• The author is grateful for financial support from the Foundation for Advanced Studies on International Development. The author would also like to express his gratitude to the editor and two anonymous referees of the journal; and Yasuyuki Sawada, Hidehiko Ichimura, and Amy Ickowitz for helpful and constructive suggestions. The paper also benefited from the comments of Yutaka Arimoto, Koichi Fujita, Masayoshi Honma, Hiro Ishise, Takahiro Ito, Hisaki Kono, Sarah Pearlman, Takeshi Sakurai, Chikako Yamauchi, Junfu Zhang, and participants at the Eastern Economic Association 2007, the International Atlantic Economic Conference 2007, and seminars at Clark University, Tsukuba University, and the University of Tokyo (Department of Economics and Department of Agriculture). Any errors and omissions are solely the responsibility of the author.


Frequent and strictly scheduled repayments and savings in microfinance often deteriorate the liquidity of members in the face of negative shocks. Previous articles suggest the introduction of a contingent repayment system that allows such members to be rescheduled, but the unavailability of a suitable dataset makes it difficult to examine how it would actually work. This study is one of the first to evaluate the impact of this repayment system on household livelihood. In employing a unique dataset from Bangladesh, I show that rescheduling reduces the possibility of binding credit constraints and borrowing from moneylenders, and may also reduce transitory poverty. However, short-term rescheduling has insignificant effects. Indebted members with less liquid assets are more likely to be rescheduled.


This study evaluates the impact of contingent repayment systems in Microfinance Institutions (MFIs). A distinction of the standard MFI loans from other credit sources is the frequent and strictly scheduled repayment structure. Ashraf, Karlan, and Yin (2006) show that this plays the role of a commitment device for borrowers. Other studies also investigate the impact of repayment frequency on repayment rates.1 On the other hand, the negative impact of frequent repayment on household livelihood is also found; it potentially deteriorates borrowers' liquidity and causes credit constraints when they are affected by negative shocks (Zeller et al. 2001). This could cause them to face further poverty. For example, Coleman (1999) finds that MFI members would rather borrow from informal moneylenders than others.

In the face of this trade-off between the pros and cons of frequent repayment, previous articles suggest the introduction of a contingent repayment system that allows rescheduling of installments and savings only for, for example, disaster-affected members (Ledgerwood 1999; Norell 2001; Meyer 2002; Park and Ren 2001). However, few previous studies empirically examine how the system actually works, partially because of a lack of available data; only a few MFIs introduce contingent repayments.

The goal of this paper, therefore, is to use a unique dataset from Bangladesh to examine whether contingent repayment mitigates the repayment burden and plays the role of a safety net during negative shocks. MFIs in Bangladesh have been introducing the contingent repayment system since 2002 (Dowla and Barua 2006; Meyer 2002), and this paper uses data collected after a nationwide flood in 2004, the first case in which most MFIs in Bangladesh rescheduled installments and savings. The data report that 39% of MFI members were rescheduled during the flood.

Rescheduling potentially plays the role of a safety net, but its impact is an empirical one: rescheduling does not change the permanent income of beneficiaries, only the inter-temporal resource allocation. Thus, the treatment effect of rescheduling is larger for credit-constrained households. Its impact may also depend on the duration of rescheduling. Given this, I examine the following three questions. First, do MFIs allow rescheduling particularly for members suffering from poor liquidity? Second, how much does rescheduling reduce the possibility of binding credit constraints? Third, how much does the impact change with the duration of rescheduling?

To preview the results, I find that rescheduling mitigates the possibility of credit constraint by 20–27 percentage points. However, short-term rescheduling has insignificant effects. At least two weeks of rescheduling was required to ensure a significant impact, but one-third of rescheduling beneficiaries were allowed rescheduling for only one week. Finally, members indebted and holding less liquid assets were more likely to be rescheduled.

In a previous study Shoji (2010) also focuses on evidence from the 2004 flood to show that rescheduling helped affected individuals ensure food availability during the disaster. The present study has three distinctions. First, it uncovers how rescheduling plays the role of a safety net, while Shoji's (2010) study evaluates how much it does so. Rescheduling can, first of all, improve welfare by mitigating credit constraints and helping MFI members smooth consumption. Although the present study focuses on this channel, there could be other channels as well.2 The second distinction is that this study pays attention to the changes in the impact with the duration of rescheduling. These are important because they are directly related to policy implications. Third, this study employs the difference-in-differences matching estimator of Heckman, Ichimura, and Todd (1997) and the matching estimator with multiple treatments of Imbens (2000), while Shoji (2010) uses the maximum likelihood model. The matching estimation is preferable for a number of reasons in this context.

Furthermore, this study attempts to make three original contributions to the literature. It is one of the first to examine the contingent repayment system in MFIs. Second, it utilizes a direct indicator of credit constraint following Jappelli (1990) and Boucher, Guirkinger, and Trivelli (2009). The use of this indicator addresses concerns regarding the approximated indicators used in previous studies, such as Zeldes (1989) and Foster (1995). Finally, although some studies find an insignificant poverty reduction effect of MFIs based on the standard repayment system (Coleman 1999), the findings in this study imply that the introduction of contingent repayment may alleviate transitory poverty by reducing credit constraints.3

This paper is organized as follows. The first part of Section II describes the contingent repayment structure of Bangladeshi MFIs and floods, and the second part describes the dataset. In Section III, the empirical methodology is introduced, while Section IV discusses the findings. Section V examines the changes in the rescheduling effect with the duration. Finally, Section VI concludes the paper.


