital to make lumpy investments. In the absence of direct costs of education, all households who are not running a business and for whom the return to education exceeds the interest rate will send young adults to school. However, things are dierent in the presence of a household business. If hiring outside labor is expensive because of regulation and if such regulatory costs exceed the dierence between the return to education and the interest rate, then the business will employ the internal labor of young household members. When a household starts a business, and when existing businesses grow, households will employ more of this internal labor. 13

15 All in all, the model generates four main predictions for us to test. For households that receive access to loans, we predict that: 1. Consumption increases for households with an existing business as their liquidity constraint is relaxed; 2. Savings will decrease; 3. Consumption will decrease for lower-income households with a start-up; 4. Labor supply by young adults will increase (and educational participation will decline), in particular for those with lower ability and thus lower expected returns to education. 4.4 Do borrowers make lumpy investments? One of the main assumptions of our model is that investments are lumpy. To see to what extent this assumption holds true for the investments made by the borrowers in our sample, we analyze reported loan use. A limitation of this approach is that, although the information we collected on loan use is detailed when compared to many other household surveys, the categories are still quite broad. To give an example, we know whether a client used the loan for the purchase of livestock but we do not know whether she bought, say, twenty chickens or one cow. Given such limitations, this section aims to provide some indicative evidence of the lumpiness of the marginal borrowers' investments. A rst indication of lumpy investments is that on average borrowers used the EKI loan for only 1.32 dierent purposes (with a standard deviation of 0.59). Thirty per cent of the loans have been used exclusively for one single purpose. From Table 2 we can see that most loans are used for purchasing livestock of the marginal clients (24 per cent of all clients) report this use (columns 1 and 2). The average amount used for this purpose was BAM 1,636 (~USD 1,000) (column 3) or about 77 per cent of the average loan amount. The remainder of these loans were almost completely put towards buying auxiliary agricultural inputs such as seed, fertilizer, and fodder (column 6). The rst two rows of Table 2 show that investments in livestock combined with buying seed, fodder, and other agricultural items is indeed very common. A large number of clients have used practically the whole loan amount to start up or to substantially expand an agricultural activity. This suggests that borrowers had to 14

16 Amount used for Table 2: Main loan use % % of loan remaining Obs. Mean Std.Dev. clients amount amount: use & % (1) (2) (3) (4) (5) (6) (1) Purchase of livestock ,636 1, (2) 85% (2) Investment in seed, fertilizer, etc , (1) 86% (3) Purchase of engine, tools, etc ,588 1, (9) 81% (4) Investment in developing own work ,983 1, (9) 77% (5) Purchase of goods ,790 1, (9) 100% (6) Private purpose , (1) 72% (7) Investment in real estate ,133 2, (1) 75% (8) Buying and maintaining cars/fuel ,550 1, (1) 50% (9) Other ,552 1, (1) 87% This table reports the number and percentage of marginal clients that used their loan for various purposes (columns 1 and 2), the average amount invested (column 3, in BAM), the standard deviation of this amount (column 4, in BAM), the average percentage of the loan that was used for this purpose (column 5), and the main use of the remaining part of the loan (column 6) as well as the percentage of this remaining part that was used for this secondary purpose (also column 6). Categories with less than ve respondents, such as 'purchase of computer/laptop', are omitted. For variable denitions see Table 22. cover some upfront costs that are more than proportional to returns (i.e. are lumpy) to make their investment. 5 Estimating the program eects 5.1 Main outcomes of interest The theoretical predictions set out in Section 4.3 point to the following main outcomes of interest: consumption, savings, labor supply, and school attendance. Moreover, we expect impacts to be heterogeneous across the treated population. Predictions (1) and (3) suggest that households with and without a business at baseline may experience opposite impacts on their consumption patterns while prediction (4) indicates that outcomes may dier according to the learning abilities - and therefore the return to schooling - of the young adults in a household. 15 Prediction (3) additionally suggests that eects on consumption may dier depending on the income level of the household. In this section we therefore estimate separate treatment eects by splitting the sample according to whether the household had a business at baseline or not and according to the level of education of the marginal borrower. We interpret the level of education both as a proxy income measure, more reliable than a wealth indicator given that it is not aected by the program and is more stable over time, and as proxy for a 15 While we do not observe the learning ability of children directly, we can proxy it by the ability of their parents. 15

