Microfinance Can Raise Incomes: Evidence from a Randomized Control Trial in China *

Transcription

1 Microfinance Can Raise Incomes: Evidence from a Randomized Control Trial in China * Shu Cai, Jinan University Albert Park, HKUST Sangui Wang, Renmin University of China 2017 Abstract This study evaluates the impact of a randomized control trial (RCT) in China that introduced externally funded village credit funds in poor, rural villages. In contrast to recent RCT-based studies that fail to find evidence of significant income increases from microfinance interventions, we find that the Chinese program significantly raises household incomes and reduces poverty. We assess possible explanations as to why estimated impacts may be greater in China: less frequent repayment schedule, lower interest rate, and greater potential returns to off-farm employment opportunities that are credit-constrained. Keywords: microfinance, program evaluations, randomized control trial JEL Codes: D12, D22, G21, I32, O16 * We thank Bob Gregory and Joe Kaboski for helpful comments, as well as participants at the 2015 Chinese Economist Society North America Conference at Ann Arbor, US, the 3 rd Asian Bureau of Finance and Economic Research Annual Conference at Singapore, the Micro-enterprises and SMEs workshop at U Michigan, the 2016 China Meeting of Econometric Society at SWUFE, and seminars at UCSD, HKUST, Korea University, and Wuhan University. 1

2 1. Introduction There is growing evidence on evaluating the impacts of microcredit, from earlier studies employing exogenous variations in program participation (Pitt and Khandker, 1998; Coleman, 1999; Kaboski and Townsend, 2005, 2012) to more recent developments in control trials, randomized on either the regional level (Banerjee et al., 2015; Tarozzi et al., 2015; Angelucci et al., 2015; Attanasio et al., 2015; Crepon et al., 2015) or the individual level (Augsburg et al., 2015; Karlan and Zinman, 2010, 2011). However, the empirical evidence on the impact of microcredit is inconclusive. The microcredit, especially the Grameen model, has been viewed as a great success in fighting poverty (Morduch, 2005). However, a recent symposium of six randomized control trials, published in American Economic Journal: Applied Economics (2015), finds no strong evidence that microfinance significantly increases household income or consumption (Banerjee, Karlan, and Zinman, 2015). These findings cast doubt on the role of microfinance in raising income. In this study, we provides the first systematic evaluation of what is possibly the world s largest village banking program in China. We undertake a randomized control trial similar to those undertaken in the symposium referenced above and find substantial increases in income from both self-employment activity and wage employment activity in areas where the program has been introduced. More specifically, our evaluation suggests that the village banking program increases the likelihood of access to credit for production by 15%. Households from treatment villages plant more cash crops and invest more in inputs of feedstuffs in animal husbandry than those in the control villages. Their revenues from crop farming and animal husbandry increase in response to the program, as well as their profits. Of special interest, however, is the finding that the program also increases the employment income of households in the treatment villages. Further analyses suggest that households allocate more of their labor force to employment activity outside their home township, and particularly to employment activity outside their home province. Meanwhile, on average households allocate less labor to employment activity within their home township. The labor force allocated to self-employment also increases as a result of the program, but the estimated impact is insignificant. The impacts on labor employment imply significant costs for migration to foreign provinces, which may be constrained by credit. Since the wage rate in foreign provinces as a migration destination (usually coastal cities) is much higher than the wage rate in rural villages or townships, we find significant increases in employment income as a result of the program. Eventually, the program significantly raises household income and reduces poverty. Expenses on consumption increase as well, particularly for 2

3 consumption of durable goods. The improvement in quality of life is confirmed by subjective well-being, including self-assessed quality of life and satisfaction with life. We proposed several explanations for the substantial impacts of the microcredit program on raising income found in China: (1) Pre-program access to formal loans is low. (2) The take-up rate is relatively higher in the experiment in China when compared with other RTC studies. (3) The one-time rather than frequent repayment may also contribute to the substantial income impacts. (4) Finally, and probably most importantly, we found a significantly large increase in wage employment earnings in addition to increased selfemployment income. This may be because of exceptional demand for migration labor and pre-existing liquidity constraints on migration in China. The study is closely related to substantial literature on evaluating the impact of microcredit programs around the world. Early efforts to identify the impact of microcredit try to employ exogenous variations in access to microcredit to address the potential bias caused by endogenous placement and self-selection into the program (Morduch, 2005). For instance, Pitt and Khandker (1998) estimated the impact of participation in group-based credit programs in Bangladesh by comparing eligible households and ineligible households from both program villages and non-program villages. They found significant impacts on household expenditures, assets, and schooling of children. Coleman (1999) compared outcomes of village bank members in old program villages with members in new program villages who hadn t yet receive loans in northeast Thailand. He found that the impacts on business creation, self-employment, and assets are positive but insignificant. Program placement is not random in these two studies. Although both studies controlled for village fixed effects in their regression analysis, they could not rule out confoundedness of time-varied unobservable village characteristics that affected both placement and outcomes (Morduch, 1998). Kaboski & Townsend (2012) instead used variation in the number of households in the village as the instrument for credit access to each household, since the amount of credit was fixed in every village in their study. They found that the village fund had increased agricultural investments, income, and consumption two years after the intervention. More recently, researchers have turned to the use of randomized control trials to evaluate the impacts of microcredit programs. Banerjee et al. (2015) found that access to microcredit extended existing businesses and increased profit. However, this didn t lead to increases in consumption. Instead, households reduced expenditures on temptation goods and increased investments in assets. Crepon et al. (2015) also identified a boost in the profits of businesses. However, employment earnings declined. Overall, income and 3

4 consumption didn t change significantly. Similarly, they found that expenditures on festivals and celebrations decreased, while assets increased. Angelucci et al. (2015) found a positive impact on business extension, but impacts on business profits, income, consumption, and assets were not significant. Tarozzi et al. (2015) found that impacts on economic activities, income and consumption were not statistically significant. Attanasio et al. (2015) found significant impacts of access to credit on business creation and food consumption. Augsburg et al. (2015) found significant impacts on business creation. They found that self-employment increased, while employment activity decreased. Overall, they found no significant improvements in business profits or consumption. This paper complements previous studies by implementing a randomized control trial to evaluate the microcredit program in a country which has not yet been studied in the literature. Meanwhile, the scale of the program makes it a significant case study for research on microcredit. More importantly, we find that the extension of microcredit encourages households to seek employment farther away, and this effect, which has not been documented in studies of other programs in other countries, makes a major contribution to rises in income. The paper is organized as follows. In Section 2, we introduce the village banking program, the experimental design, the context of the surveyed villages before the program, and the data employed. Section 3 presents the empirical strategy for identification. Section 4 reports the estimated average intention-to-treat effects on credit, investment, income, consumption, etc. In Section 5, we conduct a meta-analysis by comparing our study with other studies that have evaluated microcredit programs. Section 6 concludes. 2. The Program and Experimental Design Village Banking Program Credit markets are far less developed in rural China, especially in poor areas. Since 1990, the Chinese government has made many attempts to improve credit access for the poor by studying the experiences of other developing countries, including Grameen Bank in Bangladesh (Yunus, 2006). However, most of these efforts proved unsuccessful. In 2006, the Chinese government (the State Council Leading Group Office of Poverty Alleviation and Development and the Ministry of Finance) initiated the village banking program. Unlike previous anti-poverty programs in China, the village banking program targets households in poor villages instead of the villages themselves. The cross-guarantees 4

