VULNERABILITY OF MICROFINANCE TO STRATEGIC DEFAULT AND COVARIATE SHOCKS: EVIDENCE FROM PAKISTAN

Authors


  • The authors are grateful to two anonymous reviewers of this journal; Lisa Cameron, Rients Galema, Niels Hermes, Yoko Kijima, Hisaki Kono, Robert Lensink, Fumiharu Mieno, Yasuyuki Sawada, Chikako Yamauchi; and participants of the 2009 Far East and South Asia Meeting of the Econometric Society, the 1st International Workshop on Microfinance Management and Governance, the University of Melbourne Microeconometrics Workshops, the Osaka University OEIO Seminar, the Keio University Public Economics Workshop, the FASID Hakone Conference on Development Economics, and the Japanese Economic Association Annual Meeting for their useful comments on earlier versions of this paper. All remaining errors are ours.

Abstract

This paper investigates the repayment behavior of borrowers of a Pakistani microfinance institution (MFI) using a unique dataset of approximately 45,000 installment records over the period 1998–2007. In early 2005, the MFI introduced reforms that included improved enforcement of contingent renewal. The reforms led to a healthy situation with almost zero default rates. We hypothesize that strategic default under the joint liability mechanism was encouraged by weak enforcement of contingent renewal and was one of the factors responsible for the pre-reform failure. To support this hypothesis, we show that before the reforms, a borrower's delay in installment repayment was correlated with other group members' repayment delays beyond the level explained by possible correlation of project failures due to locally covariate shocks. Such excessive correlation disappeared after the reforms, including the period after the 2005 Kashmir earthquake. The empirical evidence thus demonstrates the existence and seriousness of the strategic default under weak dynamic incentives.

I. INTRODUCTION

This paper empirically analyzes the repayment behavior of microfinance borrowers who were subject to high levels of idiosyncratic and covariate shocks and to subsequent reforms in the contract design concerning repayment installments and dynamic incentives. In the literature on microfinance, welfare impacts and mechanisms that led to its success in maintaining high repayment rates have been the focus of research (Kono and Takahashi 2010; Armendariz and Morduch 2010). With respect to mechanisms, first generation studies emphasize various aspects of the role of joint liability in group lending, such as peer selection (Ghatak 1999, 2000), peer monitoring (Stiglitz 1990; Varian 1990), and peer enforcement (Besley and Coate 1995; Armendariz de Aghion 1999). More recently, in second generation studies, there has been a shift in emphasis—Armendariz and Morduch (2010) argue that joint liability is only one factor contributing to successful microfinance schemes and that there are other important aspects such as sequential financing, contingent renewal, dynamic incentives, frequent repayment installments, and public repayments.1Chowdhury (2005) theoretically shows that without sequential financing, group-lending schemes may suffer from under-monitoring, which might lead to borrowers investing in risky projects. In another theoretical paper, Chowdhury (2007) outlines that contingent renewal alone can resolve the moral hazard problem only when the discount rate is relatively low; however, a combination of sequential financing and contingent renewal can be successful even when the discount rate is relatively high.

The analysis of different theoretical mechanisms in the literature notwithstanding, empirical evidence on the effectiveness of each continues to be limited,2 although it has seen an increase in recent times (Kono and Takahashi 2010; Hermes and Lensink 2007). Kono (2006) reports the results of the Vietnamese experiment with microfinance games wherein contract designs are varied, and shows that joint liability per se leads to more frequent strategic default. Using the field experiment approach, Gine and Karlan (2011) evaluate the impact of removing group liability in the Philippines and find no adverse impact on repayment as long as public and frequent repayment systems are maintained.3 At the same time, few empirical studies use nonexperimental data while attempting to establish an explicit theoretical link. Using data from Peru, Karlan (2007) finds evidence for the existence of peer monitoring under joint liability. Using data from Thailand, Ahlin and Townsend (2007) test theoretical models of joint liability lending, showing that the joint liability rate negatively affects repayment while the strength of local sanctions positively affects repayment. Interestingly, they show that the Besley and Coate model of strategic default has low explanatory power in general but its explanatory power increases when applied to poor, low-infrastructure areas. These studies mostly support the view that mechanisms such as sequential financing, contingent renewal, dynamic incentives, frequent repayment, and public repayments are more important than joint liability is to the success of microfinance lending. However, they do not provide detailed empirical evidence for each theoretical mechanism, especially regarding the intra-group repayment dynamics that play a key role in each model. More recently, Ahlin (2009) tests risk-matching and intra-group diversification of risk using Thai data and finds evidence of homogeneous sorting by risk and risk anti-diversification within groups.

To extend the empirical research in this direction, this paper investigates a unique dataset of about 45,000 installments/repayments covering 2,945 borrower households in Pakistan over the period 1998–2007, in order to offer empirical evidence that (1) strategic default is a serious problem for microfinance, and (2) contingent renewal is a prerequisite to prevent strategic default. Modern-style microfinance took root in Pakistan in the late 1990s; since then there have been several examples of microfinance schemes that have failed despite following a group lending design on the lines of Grameen Bank in Bangladesh. Our dataset is unique not only in its level of disaggregation but also in its time coverage, which includes two important events. First, in early 2005, contract designs were changed to stress improved enforcement of contingent renewal and more frequent repayment of installments. Our dataset includes this break and enables us to analyze the outcome of contract design changes regarding repayment behavior. The second event occurred in October 2005 when a disastrous earthquake known as the 2005 Kashmir earthquake, measuring 7.6 on the Richter scale, struck Pakistan, killing more than 70,000 people. An earthquake is a strong covariate and an unexpected natural shock. Therefore, our dataset enables us to analyze the impacts of covariate shocks of a high magnitude, whose impact on microfinance is rarely mentioned in the literature (Shoji 2010; Becchetti and Castriota 2007).

Our analysis of the installment-level dynamics of repayment delays suggests the existence of a type of strategic default under the joint liability mechanism, wherein if one group member is hit by a negative shock and faces difficulty in repayment, other members who, although capable of repayment, may decide to default as well, instead of helping out the unlucky member. The delay in the unlucky member's repayment of each installment was correlated with the other members' repayment delays beyond the level explained by the possible correlation of project failures due to locally covariate shocks. Although this type of strategic default has been shown to exist theoretically (e.g., Besley and Coate 1995; Bhole and Ogden 2010), this paper provides the empirical evidence lacking in the literature.

The rest of the paper is organized as follows. Section II describes the dataset and its background, providing the empirical definition of default and repayment delay adopted in this paper. The correlates of loan default and repayment delay under the old system at the borrower level are also examined in this section, in order to show that the 2005 reforms were a step in the right direction. Section III documents the bulk of our empirical analysis, in which we compare the dynamics of installment repayments for loans made under the old system and those under the new system. The analysis in this section also demonstrates that the earthquake only marginally affected the new system in terms of repayment delays. Section IV presents the conclusions.

II. DATA AND BACKGROUND

A. Microfinance in Pakistan

Pakistan is one of the low-income countries in South Asia where poverty is rampant and the majority of the poor face difficulties in getting access to efficient sources of credit (World Bank 2002). Modern-style microfinance in Pakistan began only in the late 1990s. In 1999, microfinance was declared a priority in the official Poverty Alleviation Strategy, and in 2001, a regulatory framework for the promotion of microfinance was established. The result was a massive investment of at least US$400 million in the period 1999–2005, largely from multilateral sources such as the World Bank and the Asian Development Bank. Among the various microfinance providers today, Khushhali Bank, the flagship institution established by the Pakistani government in 2000, served more clients than the collective client base of all the NGOs and rural support programs before 2001 (Montgomery 2005).

Microfinance institutions (MFIs) in Pakistan now comprise a voluntary network called Pakistan Microfinance Network (PMN). With the passage of time, the country has not only moved forward but also benefited from the experience of countries that were ahead of it on the microfinance front. However, their outreach has been limited, with many of the poor untouched by them (Montgomery 2005).

During the 1990s, the economy of Pakistan registered moderate growth; however, due to rising economic inequality, poverty did not decline proportionately (World Bank 2002). The macroeconomic scene in the country did not change substantially during the 2000s as far as macroeconomic indicators are concerned. Because of the lack of sustained high growth, the absolute level of real per capita income is still very low, standing at US$2,550 in 2011 in constant 2005 PPP prices [United Nations Development Programme (UNDP) 2011]. Against this backdrop, transferring the benefit of growth equitably across the board, especially among the poor, remains a challenge.

