Credit Lines in Microfinance: Evidence from the Mann Deshi. Field Experiment

Transcription

1 Credit Lines in Microfinance: Evidence from the Mann Deshi Field Experiment Fernando M. Aragón Alexander Karaivanov Karuna Krishnaswamy August 2018 Abstract This paper studies the effect of flexible microcredit loans on business performance. We use a field experiment which offered a credit line to female vendors in India. This new product provides flexible withdrawals and repayment, but retains standard features such as group lending and weekly repayments. We find evidence of a positive incremental effect: a credit line increases vendors profits by more than a standard term loan. While the estimated average effect is small, we document important heterogeneity consistent with costly learning: the increase in profits for borrowers with good knowledge of the new product is larger and rises over time. The results are driven by improved debt management and vendors investing in new, more profitable, goods. Our findings highlight the role of loan flexibility as a possible way to raise the impact of microfinance. 1 Introduction Ever since its origins in the Grameen Bank of Bangladesh, microfinance has followed a wellestablished and relatively standardized model of providing credit to poor people without collateral: small, often joint-liability, loans with frequent (e.g., weekly) meetings and repayments (Banerjee, We thank Xavier Giné for useful comments. Karaivanov acknowledges financial support by the Social Sciences and Humanities Research Council of Canada. Simon Fraser University, Department of Economics, faragons@sfu.ca Simon Fraser University, Department of Economics, akaraiva@sfu.ca Public Health Foundation of India 1

2 2013, Ghatak and Guinnane, 1999, Morduch, 1999). A key feature of this standard model is the use of term loans: microfinance borrowers receive a fixed amount, to be repaid in regular payments over a set period of time, usually 1-3 years. However, an important concern with this form of credit is that it may not be optimal in many situations, especially when a business faces unpredictable cash flows or recurring need for working capital, as is the case for many microcredit clients. In those situations, a term loan may be too rigid and fail to match the clients borrowing needs. Indeed, most businesses in the developed world do not rely exclusively on loans or equity to finance their daily operations, but use more flexible financing sources such as trade credit or credit lines. Surprisingly, there is not a lot of empirical evidence on the effect of credit lines on the business performance of microcredit clients. Is a more flexible financial product, like a credit line, more effective in improving business performance than a standard microcredit loan? This is an important question, due to the emphasis given to microfinance as an engine of entrepreneurship and mounting evidence that the standard microcredit contract may be unnecessarily rigid and restraining entrepreneurial behavior (Barboni, 2017, Field and Pande, 2008, Field et al., 2013, 2012, Fischer, 2013, Giné and Karlan, 2014). In this paper we examine empirically the effect of a credit line on microfinance clients business performance. Our main contribution is to show that the flexibility of borrowing and repayment offered by a credit line can have a positive effect on profits, including by encouraging some borrowers to invest in riskier but more profitable goods. We use data from a field experiment implemented by Mann Deshi Bank, a large microfinance lender in Maharasthra, India. The bank used a randomized control trial to introduce a new financial product, credit line, alongside standard term loans. Specifically, credit line clients were able to withdraw or repay anytime any amount up to a pre-approved limit. The only required minimum payment was interest accrued on the outstanding debt. The product retained the joint-liability, weekly meetings, and lending to female borrowers features of the standard microfinance loan. We focus on the credit line effect on gross profits, but also examine proxies for business size, such as sales. Our experimental design compares clients who received the credit line (CL) with 2

3 clients who received a standard term loan (TL). That is, we estimate and evaluate the incremental effect of a credit line relative to a term loan, not the effect of access to microcredit per se. We find that a credit line has a positive, albeit statistically insignificant, incremental effect on average daily profits: over the duration of the experiment, the profits of CL clients increased by 8 percentage points (INR 62) more than the profits of TL clients. We observe similar results for other measures of business performance. This average effect, however, hides two important sources of heterogeneity. First, we find that the effect of a credit line increases monotonically over time: it is slightly negative in the first weeks into the treatment, but after 18 weeks it is almost twice the magnitude of the average effect. Second, we find that the credit line effect is larger and statistically significant among clients who demonstrate a good knowledge of the new financial product, but negligible among clients with poor knowledge. We interpret these findings as suggestive evidence for costly learning: it took time for the borrowers to learn how to exploit the benefits of a credit line, and, in the 6-month duration of the experiment, many simply did not learn that. If this interpretation is correct, then our results suggest that flexible loans, such as credit lines, could strengthen the positive impact of microcredit on small businesses. Furthermore, the results highlight the need for adequate participants training for such interventions to succeed. An alternative interpretation is that the heterogeneous effects we document are picking up other differences between clients with good and bad product knowledge, such as education level or business skill. We address this concern by controlling for interactions of the treatment variable with multiple observable characteristics. Of course, there might exist unobservable differences or possible reverse causality: credit line clients who obtained higher profits may have dedicated more effort to learn about the product. If that were the case, our results still suggest that some clients are more able to benefit from the credit line product but the policy recommendation would be different: targeting instead of training. We also analyze the possible economic mechanisms explaining why profits increase more using a credit line than a term loan. We develop a model which highlights two main channels. First, 3

4 credit lines allow for more flexible stock and debt management: the traders could borrow more and increase stocks when demand is strong and do the opposite when market conditions worsen. Second, the repayment flexibility associated with a credit line allows traders to invest in riskier but more profitable goods for sale. We examine these two economic mechanisms empirically and find evidence supportive of the second channel: the increase in profits appears to be mostly driven by the traders investing in different goods for sale. We also find weak evidence for the flexible debt management channel. Our work relates to and complements an existing experimental literature which examines what features of the standard microcredit contract are effective vs. not effective. One interesting finding in this literature is that increased flexibility of the loan contract appears to have a positive effect on business performance, without necessarily hindering repayment rates. For example, Field and Pande (2008) and Field et al. (2012) study the effect of changing the frequency of repayments from weekly (in the control group) to monthly (in the treatment group). They document a positive effect on business income and investment, with no significant effects on clients delinquency or default. The monthly clients also reported significantly lower rate of financial stress or anxiety. Field et al. (2013) find that a more flexible repayment schedule (two-month grace period instead of immediate repayment) has a positive effect on business profits, although there is an associated increase in default rates. 1 In contrast, we find increased profits for credit line users, with no observable impact on default rates, at least in the short run on which we have data. The authors interpret these findings as evidence that the additional flexibility of repayment encourages more profitable but riskier and illiquid investments and allowed clients to experiment with new services or products. The latter result is consistent with our findings here. Barboni and Agarwal (2018) document similar effects of introducing repayment holidays, also combined with an improvement in repayment rates. Note that while these papers examine particular aspects of repayment flexibility (reduced frequency or grace period), these aspects do not approach the full flexibility of a credit line that we analyze here. 1 In other, non-experimental work Armendáriz and Morduch (2010) report more flexible repayment as associated with higher default rates in Bangladesh, while (McIntosh, 2008) finds less delinquency among Ugandan MFI clients who choose more flexible repayment schedules. 4