5 lead to high loan payments, making it possible for borrowing among the poor, who are most likely to experience credit rationing. Meanwhile, the program aims to build up self-managed organizations to provide local credit services sustainably. Members of village banks elect managers, formulate bylaws, conduct bookkeeping, and manage funds on their own. 1 Although the village banks are highly independent, they are under the supervision of county governments. Each program village received 150,000 yuan (US$24,000) from the national government. Additionally, membership fees paid by the households comprise a portion of the village bank funds. On average, the initial fund for a village bank is about 200,000 yuan (US$32,000). Program loans are required to be used for income-generating activities. The program was first implemented among 100 trial villages in 14 provinces in 2006 and was then extended to another 270 villages in By the end of 2013, the village banking program had been implemented in 19,400 villages in 1,407 counties of 28 provinces in China (The Yearbook of China s Poverty Alleviation and Development, 2014). This study conducts the first systematic evaluation of possibly the world s largest village banking program by using the randomized control trial method. 2 Experimental Design To evaluate the program at its early stage, the State Poverty Alleviation Office appoint a research team from the Renmin University of China leaded by one of the authors Sangui Wang to conduct a randomized intervention of village banking programs in poor villages in five provinces in China. The provinces were selected for the purpose of regional representativeness. Among the five provinces, Shandong province is in Eastern China; Henan and Hunan provinces are in Central China; and Sichuan and Gansu provinces are in Western China. In each province, two counties that had not previously implemented the village banking program previously were recommended by provincial officers. In each county, five officially recognized poor villages were recommended as candidates for the program. To ensure the program villages were selected randomly, the research team randomly chose three villages as 1 Each village bank consists the generally meeting of all members, an administrative council, and a supervisory board. The general meeting is the highest authority of the village bank. It elects and removes members in the administrative council and the supervisory board. The administrative council consists 3-5 person, including the chairman, accountant, cashier, etc. The administrative council is mainly in charge of disburse and recover loans. The supervisory board usually consists 3 person who are responsible for overseeing the operation of funds and the work of the administrative council. It is required that at least one person in the supervisory board should be from the poor households. 2 Early efforts in evaluating the Chinese microfinance programs include Park, Ren, and Wang (2001) and Park et al. (2003), in which they compare the efficacy of different microfinance programs and find nongovernmental programs perform better than governmental programs in terms of targeting, sustainability, and program benefits. 5

6 treatment subjects and left the other two villages untreated. Thirty households were randomly selected in each of the treatment and control villages. 3 Figure 1 illustrates the geographic location of the selected provinces and counties. All of the counties are geographically separated. Moreover, treatment villages are also separated from control villages in each county, with an average distance of 14.5 kilometers (not shown in Figure 1). It is thus unlikely for spillover effects between the treatment and control groups. In total, 1,500 households in 50 villages, comprising 10 counties, were surveyed at baseline in August 2010, before the initiation of the village banking program. The household survey collected detailed information on demographic characteristics and labor employment of each household member 4, as well as income (including income from crop farming, animal husbandry, small business, employment, etc.), consumption expenses, assets, and each loans borrowed from January 2009 to August 2010 for every household in the survey. 5 The survey also collected detailed information about the villages through village questionnaires, such as population, area of arable land, amounts of public spending from superior governments on various village projects, etc. After the baseline survey, the village banking program was gradually implemented in the treatment villages. 6 Two years after the baseline survey, in July 2012, the research team conducted the second wave of the survey by following up with the same households in both the treatment and control villages. In addition to the questions asked at baseline, the researchers asked questions about the village banking program in both the household and village questionnaires. 7 A total of 1,351 households were successfully re-interviewed, with an attrition rate of 9.9%, which is mainly due to household migration or temporary absence of all household member for reasons such as visiting relatives according to field observation. Since the treatment villages in one county of Hunan province began to extend loans very late, we dropped this county in the following analysis. Eventually, the analysis sample consisted of 1,234 households in 45 villages, 9 counties, and 5 provinces. Among 3 To obtain random sample of households in the village, stratified systematic sampling method is used. More specifically, when come to a village, the investigators asked the village officials to provide a list of natural villages (a subunit of administrative villages that exists spontaneously) ranked by per capita income. The investigators then randomly select one or two natural villages from the list according to weighted population. The investigator only need to select a second natural village when there are less than 50 households in the first randomly selected village. Given the randomly selected natural village(s), the investigators asked the village officials to provide a list of households in that natural village(s) ranked by per capita income. The investigators then selected 30 households from the household list by randomly systematic sampling method. 4 The survey collected information on days of employment, work place and earnings of each household member in the year before the survey. 5 For each loan, the survey collected information on the month the household borrowed the loan, the amount, source, interest rate, duration, and purpose of the loan, and conditions of collateral and guarantee. 6 The village banking programs are not anticipated by villages. In most counties, the village banking program started to deliver loans during the first half year of The only exception is one county in Hunan province, where all of the three treatment villages didn t start to deliver loans until one month before the second wave of the survey. 7 The second wave of the survey recorded all of the loans borrowed from August 2010 to July 2012 for each household surveyed. 6

7 them, 493 are in control villages, and 741 are in treatment villages. To ensure the comparability of the treatment and control groups, we examine the balance of attrition between the treatment and control villages in Section 2.5. Program Implementation and Loan Characteristics The county poverty alleviation office is in charge of initiation of village banking program in the treatment villages at early stage. They organize a meeting of all cadres in the village to introduce the village banking program, including its objection, rules, management, etc. The information are passed to the villagers by the village cadres via separate meetings in each nature village. Then, general meetings of all members in the administrative village are organized to explain the program, collect feedbacks, and elect members of a preparatory group. The preparatory group is in charge of drafting charters of the village bank, approving application of being village bank members, organizing the first general meeting of all member households to elect the administrative council and supervisory board. After that, the village bank is registered at the county civil affairs bureau, and the management of the village bank is handed over from the preparatory group to the administrative council. After formal registration, each treatment village received 150,000 yuan (US$24,000) from the national government and the village bank started to deliver loans. The maximum size of loans is typically 5,000 yuan, about 45% of average total annual household income at baseline. The duration of village bank loans is usually one year. The average annual nominal interest rate of loans borrowed from the village banks is 9.4%, while the average annual nominal interest rate is 10.8 % for loans borrowed from the Rural Credit Cooperatives with the same duration. In most program villages, the loans are paid in one installment and uncollateralized. The village banking program is usually organized in the form of group lending. The group size is 5 to 7 households. Members of the group share joint liability. 8 Unlike most other group lending practices, there is no requirement on regular meetings of group members. 9 In principle, loans are lent for the purpose of income generation. Only one member of a household can become a member of village banks. To be a member of a 8 If any member default, the other group members have obligation to repay the loans, and they are not allowed to borrow from the village bank until the loans are repaid. Exceptions are allowed for reasons such as disaster. Application of such default is reviewed by the administrative council and discussed on the general meeting of all members of the village bank. 9 As documented later, the geographic size of the villages usually are small and the relationship between village members are usually based on kinship. Even though the group members are not required to meet formally, they can easily be aware of activity of other members through informal interaction. 7

