Demand for Flexibility in Microfinance Loans: Evidence from India

Demand for Flexibility in Microfinance Loans: Evidence from India Jihae John Hwang Northwestern University June 2011 Abstract Microfinance loans have expanded greatly in the recent years as a way of helping the poor gain cheaper loans, but percentage take-up of microcredit has remained lower than expectations. In an attempt to explain this phenomenon, this paper uses a randomized evaluation of the impact of introducing microfinance loans in a new urban setting in Hyderabad, India, to determine if lack of flexibility in microfinance loans can be an answer to the problem. Using wages and sales volatility as proxies of demand for flexibility, this paper explores whether people are behaving as if microfinance loans are too stringent for their needs. The findings are consistent with the fact that there is some demand for more flexibility in microcredit, especially among business owners, but there also seems to be some surprising differences in terms of stratification according to education. For the high-education people, increasing sales volatility is associated with less take-up of microfinance loans, but for the low-education people, increasing volatility is correlated with more take-up of microfinance loans. This introduces the potential that microfinance institutions are missing out on the most profitable clients in the market. Dividing the population according to a measure of risk-averseness did not have a conclusive impact. A very special thanks to Professor Cynthia Kinnan for serving as the faculty advisor for this thesis and helping me through all the difficulties. I would also like to mention Keyoung Lee for helping me identify the topic of interest and my parents for their feedback on the microfinance issues.

1 Introduction Microcredit has expanded greatly in the recent years as a way for the poor to gain cheaper loans. According to Reed (2011) in the 2011 annual report by the Microcredit Summit Campaign, as of December 2009, microfinance institutions had a total of 190,135,180 clients, with about 81.7% of them being women. Moreover, people have started to look at microfinance as a successful way of alleviating poverty in the developing countries, perhaps the most clearly evidenced by Mohammad Yunus, founder of The Grameen Bank, winning the Nobel Peace Prize in 2006. However, despite the growing number of microfinance clients, percentage take-up of microfinance (MFI) loans has not been as high as institutions expect. Karlan, Morduch, and Mullainathan (2010) show that take-up rates of financial services range between 2% and 84% of eligible individuals, and the variation in take-up of loans is actually larger than that of participation rates in other types, such as savings and insurance. In most cases, take-up of microcredit did not exceed 50%, as was the case with the data used in this paper, when Spandana offered its microfinance services in Hyderabad, India. Thus, this implies that microfinance loans are not for everybody, despite being recognized as the best remedy to the poor by a growing number of people. To attempt to explore this phenomenon in more detail, I look at the common characteristics of microcredit. Traditionally, microfinance loans have several criteria that make them rigid, including group liability and frequent, regular payment (Armendariz and Morduch 2000). Thus, it lends to the possibility that microfinance loans may not be structured in the best possible way for some people, especially those who face irregular income. Moreover, Collins, Morduch, Rutherford, and Ruthven (2009) claim that the poor are characterized by what they term the Triple Whammy : low incomes, irregularity and unpredictability, and lack of tools. Therefore, it may be that irregularity and unpredictability so often found among the poor are preventing them from taking on the rigid microfinance loans. Expanding on this idea, I focus this paper on attempting to shed light on whether rigidity in microfinance loans can be an explanation to why take-up percentages have been lower than expected. If irregularity and unpredictability in income is a widespread problem among the poor, you would expect people to be less able to pay back loans 1

on a tight schedule, and thus, perceive microfinance loans to be more expensive than the interest rate suggests, leading to less take-up of microcredit. In some cases, people might prefer the informal loans with higher interest rates, if these loans provide them with the flexibility to pay back whenever they are able to. On the flip side, case studies have shown that some people actually do prefer the rigidity provided by microfinance loans because it allows them to refrain from overspending on unnecessary goods. For these people, loans are substitutes for saving; since they find it hard to save when they have the extra money, they take out loans in an attempt to keep themselves from falling into temptations. A good example is provided by Collins et al. (2009), in which a case study on a person from Bangladesh, Surjo, is introduced. In this case, Surjo was determined to find a way to save money and even opened a bank account to do so, but his attempt at saving only went as far as the first month. The next month, he excused himself from depositing because of the Eid festival, and that was the end of his deposits. However, when his mother was able to get a microfinance loan of $180, Surjo was forced to make the regular weekly repayment of about $4 throughout the year, providing him with the necessary discipline that allowed him to gain a large sum of money at once. Therefore, I try to clear up this question by looking at the demand for flexibility in microfinance loans, from the point of view of the borrowers. Using data from a randomized experiment of the canonical group-lending microfinance model, I look at the households in the treatment group to see how they reacted to the increased access in microfinance loans, compared to those in the control group. I use the endline survey, done 15-to-18 months after the treatment occured, to develop proxies for measuring demand for flexibility and use these variables to perform the analysis to follow in this paper. I also attempt to see if the demand for flexibility changes with different groups of households, including measures of education and risk averseness. 2 Literature Review As microcredit gains in importance, there have been a growing amount of literature about the success and impact of microfinance loans on the poor. Collins et al. (2009) 2

give examples of different individuals in their case studies, which provide us with reasons to believe the poor can benefit greatly from increased access to financial instruments. Even though the poor have very little income to live on, other case studies reveal sophisticated informal lending schemes that exist within neighborhoods, such as savings groups and arrangements that seem to blend credit and insurance (Udry 1990, Dercon 2002). Moreover, case studies such as the one mentioned above show that microfinance, structured the way it is currently, is having a positive impact for at least some of the people. However, because of the special circumstances that the poor are often faced with, each household seems to react in a different way, sometimes even in ways that may seem irrational to classical economic theory. Therefore, case studies fail to reveal the impact of microfinance on an average household, leading to the need for more quantitative approaches. Recently, there has been more research done in the quantitative aspect, including random experiments performed to examine the impact of microcredit. In one such attempt, Banerjee, Duflo, Glennerster, and Kinnan (2010) find heterogeneous effects of microcredit across different household characteristics, especially in terms of owning a business. Households with existing businesses invest more in durable goods while not changing their nondurable consumption, and households with high propensity to start a new business increase spending in durable goods while decreasing consumption in nondurable goods, possibly to pay a fixed cost needed to start a new business. Households unrelated to business tend to increase their nondurable spending with the new microfinance loans. However, despite evidence for positive impact for microfinance, there have also been concerns with the low take-up percentages of microfinance loans. Karlan et al. (2010) provide important reasons that explain why take-up and participation rates in microfinance matter. First, even though microcredit has had success providing better access to financial services for the poor entrepreneurs, the impact has been limited in terms of helping the poorest individuals. Second, take-up rates can be interpreted as interest in given products, meaning that they can be used to interpret the needs of the poor individuals and determine what they want the most. Third, take-up rates can affect the empirical methodology used to measure impact because low take-up rates can force 3

experimenters to increase the sample size for surveys to gain the necessary statistical power. Karlan et al. (2010) also illustrate that take-up in microcredit has often been less than 50%, and there is a great variation in take-up rates calculated by different studies around the world. They notice that it is difficult to discern a clear pattern in determining take-up rates, making it more important to understand the reasons behind why they are so low. In conjunction, they provide results from a survey done to gather data on why individuals did not borrow, in which they find that the most common reason is a desire not to be in debt. However, this gives rise to the subsequent question of whether the respondents were simply averse to the concept of borrowing in general or were averse to borrowing from the particular loan that they had access to. Thus, the authors conclude that it is difficult to say whether debt at a lower interest rate or with more flexible repayment structure would have resulted in higher take-up rates for the group. In a similar note, Jain and Mansuri (2003) assert that the use of regularly scheduled repayments in microfinance loans may be a reason why informal lenders still thrive in regions where microcredit has been firmly established. Instead of arguing that regular repayments are for fiscal discipline, the authors point out that this characteristic of microfinance loans is compensating for the lack of ability to monitor the borrowers actions. By introducing inflexibility, microfinance institutions are forcing the borrowers to turn to the informal sector for help in paying back on a regular basis, and thus, using the informal lenders superior ability to monitor the borrowers actions. Thus, even in areas where microfinance has a firm footing, informal lenders are able to survive by taking advantage of the information asymmetry between the microfinance institutions and the borrowers. In literature that is more specific to what I attempt to explore in this paper, Karlan and Mullainathan (2006) provide two short case studies of a farmer and a small provision shop owner to motivate the idea that more flexibility in microfinance loans may help a large group of the poor. From the demand side, they observe that rigid contracts greatly constrain loan size and keep some lucrative borrowers from taking out MFI loans. Moreover, they argue that more flexibility can also help the microfinance institutions giving out the loan, by saving loan officers time and increasing the impact 4