5 In other related work, Carpena et al. (2012) and Giné and Karlan (2014) find that relaxing the standard microcredit requirement of joint liability does not affect repayment performance, although it may affect how the loans are used and the impact of micro-credit on entrepreneurship and poverty (Attanasio et al., 2015). The rest of the paper is organized as follows. Section 2 describes the experiment and develops a simple theoretical model highlighting the economic effects of a credit line relative to a term loan. Section 3 discusses the experimental methods and the data. Section 4 presents our main empirical results. Section 5 concludes. 2 Background 2.1 The credit line experiment The credit line intervention we study was implemented by Mann Deshi Mahila Bank. Mann Deshi is a regulated urban cooperative bank run by and for women. It offers customized loans and microfinance services to female entrepreneurs in the state of Maharashtra, India. The bank was founded in 1997 and is currently serving almost 200,000 clients through seven branches. In 2013, Mann Deshi did a study aimed at better understanding of the needs of their clients, most of them street vendors. 2 The study showed that, due to volatile sales and cash flows and risks associated with trading perishable goods, the street vendors had highly variable working capital needs. These needs were not being met by the microcredit product offered by the Bank: a fixedterm loan with joint liability and weekly repayments. Instead, street vendors were relying on their own savings and expensive credit from wholesalers to fund their working capital. Based on the preliminary study findings, Mann Deshi launched an overdraft facility called Cash Credit Loan. This new product is effectively a credit line: at any time it allows clients to withdraw or pay back any amount, up to a pre-approved maximum (the drawing limit). Interest is accrued and must be repaid weekly, but only on any outstanding debt, not on the drawing limit. Similar to traditional microcredit loans, the new credit line product was made available to groups of three 2 77% of the vendors were selling fruits or vegetables, the rest sold sweets, other foods, spices, clothing, etc. 5

6 unrelated women who were jointly liable in case of default. A pilot study for the credit line product was performed in three districts of rural Maharashtra: Satara, Pune, and Solapur, between November 2014 and July The Bank solicited loan applications without disclosing whether applicants were eligible for the credit line. The applicants formed groups of three and applied for loans. All applications were evaluated using standard Bank procedures and information collected by loan officers through site visits and subsequent credit bureau checks. A list of 120 approved groups (360 borrowers) was provided to the research team who randomly assigned them to a credit line (CL) or a term loan (TL). The term loans are standard joint-liability loans with weekly repayment schedule. All applicants who were offered a loan of either type accepted the loan and participated in the study. Before loan disbursement, all clients took part in a one-hour training program conducted by Mann Deshi bank staff. The training program provided basic financial education and explained the products main features. For the purposes of the study, the CL and TL products were made as similar as possible. 3 The initial drawing limit (the initial loan size for TL clients) was set to INR 10,000 for new customers and INR 20,000 for existing customers. Both microcredit products charged annual interest rate of 24% and the Bank required/recommended similar weekly repayments (around 2.5-3% of the loan amount, or INR 300 for a INR 10,000 loan and INR 500 for a INR 20,000 loan). Payments were made to a bank officer who collected them from a designated group member in their village or locality on a fixed collection day. The loan period was set to three years for CL customers and one to two years for TL customers. Clearly, the main difference between a credit line (CL) and a term loan (TL) is the additional flexibility of the credit line product: CL clients can withdraw funds up to the drawing limit at any time and repay in part or withdraw additional funds. In contrast, TL clients can only withdraw the funds at the beginning of the loan period and must repay weekly following a rigid installment schedule. 3 See Table B.8 in the Appendix for detailed description of the credit line product terms. 6

7 2.2 Analytical framework We first develop a simple theoretical model to illustrate the key economic implications of a credit line vs. a term loan and guide our subsequent empirical analysis. Given the available data, we focus on gross profits, defined as sales revenue minus the cost of sales, as the main outcome variable. The model highlights two main channels through which a credit line can have positive effect on gross profits, compared to a term loan: (i) facilitating more flexible stock and debt management and (ii) allowing traders to invest in riskier but more profitable goods for sale. Consider credit-financed traders selling a perishable good over two time periods. We model the traders as risk-neutral and interpret their net payoff (utility) as gross profits or business cash flow. The traders have no other assets or income. Each period t = 1, 2 a trader buys stock of goods for sale s t at cost 1 per unit and sells it at price p t > 1 per unit. This implies that it is never optimal for the trader to carry cash balances over time, i.e., all cash on hand from sales or loans is fully used to purchase stock. Motivated by our empirical application, we consider two different financial products: a term loan (TL) and a credit line (CL). Both products charge the same interest rate r > 0 and have the same size or limit, L. Default is not possible. The main difference between the two products is in the timing for receiving and repaying the loan. With a term loan, the trader receives the full loan amount L upfront, before any sales are realized. In addition, the trader is required to repay a fixed installment each period. In contrast, with a credit line, the trader can withdraw any amount up to L after observing p t. This captures the main distinction between the two products: the term loan fixes the debt and repayment schedules, while the credit line offers the flexibility of varying the loan and repayment amounts in response to demand shocks. For simplicity, assume that the first-period price p 1 > 1 + r is constant/known, 4 while p 2 is a random variable with support [1 + ε, 1 + ε] where ε ε > 0. We interpret the randomness in p 2 as a stochastic demand shock, e.g., due to variation in the price or the fraction of goods sold. We start with the observation that a trader can earn higher expected profits with a credit line than with a term loan of the same size (see Appendix A for formal analysis). Intuitively, a credit 4 This assumption implies that borrowing the maximum amount is always optimal in the first period. All our results remain valid in the case p 1 < 1 + r. 7