8 village bank, the representative of a household should be regular resident in the village, above 18 years old, and with ability to work. They also need to submit an application and pay a lump-sum membership fee if application is approved. The membership fee is usually 200 yuan (US$32). It can be deducted or waived for the poor households. The village banks does not provide saving service. The members can retrieve their membership fee (without interest) when they withdraw from the village bank and have no default loans. Village bank members apply loans from the administrative council. For group lending, the application need agreement (signature) of all the other members in the group. The administrative council then make decision of the application according to the stated loan size, duration, and purpose. Application from poor households and female members are given priority. If the application is approved, the applicant and the council sign a loan contract and the loans are delivered. By the time of the second survey, on average, 57% of the households in treatment village join the village banking program and 28% of the households take loans from the village banks. At the same time point, 70% of the village bank funds were lent out, and the repayment rate is 98%. 10,11 Contexts of the Villages The surveyed villages are relatively poor. In 2009, average annual income per capita was 3,559 yuan, while average annual income per capita in rural areas nationwide was 5,153 yuan (National Bureau of Statistics of China). About 46.8% of the households in our sample would be classified as poor households according to the national poverty line. 12 Almost all of the households conducted crop farming at baseline (96%), and 75% of the households conducted animal husbandry. While agricultural production is popular in these areas, small businesses and some form of employment activity are less frequent. Some 14% of the households engaged in some small business at baseline and 48% of the households had some employment activity. In particular, only 23% of the households had some member employed in foreign provinces at baseline. The average annual employment earning per capita is 1,210 yuan, and the average annual self-employment income per capita (including income from crop farming, 10 The repayment rate is defined as the percentage of amount of loans due that are repaid on time. In addition to the crossguarantees, the high repayment rate may partly because the loans are used productively as shown later. 11 The State Poverty Alleviation Office and the Ministry of Finance published a guide booklet on the operation of village bank programs, and they organized group training for representatives from all program villages national wide in October Although, there are non-negligible heterogeneity in implementation across regions. 12 The national poverty line set in the year 2011 was 2,300 yuan (in 2010 price) in terms of per capita annual income, or $1.57 per person per day at 2005 PPP by using 3.46 yuan to a dollar as Chen and Ravallion (2010). We adjust it by rural CPI, and define a household as a poor household if its annual income per capital in 2009 is equal to or less than 2,220 yuan. 8

9 husbandry farming, and small business) is 1,226 yuan. The average per capita income from other sources, including public and private transfers, is 1,123 yuan. Table 1 shows that before the program, 61% of the households had borrowed from January 2009 to August While more than half of the households borrowed interest-free informal loans, only 13% of the households borrowed from formal financial institutions, primarily the Rural Credit Cooperative, 13 and 5% borrowed from informal loans with positive interest. These results imply that even though borrowing is very popular among households in the surveyed areas, they rely primarily on informal credit. Consumption loans are more common than production loans, especially for informal interest-free loans. More relevantly, access to production loans from formal financial institutions is very low. Only 5% of households obtain formal loans for production. Panel B reports the median amount of loans among the panel households that have borrowed each type of loan from January 2009 to August Hereafter, all monetary values are measured in 2009 RMB by adjusting national level rural CPI. 14 It shows that the median amount of loans from formal sources is much greater than interest-free informal loans, even though the latter is more popular. Within the informal interest-free loans, the median amount of production loans is also smaller than consumption loans. These results indicates that informal financial networks plays a critical role only in consumption credit, but not in production credit. There is room for a village banking program that focuses on providing production credit access to these households. Among the control villages, Table 2 reveals that 60% are in mountainous areas, 30% are in hilly areas, and only 10% are in plains areas. The average size of the village population is around 1,000. The average diameter distance of the villages is 2.62 kilometers. Some 72% of the households share the same surname as one of the top three surnames in the villages. The relative small geographic size of the villages and close kinship within each village that indicate that people who live in these villages are usually familiar with one another. In terms of resources, the area of arable land in the villages on average is 1,396 mu (around 1 square kilometer). 15 On average, about half of the village population is labor force. The educational level of this labor force is very low. Only 12% have attained an educational level of senior high school or above. Another 31% have an educational level of junior high school. The rest have educations lower than junior high school. About 40% of the labor force are migrants, having 13 The organization of Rural Credit Cooperatives is similar to a bank. They are usually located in the administrative center of the township. 14 The results are robust to the adjustment. Using provincial level rural CPI gives us very similar results. These are available upon request mu is equal to square kilometer. 9

10 worked outside their home town more than three months during the previous year. Among the migrants, half of them worked at foreign provinces, mostly in coastal cities such as Shanghai and Guangzhou. The average wage of skilled laborers in the villages is 73 yuan per day, while the average wage of unskilled laborers in the villages is 50 yuan per day. The outreach of credit from formal financial institutions was very limited before the village banking program. Only 15% of the households borrowed from the Rural Credit Cooperatives. The average amount of these loans is about 2,400 yuan. Infrastructure in these villages is poor. Only 78% of the households are accessible by telephone or mobile phone connections. Two-thirds of the nature villages have paved roads. Living conditions are bad as well. Only 60% of the households have access to clean drinking water, and only 17% are equipped with a sanitation toilet. Most of the nearest Rural Credit Cooperatives or commercial banks are located in the administrative center of local township governments, which are on average 4.6 kilometers away. The primary target of public spending financed by the superior government is education, while public spending on other aspects is relatively smaller. In addition to village characteristics, Panel B shows summary statistics of household characteristics at baseline. Among households in the control villages, most of the household heads are male. Their average age is about 52 years. The average household size is The school attendance rate for children between 17 and 19 years of age is only 65%. 16 The randomized control trial is designed to randomly place the treatment village within selected counties. In columns 3 and 4 of Table 2, we examine the randomness of village placement by comparing the baseline characteristics of the villages and households between the treatment and control groups. We don t find significant differences among almost all of the characteristics, except that the ratio of creditworthy households in the treatment villages is smaller (p-value 0.092), and households in the treatment villages have less access to electricity than households in the control villages (p-value 0.073). Households in the treatment villages are similar to those in the control villages in terms of household heads demographic characteristics and household composition. Among all of the 60 variables tested, we find two variables with differences at the 10% level. 17 This is very much within the range we would expect. To summarize, there are no statistically significant differences between the treatment and control villages, which confirms the randomness of the treatment assignment. 16 These are the groups of children who likely have finished their junior high school education, which is compulsory, and are of the age to go on to senior high school. 17 We also check the balance among the sample when we drop the aforementioned county that delivered loans very late. The results are robust and available upon request. 10