A. The 2004 Flood and Rescheduling of Installments in MFIs

One feature of standard loans in Bangladeshi MFIs is the frequent and strictly scheduled repayment system (Khandker 1998; Mondal 2002): once a MFI member borrows from his/her MFI, the amount to be repaid is divided into approximately 50 to 60 weekly installments. S/he is required to pay tightly scheduled weekly installments beginning soon after the loan disbursement.

Also, members must deposit money into a savings account of their MFIs every week, regardless of whether they are indebted, and borrowers who have defaulted on their loans are excluded from future access to credit (Armendariz de Aghion and Morduch 2005). In the standard system, the joint-liability groups make the repayment schedule more flexible because borrowers facing repayment difficulties can ask other members in the group to cover their repayment (Townsend 2003). This system, however, does not work during covariate shocks such as natural disasters.

The nature of this standard repayment structure raises the demand for a safety net available for MFI members during covariate shocks because Bangladesh is a flood-prone country. Its geographical location, deforestation, and subtropical monsoon climate cause yearly floods (Khan and Seeley 2005), the severity of which is hard to predict. The flood in 1998, for example, inundated around two-thirds of the land, negatively affected the economy, and burdened MFI members with repayment (del Ninno et al. 2001).

Learning from the 1998 flood, MFIs in Bangladesh have been introducing a contingent repayment structure since 2002. This structure allows rescheduling of weekly savings and installments during disasters without charging additional interest. Indebted MFI members are allowed to reschedule both savings and loan installments, while those who do not have debt postpone only savings. MFIs have also switched loan contracts from joint liability to individual lending (Dowla and Barua 2006).

The first nationwide flood since the introduction of the contingent repayment occurred in July 2004, inundating 39 out of 64 districts of the country. Since the flood started during the planting of the main crop in the rainy season, it affected the harvest that was expected in December. Consequently, households became worried about earning income persistently, even though the floodwater had receded by the end of September.

MFIs postponed collecting weekly savings and debt installments when the flood started. Rescheduling was targeted on members who had difficulty in attending the member meetings, and in paying for weekly savings and installments. However, MFIs did not use any concrete criteria, such as asset holdings, to choose beneficiaries of rescheduling. Officers in affected branches visited each member's residence and determined whether rescheduling should be applied. Where the flood damage was severe and it was dangerous for officers to visit, they abandoned efforts to visit the members and allowed them to reschedule (Shoji 2010). This approach makes better use of the limited financial resource of MFIs rather than rescheduling all loans in affected areas, but it requires officers to visit all affected MFI members during disasters to assess flood damage, which incurs significant administrative and monitoring costs.

Rescheduling was important for MFI members, particularly at the beginning of the flood. The Bangladesh government also initiated the Vulnerable Group Feeding (VGF) and Gratuitous Relief (GR) programs that aimed to provide victims with food and agricultural inputs, such as seed and fertilizer. However, mostly they were not implemented until September and October, two months after the flooding began.

B. Data Description

This study uses a unique dataset. A key feature of the data is that it includes information on rescheduling collected using MFI members' bankbooks. The use of bankbooks alleviates the possibility of recall bias, which is common in retrospective surveys. The second distinction is the availability of data on a direct indicator of credit constraint following Jappelli (1990) and Boucher, Guirkinger, and Trivelli (2009). The use of this indicator addresses concerns regarding the approximated indicators used in previous studies such as Zeldes (1989) and Foster (1995).

This dataset is a follow-up survey of a dataset of the International Food Policy Research Institute (IFPRI) conducted in 1998, 1999, and 2004 that examined the 1998 flood (del Ninno et al. 2001). The IFPRI dataset followed a multistage stratified random sampling methodology for seven districts that were selected according to their economic status and the intensity of the flood in their region: Chadpur, Manikganj, Magura, Barisal, Sunamganj, Narsingdi, and Madaripur. In the second stage, IFPRI randomly sampled one Thana from each district and three unions from each of those Thanas.4 In the next stage, about six villages from each union and two clusters from each of the villages were randomly selected. Approximately three households from each cluster were chosen depending on the village size.

The data in this paper was collected in December 2005 from three out of the seven IFPRI-survey districts based on flood severity, poverty level, geographical properties, and the MFIs' diffusion: Chadpur, Manikganj, and Magura. This survey succeeded in interviewing 326 out of the 335 households that IFPRI interviewed in these three districts in 2004.5 In the December 2005 survey, retrospective information was collected, based on recall, for four subperiods preceding December 2005: mid-January to mid-July 2004, mid-July to mid-November 2004 (during the flood), mid-November 2004 to mid-July 2005, and mid-July to December 2005.6 From this retrospective information, a pseudo-panel dataset was compiled. This paper uses only observations that include a MFI member in the household. The questionnaire was designed to collect data on flood intensity, demographics, labor and nonlabor income, asset holdings, savings, credit constraints, MFI membership, rescheduling, and food consumption.