8 line financed trader can always choose to replicate the actions of a trader using a term loan of the same size. However, as we show formally in Sections and 2.2.2, a CL trader has additional flexibility in debt and sales stock management as well as in the choice of goods for sale. The credit line flexibility implies a positive effect on traders profits compared to a term loan Stock and debt management A term loan (TL) financed trader borrows L at the beginning of t = 1 and repays equal installments, F at the end of each period. 5 Since p 1 > 1, carrying money balances across periods is returndominated by buying goods for sale and thus the trader uses the full loan balance, b T L 1 = L to purchase stock s T L 1 = L. At the start of period 2, the TL trader has debt b T L 2 = (1+r)L F = (1+r)L 2+r and available cash p 1 s T L 1 F which, since p 2 > 1, he uses to purchase stock s T L 2 = p 1 L F. Overall, this yields expected profits: Π T L = E(p 2 )(p 1 L F ) F (1) We also assume that, for any possible p 2, the TL trader always has enough funds to repay the loan installment, that is p 2 (p 1 L F ) F, for all p 2 (2) A credit line (CL) financed trader can choose any debt level b t L at t = 1, 2, after observing p t. At the end of each period, the CL trader is only required to pay interest on her current debt, rb t. 6 The trader can otherwise increase or decrease her debt level as she finds optimal. All outstanding debt must be repaid by the end of period 2. Since we assume p 1 > 1 + r, the CL trader initially borrows the maximum amount b CL 1 = L and purchases stock s CL 1 = L. 7 At the end of the first period, she must repay rl and has cash on hand (p 1 r)l. We focus on the second period. Suppose the realized second-period price p 2 satisfies p r ( high price ) sales profits per rupee exceed the cost of carrying debt. Then it is easy to show that the CL trader optimally rolls 5 Using the compound interest formula, the installment size is F = (1+r)2 L. 6 2+r This assumption matches the lender s credit line regulations in our data. 7 Note, however, that if we had p 1 (1, 1 + r) then the CL trader would not borrow in the first period, that is, b CL 1 = s CL 1 = 0 unlike the term loan, the credit line provides flexibility when the return on sales is low. 8

9 over her debt, that is, b CL 2 = L and buys the maximum possible amount of stock, s CL 2 = (p 1 r)l. 8 Note that b CL 2 > b T L 2 and s CL 2 > s T L 2. That is, in good times (high p 2) the credit line allows the CL trader to borrow more and purchase more stock than a TL trader. Suppose now that the realized p 2 satisfies p 2 (1, 1 + r) ( low price ) sales profits per rupee are lower than the cost of carrying debt. Now it is optimal for the CL trader to bring her debt to zero, i.e., b CL 2 = 0 and purchase stock s CL 2 = (p 1 1 r)l (since p 2 > 1, purchasing stock dominates carrying cash). That is, in bad times (low p 2 ) the credit line allows the trader to borrow less (note that b CL 2 < b T 2 L ) and purchase less stock (note that scl 2 < s T 2 L ) compared to a TL trader. To summarize, a TL trader s debt schedule, {b T L 1, bt 2 L } and stock purchases, {st L 1, st L 2 } are rigid and do not depend on the realization of p 2 (the demand shocks). In contrast, a CL trader s debt and stock purchases are flexible and optimally vary with the realization of the sales shocks. As a result, the CL trader always achieves larger profits Safe and risky goods We now describe a second mechanism that can enable CL traders to achieve larger profits compared to TL traders, by choosing different goods for sale. This mechanism can operate in addition to the more flexible stock and debt management under credit line discussed in the previous section. Hold everything else the same as before, but now suppose that there are two possible goods for sale in period t = 2: a risky, high-return good and a safe, low-return good. The safe good yields p s per unit of stock with certainty. The risky good has stochastic yield, p r with larger expected return than the safe good, E(p r ) > p s and support p r [1 + ε, 1 + ε] where ε > ε > 0. Assume also p s > 1 + ε. We show that it is possible that the rigidity of the term loan constrains TL traders to stock the safe good while, for the same parameters, CL traders invest in the higher-return risky good and obtain larger profits. Indeed, suppose that the repayment feasibility condition (2) is not satisfied at p 2 = 1 + ε but is satisfied at p 2 = p s. Then, TL traders must invest in the safe good because 8 For p r and p 1 > 1 + r it is easy to check that the debt can be fully repaid at the end of period 2, that is: p 2(p 1 r)l (1 + r)l. 9 For p r the CL trader s expected profits are Π CL = E(p 2)(p 1 r)l (1 + r)l. For p 2 (1, 1 + r) the CL trader s expected profits are Π CL = E(p 2)(p 1 1 r)l. Using (1) it is easy to verify that Π CL > Π T L in both cases. 9

10 investing in the risky good can put them in a position of being unable to repay F when sales are weak (low p 2 ). On the other hand, since E(p r ) > p s, a CL trader expected profits from stocking the risky good are always larger than her profits from stocking the safe good, 10 which in turn are larger than a TL trader profits from stocking the safe good: Π CL risky > ΠCL safe > ΠT L safe The first inequality captures the increase in expected profits for a CL trader from the choice of good for sale and the second inequality captures the increase in expected profits from more flexible debt and stock management Testable implications The stylized model suggests two possible economic mechanisms through which borrowing via credit line can increase traders gross profits relative to borrowing via term loan. First, the credit line gives flexibility to the traders to adjust their stock levels to market conditions: borrow and buy more stock when conditions are favorable or carry lower stock and debt levels when conditions worsen. This implies a higher variation in stock and debt balances under a credit line. Second, access to a credit line can encourage traders to invest in riskier and more profitable goods. The reason is that the credit line gives traders the ability to borrow and repay when market conditions are favorable. In contrast, term loan traders always borrow upfront and are required to repay the same fixed installment in both good and bad conditions. Hence TL traders may be reluctant to stock riskier goods if there is a possibility that they cannot meet the required repayment ex-post. In the empirical section which follows, we quantify the effect of the credit line on gross profits. We also examine our model s testable implications regarding the variability of stock and debt and the type of goods sold and evaluate and distinguish the possible economic mechanisms behind our findings. 10 Compare the expressions for Π CL given in footnote 9. 10