B. Primary Data of Microcredit Borrowers and Their Repayment

We utilize census data of microcredit borrowers maintained for the purpose of financial monitoring of microcredit intervention. The information was obtained from a member MFI of the PMN. The names and identities of the participating households have been replaced by computer-generated numbers to safeguard privacy. The dataset covers a whole district in the North-West Frontier Province (NWFP),4 Pakistan, which was severely affected by the earthquake. The dataset includes four components: borrowers' data; data on borrowers' community organizations; data on installments; and data on repayment receipts.

1. Borrowers' data

There are 2,950 borrowers, comprising those who borrowed between May 1, 1998, and July 8, 2006. Since installment and repayment records are missing for five borrowers,5 we analyze 2,945 borrowers in this paper. Approximately 30% of the borrowers are women. Information regarding the household to which the borrower belonged was collected and made part of the borrower dataset.

2. Data on borrowers' community organizations

To be eligible for microfinance loans in the study area, borrowers need to form a community organization (CO) with joint liability. Information pertaining to COs to which the above borrowers belonged was collected. There are 870 COs in the CO dataset, comprising those established between March 11, 1997, and April 10,2006. Female borrowers usually belonged to COs designated as “female COs,” which accounted for 22.4% of the sample COs. COs maintain their own savings account, with the balance ranging from Rs. 0 to Rs. 99,000 and the average being Rs. 6,348.6

3. Data on installments

For each borrower, records were kept for each installment, such as due date, amounts due as principal and as service charges or process fee,7 installment repaid, and outstanding debt after repayment. For the 2,945 borrowers in the borrower dataset, there are 44,931 installment records. Their due dates range from June 1, 1998 to October 2, 2007.

4. Data on repayment receipts

When a borrower repays, a receipt is issued and relevant information such as the time and amount of repayment and the amount of penalty is recorded in this dataset. To the 2,945 borrowers in the borrower dataset, 32,695 receipts were issued. The receipts are dated between June 5, 1998 (the first repayment in our sample was made four days after the due date) and October 22, 2007 (the last repayment in our sample was made 20 days after the due date).

The number of receipt-level observations is smaller than that of installment-level observations for two reasons. First, some installments were not repaid, and considered clear cases of default. Second, when the borrowers paid amounts more than the amounts due in an installment, one receipt was issued. To reduce the transaction costs of monthly repayments, many borrowers preferred to pay a single installment worth several months' amount in one go. On the other hand, when the borrower could not pay a monthly installment, he/she often repaid the amount worth two months' dues in the following month. All installment dues associated with loans provided after January 2005 were already repaid by the time the dataset was updated in July 2009. Therefore, the right-censoring problem for installments is not a cause for concern.8

There are two important time breaks in our data. First, from January 2, 2005, the MFI began to provide microcredit under a new system offering improved accountability. The last loan under the old system was issued on October 4, 2003. Following this, no new loans were issued for more than 13 months, during which period the MFI continued to collect dues for old loans while it formulated a new system. The second time break occurred on October 8, 2005, when Pakistan was hit by the earthquake.

C. Reforms in January 2005

There are a number of differences in the contract characteristics of microfinance loans issued under the old and new systems (Table 1). First, the average loan size was almost halved. It was Rs. 16,300 (about US$270) under the old system and Rs. 9,000 (about US$150) under the new system.

Table 1.  Characteristics of Microfinance Loans
 Old SystemNew System
Borrowers Who Borrowed before the EarthquakeBorrowers Who Borrowed after the Earthquake
  •  NIC stands for the “National Identity Card” issued by the Government of Pakistan.

  •  “Savings” and “Number of CO members” were not reported for some of the sample COs under the old system. The reported averages consider only those COs with complete information.

Characteristics of credit:   
 First date of credit issued1998-05-012005-01-022005-10-16
 Last date of credit issued2003-10-042005-10-082006-07-08
 Amount of credit in Rs.:   
  Average16,3248,2659,632
  (SD: standard deviation)(9,427)(2,384)(2,945)
  Minimum5005,0005,000
  Maximum50,00010,00015,000
 Number of installments:   
  Average16.1711.9012.42
  (SD)(8.98)(1.07)(1.04)
  Minimum1112
  Maximum301215
 Credit duration in months:   
  Average17.6311.9012.42
  (SD)(7.99)(1.07)(1.04)
  Minimum1112
  Maximum311215
Characteristics of borrowers:   
 Ratio of borrowers with NIC information recorded78.5%100.0%100.0%
 Ratio of female borrowers24.2%50.5%47.9%
 Ratio of borrowers who are chairman or secretary of the CO4.5%18.0%20.4%
 Average number of income sources of the household1.623.162.97
 (SD)(0.86)(1.56)(1.51)
 Ratio of borrowers who had income sources outside the region18.5%21.1%100.0%
 Ratio of borrowers from female-headed households8.0%0.0%13.9%
 Ratio of borrowers from joint families47.6%36.6%20.4%
Characteristics of COs (community organizations):   
 Average CO savings (in Rs.100,000)0.1730.0560.041
 (SD)(0.160)(0.105)(0.075)
 Average number of CO members36.024.322.1
 (SD)(12.4)(9.0)(8.2)
 Ratio of COs with missing CO records0.1650.0320.014
 Average CO age in days at the time of loan issue496.11,204.21,234.4
 (SD)(418.3)(627.0)(740.2)
Purpose of borrowing (total = 100%):   
 Agricultural crop5.3%0.3%0.3%
 Livestock61.4%6.9%6.8%
 Shops, business, workshops30.4%52.1%58.1%
 Others0.6%0.6%0.9%
 Domestic needs (consumption, education, housing, etc.)2.3%40.1%34.0%
 Number of sample borrowers2,275317353

Second, the number of installments and length of credit duration decreased by 25%. Under the old system, there were several types of loans on offer. The most common was a 19-month loan to be repaid in 18 monthly installments (42.3% of the sample), followed by a 31-month loan to be repaid in 30 monthly installments (18.3%). Approximately 8.3% of the sample loans were not to be repaid monthly but in a single installment, with the loan period being 6, 9, 12, or 18 months. Under the new system, only two standard types of loans were offered: a 12-month loan to be repaid in 12 monthly installments or a 15-month loan to be repaid in 15 monthly installments. The first type accounted for 92.3% of the sample, and the second type 7.2%.

The third change between the two systems was that interests (service charges) for all loans were collected in advance as a processing fee under the new system, whereas under the old system, service charges were collected over the repayment period along with the principal in each installment. In addition, under the old system, there was a grace period of one month before repayment began. This one-month grace period was abandoned under the new system.

The fourth change pertained to the enforcement of penalty against nonrepayment. Under the old system, there was a rule that no new loans would be given to a borrower unless his/her entire group repaid. However, the rule was not enforced stringently. The new system promised better accountability, which called for stricter enforcement of the contingent renewal rule. This positive change is reflected in Table 1: Under the new system, the ratio of borrowers with a national identity information record (NIC information) is 100%. Additionally, the ratio of female borrowers also increased under the new system.

The fifth change involved linking development projects at the community level with repayment rates of COs in the region. In Pakistan, several initiatives were introduced in the 2000s to strengthen community-based development (Kurosaki 2005). Under the new system, the implementation of infrastructure or human resource development projects became contingent on the repayment record of the community. This change is another example of improved dynamic incentives for repayment.

D. Earthquake in October 2005

The second time break in our dataset is demarcated on October 8, 2005. On this day, an earthquake measuring 7.6 on the Richter scale rocked Kashmir. The calamity caused widespread destruction and heavy loss to human life, killing at least 73,000 people, severely injuring another 70,000, and leaving 2.8 million people homeless. Social service delivery, commerce, and communications were either severely weakened or destroyed. Sample borrowers in our dataset lived in a radius ranging from between 40 and 110 km from the epicenter.

The characteristics of borrowers and loan contracts under the new system are shown in Table 1, where figures corresponding to borrowers who were issued loans before the earthquake are indicated separately from those who were issued loans after. Overall, both sets of figures are similar. None of the variables in Table 1 show statistically significant differences except for the ratio of borrowers who had income sources outside the region. Table 1 thus proves that the critical change pertaining to the characteristics of loan contracts and borrowers was the system change in January 2005, and not the earthquake in October 2005.