of the loans. Rigid loan contracts may force the clients to take actions that may actually be detrimental to returns on investments; for example, they might decide to sell off an asset to not default on their loan. Thus, by increasing flexibility, clients will have more freedom to gain the best returns on their investments, possibly increasing income and allowing the MFIs to increase loan size. In a more quantitative study, Kaboski and Townsend (2005) assert that even though microfinance institutions have a positive effect as a whole, differences in repayment frequency did not have measured impacts. However, they also note that there was very little variation in requiring frequent payments, which may be the reason why they could not find significant difference. There have also been some recent studies specific to measuring the impact of flexibility in a randomized setting. In one such paper, Field and Pande (2008) observe that lower frequency repayment schedules do not result in higher default rates, meaning that MFIs could actually save money due to lower transaction costs. In a similar study, Field, Pande, and Papp (2010) argue that delaying the onset of repayment significantly increases both business investment and default, presenting a trade-off for banks and clients. Moreover, they notice that rigidity in loans deters risk-averse clients the most from taking on illiquid investments, meaning that they might be the ones who would be the most interested in increased flexibility. In a paper that looks at the flexibility issue from the borrowers side, Pearlman (2010) provides some evidence that repayment flexibility in microfinance loans matters in determining demand for microfinance loans. The author argues that flexibility is important for the poor who face high income risk and limited means of smoothing out consumption and costs, based on observations that these poor households rely more on flexible, informal loans, despite the fact that they have access to microcredit. Thus, she argues that lack of flexibility may explain low take-up rates of MFI loans and continued survival of the informal sector in the credit market. Moreover, in the appendix of her paper, she outlines the details behind a proposed experiment to test whether introducing more flexibility will increase the demand for MFI loans. In this proposal, she introduces a new type of microfinance loan, in which the borrowers are given the option of reducing the next payment and pushing the balance to 5

any future period, including the last one (Pearlman 2010, 31). By handing out this new loan to the treatment group, while providing the control group with the standard MFI loan that has a regular repayment schedule, the author hopes to use this randomized experiment to determine the effect of increased flexibility. These existing literature suggest that while there have been general studies done on the impact of microfinance as a whole, there have been limited work done in measuring the specific characteristics behind MFI loans. Case studies have shown that there is good reason to argue for both sides of the debate; Collins et al. (2009) show us the case in which flexibility has had a positive impact on the household while Karlan and Mullainathan (2006) motivate the idea that more flexibility may help others. Moreover, the recent randomized experiments done on the topic are focused more on the result of increased flexibility from the point of view of the MFI institutions, rather than from the demand side of the borrowers. Even in the paper that looks at the demand side, Pearlman (2010) uses a more general approach using high-level data to support her argument, instead of using household-level data from a randomized experiment, as I have done here. Thus, with microfinance take-up being lower than expected, looking to see if there is a strong demand for flexibility in microfinance loans would provide interesting insights into whether the structure of these loans should be changed to accommodate more people and speed up the growth of the microcredit campaign. 3 Theoretical Model 3.1 Flexibility of Loans To predict how increasing flexibility in microfinance loans will affect borrowers, I look at two different types of loans: a loan with a weekly repayment schedule (weekly loan) and a loan with a monthly repayment schedule (monthly loan). To simplify the calculations, I assume that there are four weeks in a month and that there is no time value of money. Now, suppose that for the weekly loan, you need to pay back x dollars a week, where x > 0, and for the monthly loan, you need to pay back 4x dollars a month, to make the loans equal in payment. Moreover, assume that for the weekly loan, a default is defined to be a state in which the person cannot pay back the loan in at least one of the weeks; 6

in other words, to not default on the weekly loan in an entire month, the person has to be able to pay back the loan in all four of the weeks. For the monthly loan, a default is when you are not able to pay back the loan at the end of the month. Now, suppose the weekly profit function for a person is given by the normal distribution with mean µ and standard deviation σ, where µ > 0: π N(µ, σ 2 ) Assuming that profits in any given weeks are independent of each other, the monthly profit function is given by: 4π N(4µ, 4σ 2 ) Moreover, for simplicity, assume that all of the profit can be used to pay back the loan, meaning that a default occurs only if the profit is less than the amount of loan that needs to be paid back. Under this assumption, the probability of default on the weekly loan and that on the monthly loan can be calculated. First, for the weekly loan, the probability of default (denote d w ) is given by: d w = 1 [Pr (π x)] 4 = 1 [1 Pr (π < x)] 4 Similarly, for the monthly loan, the probability of default (denote d m ) is given by: d m = Pr (4π < 4x) Now, using the distribution of π and 4π, the probability of default can be simplified to: [ ( )] 4 x µ d w = 1 1 F σ ( ) 2x 2µ d m = F σ Here, F is the cumulative distribution function of the standard normal distribution. Now, under these assumptions, it can be proven that the monthly loan is always at least as good as the weekly loan, or that d w d m for any x, µ, and σ. However, case studies have shown that there are people who would be better off with a weekly loan because of the discipline provided by frequent repayment. To capture 7

this dimension, I introduce a probability of being tempted to buy a good, denoted by p. Again, to simplify calculations, I assume that a person can only be tempted during the first week and that if one is tempted, one will use all of the first weeks profit to buy this good. Moreover, the cost of default is higher than the increased utility from buying this good, and thus, if a person is on a weekly loan, one will never use the profit to buy this good even if tempted, if this does not allow one to pay back the profit. Therefore, even with this newly introduced probability, d w does not change and is equal to what we calculated before: d w = 1 [ 1 F ( )] 4 x µ (1) On the other hand, if a person is on the monthly loan, one will use up all of the first weeks profits if tempted to buy the good. This means that in this case, one will have to use only three weeks worth of profits to pay back 4x. Again, assuming independence of profits, we have: σ 3π N(3µ, 3σ 2 ) Therefore, the new probability of default on the monthly loan is calculated to be: [ ( )] [ ( )] 4x 3µ 2x 2µ d m = p F + (1 p) F 3σ σ (2) Now, using these new probabilities of default, define the difference between the two equations given by (1) and (2): D = 1 [ 1 F ( x µ σ )] 4 [ ( )] [ 4x 3µ p F (1 p) F 3σ ( )] 2x 2µ To compare between the two loans, we can now look at the sign of the difference defined above in equation (3). More specifically, if D > 0, the probability of default on the weekly loan is higher, meaning that the monthly loan is better; and if D < 0, the probability of default on the monthly loan is higher, meaning that the weekly loan is better. First, we make the observation that D is a monotonically decreasing function in p. We can see this by taking the first-order partial derivative with respect to p: ( ) ( ) D 4x 3µ 2x 2µ p = F + F 3σ σ σ (3) 8

Since F is the cumulative distribution function of the standard normal distribution, it is an increasing function. Thus, to determine the sign of the above equation, we can just look at the sign of the difference between the two components inside F : 4x 3µ 2x 2µ 3σ σ = ( 4 2 3 ) x + ( 2 3 3 ) µ 3σ (4) Equation (4) is greater than 0 because of the assumption than the loan size and the mean profit is both greater than 0. Therefore, we have: ( ) ( ) 4x 3µ 2x 2µ F > F 3σ σ and ( ) ( ) D 4x 3µ 2x 2µ p = F + F < 0 (5) 3σ σ Therefore, from equation (5), we now know that D is decreasing in p, meaning that as the probability of being tempted increases, the weekly loan becomes a better option. Thus, we can assume that for the right parameters (x, µ, σ), there will be a probability of being tempted, p, such that the difference is 0, i.e., the person will be indifferent between taking the weekly and the monthly loan. This p can be found by setting D in equation (3) equal to 0: p = 1 [ 1 F ( )] x µ 4 ( σ F 2x 2µ ) ( ) σ 4x 3µ F 3σ F ( ) (6) 2x 2µ Moreover, before the introduction of the new parameter, p, it can be easily seen that as volatility in profits (σ) increases, the monthly loan becomes a better option. Since the probability of being tempted makes the weekly loan a better option, this means that as the volatility measure increases, p must also increase to make the person indifferent, as illustrated in Figure 1. Summing things up, if we fix parameters (x, µ), there is a probability of being tempted that makes the person indifferent between the weekly and the monthly loan, p, for each σ, as given by equation (6). Moreover, p σ > 0, or p is an increasing function of σ. Thus, there are two dimensions at work here in determining whether a person prefers the weekly or the monthly loan. σ At the one extreme, those with a high probability 9