In this study, the term “credit constraints” refers to the excess demand for consumption and investment credit with respect to the overall market, including formal and informal lenders. Rescheduling is expected to reduce the demand for credit, mitigating the credit constraints. The questionnaire for credit constraints is summarized in Figure 1. Households were defined as facing credit constraints either if they borrowed money but could not borrow as much as they wanted, or if they did not borrow from any sources because of rejection of credit applications, fear of default, or lack of available credit sources. Households were credit unconstrained when they borrowed the required amount, or when they did not borrow because they did not have to. While such a module is desirable, it is not available in usual household surveys (Scott 2000). Therefore, previous studies use the amount of landholding or the income–assets ratio to approximate the extent of credit constraint (Zeldes 1989; Foster 1995). However, it is unlikely that a single variable can sufficiently approximate consumers' access to credit (Garcia, Lusardi, and Ng 1997).

Figure 1.

Questionnaire Design for Credit Constraint Module

C. Summary Statistics

Table 1 illustrates the change in livelihood of MFI members through the survey periods. First, MFI members in the sample did not drop out from their MFIs after the flood, implying no incidences of default. The flood did not severely affect the solvency of the borrowers. Furthermore, the number of MFI members in the sample households increased after the implementation of rescheduling. Second, labor income during the flood was lower by 25% than during nonflood periods. Food consumption also declined but was relatively smooth when compared to income fluctuation. Third, more than 70% of members faced binding credit constraints during the flood, and even more were constrained after the disaster. This is probably because of the persistent impact of the flood: demand for nonfood expenditure such as housing repairs may have increased after the flood, although the dataset does not include the information on nonfood consumption. Fourth, people borrowed from moneylenders during the flood more than during other periods, while interest-free informal credit was not remarkably high. This presumably reflects the unavailability of credit from sources other than moneylenders during the flood.7

Table 1.  Summary Statistics of MFI members by Period
PeriodJan–July 2004July–Nov 2004 Flood PeriodNov 2004–July 2005July–Dec 2005
  1. Note: Standard deviations are in parentheses.

Labor income (103 Tk/month)3.202.363.003.17
Food consumption (103 Tk/month)2.612.442.662.87
Dummy if binding credit constraint0.630.710.820.90
Loan from moneylenders (Tk/month)4.73108.11105.0384.92
Interest-free informal credit (Tk/month)0.0025.6864.59100.89
Dummy if rescheduling installments/savings0.000.390.020.09
No. of MFI members141148174179
Amount of rescheduling (Tk/period)0.00489.53311.50263.00
Duration of rescheduling (weeks)0.002.721.001.20
No. of rescheduled members058217

Table 1 also indicates that 39% of MFI members were allowed to reschedule savings and installments during the flood. In addition, the average duration and amount of rescheduling were 2.72 weeks and Tk 490, respectively. The duration ranged from one to eight installments in the sample areas. The amount of rescheduling was approximately 5.2% of labor income, given that the seasonal labor income during the flood period was Tk 9,436. Finally, only a few households rescheduled at the third and fourth periods, probably because some minor MFIs allowed them to do so.

Table 2 compares household characteristics between rescheduled and nonrescheduled members. I use the observations obtained during the second and fourth periods when MFIs implemented rescheduling. Only two households were rescheduled at the third period as the result of a religious festival. These members were asked to repay and save double at the next meeting. It appears that the rescheduled members were more disaster-affected and poorer in terms of asset holdings and income. The differences between the two groups are statistically significant. It is also reported that 91% of rescheduled members experienced a binding credit constraint, which was significantly higher than the corresponding statistic for the nonrescheduled members. Finally, rescheduled members borrowed from moneylenders less than nonrescheduled members, implying that rescheduling had an impact on decreasing credit or the low creditworthiness of the rescheduled.

Table 2.  Summary Statistics by MFIs Membership and Rescheduling Treatment
 RescheduledNonrescheduledMean Difference
  1. Notes: 1. The observations of the second and fourth periods are used.

  2. 2. Standard deviations are in parentheses.

  3. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively.

Labor income (103 Tk/month)1.61(1.55)3.16(5.55)***
Food consumption (103 Tk/month)2.18(0.73)2.82(1.33)***
Dummy if binding credit constraints0.91(0.29)0.79(0.41)***
Loan from moneylenders (Tk/month)32.00(172.74)114.29(688.95)*
Repayment to MFIs (Tk/month)107.79(287.99)202.81(447.27)**
Amount of rescheduling (Tk/month)110.58(114.47)   
Grain storage (103 Tk)0.58(1.03)1.91(3.74)***
Jewelry (103 Tk)2.24(3.36)6.02(10.36)***
Livestock (103 Tk)3.12(6.15)8.25(12.21)***
Owned house (103 Tk)20.09(20.31)31.44(45.26)***
Owned field (106 Tk)80.85(152.40)141.94(298.80)**
Other productive assets (103 Tk)2.31(4.17)6.42(23.94)***
Dummy if own house is inundated0.12(0.33)0.01(0.09)***
Covariate shock indicator5.66(2.51)7.10(3.51)***
Males over 161.69(1.01)2.09(1.21)***
Females over 161.84(0.84)2.12(1.11)**
Children under 162.12(1.26)2.14(1.58) 
Age of head46.72(10.42)48.48(12.14) 
Female head dummy0.11(0.31)0.09(0.29) 
Distance to member meeting0.22(0.21)0.22(0.22) 
Distance to river2.32(2.44)1.87(1.93) 
Distance to school0.70(0.62)0.67(0.44) 
Distance to paved road0.74(0.75)0.62(0.62) 
Dummy if evacuation shelter is available0.19(0.39)0.11(0.31) 
N75 252  


This section describes the methodology used to estimate the rescheduling impact on credit constraints and credit from moneylenders. Credit from moneylenders is considered an indicator of high demand for and poor access to credit: moneylenders charge high interest rates and people borrow from them as a last resort when other credit sources are unavailable (Khan and Seeley 2005; Shoji 2008).