E. Data Compilation for Loan Default and Repayment Delay

From the primary data described above, we compile several variables for loan default and repayment delay, since these are the two pressing concerns of the MFI (Table 2). Aggregating the number of loans fully repaid by the time of our survey, we obtain the borrower-level variable Default, which is plotted in Figure 1. As can be seen clearly from the figure, all 670 borrowers under the new system repaid fully (Default= 0). In contrast, out of 2,275 old borrowers, only 1,119 had completed the repayment by the time of our survey, indicating a default rate of 50.8% (Table 2).9

Table 2.  Defaults and Delays in Repayment of Microcredit
 Old SystemNew System
Borrowers Who Borrowed before the EarthquakeBorrowers Who Borrowed after the Earthquake
  1. Note: NOB = Number of observations for which this variable can be defined.

Borrower-level variables:   
 Total number of observations2,275317353
 Default (dummy for nonrepayment):   
  NOB2,275317353
  Average (ratio of defaults)0.508100
 Avg_delay (average delay in repayment):   
  NOB1,119317353
  Average (in days)100.0−1.32.1
  SD144.812.610.7
  Minimum−552.0−93.8−68.7
  Maximum1,014.015.921.1
Installment-level variables:   
 Total number of observations36,7773,7714,383
 No_repay (dummy for nonrepayment):   
  NOB36,7773,7714,383
  Average (ratio of nonrepayments)0.207800
 Delay (delay in repayment):   
  NOB29,1343,7714,383
  Average (in days)101.1−1.42.3
  SD176.118.816.1
  Minimum−552−249−208
  Maximum1,5602877
 Problem (dummy for nonrepayment until 31 days after the due date):   
  NOB36,7773,7714,383
  Average0.65490.00000.0046
Figure 1.

Borrower-Level Default Rates

In Figure 1, the predicted probability from a probit model is also plotted for samples under the old system, in which Default was regressed on the date of loan issue and its polynomials. Polynomials of the fifth order and above did not havestatistically significant coefficients. The default rates show a highly nonlinear pattern over the period.

The default rate for the old loans, nevertheless, underestimates the cost for the MFI because some bad loans were repaid much after the last due date. For instance, in our sample, a borrower who obtained a credit of Rs. 25,000 on May 7, 1998, had paid only about two-thirds of his/her due by November 7, 1999 (the due date of the last installment) and paid the rest on July 19, 2001. Such cases are not counted as defaults.

Therefore, in order to accurately measure the quality of repayment by the 1,119 borrowers with Default= 0, a variable named Avg_delay, i.e., the average delay in days of repayment relative to the due date for each installment, was calculated. Table 2 shows the mean of Avg_delay for the 1,119 borrowers at 100.0 days, ranging from −552.0 to 1,014.0 (standard deviation [SD] at 145). If we exclude five outliers who made early repayment (Avg_delay < −200) and four outliers whomade very late repayment (Avg_delay > 1,000), the mean of Avg_delay becomes 98.5 (SD = 131.6).10

Similar variables can be defined at the installment level. The first one, No_repay, is a dummy variable that takes the value of 1 when the installment is not paid. For installments that are paid, another variable Delay (delay in days of repayment from the due date for the installment) is computed. In addition to these two, a third variable, Problem, is a dummy variable assuming the value of 1 if the installment is not repaid within 31 days from the due date. Since early repayment is not a concern for MFIs, the variable Problem ignores early repayments. Since late repayment and no repayment are concerns for MFIs, Problem aggregates the late and no repayment information into one variable. As shown in Table 2, all three variables have very high average values for loans provided under the old system. About 21% of all installments were never repaid, the average delay of repaid installments was about 101 days, and 65% of all installments suffered from late repayment or nonrepayment.

In sharp contrast, for all 670 borrowers under the new system, Default= 0 and No_repay= 0 (Table 2), whether the loan was made before or after the earthquake. We calculate Avg_delay, Delay, and Problem for these borrowers as well. The means of Avg_delay and Delay are slightly negative before the earthquake and slightly positive after the earthquake, suggesting the possibility that the earthquake led to more delays (3.4 days on average) in repayment. However, a simple t-test of the equal means cannot reject the null hypothesis that the means are the same. Therefore, it appears that no serious delay occurred under the new system and the adverse impact of the earthquake was not statistically significant,11 which will be confirmed in a multivariate regression in subsection III.D.

F. Correlates of Defaults in the Old System: A Borrower-Level Analysis

Before the main empirical analysis, we briefly investigate the correlates of default and repayment delay under the old system. We employ two dependent variables—Default and Avg_delay—and four groups of explanatory variables—borrowers' characteristics; credit characteristics (amount of credit, number of installments, period of credit in months, etc.); CO characteristics (number of CO members, value of CO savings, age of the CO when the loan was issued, etc.); and location and time effects.12

This is a borrower-level analysis, motivated as a description of our data. There are two selection problems. First, Avg_delay is defined only when Default= 0. To control for this selection bias, we adopted Heckman's two-step procedure. Second, contract type is not randomly chosen but is partly decided by the borrowers themselves. Instead of correcting the self-selection bias corresponding to the contract type, we first report noncorrected results with explanatory variables of credit characteristics and location/time controls only, and then report results with explanatory variables where credit characteristics are replaced by borrower and CO characteristics. The motivation of the first specification is to capture the net effects on repayment, including the causal effects of contract design and selection effects.

Table 3 reports estimation results with Default (dummy for borrower-level default) or Avg_delay (average delay in days in installment repayment by a borrower with Default= 0) as dependent variables. In the regression results for Default, the period of credit duration and the dummy for nonmonthly installments show positive and statistically significant coefficients. The regression results for Avg_delay show a large coefficient on the dummy for nonmonthly installments as well. These results demonstrate the appropriateness of the reforms introduced in early 2005—making all installments monthly and reducing the maximum loan duration to 15 months.13

Table 3.  Borrower-Level Defaults/Delays and Credit Contract Types
 Probit: DefaultTwo-Stage Heckman: Avg_delay
CoefficientdF/dxCoefficient
  1. Notes: 1. The average of “Dummy for loan size larger than Rs.15,000” is 0.592 and that of “Dummy for nonmonthly instalments” is 0.129. For other variables, see Tables 1 and 2 for summary statistics.

  2. 2. For the two-stage Heckman estimation, the first-stage model is the one reported in Table 4 for Default using 2,275 observations. In other words, borrowers' and COs' characteristics are used as identifying instrumental variables. The number of uncensored observations is 1,119.

  3. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors are in parentheses.

Model 1:
Dummy for loan size larger than Rs.15,000−0.0182−0.0073−24.19*
(0.0952) (12.76)
Credit duration in months0.0774***0.03090.63
(0.0072) (1.01)
Dummy for nonmonthly installments0.5414***0.207984.16***
(0.1340) (18.83)
Inverse Mills ratio  53.09***
  (16.31)
Union fixed effectsYes Yes
Date of credit issued: linear, quadratic, cubic, quarticYes Yes
Number of observations used in the estimation2,275 1,119/2,275
χ2 (16), χ2 (26) for zero slope394.5*** 408.0***
Pseudo R20.1251  
Model 2:
Dummy for loan size larger than Rs.15,000−0.0403−0.0161−31.46**
(0.1007) (13.03)
Credit duration in months0.0697***0.02780.57
(0.0072) (1.02)
Dummy for nonmonthly installments0.3639**0.142474.29***
(0.1523) (20.32)
Inverse Mills ratio  48.04***
  (17.50)
Union fixed effectsYes Yes
Year fixed effectsYes Yes
Number of observations used in the estimation2,275 1,119/2,275
χ2(17), χ2(31) for zero slope443.2*** 448.2***
Pseudo R20.1406  

The coefficient on the dummy for loan size larger than Rs.15,000 is negative, which is contrary to our expectations. It is negative in all four models reported in Table 3, with the coefficient on Avg_delay being statistically significant. This appears to reflect selection bias. Only those borrowers with characteristics associated with higher probability of on-time repayment applied for and were granted large loans. Given the strong selection effect, the potential negative effect of loan size on repayment cannot be observed in Table 3.

Significantly positive coefficients of the period of credit duration and the dummy for nonmonthly installments could also reflect selection effects. The regression results show that loans with longer repayment periods and fewer installments are detrimental to repayment because of the negative causal effects on repayment behavior and the selection effect of such contracts attracting more risky borrowers.