of being tempted and low profit volatility prefer the weekly loan, while those with a low probability of being tempted and high profit volatility prefer the monthly loan. In this sense, profit volatility is acting as a proxy of demand for flexibility in loans for a particular person because ceteris paribus, a person with more volatile profits is more likely to take out the monthly loan. 3.2 Importance of Outside Loan Options In addition to the above model, there is the possibility that high profit volatility generally increases the need to take out a loan. Those with high profit volatility would have the need to smooth out consumption and other costs, and taking out loans is one of the ways in which the poor manage to do this (Collins et al. 2009). Thus, there is the possibility that these people will be more likely to take out loans in general despite the fact that they are also the more likely people to default on the loans, which creates complications to testing the model above. Moreover, since we are testing the demand for one specific type of loans in microcredit, access to substitutes in the form of other outside loan options becomes very important to testing the model above. For example, consider two people with the same distribution of profit. Now, suppose that these two people differ in the number of outside loans options that they have, in the sense that the first person only has access to an inflexible, microfinance loan, and the second person has access to other, more flexible loans. In this case, the first person might have no choice but to take a microfinance loan to smooth out consumption, while the second person can use the greater access to his advantage to take out a more flexible loan that fits the profit schedule better. If such relationships hold, this means that for those people with access to just microcredit, sales volatility will actually be a measure of their demand for loans in general, instead of demand for flexibility. Only if people have access to other substitutes, such as the monthly loan in the first model, will I be able to say that sales volatility is a proxy for demand for flexibility in loans, which is the variable of interest in this paper. 10

3.3 Predictions The existence of two dimensions in the first model suggests that it will be difficult to say whether there will be a general demand for more flexibility in MFI loans. For example, if it was the case that there were a lot of people concerned about being tempted by unnecessary goods, it might be the case that there is no demand for further flexibility in loans, and people like the MFI loans the way it is constructed right now. However, I predict that there will be a general demand for more flexibility because low take-up rates of microcredit signal that these loans are not for everybody. It is likely that the rigid structure of MFI loans is holding some people back from making free use of them, and thus, I would expect there to be at least some level of demand for flexibility. Moreover, I would expect the demand to be higher among the people who have a small probability of being tempted and are more risk-averse. If the probability of being tempted is not a concern, then one of the advantages of microcredit disappears, and it is more likely that the person will demand more advantages from the loan to take advantage of it. There would be two types of people who fall under this description: (a) those who are less likely to be tempted by unnecessary goods and (b) those who do not realize the risk of being tempted. In other words, it is the perceived probability of being tempted that is important for the model above, not the real probability. Thus, interpreting the second condition in another way, people have to be sophisticated and recognize the risk of being tempted by unnecessary goods for them to realize the benefits of additional commitment that microfinance loans bring. In addition, if a person is more risk-averse, one will most likely be very much concerned with the possibility of default. Thus, this person will probably demand more flexibility than the average person and ceteris paribus, be less likely to take out a microfinance loan. 11

4 Description of Data and Experimental Design 4.1 Product The data is from a randomized experiment, with cooperation from Spandana, one of the largest and fastest growing microfinance organizations in India. Spandana mostly offers the canonical group loan product, in which a group is comprised of 6 to 10 women and is jointly responsible for the loans of the group members. The first loan is for Rs. 10,000, and it takes 50 weeks to pay back principal and interest, at an APR of 24%. If all members repay, they are eligible for second loans of Rs. 10,000 to 12,000, and the loan amounts increase up to Rs. 20,000. Moreover, Spandana does not require clients to start a new business to take out a loan, meaning that they are free to make the best use of money, as long as they are able to repay the loan (Banerjee et al. 2010). Eligibility for a Spandana loan is determined by using the following criteria: (a) be female, (b) be aged 18 to 59, (c) have resided in the same area for at least one year, (d) have valid identification and residential proof (ration card, voter card, or electricity bill), and (e) at least 80% of women in a group must own their home. Groups are formed by women themselves, [and] Spandana does not determine loan eligibility by the expected productivity of the investment (Banerjee et al. 2010, 5). 4.2 Experimental Design Spandana first identified 120 neighborhoods in Hyderabad, India as places in which they were possibly interested in expanding their microfinance business. Spandana selected these areas based on the following criteria: (a) having no pre-existing microfinance presence, (b) having residents who are desirable as potential borrowers, meaning that they fall into the poor category but not the poorest, and (c) not having a high concentration of people who move frequently, such as construction workers. In choosing these neighborhoods, the largest areas were excluded from the study as Spandana wanted to expand operations there, and the population in the neighborhoods selected ranged from 46 to 555 (Banerjee et al. 2010). A baseline survey was conducted for each of these areas in 2005. Households surveyed were selected conditional on having a women aged between 18 and 55, and information 12

was collected on various individual and household characteristics, outstanding loans of the households, and businesses currently operated or stopped within the year by the households. A total of 2,800 households were surveyed for this portion of the study (Banerjee et al. 2010). Before randomization, sixteen areas were dropped from the potential neighborhood list because of a large number of migrant-worker households. Like most other microfinance institutions, Spandana only provides loans to residents who have lived in the same area for more than a year because the incentive of providing future credit works better for these households. Then, the 104 neighborhoods were paired based on minimum distance according to per capita consumption, fraction of households with debt, and fraction of households who had a business, and one of each pair was randomly assigned to the treatment group (Banerjee et al. 2010, 6). Since randomization took place at the neighborhood level, all standard errors in the regressions run for the purposes of this paper were clustered to the neighborhood level, to adjust for correlated outcomes among the residents of the same neighborhood. Between 2006 and 2007, Spandana began operating in the 52 areas assigned to the treatment group. Even though other microfinance institutions also began operating in both control and treatment areas during this time frame, data shows that there was a significant difference in microfinance borrowing in treatment and control groups. Moreover, in early 2007, a comprehensive census of each area was undertaken to determine the percentage take-up of microfinance loans and to establish a sampling frame for the subsequent endline survey (Banerjee et al. 2010). The endline survey took place between August 2007 and April 2008. It was conducted at least 12 months after Spandanas operations began, and more generally, 15 to 18 months after. Because the census suggested low rates of microfinance borrowing in the areas, the endline sample focused on households with high propensity to borrow based on the following conditions: households who (a) had stayed in the area for at least 3 years and (b) had at least one woman aged 18 to 55 living in the household. Because of low rates of borrowing, the sample size was increased for the endline survey, and households who were in the baseline survey were not purposefully resurveyed, although there still might be overlaps (Banerjee et al. 2010). 13

The tables in the appendix summarize some key statistics for the control and the treatment groups. First, Table 1 shows that the areas did not differ in their baseline levels of number of households, maximum education, consumption per capita, number of loans, total household indebtedness, and number of businesses. These results are not surprising given the stratification done to pair the areas before randomization (Banerjee et al. 2010). Table 2 shows some key household-level characteristics in the endline survey for control and treatment groups. It shows that there was no difference between the areas in terms of maximum education, having at least one person working for a wage, number of people working for a wage, total wage earnings per adult equivalents, sales volatility in businesses, owning an old business, and adult equivalents. Again, this is not a surprising result because treatment was randomized within a stratum (Banerjee et al. 2010). Thus, these characteristics are not outcomes of increased access to microfinance loans and will be used as proxies of demand for flexibility and control variables in the regressions to follow in the paper. Moreover, as mentioned briefly above, there were other microfinance institutions that started to operate in the areas after the Spandana intervention. Thus, in order to interpret the differences between the control and the treatment as groups as being due to the increased access in microcredit, it must be the case that MFI borrowing is higher in the treatment group compared to the control. Table 3 shows this to be the case. Households in the treatment area were 10.2% more likely to report having a MFI loan, as 18.2% of the households in the control group had microcredit, while 28.4% of the households in the treatment group had an outstanding microfinance loan. Even though the absolute levels were not high, the number of households using microcredit is about 50% higher in the treatment areas, compared to the control. In terms of the amount borrowed, households in the treatment areas report having about Rs. 1,541 more MFI borrowing than those in the control. Thus, we can safely say that the differences between the control and the treatment are due to the random assignment of microcredit access (Banerjee et al. 2010). 14