11 3 Methodology 3.1 Randomization and data collection The unit of randomization is the joint liability group. To ensure a balanced sample, the research team stratified the applicants by market (n=13). Within each stratum, we randomly assigned half of the groups to receive the CL treatment while the other half received a term loan (TL). 11 Overall, 61 groups were assigned to a CL (the treated group) and 59 groups were assigned to a TL (the control group). 12 The randomization was done in November 2014, and all term loans and credit lines were distributed between December 2014 and March We collected data from three main sources: a baseline survey, financial diaries, and an endline survey. 14 The baseline survey was performed one to two weeks before the participants received the term loan or credit line. We also collected financial diaries, that is, short surveys about daily outcomes of the participant s business. Collection of the financial diaries started immediately after the baseline survey and was done every two weeks until the last week of July The financial diaries are the main data source used in the empirical analysis. Finally, we did an endline survey (from the end of July to mid August 2015) with questions regarding business practices, knowledge of the financial products, and customer satisfaction. In robustness checks we also complement these three datasets with administrative data on debits, credits and loan balances provided by Mann Deshi Bank. The CL product proved to be very popular and, by the end of July 2015, most term loan clients were switched to a credit line, which effectively concluded the experiment. This raises two important caveats. First, we only observe most program participants for up to 6 months into the treatment, although the financial services still continue for 0.5 to 2.5 additional years. Second, we are unable to observe end outcomes, for example, default rates. Thus, we can only examine the 11 In practice, the treated participants were invited to take up a credit line instead of term loan. All participants accepted the invitation. 12 The number of treated groups is not 60 because of rounding in the strata with odd number of groups. 13 Each month around 25% of the participants received their loans. 14 All questionnaires used in the surveys are available upon request. 15 The data collection was performed by a team of enumerators. The interviews took place on a randomly chosen day of the week shortly before participants closed business for the day. 11

12 short-term effects of the intervention. 3.2 Integrity of the experimental design We first examine the integrity of the experimental design with a balance test, comparing the means of all variables of interest between individuals assigned to the control and treated groups (see Table 1). The results show that the two groups were very similar before the treatment (no statistically significant differences). This suggests a well-randomized experiment. There are, however, two sources of non-random variation in the experiment. First, individuals from six joint liability groups initially assigned to a term loan were switched to a credit line. This non-compliance does not seem to be empirically important. The balance test and regression estimates using actual treatment instead of the initial assignment produce similar results. However, due to this non-compliance, our baseline results using the initial assignment should be interpreted as intention to treat (ITT) estimates. 16 Second, approximately two months into the program, the loan limit was raised from INR 10,000 to INR 20,000 for some borrowers, based on repayment performance and the discretion of the loan officer. These loan upgrades were granted to 34% of the CL clients and 20% of the TL clients. A relevant concern is that these loan terms changes may confound the estimated effect of the credit line. However, we show that our results remain robust to controlling for obtaining a loan upgrade (see Tables B.3 and B.4 in the Appendix). 3.3 Estimation strategy Based on the analytical framework discussed in Section 2.2, we are mainly interested on the effect of the credit line on traders gross profits. We also examine other outcomes such as sales, stock value, and business practices. 17 The empirical analysis compares clients who received a credit line to clients who received a term loan. Note that since each group received a loan, this comparison cannot provide experimental 16 In Table B.5 in the Appendix, we present estimates of local average treatment effects (LATE) using the initial assignment as an instrument. 17 We define gross profits = sales cost of sales where cost of sales = initial stock value final stock value. All variables are constructed from the traders financial diaries and refer to the outcome in a given day. 12

13 Table 1: Baseline characteristics Initial assignment p-value Variable TL (control) CL (treated) (1)=(3) mean s.d. mean s.d. (1) (2) (3) (4) (5) A. Profits, sales and expenses Gross profits Sales Cost of sales Other expenses B. Initial and final stock Initial stock Final stock C. Business characteristics Monthly business income Sells perishables Years in business Balance in savings account D. Sources of working capital Bank or microfinance Savings or business profits Wholesaler E. Demographics Household size Can read and write Is married No. Joint liability groups No. Individuals Notes: Monetary outcomes are measured in Indian Rupees (INR). Data are collected in baseline survey or pre-treatment financial diaries. Groups are based on intention to treat. Variables in Panels A and B are trimmed of top and bottom 1% values. Column 5 displays p-value of a mean comparison test. TL: Term Loan. CL: Credit Line. 13

14 evidence on the effect of receiving microcredit per se, but instead evaluates the incremental effect of receiving a credit line compared to a term loan. We estimate the following baseline regression: y ij = βcl ij + ɛ ij, (3) where the unit of observation is client i in joint-liability group j, y ij is the outcome variable, and CL ij is an indicator of the initial assignment (1 for CL and 0 for TL). We do not include any additional covariates in the baseline specification. 18 Since the randomization units are not individuals but joint liability groups, we cluster the errors at the group level. To reduce the influence of outliers, we also trim the top and bottom 1% observations based on the value of the outcome variable. 4 Results 4.1 Main findings We start by illustrating the main features of our data. Figure 1 displays the average daily profits for TL and CL groups before and after receiving the loan. There are two important observations. First, before receiving a loan, both groups had daily profits of around INR 550. After receiving a loan, profits increase to around INR This represents an increment of almost 40 percent. 19 The increase in profits after a loan is received is statistically significant, but does not necessarily represent a causal effect of access to credit: it could be over- or under-estimating the true effects. However, this finding echoes the results of other experimental studies that find statistically and economically significant effects of microcredit on business size and profits (Banerjee et al., 2015a,b, Crépon et al., 2015). Second, the average increase in profits for CL clients is greater than that for TL clients. This experimental evidence suggests that access to a credit line has a positive incremental effect on 18 This is a conservative approach. Including covariates such as baseline outcome values and socioeconomic characteristics produces similar and more precise estimates (see Table B.1 and B.2 in the Appendix). 19 This result is robust to including additional covariates and clustering by joint-liability group. 14