As a more reduced-form approach, Default or Avg_delay is regressed on explanatory variables, including borrowers' individual characteristics, borrower households' characteristics, CO characteristics, and location and time effects (Table 4). The explanatory variables are jointly significant and those with individual statistical significance show expected signs. First, borrowers who reported their NIC information and female borrowers were less likely to default. When female borrowers repaid their loans, they did so on time. The finding of the higher repayment rate by female borrowers is consistent with those of other empirical studies on microfinance in South Asia (Armendariz and Morduch 2010). This may reflect difference in preferences and alternative sources of credit availability. The result pertaining to NIC information shows that NIC information made it more difficult for a defaulting borrower to reapply for a loan in the future. In other words, the enforcement of the contingent renewal rule was imperfect under the old system.

Table 4.  Borrower-Level Defaults/Delays and Borrowers' Characteristics
 Probit: DefaultTwo-stage Heckman: Avg_delay
CoefficientdF/dx
  1. Notes: 1. “Savings” and “Number of CO members” were not reported for approximately 15% of the sample COs. In such a case, “Dummy for missing CO records” takes the value of 1, and the means of CO's savings and the number of members are included. The reported standard deviations for “Savings” and “Number of CO members” are based on the subsample for which these two variables were available.

  2. 2. For the two-stage Heckman estimation, the first-stage model is the one reported in this table for Default. In other words, the model is identified only through the nonlinearity of the inverse Mills ratio.

  3. ***, **, * represent statistical significance at the 1%, 5%, and 10% levels, respectively. Standard errors are in parentheses.

Borrowers' individual characteristics:
Dummy for the availability of NIC information−0.6548***−0.2513−142.41***
(0.0901) (36.15)
Dummy for a female borrower−0.4779***−0.1879−78.40***
(0.0772) (25.86)
Dummy for CO chairman or secretary−0.1922−0.0764−3.52
(0.1383) (26.72)
Borrower households' characteristics:
Number of income sources of the household−0.3898***−0.1555−49.82***
(0.0464) (17.32)
Dummy for income sources outside the region−0.3664***−0.1447−47.51*
(0.1154) (27.49)
Dummy for a female-headed household−0.8122***−0.3004−7.66
(0.1757) (49.25)
Dummy for a joint family−0.0632−0.025219.76
(0.0714) (14.72)
CO characteristics:
CO's savings (in Rs.100,000)−0.4560*−0.1818−40.51
(0.2430) (51.90)
Number of CO members0.00100.00041.57***
(0.0027) (0.51)
Dummy for missing CO records0.2393**0.094829.48
(0.1026) (23.74)
CO's age in days at the time of loan issue0.00064***0.000260.080**
(0.00010) (0.038)
Inverse Mills ratio  −212.31**
  (82.86)
Union fixed effectsYes Yes
Date of credit issued: linear, quadratic, cubic, quarticYes Yes
Number of observations used in the estimation2,275 1,119/2,275
χ2 (26), χ2 (49) for zero slope531.4*** 603.2***
Pseudo R20.1685  

Second, to control the resource availability of a borrower, several household-level variables characterizing household income sources are included and all have significant coefficients. Households with more income sources and households with income originating outside their residential areas were less likely to default, and when borrowers from these households repaid loans, they did so on time. This shows the importance of income diversification in avoiding default.

Last, several variables characterizing the CO to which the borrower belonged have significant effects on repayment. COs' savings have a negative coefficient, implying that either the savings accumulated at the CO serves as a buffer against default or borrowers in COs with higher savings tend to be safer investors. COs with incomplete records were more likely to default. The age of the CO at the time when the loan was issued has a positive coefficient, suggesting that older COs are less successful at repayment, which in turn points to group fatigue (Sharma and Zeller 1997). Among the CO characteristics, the number of CO members has a positive coefficient and is statistically significant in the Avg_delay regression. We attribute this result to the difficulty faced by larger COs in coordinating within the group and the need to avoid the free-rider problem.14

The regression results in Table 4 thus appear to indicate a limited role played by COs in preventing default and repayment delays; instead, our field observations indicate that COs functioned as a collusive forum that was counterproductive under the old system. This inference will be further investigated in the next section.

III. REPAYMENT DYNAMICS, INTRA-GROUP INTERACTIONS, AND STRATEGIC DEFAULT

A. Example

To motivate our empirical strategy to investigate the mechanism underlying frequent defaults under the old system, we first show an example of repayment dynamics within a CO. Figure 2 plots the pattern of repayment for five sample borrowers who belonged to the same CO and borrowed under the same conditions. They borrowed Rs. 20,000 on November 7, 2000, which was to be repaid in 18 monthly installments after a grace period of one month. In the figure, the cumulative amounts of repayment are plotted for the borrowers, and the scheduled repayment is traced in a line with solid diamond dots. For each borrower, a marker on the curve represents the cumulative amount repaid at that date.

Figure 2.

Examples of Repayment Dynamics in a CO (#415)

If the curve lies below the scheduled line, it implies a delay in repayment. The figure shows that borrowers 401, 403, and 405 made delayed repayments, while borrowers 402 and 404 paid before the due date (Avg_delay for these two borrowers is negative). If a curve plateaus before reaching the level of Rs. 23,490 (the total amount that should be repaid), it implies that the loan was defaulted. From the figure, borrowers 401, 403, and 405 can be said to have defaulted (Default for these three borrowers is 1; hence, Avg_delay is not defined). Borrower 405 repaid the amounts corresponding to the first 10 installments, although the repayments were delayed. Therefore, the installment-level variables for borrower 405 for the first 10 installments are calculated as No_repay= 0 and Delay > 0, and for the last eight installments as No_repay= 1, and therefore Delay is not defined for the last eight installments for borrower 405.

Figure 2 provides two interesting insights. First, borrowers with repayment problems have linear or slightly concave curves. This implies that once a borrower makes a delayed repayment, it becomes difficult for him/her to catch up with the scheduled repayment by the next installment; instead, it is more likely that the borrower will be delayed again by a similar or longer margin. In other words, Delay appears to be positively autocorrelated.

Second, the curves for some of the borrowers closely resemble each other. When borrower 402 repaid more than the required amount for the third installment, borrower 404 followed suit. Further, borrowers 401 and 405 missed the first two due dates but paid the amount due for one installment on the due date of the third installment, and while they repaid an amount when the fourth installment was due, it was not enough to clear their outstanding dues. In other words, Delay appears to be positively correlated among the CO members (we call this peer correlation).

Since members of a CO reside in the same village, they are subject to village-level covariate shocks. If it can be shown that borrowers 401 and 405 (or borrowers 402 and 404) suffered from such covariate negative shocks, peer correlation between them is to be expected. That is, it can be expected that microenterprise project returns will be more correlated if borrowers invest in similar investments. However, the actual investments do not appear to match our expectations if we use the adoption of similar projects as the crude proxy for higher covariate risk: borrower 402 invested in a small grocery shop, while borrower 404 invested in buffaloes; further, borrower 401 invested in buffaloes, while borrower 405 invested in a small grocery shop. This does not explain the observed peer correlation by village-level covariate shocks.

B. Empirical Strategy

By analyzing more cases and holding discussions with program participants and MFI officers in the field, we interpret that the mismatch mentioned above may be reflective of a kind of strategic default. Because the dynamic incentives were weak under the old system, if one group member was hit by a negative shock and faced difficulty in repayment, the other members who were able to repay might have decided to default themselves as well, instead of helping the unlucky person make his/her repayment.15

Using our detailed dataset at the installment level, we demonstrate indirect evidence for this interpretation. More concretely, we compare the extent of peer correlation associated with CO-level shocks for loans made under the old system and under the new system. If repayment follows the individual liability principle, most of the peer correlation is attributable to CO-level shocks. On the other hand, if repayment follows the imperfect joint liability principle, a significant part of the peer correlation should remain even after controlling for CO-level shocks. In other words, we examine whether a borrower's delay in installment repayment was correlated with other group members' repayment delays beyond the level explained by possible correlation of project failures due to locally covariate shocks.

We begin with an installment-level model for Delay under the old system, because most cases of No_repay= 1 occur after several instances of repayment delay. By definition, for each borrower, the variable No_repay switches only once from 0 to 1, after which it continues to take the value of 1. Therefore, after analyzing the determinants of Delay, we analyze the determinants of transition probability from No_repay= 0 to No_repay= 1.