11 Attrition One concern regarding analysis based on panel data is sample attrition. This is particularly important in program evaluation, since attrition may affect the internal validity of the estimated impacts (Karlan and Zinman, 2010). We examine whether attrition is independent of village treatment status by testing if the coefficients of village treatment status are insignificant. In the whole 1500 households, the attrition rate is 9.5% in control villages, comparing to 10.2% in treatment village which is higher. However, the difference is not statistically significant (p-value is 0.660). Table A1 reports the estimation results of a probit model of household attrition status in the follow-up survey. In all of the regressions, we cluster the standard errors by village. The result in the first column confirms the probability of attrition in the treatment and control groups is, on average, not statistically different. In columns (2) to (4), we further control for variables of village characteristics, characteristics of household head and characteristics of the household. Overall, we find no significant difference in attrition between the treatment and control villages in all of the specifications, suggesting the probability of attrition is the same in the treatment and control villages by fixing these observed characteristics. The estimation results indicate that some of the covariates are correlated with the likelihood of attrition (not reported in Table A1). For instance, households in villages located in mountain or hill areas are more likely to be attrited than those from plain area. The household attrition rate increases with log arable land and decreases with log labor force in the village. The probability of attrition is also lower among households in which the heads are older, married, or less educated. The correlation between the probability of attrition and having outstanding loans at the time of the baseline survey is not significant. However, households with less assets are more likely experience attrition. Statistics on testing the joint significance of all of the explanatory variables, excluding village treatment dummy suggests these characteristics are jointly significant in predicting attrition of the household. 18 We further check the balance of household characteristics among the panel sample (dropping or without dropping the aforementioned county). All the household level variables listed in Panel B of Table 2 are not statistically different between treatment and control groups of the panel households at the significant level of 10%. Results are available upon request. 18 The results reported in Table A1 are robust if we dropping the aforementioned county from the analysis sample. 11

12 3. Methods Empirical Strategy We present the empirical strategy used for the estimation in this section. We employ the differencein-difference (DID) method to estimate the intention-to-treat effects of access to the village banking program by utilizing the following regression equation y = α + βt + γv + ε, (1) where i indicates household and j indicates village. y is the change in outcome variables of interest for household i in village j between the end-line and baseline surveys; T is a dummy of program treatment status of village j; V is the control variables of village j at baseline (include type of geographic feature, population, area of arable land, public spending financed by superior governments on various village projects), to take account of village-specific trends; and ε is the error term, which is clustered by village. β is the parameter of interest, which reflects the average intention-to-treat effect of the village banking program on outcome y. Since T is independent of the error term, the parameter β can be identified by ordinary least squares estimation in the above first-difference equation. It compares the difference in changes in outcomes between households in the treatment and control villages. 19 Multiple Inference Since we examine many outcome variables in the study and some of them belong to the same group or family (such as credit access), the type I error will increase with the number of individual hypotheses we are testing. One method to address the multiple inference problem is to construct an summary index among the same outcome family and then implement a global hypothesis test on this index (Kling, Liebman, and Katz, 2007). This method ignores impacts on individual outcomes and the estimated impacts on the index is hard to interpret. Since most outcomes we examined are of interest by their own 19 In addition, we examine an alternative specification for all outcomes in Table 3 to Table 8 by further controlling for lagged dependent variable in equation (1). The results are robust in sign, except for the variables of access to interest-free informal loans in Table 3, total sown area in Table 5, expenses on nondurable and total consumption expense in Table 8. The sign of these variables turn to the opposite, although none of the coefficients are statistically significant. Following variables turn to be insignificant in the alternative specification: access to formal loans, amount of any loans in Table 3, household employment income and pre-transfer income in Table 4, sown area in cash crops in Table 5, employment working days at home village and employment working days at foreign province in Table 7, total income and expense on durable service in Table 8. Impacts on some other outcomes turn to be significant, including expenses on irrigation, herbicide etc. in Table 5 and profit from animal husbandry in Table 6. Results are available upon request. 12

13 and we want to compare them with other studies, we thus implemented the alternative approach by controlling the family-wise error rate (FWER) for each individual hypothesis. More specifically, we treat variables in the same table reported below as an outcome family and use free step-down resampling methodology of Westfall and Young (1993) to adjust p-values for testing each individual outcome. 20 Furthermore, we examine outcomes by following the template used in the aforementioned six studies to avoid selection on significant results. Generally, the significance of individual tests don t change by correcting multiple inference. 4. Results In this section, we present estimation results of the intention-to-treat effects of the microcredit program in China. We start from the first order effect on access to credit from various sources. We then examine program impacts on total income and income by components. To detect reasons for change in income, we examine more details on income-generating activities, including self-employment activity (crop farming and husbandry farming) and employment activity. Following this, we investigate program impacts on poverty rate, expenses on consumption, and subjective well-being. Borrowing We begin by examining the program impacts on access to credit, because if the program has any impact on investment or income, it should first affect credit access of the households. Since the duration of the village bank loans is usually one year, we focus on the impacts on short-term credit market, namely the credit with duration no more than one year. 21 Panel A in Table 3 shows, on average, that 23% of the households in treatment villages borrowed from the village bank for production purposes There are other methods to control FWER. For instance, the Bonferroni method simply multiplies the unadjusted p-values by the number of outcomes in the family. The Holm (1979) method multiplies the original p-value by its rank in the family (in decreasing order, thus lower p-value has higher rank) and reject the individual hypotheses in a step-down sequence, while Hochberg (1988) modified the procedure and reject the hypotheses in a step-up sequence. The disadvantage of the above methods is that the power for rejecting the null hypotheses decreases with the numbers outcomes in the family. In addition, these methods are based on the assumption that the individual tests are independent of each other, which is violated mostly in our study. The algorithm of Westfall and Young (1993) takes account of dependence structure of outcomes and is more powerful than the algorithm of Bonferroni and its modifications, especially when outcomes are highly correlated (see Anderson 2008 for details). In practice, we use the Stata command wyoung (Jones et al., 2018) and use 10,000 bootstraps to adjust p-values for multiple hypotheses test. 21 Over the period from January 2009 to August 2010, about 2.6% of the new credit were with duration over one year (mostly two or three years). The share of new credit with duration over one year among all new credit other than village bank loans is 1.8% from September 2010 to July In principle, the village bank lent loans for production purposes. But in practice, the rule is not strictly enforced. There are 5% of the households borrowing from the village bank for consumption. We find the program on average decreases the incidence of borrowing for consumption, particularly for informal interest-free loans. Overall, we found no impact on the 13