4.3 Variable Descriptions For the purposes of this paper, individual-level data were collapsed down to the household level because borrowing decisions are most likely made with the entire household in mind. Moreover, since the endline survey does not match up with the households in the baseline survey, the baseline data was used only to control for various neighborhood characteristics. To pick out good variables to use as proxies of profit volatility (which is being used as proxy of demand for flexibility in the first theoretical model), I first went through the endline survey to find related questions and settled on two main categories: wage and sales volatility. I was motivated to use wages as a proxy variable because of the findings by Collins et al. (2009). In their book, the authors divide earnings activity into three definitions: regular wages, casual work, and self-employment. They also show that India has a lot of variance among the three definitions, which means that it would be a good variable to use in the regression. Moreover, earning wages would provide the household with a more stable source of income, leading to less volatile profit functions and less demand for flexibility. It is more intuitive to use sales volatility measures as a proxy. More variance in sales numbers means that the related profit functions would also be more volatile; and this leads to the conclusion that the household may find it difficult to pay back loans on a regular schedule. In the extreme example of earning zero in a bad week, you would not be able to pay back a loan that is on a weekly repayment schedule. Thus, having more sales volatility would increase the demand for flexibility in loans. Furthermore, since the question of interest is about the take-up of microfinance loans, I was careful to use variables that are not likely to change due to increased access to microfinance loans. For an already existing business, sales volatility is something that is not likely to change unless the business itself changes, and having wage earnings is also something that is unlikely to change because of accepting a microfinance loan. Thus, even though I use the endline survey taken about a year after the Spandana operations, the variables used should not be so different from the time that the household decided to take out the loan. To test for the theoretical model incorporating the idea about probability of temp- 15

tation, I use education as a proxy and divide the households into groups by education level. To define education as a variable, I used the maximum education level in the household. Education can work as a good proxy because you might expect people with higher education to be more financially literate and thus, more knowledgeable about the risks around them. Therefore, we would expect high-education people to be not as easily tempted by the luxury goods, compared to the low-education people. Table 4 reiterates this result. The regression uses the percentage of household consumption spent on temptation goods, as identified by the household, as the dependent variable and education as the independent variable. The baseline data was used for the purposes of this regression to make sure that the relationship is not due to increased microcredit access. The results show that even though the coefficient is a small negative number, it is statistically significant at the 5% level, lending some evidence to the fact that high-education people consume less temptation goods than the low-education people. Going back to the first theoretical model, this means that low education people are more likely to prefer microfinance loans because their relatively high probability of being tempted makes the discipline provided by the inflexible loan more attractive. In addition to being a proxy for the probability of being tempted, education also brings another confounding variable into the mix in the form of increasing outside loan options. To see this, I use a multinomial probit regression to predict the type of loans taken out by different groups of households. For the purpose of this model, loan types are divided into three categories: informal loans, MFI loans, and formal loans. Then, the households are categorized into eight different cases: a household (a) has no outstanding loans, (b) has only informal loans, (c) has only MFI loans, (d) has only formal loans, (e) has informal and MFI loans, (f) has informal and formal loans, (g) has MFI and formal loans, and (h) has all three types of loans. Control variables to be used in subsequent regression models are used as the independent variables, along with the indicator variable for treatment. The base case in the regression is the household with just an informal loan. Table 5 shows the results from the multinomial probit model. It predicts that higheducation households take out more formal loans, while you do not see a significant impact of education on a household having a microfinance loan. Therefore, it is plausible 16

to think that higher education can be linked to having access to more variety of loans, giving them more substitutes for microfinance loans in the case that these loans are not as attractive to the high education borrowers. Thus, going back to the second theoretical observation, higher education households, by virtue of having more outside loan options, can choose which kinds of loans to take out, while lower education households are bound to take out microfinance loans because they do not have other substitutes to turn to. In this sense, low education households might be inherently more likely to take out microfinance loans because they do not have a better option at hand. In addition to dividing the households by education, I try a measure of risk averseness as another way of grouping the households. To do this, I create another variable, called Risk averse. To do so, I use two survey questions, which asked the household to identify whether they had a bank account and an insurance. The variable then takes one of three values: a) 0, if household had none of the two, b) 1, if household had exactly one of the two, and c) 2, if household had both. The key control variables used in the regression models are five baseline neighborhoodlevel average control variables (education, consumption per capita, number of loans, number of businesses, and amount of loans), number of households in the neighborhood, total monthly household consumption per adult equivalents 1, an indicator variable for whether the household had a pre-existing business (business more than a year old), amount of pre-existing non-mfi loans (more than a year old), an indicator variable for whether the household had taken out a MFI loan in the past that is no longer outstanding, and an indicator variable for whether the household has a pre-existing loan that is flexible (allows irregular payment cycles). Moreover, even when education is not being used to divide the population into groups, I included it as a control variable in the model. As mentioned above, the key variables of interest are in two categories: existence of wage jobs in households and sales volatility measures of household businesses. Variables related to wage jobs are the following: (a) Workwage, an indicator variable for whether the household has at least one person working for a wage and (b) Wagetot, total number 1 Used the conversion to adult equivalents by Townsend (1994), in which the weights are: for adult males, 1.0; for adult females, 0.9; for males aged 13-18, 0.94; for females aged 13-18, 0.83; for children aged 7-12, 0.67; for children aged 4-6, 0.52; for toddlers 1-3, 0.32; and for infants, 0.05. 17

of people in the household working for a wage. Variables related to sales volatility of businesses are the following: (a) Sales dif, the reported difference in sales between the best month and the worst month, (b) Week dif, the absolute difference between sales in the past week and sales in the past day multiplied by 7, (c) Month dif, the absolute difference between sales in the past month and sales in the past week multiplied by 365 12 7, and (d) Mon2 dif, the absolute difference between sales in the past month and sales in the past day multiplied by 365. The dependent variable in all of the regressions is 12 No mfi, the number of MFI loans that the household has outstanding. 4.4 Context: Summary Statistics of Key Variables Before I go into the analysis, I provide selected summary statistics of key variables in Table 6. This is to set the context for the models to follow, and I will be referencing back to this table to give more concrete interpretations of the coefficients. All of the summary statistics are from the endline survey data, which took place after treatment. First, column 1 in the table shows that less than 25% of the households have a MFI loan outstanding, and most of the households that have a MFI loan has only one. In terms of maximum level of education in the household, the average is about 11 years, which translates to 10th grade study. The median level of education is also 11, meaning that education seems to be pretty balanced in terms of households in the lower-level and the higher-level. Moreover, over half of the households had at least one of the two financial instruments used to create the risk averse measure: bank accounts and insurance, as shown by the median number in column 3. In columns 4 through 9, key statistics for proxy variables of flexibility demand are given. In terms of having wage jobs, about 72.17% of the households had at least one working for a wage. When looking at the total number of people, the mean was about 1.19, meaning that there were quite a few households in which more than one person was working for a wage. In terms of sales volatility, all four measures have different number of observations, and this is because the variables were created using self-reported measures. Thus, not all people responded to the questions asked in the survey about the sales numbers in business, and this means that there is no data available for some of the sales measures. 18