Given the unavailability of the data on a randomized experiment of rescheduling during natural disasters, I use the difference-in-differences matching estimator (DIDM), a type of the propensity score-matching (PSM) model, to control for endogeneity of rescheduling (Rosenbaum and Rubin 1983; Heckman, Ichimura, and Todd 1998). The DIDM particularly controls for time-invariant factors using the prerescheduling period of panel dataset (Heckman, Ichimura, and Todd 1997; Smith and Todd 2005). Another possible approach is the maximum likelihood estimator (MLE) with the endogenous rescheduling treatment variable. However, I use the PSM estimator for three reasons. First, the PSM requires a weaker set of assumptions. This is important particularly when the outcome variables are censored or binary as in this case (Wooldridge 2002). A second concern regarding the MLE is the incidental parameter problem (Lancaster 2000). There also exist types of MLEs that control for fixed effect, such as Honoré (1992). However, I do not use the model because of the endogeneity of the rescheduling treatment. Finally, Shoji (2010), which uses the same dataset to evaluate the rescheduling impact, employs the MLE. Since the small sample size of this data makes the estimation result unstable, it is important to use a different identification strategy as well to confirm the robustness of the findings.

The goal of the PSM estimators is to quantify the average treatment effect to the treated (ATT); how the outcomes—credit constraint and credit from moneylenders—of rescheduled members would have changed if they had not been rescheduled. It compares the outcome of each rescheduled member to members who had “similar” characteristics to the rescheduled member but did not reschedule. PSM employs similarity in terms of the conditional probability of being rescheduled given various household characteristics.

Define Ri as a dummy variable which takes the value of 1 if observation i had an opportunity of rescheduling and 0 otherwise, and Yir denotes the outcome of observation i when the rescheduling regime is r. Therefore, observation i's observed outcome, Yi, is described as Yi= RiYi1+ (1 −Ri)Yi0, because Yi= Yi1 for Ri= 1 and Yi= Yi0 for Ri= 0, respectively. Given the notations, ATT is defined as E(Y1Y0|R = 1), where E denotes the expectation operator.

Since Yi0 is not observable from the data when Ri= 1 (the counterfactual), the PSM assumes the selection on observables,


where X denotes time-variant and invariant observable determinants of rescheduling. This assumption means that the nonrescheduled outcome Y0 is independent of rescheduling treatment, R, conditional on X. If this assumption is valid, it implies that there is no omitted variable bias once X is included in the regression. Another assumption to implement PSM is the overlap assumption:


This assumption ensures that for each rescheduled individual there is another matched nonrescheduled individual with a similar X. Under these conditions, the estimated propensity score should satisfy:


This is referred as to the balancing score condition and is used to test the validity of the estimation specification (Dehejia and Wahba 1999, 2002). Finally, these assumptions provide the following arrangement of ATT (Rosenbaum and Rubin 1983):


In particular, this study uses the DIDM model. DIDM controls for observable and hard-to-observe time-invariant factors using the data of the prerescheduling period. In this study, MFIs rescheduled installments and savings only during the flood period, but members were not allowed to reschedule even in the face of idiosyncratic negative shocks in other seasons. Indeed, Table 1 shows that MFIs rescheduled mainly in the rainy seasons of the second and fourth periods, but only two members had the opportunity to reschedule during the dry seasons of the first and third periods.8 Given the nature of rescheduling, I consider the first and third periods as pre-treatment periods, and the second and fourth periods as the treatment periods. Therefore, the ATT in DIDM estimator is obtained by the following specification:


where, ΔYt ≡ YtYt−1, and t = 2, 4.

As Jalan and Ravallion (2003) mention, an important process in conducting the PSM is the estimation of Pr(Rt= 1 |Xt). MFI officers allowed rescheduling mainly for poor and disaster-affected members who had difficulties in attending member meetings and paying for installments and savings on time.9 Therefore, this study considers the following determinants as the covariates: flood intensity, poverty level, distance to the meeting place and other geographic characteristics (approximation of access to the meeting place), income correlation among villagers, debt, and other household characteristics.

1. Flood intensity

The 2004 flood caused various losses to households, such as income and assets. A concern regarding the PSM estimation assuming the selection on observables is the possibility of bias caused by unobservable determinants of the treatment (selection on unobservables). Unobservable flood damage would bias estimation. To address this possibility, this study reports the list of self-reported flood damages obtained from the open-response question in Table 3.