14 Access to the village bank crowded out production loans from other sources to some extent, especially for those who borrowed from formal financial institutions. 23 This may be because the interest rates of loans from village banks are generally lower than rates of formal financial institutions, or borrowing from the village bank doesn t require collateral and is much easier. The last row shows that the village banking program significantly increases the total credit access among households in treatment villages. The share of households with production loans increases by 16 percentage points, a sizeable impact comparing with an average 20 percent of control households measured at baseline. Turning to amount of credit, households in the treatment villages on average borrowed 957 yuan from the village bank. Considering that only 23% of households in the treatment villages borrow program loans for production, this indicates that the average size of program loans is 4,160 yuan. The programs crowd out interest-free informal loans by amount of 415 yuan, while they also crowd in informal loans with positive interest rate by amount of 506 yuan, though neither is statistically significant. Overall, the total credit amount of households in treatment villages increases by 1018 yuan with significance at 10 percent level. The positive impacts on total credit amount indicates, in total, the injection of village banking loans doesn t simply crowd out loans from other sources. Columns (1) and (2) in Table A2 in the appendix examines the impact of village banking program on the interest rate, by regressing annual real interest rate of loans on dummies that indicating village treatment status, post the time of first survey and their interaction term among short-term production loans. While the sample in the first column includes loans from all sources, the sample in Column (2) excludes program loans. Coefficients of the interaction term suggest the interest rates of production loans increase with the injection of village bank loans, though insignificant in statistic. The results are robust by using regressions on household level, where the outcome variable is defined as household average interest rate by following Kaboski and Townsend (2012). There results confirm that the positive program impacts on total credit is not a result of possible decrease in interest rate, rather evidence of credit constrained among the households in financing production before the program (Banerjee and Duflo, 2014). likelihood of any loans for either production or consumption. Further analysis suggests the incidence of any loan increases for program borrowers in the treatment villages, while the incidence of consumption loans decreases for both borrowers and nonborrowers from the program in the treatment villages. These results suggest that while the village bank increases overall credit access for borrowers, it also has negative spillover effects on non-borrowers. Results of the impacts on borrowing for consumption purposes and impacts on any borrowing are available upon request. 23 Park et al. (2003) find a similar crowd-out effect on formal financial institutions by increased competition in the financial markets. 14

15 Income We examine the program impacts on income in this section, including self-employment income and employment income. Columns (1) to (3) in Table 4 report the estimated program impacts on household income. On average, the village banking program increased self-employment income by 2,417 yuan and increased employment income by 1,736 yuan among households in the treatment villages. Total pretransfer income increased by 4,153 yuan, which is about 52% of the average income among households in the control villages at baseline. Columns (4) to (6) report the estimated impacts on income per capita, which also increased substantially among households in the treatment villages. Per capita income from self-employment activity increased by 564 yuan. Per capita income from employment activity increased by 755 yuan. In total, per capita per-transfer income increased by 1,319 yuan for households in the treatment villages. This is about 55% of the control mean at baseline. Unlike the studies of Crepon et al. (2015) and Banerjee et al. (2015), we found that employment income increased substantially as a result of the program rather than decreasing. We ll discuss this in more detail in Section 5. In addition to the pre-transfer income, Table A4 in the Appendix examines the program impacts on other income, including public transfers, private transfers, sales of assets, insurance claims, and others. The results show that there are no significant impact on public transfers, suggesting that estimated impacts didn t crowd-in or crowd-out public transfers. In addition, there is no significant impact on private transfers. For households in the treatment villages, income from selling assets on average decreased as a result of the program, indicating increased financing ability of the households. In total, we find no significant impacts on other income. Self-Employment Activity To assess how the village banking program affects self-employment activity in the treatment villages, we start by investigating the impacts on investments in crop farming and animal husbandry, and then examine whether these investments are profitable. Table 5 shows that the sown area of cash crops of households in the treatment villages on average increases by 0.49 mu. The sown area of grain crops of these households decreases, but the drop is not statistically significant. The total sown area increases, but the increase is insignificant. These results suggest that households in the treatment villages tend to switch their planting from grain crops to cash crops. This could be because of following reasons. First, cash crops are riskier than traditional grain planting. 15

16 The microcredit program may play a role as a form of insurance (Bryan et al., 2014). Second, investments in cash crops may be more costly than grain crops, and are thus more likely to be constrained by liquidity. Columns (4) to (7) show that inputs in crop farming increase by 164 yuan as a result of the program, which is about 12% of the mean value in the control group at baseline. Among the inputs, increase of inputs in irrigation and herbicides is the highest, although it is not statistically significant. Are the investments in crop farming profitable? Columns (8) and (9) examine the program impacts on revenue and profit for crop farming. The results show that revenues from crop farming on average increase by 1,197 yuan as a result of the program. The village banking program eventually leads to a net profit of 1,033 yuan from crop farming, which is about 50% of the average profit in the control villages at baseline. Table 6 shows the estimated impacts on animal husbandry. On average, investments in animal husbandry in treatment villages increase by 1,165 yuan, which is primarily driven by increased inputs in feedstuffs. The impact on expenditures for hired labors is nearly 0. Meanwhile, we don t find significant changes in investments in the purchase of new animals. This suggests that the increased investment in animal husbandry is primarily on the intensive margin (increases in investments in working capital of existing business), rather than the extensive margin (expanded business size). Revenue from animal husbandry also increases significantly by 2,130 yuan for households in the treatment villages. The effect is nearly equivalent to the revenue in the control villages at baseline. In terms of profits from animal husbandry, the impact is positive. But the standard error of the estimate is high, suggesting that the profitability from animal husbandry varies over the sample. In addition to crop farming and animal husbandry, we examine the impacts on small business. The results are reported in Table A3 in the Appendix. We don t detect any significant impacts on the establishment of small businesses. The profit of businesses outside home villages on average increases by 1,002 yuan in the treatment villages but is not statistically significant. Overall, among households in the treatment villages, profit from small businesses increased by 418 yuan. As stated, the program increased profit from crop farming and animal husbandry by 1,033 yuan and 965 yuan, respectively. These sum up to a total increase in profits from self-employment income of 2,417 yuan (76% of the mean of control groups at baseline). As shown in Tables 5 and 6, expenses from investments in crop farming and animal husbandry have increased on average by 164 yuan and 1,165 yuan, respectively, for households in the treatment villages. These add up to a total increase in agricultural investment of 1,329 yuan. By taking account of possible 16