First, looking at the mean level of volatility, it is the highest in column 6, followed by columns 9, 8, and 7. This is due to the construction of these variables. Because Sales dif measures the volatility between the best and the worst month, you would expect the difference to be the highest. This is followed by the Mon2 dif variable, which measures the volatility between month and day sales, and the Month dif variable, which measures the difference between month and week sales. You would expect volatility to be higher when comparing month sales with a smaller time frame (day sales), compared to a larger one (week sales). Finally, the Week dif variable measures volatility between week and day sales, which is at the week level and not the month level as the other three variables were. Therefore, you would expect this number to be the lowest. All four sales volatility measures also have very high standard deviations and distributions that seem to be skewed to the right. Comparing the median and the mean levels, the mean is always to the right of the median, signaling that there may be extremely high values of volatility. The high standard deviation numbers mean that most of the data is within a single standard deviation from the median, especially for smaller values. Thus, it might be more useful to look at the quartile numbers in interpreting the coefficients on sales volatility. 5 Analysis and Implications To look at the demand for flexibility, I fix a dependent variable, the number of MFI loans outstanding in the household (No mfi), and observe the impact of the variables of interest listed in the previous section. In the first part of the analysis, I just use the proxy variable interacted with treatment to gain an overall sense of whether there is some demand for flexibility. Then, in the second and third parts, I go deeper into the analysis, stratifying the population in groups to calculate the group-specific impact of these proxy variables on the take-up of microfinance loans. This would be a test to show if various household groups have different demands for flexibility in MFI loans. In the second part, I stratify the population according to education, and in the third part, I use a measure of risk averseness to divide the population into groups. In addition, as column 1 in Table 6 shows, most of the households have either zero or 19

one MFI loan. Because this is the case, the models below can approximately be thought of as linear probability models, in which the coefficients have the interpretation that if the independent variable increases by 1, the probability of taking out a microfinance loan increases by the number given by the coefficient. Moreover, all standard errors in the regressions are cluster-robust standard errors, in which the neighborhoods were used as the clusters. As mentioned briefly above, this is to account for the fact that randomization took place at the neighborhood level and not the household level. 5.1 General Analysis For a general analysis of whether there is demand for flexibility in the population, I use the regression model of the following form: y i = α + γ Controls i + β 1 t i + β 2 F lex i + β 3 (t i F lex i ) + ɛ i Here, y i is the dependent variable given by the number of MFI loans outstanding in the household (No mfi), t i is an indicator variable for whether the household is in a treated area, and F lex i is the proxy variable of interest. I am most interested in the coefficient on the interaction term, β 3, which measures the impact of demand for flexibility on takeup of microcredit in the treatment group. In other words, this coefficient measures how households with higher demand for flexibility reacted to increased microcredit access. Table 7a presents the results of the above model for the entire sample, for those proxy variables related to wage, while Table 7b presents the results of the above model for business owners. Although not all of the proxy measures provide statistical significance, results still point to some evidence that there is demand for flexibility, especially among business owners. First, looking at Table 7a, the first column uses the Workwage variable, which is an indicator variable for whether the household has at least one person working for a wage. The β 3 coefficient on the interaction term here is 0.0193, which suggests that having at least one person working for a wage increases the likelihood of the household taking out a microfinance loan by about 1.9%. However, the coefficient here is not statistically significant. Looking at the second column in Table 7a, the model uses the variable Wagetot, measuring the total number of people in the household working for a wage, as the proxy 20

variable. The β 3 coefficient here is 0.0254, which means that if you increase the number of people working for a wage in the household by 1, it is 2.5% more likely to take out a MFI loan. Although still not statistically significant at the 10% level, the β 3 coefficient on the interaction term has a t-statistic of 1.3, suggesting a positive relationship between having more people working for a wage in the household and the number of MFI loans taken out. Providing a more concrete evidence is the fact that if we look at only the business owners from the sample, as in the second column of Table 7b, the same coefficient using Wagetot becomes statistically significant at the 10% level. The coefficient also becomes larger, to 0.0802, which means that for the same increase of 1 in the total number of people working for a wage, a business-owning household is about 8% more likely to take out a MFI loan. This is a 5.5% increase from the number using the entire sample above. Even when you look at the first column of Table 7b, you can see that the t- statistic on the interaction term between treatment and Workwage (indicator variable on whether there is at least one person working for a wage) is 1.4, which is very close to becoming statistically significant at the 10% level. These results are consistent with the observation that having a wage paying job would provide a more stable source of income, enabling the household to take out an inflexible loan. Now, looking at the sales volatility measures from columns 3 through 6 in Table 7b, we can see that the β 3 coefficients on the interaction term between treatment and sales volatility measures are all negative. Especially for the the Sales dif measure, which is the difference between sales in the best and the worst month, the coefficient was statistically significant at the 5% level. The size of the coefficient is 1.1E 06, and to put this number into more perspective, I use the quartile numbers of the Sales dif variable to see how the probability changes with jumps in quartiles. From column 6 of Table 6, the 25% quartile is given by 700, the 50% quartile is given by 2500, and the 75% quartile is given by 6000. Thus, if sales volatility of a household business increases from the 25% quartile to 50%, it is about 0.2% less likely to take out a microfinance loan, and if sales volatility increases from 50% to 75%, it is about 0.4% less likely to use microcredit. Moreover, because of the high standard deviation number, one standard deviation increase in sales volatility suggests that the household is 21.4% less likely to take out a MFI loan. Taken 21

together, this suggests that households with high sales volatility measures reacted to the increased access in microcredit by taking out less MFI loans compared to others. This is consistent with the observation that higher sales volatility measures would mean more demand for flexibility, leading to less take-up of inflexible loans, and provides evidence that there is a level of general demand for more flexibility in microcredit. 5.2 Education Having gone through the general analysis, I now attempt to stratify the population according to education. To do this, I introduce a triple interaction term, between treatment, education, and the proxy variable of interest, to see if educational disparities result in different demands for flexibility. Thus, the regression model is now of the form: y i = α + γ Controls i + β 1 t i + β 2 Edu i + β 3 F lex i + β 4 (t i Edu i ) + β 5 (t i F lex i ) + β 6 (Edu i F lex i ) + β 7 (t i Edu i F lex i ) + ɛ i Here, the variables are the same as in the general analysis model above, and Edu i is the variable that measures the maximum level of education attained by someone in the household. The main coefficients of interest are β 5 and β 7, where β 5 measures the impact of demand for flexibility for the lower-education households, while β 7 measures the same impact for higher-education households. Table 8a summarizes the result for the business owners in the sample. I have chosen to focus on the business owners because they seem to be the ones who are more responsive to demand for flexibility and thus, are more interesting for the purposes of this paper. First, looking at columns 1 and 2 which correspond to variables related to wage, the sign of the β 5 coefficient is negative and the sign of the β 7 coefficient is positive, although in both cases, none of the coefficients are statistically significant. However, the signs point to the observation that there is some difference in the way low- and high-education households are behaving with regards to flexibility demand; for the loweducation households, working for a wage has a negative relationship with taking out more microfinance loans, while for the high-education households, working for a wage has a positive relationship. In other words, low-education households are acting in the 22

opposite way of what was predicted in the general analysis above, in the sense that those without wage jobs are taking up more microcredit. This difference in behavior still persists even if sales volatility measures are used as proxies of demand for flexibility. If you look at columns 3, 5, and 6 in Table 8a, you can see that the sign of the β 5 coefficient is positive and the sign of the β 7 coefficient is negative. For Sales dif in column 3, both of these coefficients are statistically significant, with β 5 being statistically significant at the 10% level and β 7 at the 5% level. Again, the observation is the same as before with variables related to wage; low-education households are acting in the opposite way of what was predicted in the general analysis, in that increasing sales volatility is associated with more take-up of microfinance loans. The only place where the coefficient changes signs is in column 4, which uses the Week dif variable, but the coefficients are not statistically significant for this regression. Looking more closely at the Sales dif variable in column 3, consider the extreme case of a household with a maximum education level of 0. Again using numbers from column 6 of Table 6, we can see that for such a household, increasing sales volatility from the 25% quartile to the 50% increases the probability of taking out a MFI loan by 2.9%. If a household moves from 50% quartile to the 75%, it is 5.6% more likely to take out a MFI loan. Now, consider a household at the 90% quartile of education. From column 2 of Table 6, you can see that such a household has a maximum education level of 16. Here, the triple interaction term comes into the equation, and you would need to use both β 5 and β 7 coefficients to calculate the effect on the dependent variable. If you do this for an increase of sales volatility from the 25% quartile to the 50%, you can see that the probability of taking out a MFI loan decreases by about 1.2%. For an increase from the 50% quartile to the 75%, the household becomes 2.2% less likely to take out microcredit. In addition to the above analysis, even though the coefficients on Week dif variable is not statistically significant, I still delved deeper into the model to explain what might be going on with the signs being switched. In one such attempt, I included both Sales dif and Week dif variables in the same regression. The results, presented in Table 8b, show that the Sales dif variable is the dominant force at work, as both coefficients on the Sales dif variable are statistically significant, compared to the coefficients on 23