17 borrower's know-how and entrepreneurial ability Overall program eect We estimate the eects of the program through a simple comparison of means, namely Y MBf Y Cf, where Y is the outcome of interest, f stands for follow-up survey, MB for marginal borrower (treated individuals) and C for control (untreated individuals). To improve precision, we include baseline covariates and estimate the following equation using OLS: Y = α 0 + α 1 T + α 2 X b + u (11) where T is the treatment indicator (T = 1 if the individual received a loan and T = 0 if not), so that α 1 is the average treatment eect. X b is a set of baseline covariates that includes the respondent's age, gender, and marital, educational, and economic status as well as household characteristics such as composition and the economic status and income level of the household members. u is the error term. To estimate how the eect may vary with observable characteristics, we repeat the calculation on suitably dened sub-samples, such as those who did not have a business at baseline. 6 Results In this section, we discuss our empirical ndings on the impact of microcredit on four main outcome indicators: consumption and expenditures, savings behavior, the labor supply of household members, and school attendance. Before we do so it is useful to analyze whether the microcredits provided by EKI have indeed contributed to new business creation and the expansion of existing small enterprises. 6.1 Business creation and development EKI loans were meant to nance investment in enterprises in the form of either working capital or xed assets. In this sub-section we look at the extent to which this objective was achieved. The rst two variables in Table 3 give information on how the employment situation changed at the household level. Column 1 indicates that while overall unemployment of household members did not change, self-employment increased. Households of marginal borrowers are 6 per cent more likely to receive income from self-employment than households in the control group. We also nd that marginal clients are 6 per 16

18 Unemployment Self-employment Business ownership Business in services Business in agriculture Ownership of inventory Business prot (BAM) Business expenses (BAM) Business revenue (BAM) Table 3: Eect on business creation and development ESTIMATED EFFECT Overall Business at baseline By education Yes No Low High (1) (2) (3) (4) (5) (0.029) (0.036) (0.048) (0.053) (0.035) 0.060** * (0.029) (0.034) (0.052) (0.050) (0.036) 0.058* * (0.031) (0.038) (0.050) (0.055) (0.038) ** (0.025) (0.033) (0.036) (0.042) (0.030) * (0.028) (0.037) (0.041) (0.053) (0.032) 0.053** 0.057** 0.041* 0.072** 0.044* (0.020) (0.029) (0.024) (0.032) (0.026) (541) (812) (461) (979) (667) * (593) (879) (331) (530) (811) 1,384 1,547 1, ,780 (981) (1,464) (717) (1,296) (1,298) This table shows estimated coecients for the treatment impact on business creation and development (standard errors in brackets). The rst ve variables as well as the last are expressed as probabilities and the business prot, expenses and revenue variables are in BAM (USD). Column 1 describes the outcome variable; column 2 gives estimated eects for the whole sample; columns 3 and 4 show heterogeneous eects based on whether respondents had a business at baseline ('Yes') or not ('No); and columns 5 and 6 show heterogeneous eects by whether respondents only had primary education at baseline ('Low') or were more highly educated ('High'). Estimations include covariates. * signicant at a 10% signicance level; ** at the 5% level. Business prot, expenses, and revenues are expressed in Bosnia-Herzegovina Convertible Mark (BAM). The exchange rate at baseline was USD 1 to BAM For variable denitions see Table 22. cent more likely to own a business, indicating that it is the clients themselves that are now contributing to overall household income. The subsequent columns provide information on how this eect varies by certain baseline characteristics. The impact on earning income from self-employment and the impact on business creation is mainly driven by the highly educated, whereas there is no signicant dierence between those that did and those that did not have a business at baseline. Those with higher education are 7 per cent more likely to own a business at follow-up than the control group. We also observe some interesting heterogeneity by education level in terms of the types of businesses that are created. Those with not more than primary education are more likely to start up agricultural activities than the control group (even though they are not more likely to own a business at follow-up). In contrast, those with 17