17 increases in investments in small business and financing migration, the magnitude of impacts on investment seems greater than the estimated impacts on production loans from the program, 1,089 yuan, as shown in Table 3. These may imply some co-founding from saving for investments. Labor Employment The impact of access to credit on employment activity is ambiguous from a theoretical point of view. On the one hand, increased investments in self-employment activity as a result of financial inclusion, such as fertilizer and feedstuffs, may increase demand for labor in self-employment activity, thus decreasing labor related to employment. On the other hand, access to credit may increase employment when initiation of employment is costly and requires a lump-sum investment before receiving the returns (Evans, 1989; Nelson, 2011; Angelucci, 2015). Therefore, how the microcredit program affects labor employment is an empirical question. We examine the impact on employment in Table 7. Since we only have information on working time for self-employment activity (including farm and non-farm activity) from the follow-up survey, in columns (1) to (3), we replace the difference in outcomes in equation (1) by the outcomes in follow-up wave. 24 As shown, on average, the number of employed working days for households in the treatment villages increased by 9 days (about 7% of the average employed working days in control groups at baseline). Self-employed working days decreased by 3 days. As a result, total working days increased by about 6 days for households in the treatment villages. To get a sense of whether the estimated impacts on employment income are reasonable, we decompose impacts on labor employment according to work location and account for wage differences. Columns (4) to (8) in Table 7 report impacts on labor employment of households by work location, which we define as L. On average, the program increased employment in foreign provinces by 25 days for households in the treatment villages. It also increased employment in the home county and home province by 5 days and 7 days, respectively, although the impacts are not statistically significant. At the same time, the program decreased employment in home villages and home townships by 9 days and 3 days (not significant), respectively. The bottom of Table 7 shows the average wage by work location, which we define as w. The data demonstrates 24 We conduct the same exercise for all the other outcomes in Table 3 to Table 8. Only 8 of them are of the opposite sign among the total 46 outcome variables, comparing with results using specification in equation (1). This makes us more confident about the results reported in column (1) to (3) in Table 7 and columns (1) and (2) in Table 9 as discussed later, for which we lack information at baseline. 17

18 that the wage differs by destination. Generally speaking, the wage increases in accordance with distance from the home village. More specifically, the monthly wage at the home village is, on average, 1,078 yuan at baseline. The average monthly wage at the home township, home county, and home province is similar at about 1,200 yuan. The monthly wage at foreign provinces is, on average, 1,510 yuan. A back-of-the-envelope calculation suggests that the impact on earnings from employment of a household in treatment villages is L w, which is 1,363 yuan. This is similar to the estimated impact of the program on employment income, i.e., 1,736 yuan (as shown in Table 4). The impact is about 36% (=1736/4847) of its mean value at baseline. To summarize, these results suggest that households in the treatment villages tend to allocate more of their labor force to more remote work locations, where they can earn a higher wage. This is consistent with the findings of increased income from employment for households in the treatment villages. Consumption and Poverty Reduction One aim of the village banking program is to alleviate poverty. In Table 8, we examine impacts of the village banking program on income, poverty reduction, consumption, and saving. As shown, among households in the treatment villages, income (the sum of pre-transfer income and other income, which is primarily income from transfers) increased by 4,449 yuan, which is about 40% of the income for households in the control villages at baseline. To calculate the poverty line in 2009 RMB, we adjusted the official poverty line by rural CPI and arrived at the adjusted poverty line, which is 2,220 yuan in 2009 RMB. A household is then defined as under the poverty line if its per capita income is below the adjusted poverty line. The estimate in column (2) suggests the poverty rate for households in the treatment villages declined by 17 percentage points. 25 These results suggest that the village banking program significantly reduces poverty of households in the treatment villages. Columns (3) to (6) examine the program impacts on consumption expenses. The expenses on durable services and housing services are constructed by multiplying the current value of durable goods or housing by d/(1-d), where d is the depreciation rate. We assume the life period of durable goods is seven years, and the life of housing is 20 years. The estimates suggest that expenses increase for nondurable items, durables services, and housing services among households in the treatment villages, 25 According to the old criteria of poverty line of China, namely 1067 yuan in 2008 price, the adjusted poverty line in 2009 is 1196 yuan. Using this old poverty line, the poverty rate at baseline in control group is 24%, and the program reduce the poverty rate by 0.15 (with a standard error 0.05). 18

19 although the impacts on non-durable items and housing services are insignificant. Total yearly expenditures and savings increase for households in the treatment villages, but neither is statistically significant. Table A5 in the Appendix shows details of the impacts on expenditures for non-durables. Among households in the treatment villages, expenditures on daily necessities and medical needs increase substantially, but none of the effects are statistically significant. To summarize, among households in the treatment villages, income increases and the poverty rate falls. Expenditures on non-durables increase, but the increase is insignificant, while expenditures on durables increase significantly. The results are consistent with the findings of Ravallion and Chen (2005) that development projects in China result in income gains, although the impact on consumption is not as strong. Subjective Well-Being In addition to the objective measures of life quality, we evaluate program impacts on subjective life quality in Table 9. The measurements of subjective well-being examined include self-assessment of quality of life, self-reported life satisfaction, and assessments on quality of life of households by village officials. In the 2012 survey, the households were asked, How about the current quality of life of your family compared with two years ago? and Are you satisfied with your life currently? The answers to these questions were increasingly coded with results demonstrating better life quality or life satisfaction. In addition, the survey asked the village officials, How about the economic condition of this family in 2006? Similar retrospective questions were posed regarding the economic condition of the family from 2007 to We calculate the average score for assessment of economic condition before 2011 and compare it with the score in the year after the program. The estimates in columns (1) and (2) suggest that the self-assessed quality of life and satisfaction with life on average are higher among households in the treatment villages than those in the control villages after the program. One concern about the self-assessment of quality of life or satisfaction with life is that people may be unhappy about not receiving loans. On the one hand, this may cause an underestimation of the ITT effect if non-borrowers in the treatment village are aware that they are not receiving loans. On the other hand, if those in the control villages were also aware of their non-treatment, it could overestimate the ITT effect. To address these concerns, we examine the program impacts on household quality of life through assessments of village officials. Column (3) shows there is no significant difference in the 19

20 quality of life between households in the treatment and control villages before the program. Column (4) shows that quality of life among households in the treatment villages is significantly higher than that in the control villages after the program. The difference-in-difference estimate in column (5) suggests the program significantly increases quality of life among households in the treatment villages Comparison of Microcredit Programs Banerjee, Karlan, and Zinman (2015) identify several potential explanations for the lack of income effects of the microcredit program, as found in the six studies using randomized control trials. We summarize them here: (1) low take-up rates or small differential in take-up rates between the treatment and control groups; (2) lack of demand for credit (availability of credit from other lenders); (3) multiplicity of loan use purposes (e.g., consumption smoothing, productive investment); (4) increases in business income are sometimes offset by reductions in wage income; (5) features of loan products, such as frequency of repayment, reduce risktaking; and (6) spillovers from treatment to control groups. In this section, we compare the results of the present study with those of the other six RCT-based studies published in the American Economic Journal: Applied Economics (2015) and the study by Kaboski and Townsend (2012) using meta-analysis. The results are summarized in Table 10. We focus on the following four aspects, which are comparable across the eight studies referenced in this paragraph: (1) availability of formal credit before the program; (2) take-up rate; (3) repayment frequency; and (4) return to off-farm employment. Availability of Formal Credit. As shown in Table 10, 13% of the households had outstanding loans from formal financial institutions at baseline in the sample of China. This is relatively smaller than those in Mexico (28.8%), Mongolia (47.7%), and Bosnia and Herzegovina (30%). However, it is higher than those in Ethiopia (2.6%), India (3.6%), and Morocco (6%). As mentioned previously, a formal financial institution is rarely accessible to households in poor rural China. Financial regulations strictly restrict development of microfinance organizations in China, rendering credit rationing much more severe. However, the availability of loans from formal financial institutions in China is not the lowest in this group. The share of households 26 One may speculate that village officials in treatment village may have incentive to under-report quality of life of the households before the program and over-report it after the program. The results of insignificant difference in villager officials assessments on quality of life between treatment and control group suggests there is no clear under-report on quality of life. Moreover, we examine correlation coefficients between expenses on durable service, which we considered as objective measure of quality of life, and village officials' subjective assessment. The correlation coefficients are and in two years, respectively. These are high, especially for correlation in the second year (remember the retrospective property may lead to more error in measurement for village officials' assessment on baseline economic conditions). For reference, the autocorrelation of expenditure on durable consumption is