To focus on the installment-level dynamics and to avoid additional complications caused by differences in repayment schedules, we use the subset of our installment-level data with 18 monthly installments. Out of 36,777 installments under the old system (see Table 2), 17,316 installments comprise the subset used in the analysis below. A larger subset is used in the robustness check. Each credit contract is denoted by subscript i and the order of its monthly installment isdenoted by subscript t (t= 1, 2, . . . , 18). The following model is estimated to investigate the determinants of Delayit:

image(1)

where DelayCit is the average over j of Delayjt, and j is a member of the CO to which borrower i belongs, is a borrower whose due date is the same as that of borrower i, and has not defaulted by t. In calculating DelayCit, borrower i is excluded to avoid the spurious correlation between Delayit and DelayCit due to construction. This variable captures the extent of peer correlation. ui is a borrower-specific unobservable factor, ut is an unobservable factor specific to the installment order, eit is an independently identically distributed (iid) error term, and a1 and a2 are parameters to be estimated.

Equation (1) is estimated by a two-way fixed effect (FE) panel regression model, with individual credit contract as “group” and the installment number as “time” for the FE specification. The borrower-level determinants of repayment such as gender, NIC information, income sources, CO characteristics, and others, which have already been investigated in the previous section, are jointly controlled by ui. In addition, ui controls the macroeconomic factors because each credit contract is associated with the date of the credit issue so that the i-fixed effect controls unobserved factors associated with the date. The common dynamics in a repayment cycle (e.g., a borrower may have a higher motivation to repay on time for the first and the last installment, but may have lower motivations for payment of installments in between) is controlled by ut.

Since DelayCit does not fully reflect the information of No_repayjt, a variant of equation (1) is also estimated:

image(2)

where ProblemCit is defined in a way similar to DelayCit, but using Problemjt (dummy for nonrepayment until 31 days after the due date; see Table 2). To investigate the determinants of transition probability from No_repay= 0 to No_repay= 1, we compile a variable No_repayCit, which is calculated from No_repayjt in a way similar to DelayCit. We then estimate the following equation for the subsample with No_repayi,t-1= 0:

image(3)

Equations (2) and (3) are also estimated by a two-way FE model, with individual credit contract as group and the installment number as time. In all three equations, parameter a1 captures the extent of autocorrelation, and parameter a2 captures the extent of peer correlation.

If the estimate for a2 is significantly positive, it implies the existence of peer correlation. It is possible, however, that peer correlation simply reflects the ill-effects of covariate shocks that hit microenterprise projects run by microcredit borrowers. To examine this possibility, we divide the sample observations into those belonging to a CO with homogeneous microenterprise projects and those belonging to a CO with heterogeneous microenterprise projects. If the main reason for the significance of a2 is the covariance of microenterprise project returns, we expect the estimate for a2 among homogeneous groups to be larger than that among heterogeneous groups. If the estimate for a2 is similar for both groups, we interpret that peer correlation stems mainly from strategic default under imperfect joint liability.

For this investigation, we compile a CO-level variable, Homogeneity, which is a dummy variable for the same project purpose.16 It is an indicator of the homogeneity of microenterprise projects. If Homogeneity= 1, the borrowers are presumably subject to greater covariate risk in microenterprises than if Homogeneity= 0. We split the sample on the basis of Homogeneity and estimate equations (1), (2), and (3) using smaller subsamples. We then compare the estimates for a2, testing the statistical significance of the difference using regression based on the pooled sample with a full set of cross-terms of explanatory variables with Homogeneity.

We estimate an installment-level model similar to equation (1) under the new system as well.17 Because there was no instance of default under the new system, equations (2) and (3) are not applicable. We again compare the estimate for a2 among borrowers belonging to a homogeneous CO and that belonging to a heterogeneous CO. Finally, we compare the comparison result under the old system and the comparison result under the new system. This is the basic empirical strategy. If our interpretation is correct, the double comparison results will show that the excess peer correlation existed under the old system and disappeared under the new system. The key identifying assumption of our approach is that there was no change in the covariance structure among microenterprises run by borrowers belonging to the same CO. In other words, a caveat of our analysis is that we ignore the possibility that there could be CO-level covariate shocks that were large and covariate enough that they affected borrowers with different businesses equally, and the extent of such covariate shocks may have changed across two periods. We will come back to this assumption in interpreting our results.

Another caveat of our empirical strategy is that since the MFI made various changes in January 2005, we cannot claim that our interpretation is the only interpretation that explains the change in default rates. Based on our field observations, we present our interpretation as the most likely candidate and admit that the econometric evidence is highly indirect. In other words, our aim is to show the existence of such strategic default problems and not to quantify their impact relative to the impact of other factors. Quantifying the impact of other innovations adopted in January 2005, such as the linking of development projects at community level with repayment rates, tighter demand favorable for the MFI due to the credit suspension during the transition period, and the accumulation of borrowers' information leading to the exclusion of risky borrowers, is left for further research.

C. Estimation Results for the Old System

The FE estimation results for equations (1), (2), and (3) are reported in Table 5. Coefficient a1 on Delayi,t-1 is positive and statistically significant in all three equations. The size of the autocorrelation coefficient in equations (1) and (2) is smaller than unity with statistical significance at the 1% level, allowing us to conclude that the dynamic path is stable. Coefficient a2 on the peer-average variable is also positive and significant in all three equations. This means that peer correlation exists. The size of these coefficients is rather large. A 10-day delay among the peers led to a delay of 3.19 days; a 10 percentage-point increase in the number of nonrepaying borrowers until 31 days after the due date led to a delay of 3.06 days; and a 10 percentage-point increase in the number of fellow borrowers with No_repay= 1 increased the probability for a borrower to reach the status of No_repay= 1 by 4.89 percentage points. The regression results thus confirm the existence of strong auto and peer correlation.

Table 5.  Installment-Level Dynamics of Delays and Defaults
 Determinants of DelayProbability of Transition from No_repay= 0 to No_repay= 1: Equation (3)
Equation (1)Equation (2)
  1. Notes: 1. All three models are estimated by a two-way fixed effect panel regression model, with individual credit contract as “group” and the installment number as “time” for the fixed effect.

  2. 2. F-statistics for zero slope have degrees of freedom (dof) at F(18, 11714) for the determinants of Delay and F(18, 12157) for the probability of transition. F-statistics for all ui= 0 have dof at F(897, 11714) for the determinants of Delay and F(916, 12157) for the probability of transition. F-statistics for all ut= 0 have dof at F(16, 11714) for the determinants of Delay and F(16, 12157) for the probability of transition.

  3. 3. The subsample of installment-level data of borrowers associated with 18 monthly installments is used.

  4. *** represents statistical significance at the 1% level. Standard errors are in parentheses.

Lagged value of Delay: Parameter a10.736***0.842***0.00019***
(0.007)(0.006)(0.00001)
Peer effects: Parameter a2
Peer average of Delay0.319***  
(0.008)  
Peer average of Problem 30.571*** 
 (2.133) 
Peer average of No_repay  0.4886***
  (0.0091)
Total number of observations12,63012,63013,092
Total number of borrowers898898917
R2 within0.7120.6810.260
R2 between0.9140.9460.384
R2 overall0.8170.8130.219
F-statistics for zero slope1,607.88***1,389.62***237.02***
F-statistics for all ui= 02.48***1.88***2.39***
F-statistics for all ut= 01.355.57***2.21***

The size of the peer correlation parameter can indicate the upper limit of the positive correlation among borrowers who delay/default owing to covariate shocks. To examine the extent to which the positive peer correlation can be attributed to strategic default, we split the sample by the value of Homogeneity and reestimate equations (1), (2), and (3) using smaller subsamples. The summary regression results in Table 6 show that the estimates for a2 are statistically significant in all specifications and their coefficients are very similar regardless of the choice of the sample. Coefficient a2 in equation (1) is 0.320 for heterogeneous COs where some members invested in projects different from others, while it is 0.337 for homogeneous COs where all members invested in the same project. The difference is statistically insignificant. Similarly, the value of coefficient a2 in equation (2) is 29.7 for heterogeneous COs and 30.7 for homogeneous COs. The difference again is statistically insignificant. In the No_repay regression (equation 3), a2 is 0.455 for heterogeneous COs and 0.526 for homogeneous COs. Here, even though the difference is statistically significant, the absolute size of the difference is small enough for us to view the difference as economically insignificant.