15 profits relative to a term loan. The magnitude of this effect, however, is relatively small. Figure 1: Average daily profits before and after treatment, in INR We formally examine this incremental effect of receiving a credit line by estimating the baseline regression (3). The results, reported in Column 2 in Table 2, confirm a positive, albeit statistically insignificant, effect on gross profits and other business outcomes (sales, costs of sales and expenses). On average, access to a credit line increases traders profits by INR 62 more than receiving a term loan. This increase is economically meaningful: it is equivalent to 8 percent of TL traders profits. Note also that this increment is in addition to any gain in profits due to access to credit, and represents the average effect over the whole duration of the experiment. Examining the effect of CL at different times after the intervention provides a richer picture (see columns 3 to 6 in Table 2). We observe a lag in the effect and a trend for it to increase over time. For example, in the first five weeks into the treatment, the effect on profits of receiving a credit line is slightly negative and insignificant. However, after 18 weeks into treatment, access to a credit line increases profits by INR 116 compared to under term loan, that is, almost twice the average effect. Similar pattern is observed for the other outcome variables. A possible explanation for the delayed effect of the credit line is costly learning: traders appear 15

16 to need some time to learn how to use the new features of the CL product. Moreover, costly learning may also help explain why the magnitude of the average effect is small over the relatively short (6-month) duration of the experiment. We examine this issue in more detail in the next section. Table 2: Effect of CL on profits, sales, and expenses Outcome Mean value Regression estimates control group (1) (2) (3) (4) (5) (6) Gross profit (51.7) (60.5) (53.4) (75.2) (74.5) Sales (317.1) (285.7) (328.0) (384.4) (455.0) Cost of Sales (262.7) (253.4) (276.1) (313.1) (376.5) Other expenses (26.4) (33.5) (30.1) (28.4) (33.9) Weeks into All Weeks Weeks Weeks Weeks treatment 1 to 5 6 to to Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions use data from financial diaries. Columns 2 uses all the sample, while columns 3-6 use sub-samples based on the number of weeks into treatment. 4.2 The role of learning We assess the clients knowledge of the CL product using data from the endline survey. The endline survey includes questions about general characteristics of the loan products and a practical question designed to evaluate the traders knowledge on the key feature of the CL product, namely the flexibility to repay and borrow up to the drawing limit. 20 Most clients recognized some of the main features of the CL product, for instance, 97.1% report knowing that they are allowed to withdraw the full amount, while 89.4% know that interest is only charged on outstanding debt. However, knowledge on the flexibility of the credit line was mixed. Only about half of CL clients 20 The specific question we use is: Suppose your credit limit is Rs 20,000 and you have withdrawn Rs 10,000 and repaid Rs Now, what is the maximum you can withdraw? The correct answer is: Rs 14,

17 (52%) responded correctly to the practical question. This suggests that, even at the end of the experiment, some participants may have still not learned how the new financial product could be used. To examine the possible role of learning and product knowledge, we split the sample of CL clients into two groups: those who responded correctly to the practical question and those who responded incorrectly. We label these groups good knowledge and poor knowledge clients. We then re-estimate the baseline regression equation in each sub-sample. The results suggest that there exists significant heterogeneity in the effect of CL across the two groups of clients (see Table 3). The effect of CL on business outcomes among poor knowledge clients is negligible: their profits are indistinguishable from those of the borrowers who received a term loan. In contrast, the average profits of CL clients with good knowledge increased by about INR 98, or 13 percent, relative to the control group. The estimate is statistically significant at the 5% confidence level. The good knowledge group also exhibits significant increases in sales and cost of sales relative to the poor knowledge group, but no significant difference in other business expenses (see column 5). To further examine these findings, we plot the average profits of TL clients and those of CL clients with good vs. poor knowledge, before and after receiving the loan (see Figure 2). Note that all three groups start with similar average profits prior to the program and report a similar initial increase in profits after receiving a loan. In the first five weeks into treatment, their profits remain almost identical. However, after week 6, we observe a divergence: the profits of poor knowledge CL clients track the performance of TL clients, while the profits of CL clients with good knowledge of the product continue to growth. We formally confirm this observation by splitting the sample into different periods of the treatment and estimating regression (3) with interaction terms of the treatment dummy (CL i ) and the indicator of having good knowledge. 21 We observe no statistically significant difference between poor knowledge CL clients and the control group over the length of the experiment. However, from week 6 onwards, there is statistically significant effect (at 10% and 5% levels) of CL among good 21 Figures B.1a and B.1b in the Appendix present the point estimates and confidence intervals of the effect of CL for each subsample. 17

18 knowledge clients. Table 3: Effect of CL on profits, sales, and expenses, by knowledge of product Mean value Effect of Credit Line control group All CL Poor Good p-value Outcome Knowledge Knowledge (3)=(4) (1) (2) (3) (4) (5) Gross profits ** (51.7) (60.0) (56.2) Sales (317.1) (327.5) (355.9) Cost of sales (262.7) (269.9) (294.5) Other expenses (26.4) (28.9) (31.0) Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions use data from financial diaries. Column 5 presents p-value of test of equality of estimates in columns 3 and 4. We interpret the above findings as suggestive evidence that some participants did not learn, within the time frame of the program, how to effectively use the credit line. In addition, we confirmed this by visual inspection of the loan administrative data: many CL clients used the credit line as if it was a term loan by withdrawing the full amount in the beginning and making regular equal repayments. 22 Costly learning thus may explain why we find small average treatment effects in the whole sample, and why the effects increase over time. An alternative explanation could be that our indicator of good vs. poor knowledge might be picking other differences between the participants. For example, better educated participants may have been more likely to answer the practical question correctly and also better able to exploit profitable opportunities. To examine this hypothesis, we compare the baseline characteristics of the poor knowledge and good knowledge groups (see Table B.6 in the Appendix). Before obtaining the credit, both good and poor knowledge participants look alike in most aspects: they have similar 22 Most clients are observed only up to six months into treatment. Thus, we cannot examine whether these effects persist or would decrease in a longer term. 18