21 with loans from formal financial institutions is much lower in Ethiopia, India, and Morocco than in China. This comparison suggests that the rare availability of formal credit may not be the primary reason for the success of the microcredit program in China. Take-up Rate. All of the RCT studies in Table 10 are based upon randomization on the regional level, except for Augsburg et al. (2015), which randomly assigned marginal clients a 50 percent chance of receiving a loan. 27 Randomization on a regional level solves the endogeneity problem of placement. The identification is valid if we are only interested in the intention-to-treat effects. However, the small intention-to-treat effect doesn t mean the effect of the treatment-on-the-treated (TOT) is small as well. The high effect of the treatment-on-the-treated could be diluted by a low take-up rate when there is no spillover effect. Table 10 shows that the take-up rate in Mongolia, Bosnia and Herzegovina is 57% and 100%, respectively, in the treatment villages. This is because the respondents in the two studies are marginal clients, who are very likely to take loans from the programs. 28 Among studies using a representative sample (instead of marginal clients), the take-up rate is usually very low. China has a relatively higher take-up rate (28%). We thus compare the effects of TOT by assuming no spillover or a general equilibrium effect between borrowers and nonborrowers in the treatment groups in Table To render the results comparable across studies, they were divided by control mean at baseline in each study. The results suggest that the effects of TOT for employment income, as well as the impact on labor supply of employment, are the highest in China. The effect of TOT on self-employment income is also higher in China than many other countries. These results suggest that the higher ITT effect on income found in China is primarily driven by the much greater income effect of TOT, particularly of employment income. Repayment Frequency. Repayment frequency may affect households liquidity flow and investment type. Field et al. (2012) provide experimental evidence that clients of monthly repayment reported higher business investment and income than clients of weekly repayment. We also compare the repayment frequency 27 Individual-level randomization among marginal clients usually generates a high difference in take-up rates between treatment and control groups (Kalan and Zinman, 2010; Kalan and Zinman, 2011). Ideally, the estimates identify the effect of treatment-on-the-treated. But, it could be biased as a results of possible spillover effects on the control group, especially when treatment and control clients live in the same areas (Augsburg et al., 2015). 28 Sampling on potential borrowers may not necessarily increase the differences in take-up rates of loans between treatment and control groups. Angelucci et al. (2015) used the same sampling strategy by focusing on high potential borrowers. The differences in take-up rates between the treatment and control groups in their studies is only 13.1%. 29 A more fundamental question is why the take-up rate in treatment groups is so low in most of the studies. Theoretically, take-up of credit depends on the interest rate charged, potential average return to capital, risk of the return, and preference. Interest rates of the microcredit product are usually similar to or cheaper than the market interest rates. Empirical studies verify return to capital in these less developed areas is typically high (De Mel et al., 2008; Urdy and Anagol, 2006). But the high average rate of return is usually associated with a large variance. Angelucci et al. (2015) found the distribution of profit from business is more dispersed in treatment groups than that in control groups. Risk in production and incomplete insurance play an important role in hindering loan take-up and investment (Kalan et al., 2013). 21

22 among all of the eight studies in Table 10. As shown, loans in China are paid yearly, while in other countries loans are paid either monthly or weekly. The relatively longer periods for repaying the loans allow borrowers to invest in more profitable projects but require more time to realize returns. Such projects may include planting cash crops or migration to foreign provinces. Return to Off-farm Employment. Unlike other studies, we found a much more substantial increase in employment income in China. Table 10 shows that the effect of TOT on employment income (normalized by control mean at baseline) is the highest in China among all of the studies. For most other studies, program impacts on employment income are either negative or insignificantly greater than 0, except in Thailand (Kaboski and Townsend, 2012). In addition, the effect of TOT on labor employment is also the highest in China. The impacts are positive and significant in China, while the impacts are negative in other countries, such as Morocco, India, and Mongolia. This may be the result of a relatively higher rural-urban wage gap in China. Table 10 shows that the ratio of wages in employment and self-employment activity at baseline is 3.84 in China, which is higher than most of the other countries, except for Mongolia. Meanwhile, the ratio of earnings from employment activity and self-employment activity in China is relatively lower among all of the countries. These results suggests that the surveyed households in China are likely being constrained from engaging in employment activity, although the return to off-farm employment is high. Of the other two reasons for lack of income effects, as identified by Banerjee et al. (2015), we believe neither would be a serious problem in the case of China. First, by design, village banking loans are lent for production purposes in China. Peer monitoring of group lending would reduce the misuse of loans to a large extent. In practice, among the cases of village banking loans reported in the household survey, most are used for productive investment (81%). Second, the average distance between the treatment and control villages in the same county is about 15 kilometers, which is about six times the average diameter distance of the villages in the sample. Spillovers from treatment to control villages is unlikely to happen in the study of China. The program impacts could be affected by other loan characteristics as well. For instance, the nominal interest rate of the program loans in China is among the lowest in all of the eight studies (as shown in Table 10). The microcredit program in China is based on a self-organized committee for management, which helps to save on implementation costs. The program thus can be sustainable under a relatively low interest rate. This makes the program loans more attractive for poor households and leads to higher net return. 22

23 To summarize, the meta-analysis suggests that a less frequent repayment schedule, lower interest rate, and higher return to migration are relatively important for China to realize strong impacts of the microcredit program on raising income and reducing poverty among all of the potential explanations. Meanwhile, microcredit programs that are not so successful in other countries may be the result of different factors. Table A6 summarizes the specific factors that potentially explain small and insignificant income impacts for each of the six RCT studies based on results from the meta-analysis above. For Bosnia and Herzegovina, the program loan size is relative smaller; the bank loans were already very popular before the program; and the relative return to off-farm employment is small. For Ethiopia, the relative return to off-farm employment is smaller than in China. In India, the program takeup rate difference between the treatment and control groups may be too small to observe any effect; repayment is too frequent to allow for efficient investment; and the relative wage from employment and self-employment activity was smaller than in China as well. With regard to Mexico, the country has a lower take-up rate, higher repayment frequency, shorter loan term, smaller loan size, and greater availability of formal loans before intervention. Mongolia has a shorter loan term length and is more likely to receive bank loans before the intervention. Morocco has a lower take-up rate, higher repayment frequency, and lower relative wage in employment activity. The reasons for small and/or insignificant income impacts differ by countries, but all have definite potential reasons. This emphasizes how different contextual factors may matter in different settings. 6. Conclusion In this study, we conduct a systematic evaluation of the largest village banking program in the world by randomized control trial. We examine the average intention-to-treat effects on all households in treatment villages. Randomized access to credit significantly increases the likelihood of borrowing among households in the treatment village, as well as the amounts of loans, indicating that households are on average credit constrained. Their investments in crop farming and animal husbandry substantially increase, while they also plant more cash crops. The increased investment leads to substantial boosts in revenue and profit. Unlike the study by Crepon et al. (2015), we find that increases in self-employment income are not offset by any decrease in employment income. The village banking program increases income for both self-employment and employment activities in China. As a result, total income increases, as well as expenses on consumption, particularly for durable goods. The improvement in quality of life is confirmed by strong positive impacts on 23