Table 6.  Installment-Level Peer Effects and Homogeneity of Microenterprise Projects
 Coefficient a2 (Extent of Peer Correlation) or Its Difference (Addition for Homogeneous Groups)
Separate Regression Results Using Two SubsamplesRegression Results Using Pooled Sample with Cross-Terms of All Explanatory Variables with Homogeneity
Borrowers in a CO with Less Homogeneous Projects (Homogeneity= 0)Borrowers in a CO with More Homogeneous Projects (Homogeneity= 1)
  1. Note: All nine models are estimated by a two-way fixed effect panel regression model, similar to the one in Table 5.

  2. *** represents statistical significance at the 1% level. Standard errors are in parentheses.

Delay on DelayC (eq. 1)0.320***0.337***0.018
(0.011)(0.013)(0.019)
Delay on ProblemC (eq. 2)29.684***30.706***1.022
(2.788)(3.126)(4.591)
No_repay on No_repayC (eq. 3)0.455***0.526***0.071***
(0.012)(0.015)(0.019)
Number of observations8,4934,13712,630

The results in Table 6 do not support the view that the observed peer correlation was mainly due to the covariate shocks that hit borrowers; instead, they appear to corroborate the view that strategic default was (at least partially) responsible for peer correlation. To check the robustness of this finding, we attempt several different specifications. First, instead of splitting the sample by Homogeneity, we compile the Herfindahl index for project purpose (the sum of shares squared) and add its cross-terms with the peer-average variables to the three equations. The Herfindahl index takes the maximum value of 1, and if all borrowers in a CO invest in different projects, the index takes the minimum value of 1/n, where n is the number of members. If peer correlation is attributable to covariate shocks to microenterprises, we expect the coefficient on the cross-term to be positive. The regression results, shown in the last block of rows of Table 7, do not support this expectation. The cross-term has a negative and insignificant coefficient when the dependent variable is Delayit, and a positive and marginally significant coefficient, but of an economically insignificant magnitude, when the dependent variable is No_repayit.

Table 7.  Robustness of the Estimation Results Regarding Peer Effects
Split the sample by dummy variable Homogeneity:Coefficient a2 (Extent of Peer Correlation)Test for the Hypothesis That a2 Is the Same
Less HomogeneousMore Homogenous
Homogeneity= 0Homogeneity= 1
  1. Notes: 1. All models include the lagged value of delay, borrower fixed effects, and time controls (time fixed effects for cases 1, 3, and 4) as explanatory variables in addition to the peer variable.

  2. 2. Case 1: The total number of observations is 25,818 (eq. 1 or eq. 2) and 26,787 (eq. 3). The total number of borrowers is 1,836 (eq. 1 or eq. 2) and 1,881 (eq. 3).

  3. 3. Case 2: Estimated by the system-GMM method proposed by Blundell and Bond (1998). Because of memory problem, the full list of time fixed effects were not included. Instead, the relative position of the installment and its higher order polynomials (to the fourth order) were included. This replacement did not affect the structural parameters for cases 1 and 2 reported in this table. In all specifications, Hansen's J test indicates that the overidentifying restrictions implied by this GMM procedure are not rejected. The AR(2) test for autocorrelation of order 2 indicates that the null hypothesis of no autocorrelation is not rejected.

  4. 4. Case 3: In the instrumental variable estimates, to avoid the simultaneity bias within a borrowers' group, the lagged values of the peer variables and the union-level shock indicators are employed as identifying instrumental variables for the peer variables.

  5. 5. Case 4: Coefficient in the first column shows the one corresponding to the mean level of Herfindahl. Coefficient in the second column shows the one on the cross-term. The average of Herfindahl is 0.697 and its standard deviation is 0.297.

  6.  The null hypothesis is rejected at 1% =***, 5% =**, 10% =*, and not rejected at 10% =“n.s.” When a2 is larger in homogeneous COs than in heterogeneous COs, these are shown without parentheses; when a2 is smaller in homogeneous COs, these are shown in parentheses.

0. Default (see Table 6):
  Delay on DelayC (eq. 1)0.3200.337n.s.
  Delay on ProblemC (eq. 2)29.68430.706n.s.
  No_repay on No_repayC (eq. 3)0.4550.526***
1. Larger sample whose installment number is more than 5:
  Delay on DelayC (eq. 1)0.3180.331*
  Delay on ProblemC (eq. 2)30.45633.989n.s.
  No_repay on No_repayC (eq. 3)0.4590.471**
2. System-GMM estimates treating lagged Delay as endogenous:
  Delay on DelayC (eq. 1)0.2500.209(**)
  Delay on ProblemC (eq. 2)128.61379.677(**)
3. Instrumental variable estimates treating peer variables as endogenous:
  Delay on DelayC (eq. 1)0.0940.103n.s.
  Delay on ProblemC (eq. 2)10.522−3.940(n.s.)
  No_repay on No_repayC (eq. 3)0.2140.204(n.s.)
4. Cross-term with Herfindahl to identify the difference in a2:Linear termCross-term 
  Delay on DelayC (eq. 1)0.338−0.021(n.s.)
  Delay on ProblemC (eq. 2)30.832−1.223(n.s.)
  No_repay on No_repayC (eq. 3)0.3790.146*

Second, instead of limiting the sample to observations associated with 18 monthly installments, we employ a larger subset of observations associated with more than five monthly installments. The analysis may suffer from specification errors owing to differences in repayment schedules, but it can gain in statistical efficiency from a larger number of observations. The results in Table 7 show the same results qualitatively as those in Table 6: Coefficient a2 is slightly larger when observations associated with homogeneous COs are used than when those associated with heterogeneous COs are used, but the difference is not economically significant.

Third, we reestimate equations (1) and (2) using the system generalized method of moments (GMM) approach proposed by Blundell and Bond (1998). In these two equations, the lagged variable Delayi,t-1 is included in the right-hand-side, creating a dynamic panel data (DPD) structure. To avoid the potential endogeneity bias caused by the DPD structure, we employ the system GMM estimation method. The system GMM results in Table 7 show that coefficient a2 is smaller (not larger) when observations associated with homogeneous COs are used than when those associated with heterogeneous COs are used, and the difference is statistically significant. Although a clear inference does not emerge from the results, it is beyond doubt that they do not support the view that the observed peer correlation was mainly due to covariate shocks that hit borrowers' microenterprises.

Fourth, considering the possibility of the reflection problem (Manski 1993), we reestimate the three equations using instrumental variable specifications, treating the peer-average variables as endogenous. Coefficients on the three variables proxying the peer effects, i.e., DelayCit, ProblemCit, and No_repayCit, are identified by the following instrumental variables: their lagged values (DelayCi,t-1, ProblemCi,t-1, and No_repayCi,t-1) and the lagged value of variable proxying union-level repayment problems.18 The regression results in Table 7 show that the extent of peer correlation is reduced and the difference in a2 across the two subsamples distinguished by Homogeneity is insignificant in all three equations. This reconfirms the view that strategic default contributed to peer correlation.

In addition to conducting the robustness checks reported in Table 7, we reestimate the model using Delay as the dependent variable after redefining it as truncated at zero for payments made earlier than their due dates. This is because early payments do not damage MFIs. We also reestimate the instrumental variable specification using two-period lags as the instrumental variables, considering the possibility that a borrower watches not only his/her peer's contemporary behavior but also his/her peer's one-time lagged behavior when deciding on the repayment decision. As a simpler exercise, we also compare peer correlation at the borrower level between those borrowers belonging to a homogeneous CO and those borrowers belonging to a heterogeneous CO. The results are qualitatively the same as those reported so far.19

D. Estimation Results for the New System

Given the severity of the 2005 earthquake in Pakistan, an examination of whether the natural calamity affected repayment patterns is of particular interest. To identify the impact of the earthquake, we adopt a standard difference-in-difference (DID) approach. First, we determine the distance between the borrower locality and the epicenter of the earthquake from the website of the Earthquake Reconstruction and Rehabilitation Agency (ERRA). We then create a borrower-level dummy variable, D_eq, which takes the value of 1 if the household is located within a radius of 75 km from the epicenter of the earthquake, and 0 otherwise. The threshold radius of 75 km is chosen after consultation with seismologists at the Pakistan Meteorological Department. We then compile an installment-level variable, D_t1, which is a dummy variable for installments due after the earthquake, and D_t2, which is a dummy variable for the loan made after the earthquake. Coefficients on the cross-terms of D_eq*D_t1 (identified under both FE and random effects [RE] specifications) and D_eq*D_t2 (identified under RE specifications only) show the impact of the earthquake.