the Week dif variable, which are not significant. This result hints that households are weighing the Sales dif measure more heavily than the Week dif measure in choosing whether to take out a microfinance loan. Moreover, one explanation of this change in signs might be that we are measuring a different time horizon with the two variables. More specifically, the Week dif variable is measuring sales volatility within a week, since we are using day sales and week sales numbers to calculate the discrepancy; and the Sales dif variable is measuring sales volatility within a year, since we are taking the difference between sales in the best and the worst months in a year. Since microfinance loans usually have a weekly repayment schedule, sales volatility in a week might not mean as much as sales volatility over a longer time horizon. In addition, since low-education households are less likely to be financially sophisticated, they might be susceptible to errors in their perception of risk, in the sense that they might be overweighing shortterm volatility compared to the more important long-term volatility in sales. Therefore, this might cause the low-education households to react more to short-term volatility measures when they decide to take out a loan, leading to more demand for flexibility, and thus, less take-up of inflexible loans, for those with more short-term volatility. Now, going back to the theoretical models and using education as a proxy for probability of being tempted and access to more outside loan options, the difference in signs of the coefficients seems to make sense. Because low-education households are more likely to be tempted and have access to less outside loan options, you would expect these people to take out more microfinance loans. Moreover, in the case of the low-education households, lack of wage jobs and increasing sales volatility may be proxies of demand for loans in general because more volatile sources of income lead to more needs to smooth out costs and consumption, and lack of access to outside loans force them to turn to MFI loans for help. Thus, the treatment effect of increasing access to microcredit is that low-education households with more income volatility actually take out more MFI loans. On the other hand, since the high-education households are less likely to be tempted and have access to more outside loan options, income volatility is acting as a proxy of demand for flexibility in loans; and thus, those with high income volatility are responding to the increased microcredit access by not taking out MFI loans. This suggests that there is demand for more flexibility in microfinance loans among the high-education households. 24

Moreover, because lack of access to substitutes for microcredit may mean that sales volatility and wage measures are not working as proxies of demand for flexibility in loans, we cannot conclude for sure that low-education households are not demanding more flexibility in loans. For the low-education households, they might not have a choice but to take out MFI loans because these are the best sources of financial instruments that they can use to smooth out consumption and costs. Thus, they might be forced to reluctantly take out MFI loans even though ideally, they would like more flexibility to account for the unpredictability in income. In this sense, we are not able to measure demand for flexibility for the low-education households using the variables at hand in the model above. 5.3 Risk Averseness in this section, I attempt to see if disparities in measures of risk averseness of households result in different demands for flexibility. I again use a triple interaction term, between treatment, measure of risk averseness, and the proxy variable of interest, to set up the following regression model: y i = α + γ Controls i + β 1 t i + β 2 Risk i + β 3 F lex i + β 4 (t i Risk i ) + β 5 (t i F lex i ) + β 6 (Risk i F lex i ) + β 7 (t i Risk i F lex i ) + ɛ i Here, the variables previously used are the same as in the two models above, and Risk i is the variable that measures the risk averseness of a household, created according to the description in section 4.3. The main coefficients of interest in this model are the same as in the education model above, β 5 and β 7. β 5 measures the impact of demand for flexibility for the less risk-averse households, while β 7 measures the same impact for highly risk-averse households. Table 9 summarizes the results for the business owners in the sample. Again, I chose to focus on the business owners because they seem to be more responsive to demand for flexibility. The table shows mixed results, as most of the coefficients of interest are not statistically significant, and there is no real consistency even in the signs of the coefficients. However, it is worth mentioning that in column 5, the coefficients of interest 25

are both statistically significant at the 10% level and have the signs that are consistent with the prediction from section 3.3. More specifically, β 5 is a positive number, which hints that people who are not as risk-averse tend to take out more MFI loans with increasing sales volatility; and β 7 is a negative number, which means that people who are highly risk-averse take out less MFI loans with increasing sales volatility. Thus, people who are less risk-averse seem to be using MFI loans as another way of smoothing out costs and consumption in the face of high income volatility, while more risk-averse people shy away from taking out inflexible, microfinance loans for the fear of not being able to pay back the amount. More specifically, consider the two cases in which the household has a value of 0 for the measure of risk averseness, and a 2 for the same measure. From column 8 of Table 6, we can get the 25%, 50%, and 75% quartile numbers for the Month dif sales volatility measure. Using these, we can see that for a household with a risk averseness value of 0, an increase in sales volatility from 25% quartile to the 50% increases the probability of taking out a MFI loan by 0.5%; and an increase from 50% to 75% means that the household is 1.7% more likely to use microcredit. However, for a household with a risk averseness value of 2, an increase from the 25% quartile to 50% decreases the likelihood of the household taking out a MFI loan by 0.1%; and an increase from 50% to 75% decreases the likelihood by 0.2%. As the numbers suggest, the effects on the dependent variable are very small around the median, providing more evidence that dividing the households according to this measure of risk averseness is not having a great impact, even though the coefficients are statistically significant in the direction predicted. In addition, because the constructed measure of risk averseness is not perfect and probably has other correlations with take-up of MFI loans, one statistically significant result does not provide conclusive evidence that this relationship holds. The fact that the signs of the coefficients are not consistent cast doubts as to whether there is such a relationship, and more direct experiments involving risk averseness would need to be done before a definite conclusion can be reached. 26

6 Conclusion These findings suggest that there is some demand for more flexibility in microfinance loans, especially among the business owners. Households without wage jobs or with higher sales volatility were less likely to take on microfinance loans, which suggest that they may have found MFI loans to be too costly due to the rigid structure. Moreover, this effect differed for the low-education and the high-education groups; business owners with low education were actually more likely to take out microfinance loans with higher sales volatility, while those with high education were less likely to do so. This possibly alludes to the fact that higher education people may have more access to better financial instruments, such as commercial banks, and better understand their risks. For the loweducation people, microfinance loans may be the best loan option that they have; and because higher sales volatility also may mean the need for more loans to cover their costs during the bad times, they may be turning to MFI loans to help them smooth consumption and pay costs throughout the year. Furthermore, an analysis stratifying the population according to a measure of risk averseness proved inconclusive, but there was still some evidence that the less risk-averse households are the ones taking on more MFI loans, even in the face of high income volatility. Thus, in this sense, microfinance institutions may actually be missing out on the people who may turn out to be the most profitable: financially sophisticated, risk-averse business owners. However, there are limitations caused by the nature of the data used in this analysis. Because the survey was taken 15 to 18 months after the treatment, the variables used do not correctly measure the household characteristics at the time of the decision to take out microfinance loans. There is also a lot of measurement error within the data for the sales volatility measures because the numbers are as reported by the people, and thus, may not be accurate. Measurement errors drive the regression coefficients toward zero (attenuation toward zero), and thus, make it more difficult to conclude that a coefficient is statistically significant. Moreover, the survey was not tailored to look specifically at demand for flexibility, meaning that the variables that I have used are proxies at best. Thus, there may be hidden impacts besides demand for flexibility that are driving the regression results. Similarly for other variables, such as a measure of risk averseness, because the variable is not a perfect measure, it is impossible to conclude that these 27

characteristics are not creating differences in demand for flexibility. To gain a more accurate sense of whether increasing flexibility will increase takeup of microfinance loans, more randomized experiments similar to the study done by Field et al. (2010), or the study proposed by Pearlman (2010), are needed, in which the treatment group gains access to loans with more flexibility. Moreover, including explicit questions behind the reasons for not taking out microfinance loans may also help to identify whether MFI loans should be tailored more to meet the different needs of the poor. Direct measures of the variables in question, such as risk averseness and probability of temptation, may also help to clarify the reasons behind the different responses of the various groups in the population. Thus, while this paper shows that there exists some demand for flexibility, more follow-up studies may help to further identify the characteristics of people who demand the most flexibility, and whether changing the structure of the MFI loans would also make sense from the point of view of the institutions. 28

References Armendariz, B., and J. Morduch (2000): Microfinance beyond group lending, Economics of Transition, 8(2), 401 420. Banerjee, A. (2000): The two poverties, Nordic Journal of Political Economy, 26, 129 141. Banerjee, A., E. Duflo, R. Glennerster, and C. Kinnan (2010): The miracle of microfinance? Evidence from a randomized evaluation, Working paper. Collins, D., J. Morduch, S. Rutherford, and O. Ruthven (2009): Portfolios of the poor: How the world s poor live on $2 a day. Princeton University Press. Dercon, S. (2002): Income risk, coping strategies, and safety nets, The World Bank Research Observer, 17(2), 141 166. Field, E., and R. Pande (2008): Repayment frequency and default in micro-finance: Evidence from India, Journal of European Economic Association Paper and Proceedings, 6(2-3), 501 550. Field, E., R. Pande, and J. Papp (2010): Does microfinance repayment flexibility affect entrepreneurial behavior and loan default?, Working paper. Jain, S., and G. Mansuri (2003): A little at a time: The use of regularly scheduled repayments in microfinance programs, Journal of Development Economics, 72, 253 279. Kaboski, J. P., and R. M. Townsend (2005): Policies and impact: An analysis of village-level microfinance institutions, Journal of the European Economic Association, 3(1), 1 50. Karlan, D., J. Morduch, and S. Mullainathan (2010): Take-up: Why microfinance take-up rates are low and why it matters, Financial Access Initiative research framing note. 29