19 Figure 2: Average daily profits by weeks into treatment, in INR profits, incomes, business age, literacy level and use similar sources of working capital, among other characteristics. There are only (marginally) significant differences in sales, cost of sales, and household size. To formally assess whether our results are picking up initial differences between good and poor knowledge clients, we re-do Table 3 by adding interactions of the treatment indicators with several baseline characteristics. Specifically, we use interactions with baseline profits, sales, cost of sales, literacy and household size. We then check whether the effect of CL on outcomes still differs between the borrowers with good vs. poor knowledge. If the differences disappear after adding the interaction terms, this would suggest that a baseline difference could be driving our results. However, the results in Table 4 show that this is not the case. After adding interaction terms, we still find results very similar to those in the baseline regressions: the clients with good CL knowledge experience larger increases in profits, sales, and cost of sales. The above results give us reason to believe that the heterogeneous effects by product knowledge that we document are not driven by differences in observables. The policy implication, if our interpretation of costly learning is correct, is that additional training or providing more information 19

20 would complement and enhance the benefits from offering the CL product. Of course, we cannot completely rule out the possibility that some of the results might be driven by unobservable characteristics or reverse causality: participants that, for other reasons, experience an increase in profits may have been more willing to learn how to use the credit line. If this were the case, then our results still point out relevant heterogeneous treatment effects: some participants benefit from the program more than others. However, the policy recommendation is different - try to identify and target those borrowers. Table 4: Difference in effect of CL by knowledge of product Difference of CL effect between clients with good and poor knowledge of product (1) (2) (3) (4) (5) Gross profits 91.1* 76.6* ** 104.2** (49.0) (45.6) (46.0) (49.5) (51.1) Sales 610.0** 501.3** 469.2** 640.3** 675.6** (247.8) (211.9) (212.3) (257.9) (260.8) Cost of sales 496.0** 419.9** 393.5** 506.1** 541.6** (207.0) (179.1) (179.0) (212.9) (215.9) Other expenses (25.2) (25.3) (25.2) (25.3) (26.1) Interaction Pre-treatment Pre-treatment Pre-treatment Literacy Household of CL with: profits sales cost of sales size Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions use data from financial diaries. Table presents the estimates of the interaction of treatment variable (CL) with an indicator of having responded correctly to practical question on credit line. 4.3 Why did profits increase? The theoretical framework (Section 2.2) suggests two possible economic mechanisms for a credit line to have a larger effect on traders profits compared to a term loan. First, a credit line allows vendors to improve their stock and debt management, by allowing them to borrow more and buy more stock when market conditions are favorable and keep leaner stocks and lower debt when 20

21 conditions are infavorable. This added flexibility implies larger variation in stock and debt levels for CL traders. Second, a credit line allows vendors to change their procurement practices and purchase riskier but more profitable goods. We examine these possible mechanisms empirically. We analyze the effect of a credit line on measures of flexibility or variability of stock and debt (the level and coefficient of variation (CV) of initial stock, total withdrawals, and debt repayment) as well as indicators of procurement practices. We also exploit the heterogeneity documented in Section 4.2 and compare the effects between clients with good vs. poor knowledge of the CL product. Recall that only the former group appears to benefit from a credit line, hence this distinction can be informative for the mechanism linking CL to increased profits. Table 5 displays the main results. There are three important findings. First, regarding stock management, we find no significant effect, either on the level or on the volatility of daily initial stocks (panel A). 23 This finding does not suggest that CL traders profits have increased due to more active stock management. Second, regarding debt management, we observe that CL traders withdraw statistically significantly more funds on average compared to TL traders (see Table 5, Panel B). However, there are no significant differences in total repayments and in interest charges relative to net receipts. 24 We complement these results by using alternative measures of intensity of loan use constructed from the administrative data (see table B.7 in the Appendix). We find statistically significantly larger variation in withdrawals among CL traders. Overall, these results suggest that part of the effect of the credit line on profits can be due to more flexible debt management, specifically withdrawals. Third, we find sizable and statistically significant heterogeneous effects in procurement practices, as reported in the endline survey (see Table 5, Panel C and Figure 3). Poor knowledge CL clients report buying larger stock quantities and are less likely to purchase new types of goods. In contrast, good knowledge clients report buying more profitable goods. We interpret these results as evidence 23 We measure stock volatility using the coefficient of variation (CV). We calculate the CV using all observations in the financial diaries. 24 It could be argued that the effect of CL on profits is driven by the size of the loan: larger total withdrawals may mean that CL traders effectively borrow more than TL traders. However, this is unlikely to fully explain the observed results since an increase in withdrawals is observed for CL traders with both good and poor knowledge of the product, while the increase in profits is observed only among the former. 21

22 that changes in procurement practices are an important channel through which the new product increases traders profits: consistent with our theoretical model, the credit line flexibility allows some vendors to buy different (possibly riskier) and more profitable goods. Figure 3: Self-reported procurement practices, by treatment status and knowledge of product 22

23 Table 5: Effect of CL on indicators of stock and debt management, and procurement practices Mean value Effect of credit line control group All CL Poor Good p-value Outcome Knowledge Knowledge (3)=(4) (1) (2) (3) (4) (5) A. Stock Management Initial stock , (4,145.0) (4,657.5) (4761.7) CV initial stock (0.027) (0.034) (0.029) B. Debt Management Total withdrawals ,330.7*** 3,037.1*** *** (845.6) (1,159.6) (886.2) Total repayments (543.7) (673.0) (575.0) Interest charged *** (as % of net receipts) (0.028) (0.010) (0.047) C. Procurement practices Buy more profitable *** goods (0.073) (0.079) (0.075) Buys better quality goods (0.058) (0.073) (0.068) Buy more quantity ** (0.061) (0.062) (0.072) Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. Data from financial diaries, administrative records and endline survey. Column 5 presents p-value of test of equality of estimates in columns 3 and 4. Net receipts=total withdrawals - total payments. CV: coefficient of variation. 23