Table 8 reports the estimation results when we do not distinguish the type of COs.20 First, all variables capturing auto and peer correlation have positive and statistically significant coefficients. Second, the impact of the earthquake is not discernible in the installment-level analyses. The coefficients on D_eq*D_t1 are insignificant in both models. The coefficient on D_eq*D_t2, which is identified in the RE specification only, is positive; however, its magnitude is extremely small and statistically insignificant. The absence of the earthquake impact on repayment delay is confirmed robustly. As shown by Kurosaki and Khan (2011), when a borrower-level regression is run, the impact of the earthquake is marginally discernible through the coefficients on D_eq*D_t2, but its coefficient indicates only a delay of four to six days and the results are not robust to the specification of D_eq. When an installment-level regression is run, the DID impact is absent regardless of the specification of D_eq. Therefore, we conclude that the repayment delay was affected by the earthquake only marginally at best and the impact was not robust.

Table 8.  Installment-level Dynamics of Delays under the New System and the Impact of the Earthquake
 Fixed Effect (FE) EstimationRandom Effect (RE) Estimation
  1. Notes: 1. Both models are estimated with individual borrower as a “group” for the fixed (random) effect and with the installment number as the fixed time effect.

  2. 2. The effects of borrower-level variables including D_eq, D_t2 and D_eq*D_t2 are identified in the random effect specifications only.

  3. 3. “Statistics for zero slope” are F(14, 6156) and Gaussian Wald χ2(25). “F-statistics for all ui= 0” are F(616, 6156). “Statistics for all ut= 0” are F(10, 3105) and χ2(10).

  4. 4. The subsample of installment-level data of borrowers associated with 12 monthly installments is used.

  5. *** and ** represent statistical significance at the 1% and 5% levels, respectively. Standard errors are in parentheses.

Controls to identify the earthquake impact:
  D_eq 0.276
 (0.751)
  D_t1: Dummy for an installment due after the quake−0.060−0.259
(0.670)(0.598)
  D_eq*D_t1−0.367−0.235
(0.907)(0.831)
  D_t2: Dummy for the loan made after the earthquake 0.195
 (0.454)
  D_eq*D_t2 0.021
 (0.508)
Own and peer effects:
  Lagged value of Delay0.819***0.983***
(0.010)(0.008)
  Peer average of Delay0.375***0.233***
(0.012)(0.009)
Borrowers' characteristics:
  Dummy for a female borrower −0.549**
 (0.263)
  Dummy for CO chairman or secretary 0.062
 (0.285)
  Number of income sources of the household 0.095
 (0.077)
  Dummy for income sources outside the region 1.948
 (1.272)
  CO's savings (in Rs. 100,000) 2.181
 (1.524)
  Number of CO members 0.004
 (0.015)
  CO's age in days at the time of loan issue −0.0003
 (0.0002)
Total number of observations6,7876,787
Total number of borrowers617617
R2 within0.6220.609
R2 between0.8910.947
R2 overall0.7450.760
Statistics for zero slope722.49***21,350.79***
F-statistics for all ui= 02.09*** 
Statistics for all ut= 016.59***225.55***

This finding does not necessarily imply that the earthquake did not pose any threat to the MFI; rather, it may reflect a change in the lending strategy in that the MFI became more selective about its clients and started monitoring borrowers more thoroughly, thereby undermining the gravity of the delay/default problem. As shown in Table 1, most borrowers' observable characteristics did not change after the earthquake, except for the ratio of households with income originating outside their residential areas (21% before the earthquake and 100% after the earthquake). This suggests that after the earthquake, the MFI tended to lend only to those households with outside income in order to avoid repayment problems.21

To investigate whether significantly positive a2 (the coefficient capturing peer correlation) means that the tendency to default strategically remains unaffected under the new system, we prepare Table 9, in which we examine whether a2 is higher among borrowers who choose similar projects. When equation (1) is reestimated with the sample split by Homogeneity, a striking result emerges: a2 is considerably larger when observations associated with homogeneous COs are used than when those associated with heterogeneous COs are used. The difference is not only statistically significant (the significance level is less than 0.1%), but also economically significant. The results under the default specification show that under the new system, coefficient a2 is 0.292 in COs where some members invested in projects different from those of the others, while it is 0.656 in COs where all members invested in the same project.

Table 9.  Peer Effects and Homogeneity of Microenterprise Projects under the New System
Split the sample by the dummy variable Homogeneity:Coefficient a2 (Extent of Peer Correlation)Test for the Hypothesis that a2 Is the Same
Less HomogeneousMore Homogenous
Homogeneity= 0Homogeneity= 1
  1. Notes: 1. All models include the lagged value of delay and time controls (time fixed effects for cases 1, 3, and 4) as explanatory variables in addition to the peer variable.

  2. 2. Case 0: A pooled sample of 6,787 is divided into Homogeneity = 0 (4,752 observations) and Homogeneity = 1 (2,035 observations).

  3. 3. Case 1: The total number of observations is 7,452, divided into Homogeneity = 0 (5,308 observations) and Homogeneity = 1 (2,144 observations).

  4. 4. For cases 2, 3, 4, see notes in Table 7.

  5. † Figures in parentheses shows standard errors. The null hypothesis is rejected at the 1% level, which is denoted by ***. In all cases, a2 is larger among more homogeneous COs.

0. Default
  Delay on DelayC (eq. 1)0.2920.656***
(0.015)(0.020) 
1. Larger sample whose installment number is 12 or 15:
  Delay on DelayC (eq. 1)0.2980.530***
(0.014)(0.018) 
2. System-GMM estimates treating lagged Delay as endogenous:
  Delay on DelayC (eq. 1)0.1530.279***
(0.021)(0.012) 
3. Instrumental variable estimates treating peer variables as endogenous:
  Delay on DelayC (eq. 1)−0.0410.227***
(0.022)(0.040) 
4. Cross-term with Herfindahl to identify the difference in a2:Linear termCross-term 
  Delay on DelayC (eq. 1)−0.0480.695***
(0.031)(0.046) 

In Table 9, we employ four types of robustness checks, as we did for the borrowers under the old system. When we identify the difference in COs' heterogeneity using the Herfindahl index for project purposes (last block of rows in Table 9), the cross-term has a highly positive and statistically significant coefficient. When a larger subsample is employed, the results are very similar to those under the default specification. When we reestimate equation (1) using the system GMM approach, the coefficients become slightly smaller, although the test for the hypothesis that a2 is the same regardless of project homogeneity is rejected at the 0.1% level. When equation (1) is reestimated with the peer-average DelayCit treated as endogenous,22 peer correlation among the less homogeneous groups disappears, while that among the more homogeneous groups remains positive and statistically significant at the 1% level.

Again, in addition to the robustness checks reported in Table 9, we reestimate the model using Delay truncated at zero for the payments made earlier than their due dates, the instrumental value specification using two-period lags as the identifying instrumental variables, and the borrower-level comparison of peer correlation. The results are again qualitatively the same as those reported earlier in this section.23

E. Interpretation

The results so far robustly demonstrate the contrast between the repayment dynamics under the old system and those under the new system. Under the old system, a borrower's repayment delay for an installment was correlated with other members' repayment delays beyond the level explained by a possible correlation of project failures owing to locally covariate shocks. Under the new system, most of the peer correlation was explained by a possible correlation of project failures of members belonging to the same CO.

A counterargument could be that this reflects the changes in the CO-level covariate shock structure as discussed in subsection B. From our field observations, we received no indication that there occurred a substantial change in the covariance among microenterprises run by microcredit borrowers after early 2005. This appears to suggest that the counterargument does not have a strong appeal. We therefore interpret the contrast between the new and old systems as evidence that the tendency to default strategically reduced under the new system owing to improved dynamic incentives and more frequent repayment schedules. As a result, the new system was able to mitigate the negative impact of the earthquake, a typical example of a large covariate shock.24

Pakistani society historically has had limited experience with cooperative community-based development. It has been marked by the existence of strong local elites (e.g., see Kurosaki 2005), which appears to underlie the limited success of community-based group lending. The contrast shown in this paper indicates a possibility that strategic defaults will prevail in microfinance schemes applied to such a society if the enforcement of joint liability and contingent renewal rules is imperfect.