Table 3.  Flood Damages Based on Open-Response Questions (Multiple Answers)
 MFI MembersNonmembers
FrequencyFraction (%)FrequencyFraction (%)
Other assets73.9125.9
Death of household member00.010.5
Injury/sick member00.010.5
No damage3016.95527.0

This process alleviates the omitted flood damage because it creates a complete list of the major flood damages the victims suffered. It shows that the main losses included income, houses, and other assets such as livestock, but not health conditions. Given that the damage level of income and assets could be endogenous, this study controls for flood intensity using a binary variable that takes the value of 1 if the house was inundated. Inundation at home causes damage to houses, livestock, and other assets, and these in turn decrease income.

2. Poverty level

The covariates include six types of asset holdings such as grain storage, jewelry, owned field, livestock, other productive assets, and housing. I do not add income levels or income loss caused by the flood because these could be simultaneously determined with rescheduling. Instead, various types of productive assets and flood intensity variables control for them.

3. Geographical characteristics (access to meeting place)

The covariates also include geographical characteristics such as the availability of evacuation shelters and distance to member meeting places, rivers, schools, and paved roads. The distance to MFI meeting places and paved roads controls for the difficulty in attending the member meetings during the flood. Also, the distance to rivers and the availability of evacuation shelters approximate the flood intensity.

4. Income correlation among villagers

Where income is highly correlated among the villagers, the risk-sharing arrangement within the village does not work, and therefore MFIs might allow rescheduling more intensively. Therefore, I control for a village-level characteristic, Ev[vartvIncomeitv)], where Income denotes the labor income and index i, t, and v stand for household, period, and village, respectively. The low value of this indicator implies high-income correlation among villagers and therefore high demand for rescheduling. To address the possibility that rescheduling treatment and income-earning activities could be determined simultaneously, I generate this indicator using the data collected by IFPRI in 1998 and 1999 from the same households as this dataset.

5. Debt and demographic variables

Finally, I include a binary variable that takes the value of 1 if the observation was indebted from MFIs as of the beginning of the period. I also control for demographic characteristics such as the headcount of males aged over 16, females over 16, children under 16, and the age and sex of household head. These variables approximate the availability of risk-coping mechanisms as well as the preference shifter: households with more working-age males might have a higher ability to smooth income by increasing labor participation (Kochar 1999). Older household heads might have higher social capital, implying better access to risk-sharing arrangements (Coate and Ravallion 1993; Kimball 1988; Kocherlakota 1996). Also, it indicates a determinant of rescheduling such as costs associated with attending member meetings; households with many children might not be able to attend member meetings because of time constraints experienced as a result of household chores and childcare duties.


A. Estimation of the Propensity Score

Table 4 reports the estimation results of the propensity score using the probit model.10 I also estimate the Tobit model, whose dependent variable is the duration of rescheduling, as a robustness check. The first column reports that households whose homes were inundated were more likely to be rescheduled by 38.4%, but this is not robust to the inclusion of period and district fixed effects, shown in the second and third columns.

Table 4.  Determinants of Rescheduling Treatment: Propensity Score Estimation
MEMStd. Err.MEMStd. Err.Coef.Std. Err.
  1. Notes: 1. MEM stands for Marginal Effect at the Mean.

  2. 2. The observations of the second and fourth periods are used.

  3. 3. Standard errors are in parentheses.

  4. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively.

Dummy if own house is inundated0.384***(0.204)0.018(0.061)0.76(0.86)
Grain storage (103 Tk)−0.025*(0.012)−0.023**(0.009)−0.53**(0.23)
Jewelry (103 Tk)−0.008**(0.004)−0.005*(0.003)−0.10*(0.06)
Livestock (103 Tk)−0.007***(0.003)−0.004***(0.002)−0.09**(0.04)
Owned house (103 Tk)0.001(0.001)0.0005(0.0005)0.01(0.01)
Owned field (106 Tk)0.111(0.083)0.100*(0.056)2.22(1.39)
Other productive assets (103 Tk)−0.002(0.003)−0.0004(0.0017)−0.03(0.06)
Covariate shock indicator−0.003(0.005)−0.005(0.004)−0.12(0.10)
Males over 16−0.030*(0.016)−0.018*(0.012)−0.14(0.24)
Females over 16−0.019(0.020)−0.015(0.013)−0.21(0.30)
Children under 16−0.018(0.012)−0.006(0.008)−0.20(0.20)
Age of head0.006(0.009)0.004(0.006)−0.03(0.14)
Squared term of age−0.023(0.085)−0.021(0.056)0.34(1.41)
Female head dummy−0.047(0.030)−0.023(0.019)−0.48(0.78)
Dummy if indebted0.251***(0.052)0.224***(0.062)3.99***(0.63)
Distance to member meeting0.027(0.065)0.036(0.042)1.32(1.00)
Distance to river0.007(0.006)0.004(0.004)0.10(0.10)
Distance to school0.010(0.030)0.005(0.018)−0.29(0.41)
Distance to paved road0.022(0.024)0.005(0.016)0.35(0.36)
Dummy if evacuation shelter is available0.021(0.049)0.017(0.038)−0.13(0.71)
Flood period fixed effect  0.184***(0.061)3.81***(0.51)
Chandpur fixed effect  −0.019(0.024)0.43(0.72)
Magura fixed effect  −0.041**(0.022)−1.52**(0.62)
Constant    −3.60(3.53)
Pseudo R20.20 0.52 0.32 
N327 327 324 