24 5 Conclusion We use a field experiment coupled with survey and administrative data to study the impact of introducing a credit line within a microcredit program. We show that receiving a credit line leads to a larger increase in business profits compared to receiving a term loan. This occurs, in particular, for clients with good understanding of the new product. Our results suggest that allowing for more flexibility in the lending terms can increase the effectiveness of microcredit provided to small businesses in developing countries. The added flexibility allows the borrowers to pursue more profitable (possibly riskier) investments and to improve their debt management. These results complement and contribute to growing empirical evidence suggesting that the standard microfinance loan may be unnecessarily rigid and constrain entrepreneurship. Given the available data, we have focused primarily on business indicators such as gross profits, sales, debt, inventory and business practices. We have thus abstracted from the role of risk aversion and the additional potential gains from a credit line that can be obtained in consumption smoothing. On the other hand, we have also not discussed potential self-control issues associated with the more flexible access to credit and repayment, although this concern may be mitigated by the regular weekly meetings and interest payments. Our results suggest that future research needs to examine in more depth the role of training associated with the credit line product: better training may enhance its impact, and hence our study may be providing only a lower bound of the effects. More research is also needed on the possible impact on default rates. While we do not observe any difference in repayments between the credit line and the term loan within the short duration of the experiment, we cannot rule out a differential effect in the longer term. Related to this, further research is needed also on the longrun effects of the credit line intervention in general, including on other outcomes such as business growth and survival rates. 24

25 References Armendáriz, Beatriz and Jonathan Morduch, The economics of microfinance, MIT press, Attanasio, Orazio, Britta Augsburg, Ralph De Haas, Emla Fitzsimons, and Heike Harmgart, The impacts of microfinance: Evidence from joint-liability lending in Mongolia, American Economic Journal: Applied Economics, 2015, 7 (1), Banerjee, Abhijit, Dean Karlan, and Jonathan Zinman, Six Randomized Evaluations of Microcredit: Introduction and Further Steps, American Economic Journal: Applied Economics, January 2015, 7 (1), 1 21., Esther Duflo, Rachel Glennerster, and Cynthia Kinnan, The miracle of microfinance? Evidence from a randomized evaluation, American Economic Journal: Applied Economics, 2015, 7 (1), Banerjee, Abhijit Vinayak, Microcredit under the microscope: what have we learned in the past two decades, and what do we need to know?, Annu. Rev. Econ., 2013, 5 (1), Barboni, Giorgia, Repayment flexibility in microfinance contracts: Theory and experimental evidence on take up and selection, Journal of Economic Behavior & Organization, 2017, 142, and Parul Agarwal, Knowing what s good for you: Can a repayment flexibility option in microfinance contracts improve repayment rates and business outcomes?, Mimeo, Department of Economics, Princeton University, Carpena, Fenella, Shawn Cole, Jeremy Shapiro, and Bilal Zia, Liability structure in small-scale finance: evidence from a natural experiment, The World Bank Economic Review, 2012, 27 (3), Crépon, Bruno, Florencia Devoto, Esther Duflo, and William Parienté, Estimating the impact of microcredit on those who take it up: Evidence from a randomized experiment in Morocco, American Economic Journal: Applied Economics, 2015, 7 (1), Field, Erica and Rohini Pande, Repayment frequency and default in microfinance: evidence from India, Journal of the European Economic Association, 2008, 6 (2-3), ,, John Papp, and Natalia Rigol, Does the classic microfinance model discourage entrepreneurship among the poor? Experimental evidence from India, American Economic Review, 2013, 103 (6), ,,, and Y Jeanette Park, Repayment flexibility can reduce financial stress: a randomized control trial with microfinance clients in India, PloS one, 2012, 7 (9), e Fischer, Greg, Contract Structure, Risk-Sharing, and Investment Choice, Econometrica, 2013, 81 (3), Ghatak, Maitreesh and Timothy W Guinnane, The economics of lending with joint liability: theory and practice, Journal of Development Economics, 1999, 60 (1),

26 Giné, Xavier and Dean S Karlan, Group versus individual liability: Short and long term evidence from Philippine microcredit lending groups, Journal of Development Economics, 2014, 107, McIntosh, Craig, Estimating treatment effects from spatial policy experiments: an application to Ugandan microfinance, The Review of Economics and Statistics, 2008, 90 (1), Morduch, Jonathan, The microfinance promise, Journal of economic literature, 1999, 37 (4),

27 Appendix A Proofs Lemma 1: A trader always earns higher expected profits using a credit line compared to using a term loan of the same size. Denoting cash flow by c t, in each period a TL trader faces the budget constraint: c t + s t+1 p t s t F. (BCT) In contrast, a CL trader has the budget constraint: c t + s t+1 p t s t + b t+1 (1 + r)b t (BCC) where, on the right hand side, b t+1 b t is the change (positive or negative) in debt principal, p t s t is sales revenue and rb t is the required interest payment. Note that, if a CL trader sets b 1 = s 1 = L and b t+1 = (1 + r)b t F for all t = 1,...T, the budget constraints (BCC) become equivalent to the TL trader s budget constraints (BCT). The credit limit b t L is also satisfied at all t since b t+1 = (1 + r)b t F (1 + r)l r(1+r)t (1+r) T 1 L < L. Because the traders payoff function is the same, the above result implies that a CL trader, by using the TL trader s optimal debt and repayment schedules, can always replicate the TL trader s stock and cash flow over time {s T L t, c T L } and hence achieve the same gross profits. While this strategy t 27