Karlan, D., and S. Mullainathan (2006): Is microfinance too rigid?, Centre for Microfinance and Innovations for Poverty Action research note. Morduch, J. (1999): The Microfinance Promise, Journal of Economic Literature, 37, 1569 1614. Pearlman, S. (2010): Flexibility matters: Do more rigid loan contracts reduce demand for microfinance?, Working paper. Reed, L. R. (2011): State of the Microcredit Summit Campaign Report, Microcredit Summit Campaign report. Townsend, R. M. (1994): Risk and insurance in village India, Econometrica, 62, 539 591. Udry, C. (1990): Credit markets in northern Nigeria: Credit as insurance in a rural economy, The World Bank Economic Review, 4(3), 251 269. 30

A Figures Figure 1: Sales Volatility and Probability of Being Tempted 31

B Tables Table 1: Treatment-Control Balance, Baseline Neighborhood-level Characteristics (1) (2) (3) (4) (5) (6) Number of Maximum Consumption Number of Outstanding Number of households education per capita loans debt businesses Treatment -3.385 -.0447 59.92 -.0413-2990 -.0125 [29.78] [.221] [48] [.1431] [3522] [.0355] Control Mean 268*** 9.468*** 923.2*** 1.823*** 3.7e+04***.3226*** Control Std. Dev. [48.71] [.3614] [70.92] [.235] [5986] [.0535] N 104 104 104 104 104 104 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 32

Table 2: Treatment-Control Balance, Endline Household-level Characteristics (1) (2) (3) (4) (5) (6) (7) Maximum At least one Total number Total wage Sales Owns old Adult education person works of people earnings per volatility, business equivalents for a wage working for a adult difference wage equivalents between best and worst month Treatment -.0814.0235.018-63 -1.3e+04.007 -.0089 [.2245] [.023] [.054] [75.69] [1.2e+04] [.0213] [.0656] Control Mean 11***.6858*** 1.16*** 549*** 3.4e+04.2953*** 4.677*** Control Std. Dev. [.3531] [.0378] [.0857] [120.6] [2.5e+04] [.0321] [.0988] N 6783 6798 6798 6798 1101 6744 6798 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 33

Table 3: Access to Microfinance Loans (1) Borrows from MFI (2) MFI borrowing (Rs.) Treatment.1017*** 1541*** [.0305] [497.4] Control Mean.0803* 814.1 Control Std. Dev. [.0479] [761.6] N 6780 6780 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. Table 4: Education and Probability of Temptation Probabilty of being tempted Education -.002*** [7.2e-04] Constant.0591*** [.0075] N 1375 Note: Robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 34

Table 5: Determining Loan Type Multinomial Probit Results (1) No loans (2) MFI loans only (3) Formal loans only (4) Informal and MFI loans (5) Informal and formal loans (6) MFI and formal loans (7) All loans Consumption per adult equivalents 8.0e-05** 9.6e-05** 1.3e-04*** 9.80E-06 1.1e-04*** 1.1e-04*** 8.5e-05*** [3.3e-05] [3.8e-05] [3.1e-05] [4.1e-05] [3.1e-05] [3.7e-05] [3.2e-05] Owns old business.2365***.6861***.583***.5761***.4806***.9105***.8329*** [.0643] [.0655] [.0678] [.0721] [.0678] [.0815] [.0707] Amount of pre-existing loans -8.6e-06*** -1.6e-05*** 6.4e-07* 1.80E-07 2.0e-06*** 5.60E-08 1.4e-06*** [1.9e-06] [3.6e-06] [3.8e-07] [5.2e-07] [3.3e-07] [4.4e-07] [3.0e-07] Had past MFI loans -.2667**.4951*** 0.1345.5542***.2929***.3773***.4965*** [.1105] [.1338] [.0952] [.1102] [.0672] [.1201] [.1177] Education.0341*** 0.0033.0358*** -0.0094.0143*.0183* -0.0121 [.01] [.0115] [.0085] [.0079] [.0073] [.0101] [.0086] Constant -1.188** -1.179-1.332** -1.048-0.7107-1.69** -1.239 [.4736] [.7858] [.5886] [.6488] [.5896] [.749] [.8106] N 6691 6691 6691 6691 6691 6691 6691 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 35

Table 6: Summary Statistics of Key Variables (1) Number of MFI loans outstanding, no mfi (2) Education (3) Risk averseness (4) At least one person works for a wage (5) Total number of people working for a wage (6) Sales volatility, sales dif (7) Sales volatility, week dif (8) Sales volatility, month dif (9) Sales volatility, mon2 dif Number of Obs. 6780 6783 6791 6798 6798 1101 2160 1580 1602 Mean 0.3258112 10.8797 1.07186 0.721683 1.187261 13783.2 1548.112 8087.502 10033.06 Standard Deviation 0.6735343 3.816837 0.78451 0.448203 1.056351 194942.1 7871.671 45877.35 45822.24 10% Percentile 0 6 0 0 0 200 0 41.66667 62.5 25% Percentile 0 9 0 0 0 700 0 238.2143 312.5 50% Percentile 0 11 1 1 1 2500 140 966.1905 1318.75 75% Percentile 0 13 2 1 2 6000 790 3592.381 5568.333 90% Percentile 1 16 2 1 3 15000 2922.5 11084.82 15905 36

Table 7a: General Analysis Entire Sample (1) At least one person works for a wage (2) Total number of people working for a wage Consumption per adults equivalents -5.10E-06-2.10E-06 [7.0e-06] [6.9e-06] Owns old business.2263***.2423*** [.0319] [.0322] Amount of pre-existing loans -1.10E-07-1.3e-07* [6.8e-08] [7.1e-08] Had past MFI loans.1574***.1556*** [.0435] [.0434] Has pre-existing flexible loans -0.0211-0.0202 [.0253] [.0252] Education -.0069** -.0083*** [.0027] [.0028] Treatment.1378**.1211** [.0561] [.0489] Workwage 0.0401 [.0322] Treatment Workwage 0.0193 [.0441] Wagetot.0431*** [.0145] Treatment Wagetot 0.0254 [.0201] Constant 0.2401 0.2353 [.248] [.2466] N 6691 6691 R-squared 0.0525 0.059 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 37

Table 7b: General Analysis Business Owners (1) At least one person works for a wage (2) Total number of people working for a wage (3) Sales volatility, sales dif (4) Sales volatility, week dif (5) Sales volatility, month dif (6) Sales volatility, mon2 dif Consumption per adult equivalents -3.40E-06-4.00E-08-2.8e-05** -4.80E-06-8.70E-06-4.40E-06 [1.4e-05] [1.3e-05] [1.3e-05] [1.5e-05] [1.8e-05] [1.8e-05] Owns old business -8.40E-04 0.0027 0.0986-6.80E-04 0.0218-0.0098 [.051] [.0524] [.0987] [.0501] [.0606] [.0624] Amount of pre-existing loans -2.2e-07** -2.2e-07** -5.60E-08-2.2e-07** -2.00E-07-2.2e-07* [9.4e-08] [9.3e-08] [2.4e-07] [1.0e-07] [1.3e-07] [1.3e-07] Had past MFI loans 0.0904 0.0876-0.0221 0.0669-0.0376-0.0321 [.0594] [.0586] [.0802] [.063] [.0655] [.0648] Has pre-existing flexible loans -0.0506-0.0513-0.1035-0.0517 0.0108 0.0106 [.05] [.0492] [.0724] [.0545] [.0716] [.0707] Education -0.0054-0.0062-0.0077-0.0035-0.0046-0.0039 [.0051] [.0051] [.0067] [.0052] [.0059] [.0058] Treatment.1441*.135*.2724***.2165***.2492***.2563*** [.0744] [.0726] [.0992] [.077] [.0902] [.0891] Workwage 0.0695 [.0565] Treatment Workwage 0.1121 [.0785] Wagetot.0529* [.0286] Treatment Wagetot.0802* [.043] Sales dif -2.00E-08 [2.5e-08] Treatment Sales dif -1.1e-06** [4.6e-07] Week dif -4.10E-07 [1.1e-06] Treatment Week dif -7.10E-07 [4.1e-06] Month dif -1.20E-07 [4.3e-07] Treatment Month dif -1.90E-07 [5.9e-07] Mon2 dif 2.10E-07 [4.3e-07] Treatment Mon2 dif -5.60E-07 [7.2e-07] Constant 0.4196 0.4231 0.9435 0.4109.972* 1.011* [.4032] [.3974] [.6353] [.4246] [.5762] [.5582] N 2289 2289 1084 2128 1555 1577 R-squared 0.0403 0.0496 0.0387 0.0315 0.0392 0.0399 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 38