A robust finding from Table 4 is that liquid asset holdings were important determinants of rescheduling with grain storage being the most important; the second column shows that a Tk 1,000 increase of grain storage reduces the probability of rescheduling by 2.3%. This high marginal effect of grain storage is likely because it directly affects food consumption and the subsistence nutrition intake. It therefore is expected to be a more important determinant than other assets. On the contrary, MFIs did not target those with fewer nonliquid assets except for livestock. The coefficient of other productive assets is negative but insignificant. That of owned field is counter-intuitively positive and significant in the second column, but the statistical significance is marginal and not robust. The estimated marginal effect is smaller than those of the other assets: a Tk 1,000 increase in land holdings increases the probability by only 0.01%.

The indicator of income correlation shows expected signs. MFIs allowed rescheduling for those members whose income was correlated to other households in the village, and who were therefore likely to suffer from covariate shocks. However, the estimated coefficients are statistically insignificant. Also, households with fewer working-age males were more likely to be rescheduled; a working-age male in the household decreases the probability of rescheduling by 1.8%. It also appears that indebted members were more likely to be rescheduled by 22.4%. Debt also increases the duration of rescheduling by 3.99 weeks. Regarding the geographic characteristics, the coefficient of distance to the MFI meeting place is positive but statistically insignificant, unlike the result in Shoji (2010). A possible reason for the difference is that this study employs data from both the intensive flood period and nonflood period, while Shoji (2010) investigates only the former. During the intensive flood period, the road to the meeting place is inundated. Therefore, MFI members living far away from the meeting place would have difficulty in attending the meeting, which in turn increases the possibility of being rescheduled. In the nonflood period, however, the road is still accessible even in the face of negative events. Thus, distance may not be an important determinant of rescheduling.

Finally, I test the balancing score (equation 3) for the first and second columns. It is found that only the second column satisfies the condition. Conditional on the propensity score, household characteristics are not significantly different between rescheduled and nonrescheduled groups. Therefore, I use the propensity score obtained from the second column to implement the matching estimations.

B. Cross Section and DID Matching Estimation

This subsection implements the matching estimation to evaluate the impact of rescheduling. Table 5 reports results from eight matching models using only observations of the common support.11 I employ four types of matching methodologies: nearest-neighbor matching, Gaussian kernel matching, stratification matching, and the radius matching method.12 The first to the fourth columns are the results of the DIDM which controls for time-invariant determinants of rescheduling, while the last four columns present the results without controlling for them.

Table 5.  Estimated Rescheduling Effects on Livelihoods
 Expected SignsDIDMCross-Section PSM
  1. Notes: 1. Standard errors are in parentheses. The standard errors of kernel matching and stratification matching are estimated using the bootstrap method. The bandwidth of kernel matching is 0.06. 150 bootstrap replications are conducted.

  2. 2. The radius is 0.1 in the fourth and eighth columns.

  3. 3. In the first to the fourth columns, the dependent variables are first-differenced values.

  4. ** and * represent statistical significance at the 5% and 10% levels, respectively.

Credit constraints−0.27**−0.24**−0.20*−0.08−0.080.01−0.010.03
Credit from moneylenders−181.67−199.19−187.77**−93.70−168.00*−182.02−167.18*−103.29**
N (rescheduled/counterfactual) 75/3175/9169/9775/9175/3175/9169/9775/91

The DIDM results show that the estimated ATTs on credit constraint are negative and statistically significant except for the radius matching; the first three columns show that rescheduling reduces the possibility of credit constraint by 20 to 27 percentage points. In other words, the standard repayment structure, which does not allow rescheduling, means borrowers face credit constraints during negative shocks. The estimated impacts on credit from moneylenders are also consistent with consumption smoothing behavior, but are not statistically robust.

As described before, MFI members facing binding credit constraints cope with the repayment burden by borrowing from other credit sources, such as moneylenders, who charge high interest rates. Theoretically, since rescheduling can be considered an additional loan disbursement from the MFI, this would reduce the demand for credit from other sources, while it does not directly change the credit supply from them. Thus, my finding implies that even a small amount of rescheduling can, by mitigating the repayment burden for the MFI, decrease the excess demand for credit with respect to the overall market, including formal and informal lenders.

One might be concerned about the possibility of selection on unobservables; there might be unobservable determinants of rescheduling such as vulnerability and poverty. Omitting these variables would potentially bias estimates of rescheduling toward the positive, because rescheduled members who are poorer and more vulnerable are more likely to be credit constrained and borrow under burdensome conditions.13 On the contrary, the estimated ATTs are negative. Therefore, the potential bias caused by these unobservable factors would not affect the results qualitatively.

To further discuss the robustness of the results, the fifth to the eighth columns using the cross-section PSM are reported. These naïve specifications do not control for the household fixed effects, and therefore they should be affected by the sample selection bias more. Indeed it appears that the estimated results are biased toward the positive, compared to DIDM. This is consistent with the discussion above; controlling for unobservable determinants such as vulnerability and poverty increases the absolute value impact of rescheduling. Given that some of these unobservable determinants may also include time-variant components, the point estimate shows the lower bound (in terms of absolute value) of the actual ATTs.