28 is always feasible for a CL trader, in general it is not optimal a CL trader can choose a different debt profile {b t } and cashflow and stock paths {c t, s t } and achieve strictly higher present-value expected profits (see the two-period example in the main text). B Additional tables and figures Table B.1: Robustness of Table 2 to including additional covariates Outcome Mean value Regression estimates control group (1) (2) (3) (4) (5) (6) Gross profits * (44.0) (56.8) (50.1) (66.8) (55.6) Sales * (228.6) (213.1) (261.7) (321.6) (291.2) Cost of sales * (184.8) (188.3) (215.3) (258.5) (236.3) Other expenses * 27.2 (17.6) (23.4) (22.3) (20.7) (23.1) Additional controls Yes Yes Yes Yes Yes Weeks into All Weeks Weeks Weeks Weeks treatment 1 to 5 6 to to Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions include the following control variables: pre-treatment values of outcome variable, interview day of the week, household size, literacy, marital status, and an indicator of trading with perishable goods. Columns 2 uses all the sample, while columns 3-6 use sub-samples based on the number of weeks into treatment. 28

29 Table B.2: Robustness of Table 3 to including additional covariates Mean value Effect of Credit Line control group All CL Poor Good p-value Outcome Knowledge Knowledge (3)=(4) (1) (2) (3) (4) (5) Gross profits ** (44.0) (53.9) (46.4) Sales * (228.6) (245.1) (255.6) Cost of sales (184.8) (191.6) (212.0) Other expenses ** (17.6) (24.2) (19.1) Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions include the following control variables: pre-treatment values of outcome variable, interview day of the week, household size, literacy, marital status, and an indicator of trading with perishable goods. All regressions use data from financial diaries. Column 5 presents p-value of test of equality of estimates in columns 3 and 4. 29

30 Table B.3: Robustness of Table 2 to controlling for loan upgrades Outcome Mean value Regression estimates control group (1) (2) (3) (4) (5) (6) Gross profit (51.7) (61.0) (53.6) (74.8) (75.1) Sales (322.7) (292.6) (328.0) (384.9) (471.2) Cost of Sales (268.6) (261.6) (276.1) (314.4) (392.5) Other expenses (26.6) (34.4) (29.6) (28.7) (34.6) Weeks into All Weeks Weeks Weeks Weeks treatment 1 to 5 6 to to Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions include an indicator equal to 1 if client got an increase in initial loan from INR 10,000 to INR 20,000. Columns 2 uses all the sample, while columns 3-6 use sub-samples based on the number of weeks into treatment. Table B.4: Robustness of Table 3 to controlling for loan upgrades Mean value Effect of Cash Credit Loan control group All CL Poor Good p-value Outcome Knowledge Knowledge (3)=(4) (1) (2) (3) (4) (5) Gross profit ** (51.7) (58.2) (56.4) Sales (322.7) (330.3) (357.7) Cost of sales (268.6) (274.6) (297.0) Other expenses (26.6) (28.7) (31.2) Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions include an indicator equal to 1 if client got an increase in initial loan from INR 10,000 to INR 20,000. Column 5 presents p-value of test of equality of estimates in columns 3 and 4. 30

31 Table B.5: Effect of CL on profits, sales, and expenses using 2SLS Outcome Regression estimates (1) (2) (3) (4) (5) Gross profits * (46.2) (62.3) (55.5) (61.5) (69.5) Sales (275.8) (234.1) (310.4) (309.9) (462.9) Cost of sales (229.8) (207.3) (253.6) (256.0) (396.9) Other expenses (22.4) (24.5) (25.2) (25.8) (35.8) Weeks into All Weeks Weeks Weeks Weeks treatment 1 to 5 6 to to Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. All regressions use data from financial diaries. Columns 1 uses all the sample, while columns 2-5 use sub-samples based in the number of weeks into treatment. Regressions are estimated using 2SLS with initial treatment as an instrument for actual treatment. 31

32 Table B.6: Baseline characterists of CL participants, by knowledge of product Poor Good p-value Variable Knowledge Knowledge (1)=(2) (1) (2) (3) A. Profits, sales and expenses Gross profit Sales Cost of sales Other expenses B. Initial and final stock Initial stock Final stock C. Business characteristics Monthly business income Sells perishables Years in business Balance in savings account D. Sources of working capital Bank or microfinance Savings or business profits Wholesaler E. Demographics Household size Can read and write Is married No. Individuals Notes: Monetary outcomes are measure in Indian Rupees (INR). Data are collected in baseline survey or pre-treatment financial diaries. Column 3 displays p-value of a mean comparison test. 32

33 Table B.7: Effect of CL on indicators of loan use Mean value Effect of credit line control All CL Poor Good p-value Outcome group Knowledge Knowledge (3)=(4) (1) (2) (3) (4) (5) Decreased borrowing *** 0.129*** 0.152*** from other sources (0.043) (0.049) (0.044) Total net cash ,533.3*** 2,136.4*** *** receipts (615.9) (786.6) (647.5) No. of payments (0.681) (0.844) (0.717) No. of withdrawals *** 0.518*** 0.527*** (0.112) (0.141) (0.124) Repayment intensity * (0.012) (0.016) (0.018) Withdrawal intensity *** 0.017** 0.026*** (0.006) (0.008) (0.009) Total intensity *** 0.039** 0.053** (0.015) (0.016) (0.023) Notes: Robust standard errors in parentheses. Standard errors are clustered by joint liability group. * denotes significant at 10%, ** significant at 5% and *** significant at 1%. Column 5 presents p-value of test of equality of estimates in columns 3 and 4. Regressions use data from endline survey and administrative records. Repayment intensity = a loan use intensity measure constructed from the sum of squared percentage deviations from the default repayment in a TL loan (INR 300 for INR 10,000 loan and INR 500 for a 20,000 loan). A value of 0 for repayment intensity would mean that a CL client uses the CL exactly as a TL loan of the same size, with the borrower always making the default repayments. Withdrawal intensity = a loan use intensity measure constructed from the sum of squared unscheduled withdrawals as percent of current debt balance; Unscheduled withdrawals are defined as all credit entries except the first loan withdrawal (to be comparable with TL) and borrowing limit upgrades. For TL clients this measure equals 0 by construction. Total intensity = the sum of repayment and withdrawal intensity. This measure proxies debt management flexibility - it will be zero or close to zero for TL clients and large for CL clients that use the credit line most actively (defined as deviations from the TL schedule). 33

34 Table B.8: Credit line product description 34