24 subjective well-being. 24

25 References Anderson, Michael L Multiple Inference and Gender Differences in the Effects of Early Intervention: A Reevaluation of the Abecedarian, Perry Preschool, and Early Training Projects, Journal of the American Statistical Association 103 (484): Angelucci, Manuela Migration and Financial Constraints: Evidence from Mexico, The Review of Economics and Statistics 97(1): Angelucci, Manuela, Dean Karlan, and Jonathan Zinman Microcredit Impacts: Evidence from a Randomized Microcredit Program Placement Experiment by Compartamos Banco, American Economic Journal: Applied Economics 7(1): Attanasio, Orazio, Britta Augsburg, Ralph De Haas, Emla Fitzsimons, Heike Harmgart The Impacts of Microfinance: Evidence from Joint-Liability Lending in Mongolia, American Economic Journal: Applied Economics 7(1): Augsburg, Britta, Ralph De Haas, Heike Harmgart, and Costas Meghir The Impacts of Microcredit: Evidence from Bosnia and Herzegovina, American Economic Journal: Applied Economics 7(1): Banerjee, Abhijit V., and Esther Duflo Do Firms Want to Borrow More? Testing Credit Constraints Using a Directed Lending Program, Review of Economic Studies 81(2): Banerjee, Abhijit V., Esther Duflo, Rachel Glennester, and Cynthia Kinnan The Miracle of Microfinance? Evidence from a Randomized Evaluation, American Economic Journal: Applied Economics 7(1): Banerjee, Abhijit V., Dean Karlan, and Jonathan Zinman Six Randomized Evaluations of Microcredit: Introduction and Future Steps, American Economic Journal: Applied Economics 7(1): Bryan, G., Chowdhury, S. and Mobarak, A. M Underinvestment in a Profitable Technology: The Case of Seasonal Migration in Bangladesh, Econometrica 82(5): Chen, Shaohua, and Martin Ravallion The Developing World is Poorer than We Thought, But No Less Successful in the Fight against Poverty, Quarterly Journal of Economics 125(4): Coleman, Brett E The Impact of Group Lending in Northeast Thailand, Journal of Development Economics 60(1): Crepon, Bruno, Florencia Devoto, Esther Duflo, and William Pariente Estimating the Impact of Microcredit on Those Who Take it Up: Evidence from a Randomized Experiment in Morocco, American Economic Journal: Applied Economics 7(1): De Mel, Suresh, David McKenzie, and Christopher Woodruff Return to Capital in Microenterprises: 25

26 Evidence from a Field Experiment, The Quarterly Journal of Economics 123(4): Evans, David S., and Boyan Jovanovic An Estimated Model of Entrepreneurial Choice under Liquidity Constraints, Journal of Political Economy 97(4): Field, Erica, Rohini Pande, John Papp, and Natalia Rigol Does the Classic Microfinance Model Discourage Entrepreneurship Among the Poor? Experimental Evidence from India, American Economic Review 103(6): Hochberg, Yosef A Sharper Bonferroni Procedure for Multiple Tests of Significance, Biometrika 75(4): Holm, Sture A Simple Sequentially Rejective Multiple Test Procedure, Scandinavian Journal of Statistics 6: Jones, Damon, David Molitor, and Julian Reif What Do Workplace Wellness Programs Do? Evidence from the Illinois Workplace Wellness Study, NBER working paper No Kaboski, Joseph P., and Robert M. Townsend Policies and Impact: and Analysis of Village-Level Microfinance Institutions, Journal of the European Economic Association 3(1): Kaboski, Joseph P., and Robert M. Townsend The Impact of Credit on Village Economies, American Economic Journal: Applied Economic 4(2): Karlan, Dean, and Jonathan Zinman Expanding Credit Access: Using Randomized Supply Decisions to Estimate the Impacts, Review of Financial Studies 23(1): Karlan, Dean, and Jonathan Zinman Microcredit in Theory and Practice: Using Randomized Credit Scoring for Impact Evaluation, Science 332(1): Kling, Jeffrey R., Jeffrey B. Liebman, and Lawrence F. Katz Experimental Analysis of Neighborhood Effects, Econometrica 75(1): Morduch, Jonathan Does Microfinance Really Help the Poor? New Evidence from Flagship Programs in Bangladesh, Princeton University, Woodrow Wilson School of Public and International Affairs, Research Program in Development Studies Working Paper 198. Morduch, Jonathan The Economics of Microfinance, MIT Press. Nelson, Leah K From Loans to Labor: Access to Credit, Entrepreneurship and Child Labor, working paper. Park, Albert, and Changqing Ren Microfinance with Chinese Characteristics, World Development 29(1):

27 Park, Albert, Changqing Ren, and Sangui Wang "Microfinance, Poverty Alleviation, and Financial Reform in China," workshop on Rural Finance and Credit Infrastructure in China, Paris, France. Park, Albert, Loren Brandt, and John Giles Competition Under Credit Rationing: Theory and Evidence from Rural China, Journal of Development Economics 71(2): Park, Albert, and Minggao Shen Joint Liability Lending and the Rise and Fall of China s Township and Village Enterprises, Journal of Development Economics 71: Pitt, Mark M., Shahidur R. Khandker The Impact of Group-Based Credit Programs on Poor Households in Bangladesh: Does the Gender of Participants Matter? Journal of Political Economy 106(5): Ravallion, Martin, and Shaohua Chen Hidden Impact? Ex-post evaluation of an anti-poverty program, Journal of Public Economics 89(11-12): Tarozzi, Alessandro. Jaikishan Desai, and Kristin Johnson The Impact of Microcredit: Evidence from Ethiopia, American Economic Journal: Applied Economics 7(1): The Yearbook of China s Poverty Alleviation and Development Editorial Board The Yearbook of China s Poverty Alleviation and Development Beijing: Tuanjie Press Westfall, Peter, and Stanley Young Resampling-Based Multiple Testing: Examples and Methods for P-Value Adjustment NewYork: Wiley. Yunus, Muhammad Nobel Lecture. Norway, Oslo. 27

28 Figure 1: The Location of Sampled Provinces and Counties Note: The figure shows the sampling of five provinces (Gansu and Sichuan in Western China; Henan and Hunan in Middle China; Shandong in Eastern China), ten counties (two in each province). 28