IV. CONCLUSION

This paper analyzed an interesting case of microfinance in Pakistan in which an MFI successfully overcame the problem of frequent default by adopting a new system with strict enforcement of punishment against delayed repayments. We hypothesized that strategic default under the joint liability mechanism, which was encouraged by weak enforcement of contingent renewal, was one of the factors responsible for failure under the old system. Using a unique dataset of about 45,000 repayment installments covering 2,945 micro-borrower households over the period 1998–2007, we investigated the dynamics of repayment at the installment level. We found that a borrower's delay in the repayment of each installment was correlated with the repayment delays of other members in his/her group, beyond the level explained by possible correlation of project failures due to locally covariate shocks. Although peer correlation was evident under the new system as well, it was better explained by covariate shocks to microenterprises. Therefore, our interpretation is that strategic default occurred frequently and was a serious problem under the old system, while the new system was successful in suppressing collusion among group members. In terms of actual borrowers' data (not experimental data), this finding lends empirical support to the burgeoning literature on microfinance that insists that individual lending schemes are likely to be superior to joint-liability schemes when they are accompanied by dynamic incentives and frequent repayment installments.

Although the study area was hit by a disastrous earthquake in October 2005, the new microfinance system was affected only marginally in terms of repayment delay. This does not necessarily imply that the earthquake did not pose any threat to the MFI; rather, it indicates a possibility that in terms of lending strategy, the MFI became more selective about its clients and started monitoring borrowers more thoroughly. If borrowers in the earthquake-hit region faced stricter selection or monitoring after the earthquake, then it can be said that they suffered not only on account of the natural disaster but also from the inflexible repayment requirements of the MFI—an inference that is corroborated in a similar finding reported by Shoji (2010) in the case of floods in Bangladesh in 2004. The net impact of the earthquake on borrowers' welfare is a topic that merits further investigation.

A novel implication of these findings in the context of understanding microfinance pertains to the concept of covariate shocks. Even in cases where shocks to microenterprises caused by borrowers of the same group are purely idiosyncratic from the viewpoint of an individual borrower, their effect on MFIs may be similar to that of covariate shocks if the tendency to default strategically exists. This paper shows that widespread strategic default affects the sustainability of microfinance more adversely than a purely covariate negative shock such as the 2005 Kashmir earthquake.

Footnotes

  • 1

    Sequential financing refers to group loans that are staggered within the same round. Contingent renewal indicates the exclusion of defaulting borrowers from future access to loans. Dynamic incentives imply provision for larger loans to borrowers who successfully repay. Contingent renewal and dynamic incentives in this sense are sometimes combined as dynamic incentives in a broad sense. Public repayments refer to a system where borrowers make repayments to loan officers in the presence of other borrowers during the officers' visit to the concerned village.

  • 2

    The focus here is on empirical studies that attempt to establish an explicit link to theories. There are a large number of empirical and insightful studies that estimate the reduced-form determinants of group repayment behavior among microfinance borrowers. Among such studies, Wydick (1999) is one of the more influential ones as it employs the most extensive list of proxy variables to measure screening, monitoring, and enforcement within groups.

  • 3

    See Cassar, Crowley, and Wydick (2007), McIntosh (2008), and Field and Pande (2008) for other applications of the experiment approach to analyze the mechanism underlying the repayment behavior.

  • 4

    In April 2010, the constitution of Pakistan was amended and the former NWFP was renamed “Khyber Pakhtunkhwa.” In this paper, since all data correspond to a period before this constitutional amendment, the expression NWFP is used to refer to the current province of Khyber Pakhtunkhwa.

  • 5

    These borrowers obtained loans on October 6, 2005, two days before the earthquake.

  • 6

    In this paper, “Rs.” refers to nominal Pakistani rupees. The exchange rate between Pakistani rupees and US dollars was stable at around Rs. 60/$ during the study period. The government estimates for inflation rates inside Pakistan were in the range of 3.1% to 7.9% per annum.

  • 7

    Under Islamic law, no interest is charged on microfinance loans but borrowers have to pay service charges or process fees in addition to the principal. In effect, these charges are equivalent to interest (the interest rate charged on these loans was 20% per annum under the old system).

  • 8

    Our dataset updated in July 2009 includes information for those borrowers who borrowed after our initial data collection in August 2006 and were still in the repayment schedule in July 2009. To avoid the right-censoring problem, this paper employs the subset of our dataset that corresponds only to the borrowers included in the initial data collected in August 2006.

  • 9

    For loans provided under the old system, the due date of the last installment for the 2,275 borrowers ranged from July 7, 1999 to October 4, 2004. Given that our data were collected more than 57 months after the last of these due dates, it is safe to ignore the possibility of future repayments. Therefore, the right-censoring problem for the old loans can be ignored.

  • 10

    The borrower with Avg_delay at −552 borrowed Rs. 20,000 on September 2, 2002; this amount was due on April 2, 2004, without installments. However, since the borrower repaid the entire amount on September 28, 2002, i.e., 552 days before his due date, his Avg_delay was −552 days.

  • 11

    All five borrowers whose repayment records were completely missing belonged to the same borrower group and borrowed just two days before the earthquake. They lived in an area 106 km from the epicenter of the earthquake. With these five borrowers added to the list of repayment delays or defaults, the adverse impact of the earthquake on repayment would have been slightly more serious than what is indicated in Table 2.

  • 12

    The location fixed effects control unobservable heterogeneity that is specific to a region. In Pakistan, a union is the smallest administrative unit that comes under the jurisdiction of the union council. Usually, 10–20 villages comprise a single union council. We include 10 dummies corresponding to these unions. Unobservable macro effects are controlled by the date of issue of each loan. In one specification, the date of issue (measured in the number of days elapsed since May 1, 1998) and its higher order polynomials are adopted. In another specification, year dummies corresponding to the date of issue are included.

  • 13

    A recent study based on a randomized experiment in India by Field and Pande (2008) shows no differences between microfinance schemes with weekly and monthly repayment frequencies. Combining their findings with ours, it appears that monthly repayment is the optimal frequency that can be stipulated to avoid repayment delays and reduce the transaction costs of too frequent repayment.

  • 14

    In a different context in Pakistan, Kurosaki (2005) reports a positive effect of the size and diversity of a community-based organization (CBO) on the level of collective action. His interpretation is that while larger and more diverse CBOs may face difficulty in coordinating within the group and avoiding the free-rider problem, the advantage they enjoy in terms of technical skills required for the particular type of collective action he analyzes outweighs the difficulty.

  • 15

    See Kurosaki and Khan (2011) for an example of a mathematical model showing this. Their theoretical model predicts that when project outcomes of members in the same group are uncorrelated, there should be no peer correlation under the individual liability system, perfect correlation under the joint liability with perfect enforcement, and the correlation coefficient is between 0 and 1 under imperfect joint liability. It also predicts that the extent of peer correlation increases with the frequency of strategic default.

  • 16

    To compile this variable, project purposes are classified into 34 categories. For instance, “Livestock” shown under “Purpose of borrowing” in Table 1 is disaggregated into four categories of buffalo, cow, goat and sheep, and poultry.

  • 17

    One complication for the empirical model under the new system is the control for the earthquake effects. We explain this in discussing the empirical results.

  • 18

    The union-level repayment problem variable is defined as the fitted value of ProblemC on union fixed effects and year-month fixed effects for the repayment due dates.

  • 19

    The estimations results are available on request.

  • 20

    We use the subset of our installment-level data corresponding to observations with 12 monthly installments. Out of 8,154 installments under the new system (Table 2), 7,416 installments comprise the subset.

  • 21

    The increased share of borrowers with income sources outside the region may not only reflect the selection by the MFI but also the changes in (potential) borrowers' behavior after the earthquake. We do not, however, have detailed information on change in income sources. Based on our field observations, we interpret that selection effects dominated.

  • 22

    Identifying instrumental variables are DelayCi,t-1 and the lagged value of the union-level repayment problem variable. The union-level repayment problem is defined as the fitted value of DelayCit on union fixed effects and repayment-month fixed effects.

  • 23

    The estimations results are available on request.

  • 24

    The very small impact of the earthquake on the overall repayment delay could be attributable to the possibility that the earthquake damage was easy to verify so that relief from outside gave the MFI more ability to mitigate the earthquake's impact, while the damage due to CO-level shocks was difficult to verify so that the MFI was not able to insure the damage. This is only a conjecture, and needs to be tested empirically using different datasets from the one used in this paper.

Ancillary