19 more than primary education are more likely to start up an enterprise in the services sector. Table 4 shows that already at the time of the baseline survey there were some sectoral dierences according to education level (although these dierences were not statistically signicant). Finally, we note that the likelihood of owning inventory is signicantly higher (about 5 per cent) for treatment than for control households. This eect is the largest for marginal clients with at most primary education, who are seven per cent more likely to own inventory at the end of the experiment. We nd no clear heterogeneity in terms of whether a borrower already had a business at baseline or not. An exception is that those who did not yet have an enterprise report signicantly higher business expenses compared to the control group. This is likely driven by the use of loans for relatively large start-up related investments. Otherwise, we do not nd any signicant eects on expenses, revenues or prots as can be seen in the lower panel of Table 3. Table 4: Descriptive Statistics (Baseline) - Business DESCRIPTIVE STATISTICS Overall Business at baseline By education p-value Yes No Low High T vs.c (1) (2) (3) (4) (5) (6) Unemployment (0.442) (0.439) (0.470) (0.458) (0.434) Self-employment (0.41) (0.19) (0.50) (0.42) (0.41) Business ownership (0.48) (0.00) (0.49) (0.48) Business in services (0.38) (0.45) (0.30) (0.41) Business in agriculture (0.43) (0.49) (0.47) (0.40) Ownership of inventory (0.398) (0.452) (0.203) (0.393) (0.401) BAM USD Business prot 6,940 4,249 13,298 10,509 14,500 (16,221) (20,421) (14,547) (22,399) Business expenses 4,826 2,954 9,218 7,140 10,114 (21,489) (28,923) (14,335) (33,263) Business revenue 11,715 7,172 22,367 17,858 24,354 (33,149) (42,993) (25,503) (48,604) This table provides descriptive statistics for business ownership at the time of the baseline survey. The rst ve variables are expressed as fractions and the last three in BAM (USD). Columns 1 and 2 provide information on the mean for the whole sample (in USD and BAM respectively, if the variable is measured in currency value); columns 3 and 4 means by whether respondents had a business at baseline ('Existing') or not ('Start-up'); and columns 5 and 6 means by whether respondents only had primary education at baseline ('Low') or were more highly educated ('High'). Standard deviations in brackets. The last column provides the p-value for a test of equivalence of means of the treatment versus the control group. Business prot, expenses and revenue are expressed in Bosnia-Herzegovina Convertible Mark (BAM). The exchange rate at baseline was USD 1 to BAM For variable denitions see Table 22. Table 5 shows that while the percentage of business owners in our sample was 62.3 per cent at baseline, it decreased to 54 per cent at the time of the follow

20 up survey, most likely reecting the severe impact of the nancial crisis on smallscale entrepreneurs. Interestingly, the decline in business ownership was much more pronounced in the control group, where only 40 per cent owned a business at followup (numbers not shown). Overall, about 35 per cent of business owners in our sample closed their business between the two survey rounds, and only 14 per cent started one over the same period. Table 5: Business ownership: Baseline versus follow-up Owns business at follow-up No Yes Total Owns business No 24.0% 13.6% 37.5% at baseline Yes 21.6% 40.7% 62.3% Total 45.6% 54.3% 100.0% 6.2 Impact results - Consumption After having established that access to credit led to more investments in new or existing small enterprises, we now continue with testing the predictions that we derived from our model. We start by analyzing the impact on consumption. Our model predicts that consumption should have increased for those who had a business at baseline, as their liquidity constraint is relaxed, whereas consumption should have decreased for lower income households that started-up a new business. Table 6 shows the estimated impacts on a number of aggregate consumption variables. The rst row shows the eect on the household's overall consumption expenditure, which includes money spent on food (both inside and outside of the house), other non-durables (such as rent, bills, clothes, and recreation) and durables (large, infrequent purchases which here include educational expenses, purchase of vehicles, and vacations). 16 We do not nd a signicant increase in consumption for households who had a business at baseline. In terms of our model this would indicate that the loans were too small to suciently relax liquidity constraints. In addition, we nd that house- 16 Food expenditures were collected over a recall period of a week, other non-durables over a period of a month, and durables over a period of a year. To calculate the aggregate spending amount we assume that the week and month about which the household was asked were representative for the year. This assumption is not important in view of the impact analysis (as we compare treatment and control groups over the same period) but does play a role when we put the value of expenditures in context, for instance by comparing them to income. 19