However, there might be another concern about the use of the pooled data because, first of all, the composition of observations in terms of time period may not be the same between the treatment and control groups. In the treatment group 23% of observations are from the fourth period (17 out of 75 observations), while the corresponding ratio of the counterfactuals used to calculate ATT is 48% in the case of nearest-neighbor method (15 out of 31 counterfactuals). The second potential reason is the persistent effect of rescheduling. If rescheduling has any long-lasting effects, the treatment effect of rescheduling might depend on the experience of past rescheduling. There are potential channels through which the impact of rescheduling persists. The first is through the accumulation of human and physical capital. Therefore, this study controls for these characteristics in the estimation of propensity score. Second, MFI members may reduce savings for precautionary motives after the introduction of contingent repayment system. Therefore, I estimate the ATT by using the observations from only the second period, because MFIs allowed rescheduling during the period for the first time. The result is reported in Appendix Table 1. While the results are marginally significant because of a smaller sample size, it is qualitatively comparable.


This section investigates the changes in rescheduling effect by the duration using the DIDM with multiple treatments (Imbens 2000).14 Define Wi as the duration of rescheduling for observation i. I divide the observations into three levels: Wi= 0 (not rescheduled), Wi= 1 (rescheduled only for one week), and Wi= 2 (more than one week).15 The rearranged ATT in this model is described as follows:




This approach uses a conditional probability of belonging to the rescheduling level w given a rescheduling level of w or 0. I employ the multinomial logit model to estimate the conditional probability and the result is reported in Appendix Table 2.

Table 6 shows the change in the impact of rescheduling by the duration. I use only DIDM estimators in this section. First, there is no significant effect of short-term rescheduling. With long-term rescheduling, the estimated impacts are greater than those of short-term rescheduling and are statistically significant. The estimation for credit from moneylenders also reports a similar result. It is found that rescheduling significantly reduces credit from moneylenders when we focus on the long term, although the average rescheduling effect is statistically insignificant in Table 5. These results imply that at least two weeks of rescheduling was required during the 2004 flood. However, one-third of rescheduling treatment was only for one week.

Table 6.  Change in Rescheduling Effect with Rescheduling Level
 Expected SignsNNKernelStratificationRadius
  1. Notes: 1. Standard errors are in parentheses. The standard errors of kernel matching and stratification matching are estimated using the bootstrap method. The bandwidth of kernel matching is 0.06. 150 bootstrap replications are conducted.

  2. 2. The radius is 0.1 in the fourth column.

  3. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively.

A. Credit constraints     
Short-term rescheduling (one week): EY1tΔY0t|W = 1)−0.24−0.16−0.33−0.05
N (rescheduled/counterfactual) 25/1525/9319/5325/93
Long-term rescheduling (more than one week): EY2tΔY0t|W = 2)−0.36**−0.43***−0.39*−0.15**
N (rescheduled/counterfactual) 47/1847/7734/7047/77
B. Credit from moneylenders     
Short-term rescheduling (one week): EY1tΔY0t|W = 1)−196.00−152.46−152.637.92
N (rescheduled/counterfactual) 25/1525/9319/5325/93
Long-term rescheduling (more than one week): EY2tΔY0t|W = 2)−231.38*−263.87−283.12*−146.52*
N (rescheduled/counterfactual) 47/1847/7734/7047/77

One might be concerned, however, about the possibility of sample selection bias: household characteristics might be different between rescheduling levels 1 and 2. If the duration was determined according to the extent of credit constraint, it will show a similar result, even if the duration does not matter with the rescheduling impact. It might be straightforward to compare EY2−ΔY1|W = 1) and EY1−ΔY0|W = 1) to address the issue, but it is impossible because of the small sample size of the dataset. Instead, I test whether the coefficients of the multinomial logit model are significantly different between the two groups. It does not reject the null that the coefficients are jointly the same (Appendix Table 2), implying that the difference in household characteristics between short-term and long-term rescheduling groups is insignificant.


This study examines the targeting accuracy of rescheduling and its consequences on the liquidity of households. It is found that rescheduling was targeted to poor and indebted members. Also, rescheduling significantly reduces the possibility that members face binding credit constraints and borrow from moneylenders, which in turn may reduce transitory poverty. However, short-term rescheduling has insignificant effects.

These findings have important implications for the literature regarding the poverty reduction effect of MFIs, because previous studies show that binding credit constraints remarkably decrease the consumption level of households (Deaton 1991; Dercon 2005; Fafchamps 2003; Zeldes 1989). Furthermore, it might affect livelihood persistently (Banerjee et al. 2010; Carter et al. 2007; Dercon 2004; Hoddinott 2006; Quisumbing 2006).

This study's findings suggest the importance of further investigations into the new structure of MFIs. These findings must be interpreted with caution, however, since they hinge on the validity of my identification strategy and the small sample dataset.


  • 1

    See, for example, Basu (2010), Bauer, Chytilová, and Morduch (2010), Field and Pande (2008), Kaboski and Townsend (2005), and McIntosh (2008).

  • 2

    For instance, the introduction of contingent repayment may reduce the probability of binding credit constraints in the future, which in turn decreases savings for precautionary motives.