Table 8a: Stratification according to Education Business Owners (1) At least one person works for a wage (2) Total number of people working for a wage (3) Sales volatility, sales dif (4) Sales volatility, week dif (5) Sales volatility, month dif (6) Sales volatility, mon2 dif Consumption per adult equivalents -2.70E-06 3.80E-07-2.8e-05** -4.60E-06-6.40E-06-3.20E-06 [1.4e-05] [1.3e-05] [1.3e-05] [1.5e-05] [1.8e-05] [1.8e-05] Owns old business -0.0035 0.0028 0.098-6.30E-04 0.0228-0.0088 [.0505] [.0527] [.0988] [.05] [.0598] [.0618] Amount of pre-existing loans -2.1e-07** -2.2e-07** -7.40E-08-2.2e-07** -2.10E-07-2.20E-07 [9.4e-08] [9.4e-08] [2.5e-07] [1.0e-07] [1.3e-07] [1.3e-07] Had past MFI loans 0.0944 0.0903-0.024 0.0667-0.0385-0.0321 [.0592] [.0583] [.0806] [.0632] [.0658] [.0647] Has pre-existing flexible loans -0.0489-0.0511-0.1071-0.0512 0.0112 0.0109 [.0504] [.0494] [.0725] [.0544] [.0713] [.0705] Treatment 0.1064 0.0642 0.196 0.1036 0.1368 0.142 [.1582] [.1611] [.2105] [.1474] [.1663] [.1634] Education 3.70E-04-0.0064-0.0097-0.0081-0.0094-0.0087 [.0092] [.0098] [.0111] [.009] [.0098] [.0095] Workwage.3714* [.1899] Wagetot 0.1587 [.105] Sales dif -4.30E-06 [4.1e-06] Week dif 1.40E-05 [1.0e-05] Month dif -1.00E-06 [1.2e-06] Mon2 dif 3.10E-07 [1.7e-06] Treatment Education 0.0033 0.0065 0.0066 0.0102 0.0098 0.0102 [.0123] [.0122] [.0142] [.0105] [.0114] [.0111] Treatment Flexibility -0.1147-0.0088 1.6e-05* -9.70E-06 2.00E-06 5.90E-07 [.2511] [.1341] [8.2e-06] [1.9e-05] [1.3e-06] [1.9e-06] Education Flexibility -.0261* -0.0087 3.60E-07-1.10E-06 6.20E-08-6.50E-09 [.0144] [.008] [3.5e-07] [7.5e-07] [9.0e-08] [1.2e-07] Treatment Education Flexibility 0.0191 0.0071-1.4e-06** 6.50E-07-1.60E-07-1.10E-07 [.0196] [.0108] [6.7e-07] [1.5e-06] [1.0e-07] [1.4e-07] Constant 0.3597 0.4166 0.9866 0.4509 1.018* 1.057* [.4175] [.4062] [.6591] [.4298] [.591] [.5716] N 2289 2289 1084 2128 1555 1577 R-squared 0.0426 0.0509 0.0409 0.0327 0.0401 0.0406 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 39

Table 8b: Race between Sales Volatility Measures Sales volatility, sales dif and week dif Consumption per adult equivalents -3.4e-05** [1.4e-05] Owns old business 0.1073 [.0973] Amount of pre-existing loans 4.70E-09 [2.9e-07] Had past MFI loans -0.0451 [.0816] Has pre-existing flexible loans -0.1205 [.075] Treatment 0.2342 [.2093] Education -0.006 [.0096] Sales dif -6.90E-06 [5.1e-06] Week dif 4.30E-05 [4.5e-05] Treatment Education 0.0024 [.0143] Treatment Sales dif 3.4e-05** [1.3e-05] Treatment Week dif -7.30E-05 [5.7e-05] Education Sales dif 5.70E-07 [4.3e-07] Education Week dif -3.30E-06 [3.8e-06] Treatment Education Sales dif -2.8e-06*** [1.0e-06] Treatment Education Week dif 6.00E-06 [4.8e-06] Constant 0.9853 [.6903] N 1037 R-squared 0.043 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 40

Table 9: Stratification according to Risk Averseness Business Owners (1) At least one person works for a wage (2) Total number of people working for a wage (3) Sales volatility, sales dif (4) Sales volatility, week dif (5) Sales volatility, month dif (6) Sales volatility, mon2 dif Consumption per adult equivalents -1.50E-05-1.10E-05-4.4e-05*** -1.60E-05-1.80E-05-1.50E-05 [1.4e-05] [1.3e-05] [1.4e-05] [1.5e-05] [1.8e-05] [1.8e-05] Owns old business -0.0123-0.0091 0.0773-0.013 0.025-0.0057 [.0519] [.0532] [.0989] [.0515] [.0617] [.0637] Amount of pre-existing loans -2.9e-07*** -2.9e-07*** -2.00E-07-2.9e-07** -2.5e-07* -2.6e-07* [1.0e-07] [1.0e-07] [2.5e-07] [1.1e-07] [1.4e-07] [1.4e-07] Had past MFI loans 0.0762 0.0738-0.0299 0.0527-0.0508-0.0423 [.0582] [.0576] [.0768] [.0617] [.0643] [.0632] Has pre-existing flexible loans -0.0393-0.0411-0.0644-0.0413 0.0273 0.0313 [.0505] [.0498] [.0702] [.0555] [.071] [.0717] Education -.01** -.0109** -.0156** -.0084* -0.0084-0.0077 [.005] [.005] [.0072] [.0051] [.0059] [.0059] Treatment 0.0728 0.0542.3468***.1485**.1963**.2236** [.0902] [.0841] [.101] [.073] [.0848] [.0851] Risk averseness.1285***.1194***.2374***.1335***.1339***.1464*** [.0419] [.0414] [.0636] [.0394] [.0454] [.0467] Workwage 0.0924 [.0831] Wagetot 0.0601 [.045] Sales dif 1.00E-05 [7.3e-06] Week dif -1.90E-06 [1.5e-05] Month dif -3.70E-06 [3.3e-06] Mon2 dif 3.50E-06 [3.0e-06] Treatment Risk averseness 0.0511 0.0608-0.0508 0.0438 0.0316 0.023 [.0602] [.058] [.0792] [.0489] [.0598] [.061] Treatment Flexibility 0.1092 0.0923-1.00E-05 1.80E-05 6.3e-06* -8.80E-07 [.1132] [.0716] [7.4e-06] [2.4e-05] [3.5e-06] [3.0e-06] Risk averseness Flexibility -0.0285-0.0095-5.20E-06 3.60E-07 1.70E-06-1.80E-06 [.0508] [.0288] [3.7e-06] [7.5e-06] [1.7e-06] [1.5e-06] Treatment Risk averseness Flexibility 0.0052-0.0099 4.40E-06-1.10E-05-3.5e-06* -9.60E-08 [.0871] [.0465] [3.7e-06] [1.2e-05] [1.8e-06] [1.6e-06] Constant 0.3802 0.3858 0.8242 0.3521 0.9197 0.9179 [.4033] [.3993] [.6533] [.4312] [.595] [.5757] N 2289 2289 1084 2128 1555 1577 R-squared 0.0574 0.0662 0.0686 0.0507 0.0565 0.057 Note: Cluster-robust standard errors in brackets. * means statistically significant at 10%, ** means statistically significant at 5%, *** means statistically significant at 1%. 41