  • 3

    Reduction of transitory poverty is an important policy goal for developing countries. Jalan and Ravallion (1996), for example, find that poverty in rural China could be halved if transitory poverty is solved.

  • 4

    Thanas and unions are administrative units in Bangladesh: a union consists of some villages, and each Thana includes multiple unions.

  • 5

    The attrition is 2.7% mainly because of migration.

  • 6

    Each period corresponds to the agricultural calendar in Bangladesh.

  • 7

    This paper defines credit from moneylenders as credit from informal sources with interest. This is because the term Mohajon, which means professional moneylenders in Bengali, also means informal credit contracts with interest. According to the classification, the minimum interest rate of loans from moneylenders is 10% per year, and the average rate is 71.2%.

  • 8

    As discussed, rescheduling at the third period was due to a religious festival and therefore we do not consider this a reason for rescheduling.

  • 9

    The details of the targeting process are described in Section II.

  • 10

    Although some studies use nonparametric approaches to estimate the propensity score, they require large-sample data. See Lee (2005) for examples of methodologies to estimate the propensity score.

  • 11

    Estimations without the restriction of common support show qualitatively similar results (not reported but available on request).

  • 12

    I divide the propensity score into 20 equal parts in stratification matching.

  • 13

    This tendency is shown in the summary statistics in Table 2. Unless controlling for the endogeneity of rescheduling, the rescheduled members are more likely to face credit constraints than nonrescheduled members.

  • 14

    Lee (2005) and Frolich (2004) summarize the literature of program evaluation with multiple treatments. See also Lechner (2002), Lee (2004), and Behrman, Cheng, and Todd (2004) for other matching estimation methods with multiple treatments.

  • 15

    In the 2004 flood, the duration of rescheduling ranged from one to eight weeks depending on the severity of the flood; most were one or two weeks.


Table APPENDIX TABLE1.  Matching Estimation with Only the Second Period (Flood Period) Observations
 Expected SignsDIDM
  1. Notes: 1. Standard errors are in parentheses. The standard errors of kernel matching and stratification matching are estimated using the bootstrap method. The bandwidth of kernel matching is 0.06. 150 bootstrap replications are conducted.

  2. 2. The radius is 0.1 in the fourth column.

  3. 3. The dependent variables are first-differenced values.

  4. *** and ** represent statistical significance at the 1% and 5% levels, respectively.

Credit constraints−0.24−0.25−0.03−0.24***
Credit from moneylenders−211.21−190.69−138.80−177.22**
N (rescheduled/counterfactual) 58/1358/4422/8057/44
Table APPENDIX TABLE2.  Propensity Score with Multiple Treatments (Multinomial Logit Model Estimation)
 Short-Term (One Week)Long-Term (More than One Week)
  1. Notes: 1. MEM stands for Marginal Effect at the Mean.

  2. 2. H0: All coefficients are jointly the same between short-term and long-term: Chi2 (23) = 29.63 (P-value is 0.16).

  3. 3. The observations of the second and fourth periods are used.

  4. ***, **, and * represent statistical significance at the 1%, 5%, and 10% levels, respectively.

Dummy if own house is inundated−0.52(1.59)−0.00060.68(1.17)1.51
Grain storage (103 Tk)−1.07**(0.50)−0.0014−0.42(0.26)−0.67
Jewelry (103 Tk)−0.17(0.11)−0.0002−0.07(0.07)−0.11
Livestock (103 Tk)−0.09*(0.05)−0.0001−0.06(0.04)−0.10
Owned house (103 Tk)0.00(0.02)−0.0000040.02*(0.01)0.04
Owned field (106 Tk)4.76**(1.94)0.00640.88(1.59)1.42
Other productive assets (103 Tk)0.00(0.08)−0.000004−0.01(0.02)−0.01
Covariate shock indicator−0.28(0.21)−0.0004−0.06(0.12)−0.10
Males over 16−0.34(0.30)−0.0004−0.57*(0.32)−0.91
Females over 16−0.51(0.42)−0.0007−0.12(0.36)−0.20
Children under 16−0.29(0.32)−0.0004−0.16(0.26)−0.25
Age of head0.16(0.21)0.00020.01(0.17)0.01
Squared term of age−1.14(2.06)−0.00150.30(1.70)0.48
Female head dummy−2.00(1.37)−0.00140.02(0.94)0.03
Dummy if indebted4.92***(1.21)0.01393.62***(0.70)8.79
Distance to member meeting1.83(1.30)0.0025−0.44(1.32)−0.71
Distance to river−0.08(0.22)−0.00010.19(0.13)0.31
Distance to school0.05(0.82)0.00010.22(0.47)0.35
Distance to paved road−0.97(0.78)−0.00130.48(0.52)0.78
Dummy if evacuation shelter is available−44.28(1.91E+09)−0.31610.56(0.87)1.12
Flood period fixed effect2.20***(0.65)0.00334.61***(0.79)16.95
Chandpur fixed effect−1.25(1.12)−0.0014−0.09(0.89)−0.14
Magura fixed effect−1.68**(0.82)−0.0018−1.92**(0.83)−2.41
Constant−5.60(5.41) −6.42(4.33)