Any study that compares outcomes across microfinance clients who self-select into either the traditional weekly repayment schedule or the more flexible monthly repayment schedule potentially obfuscates the true impact of less frequent repayment on financial stress, since different types of clients are likely to sort into each repayment schedule. Our study addresses this concern by using a randomized controlled trial (RCT) experimental design. Harvard's Institutional Review Board approved the study design and protocol. Verbal informed consent was obtained from all participants (due to low education level among respondents, written consent requirement was waived).
We partnered with a large microfinance organization called Village Financial Society (VFS) in Kolkata, India. At the time of the study, VFS loans were distributed through five-member microfinance groups. Each client received an individual loan and her ability to obtain a subsequent loan depended only on her personal repayment record (individual liability as opposed to joint liability group lending). The loan was uncollateralized and the modal value was US $222 excluding interest costs, which were ten percent of the loan size. Clients were required to make periodic repayments to the loan officer beginning shortly after loan disbursal in a group meeting conducted in their neighborhoods.
In total, 213 clients participated in this study, all of whom were selected from a larger study group of 740 clients. Between January and September 2008, VFS recruited clients and formed 148 five-member groups comprising 740 clients. Loan sizes varied from Rs. 4000 to 12,000 (~$90 to $260), with a modal loan size of Rs. 10,000. Randomization was implemented using a random sequence of numbers generated with statistical software by the project research assistant. Treatment status was assigned to batches of 20 groups at a time based on the timing of group formation with a 1
1 allocation ratio.
After group formation and prior to loan disbursement, the field coordinator called the project research assistant to determine whether a group had been randomly assigned to either a five-weekly repayment schedule (from here on referred to as “monthly”) or a weekly repayment schedule. One exception is the first batch of treatment groups, which was composed of 12 groups assigned to a four-weekly repayment schedule as opposed to a five-weekly repayment schedule. The change to a five-weekly repayment schedule was made to better accommodate VFS' logistical needs. Clients on the four-weekly experiment were not selected for the study considered here. More information on client selection and randomization is available online in Text S3
. Clients were informed that repayment would be determined by lottery. Since all members of a group were restricted to have the same repayment schedule, the trial was a parallel cluster-randomized trial.
From these 740 clients, we randomly invited 105 weekly and 105 monthly clients to participate in the Daily Consumption Survey (DCS). Selection was based on a random sequence of numbers generated with statistical software by the project research assistant. Due to a major festival scheduled to occur several weeks after the start of the DCS survey, we chose the monthly clients from the 21 monthly groups with starting dates that ensured that the DCS survey could run from one repayment to the next for monthly clients without interruption by major festivals. To ensure balance across treatment arms, we selected the weekly clients from the 74 weekly groups with group formation dates that overlapped with the 21 monthly groups.
Twenty-three of the 210 initial clients dropped out, including 11 from control and 12 from treatment. Although the attrition rates were similar for both groups, it is possible that the types of clients that dropped out of the treatment group were systematically different from those who dropped out of the control group. For this selection to generate the main results we present later, clients who dropped out of the control group would have to have significantly lower average baseline stress levels than clients dropping out of the treatment group, which is not supported by a comparison of baseline stress measures across attritors in both groups.
To maintain a target sample size of 200, we randomly selected an additional six weekly and seven monthly clients from the larger study group of 740. The sample size of 200 was chosen based on budget considerations. To summarize, a total of 111 weekly clients from 45 groups were randomly assigned and received the intended treatment, while 100 clients from 42 groups were analyzed for the primary outcomes. A total of 112 monthly clients from 26 groups were randomly assigned and received the intended treatment, while 100 clients from 26 groups were analyzed for the primary outcomes. In the next section, we discuss how we compute standard errors in light of the potential correlation of outcomes within loan groups.
On average, weekly clients paid US$5.40 every week, while monthly clients paid $27.10 every five weeks. The loan duration in both cases was 45 weeks.
In order to assess financial stress levels accurately and in real time, we employed an innovative application of cell phone technology to survey clients every 48 hours for seven weeks. Clients were surveyed on average 16.5 weeks after receiving the loan. By contacting the microfinance clients in our study via cell phones, which were provided to each client for the purpose of this study, we mitigated recall bias, reduced non-response and non-participation rates, and collected 5000 surveys (200 clients surveyed 25 times each) in a cost-effective manner. In order to truly understand consumption smoothing and liquidity constraints among the poor, one needs data that accurately measures consumption levels, income, and assets of households over time. Particularly for consumption data, several potential sources of reporting error have been documented in the economics literature, the most important of which are recall mistakes, inability to capture total household consumption, and level of aggregation of consumption categories 
. In our project, we have attempted to mitigate the risks posed by each while keeping logistical demands and costs of surveying reasonably low through a novel survey implementation strategy that leverages cell phone technology available in our study region. For more details on reporting error on consumption data, see 
Each time the survey was administered, we measured clients' level of financial stress with four questions: confidence in ability to repay loan, anxiety about loan repayment, argument with spouse about finances, and time spent thinking about repayment. We construct four indicator variables to capture financial stress: 1 if they did not feel confident about their ability to repay the loan; 1 if they felt worried, tense, or anxious about paying the next loan installment; 1 if they argued with their spouse in the last 24 hours; and 1 if they spent at least five minutes thinking about repayment during the past day. The Cronbach Alpha for these measures is high, at 0.8386, suggesting that it is appropriate to think of the different questions as measuring one underlying construct. Thus, in addition to the individual variables, we report the effect of the equally weighted average across the four outcomes. We call this construct the Financial Stress Index.
Self-reported financial stress is an important measure of household well-being. Indeed, in our sample, financial stress is positively correlated with observable indicators of poverty, although not at a statistically significant level. The average value of the Financial Stress Index is higher for clients who are illiterate, report not having a savings account, had a shock within the past 30 days, do not have a household business, and are in the lower half of the asset distribution.
Evidence on the health relevance of self-reported measures of stress comes from a large literature that documents significant correlation between stress biomarkers (which bear a direct relationship with human health) and self-reported measures 
. conduct a literature review of studies documenting a correlation between blood pressure levels and self-reported measures of job strain. More similarly to the stress measure used here 
, find that responses to questions about ability to meet financial obligations, such as food, clothing and medical care, correlate with measures of blood pressure and cortisol response. Similarly 
, find herpes antibody levels and self-reported measures of stress are correlated in a sample of low-income women. However, the same study does not find an association between measures of salivary cortisol response and self-reported stress measures. The absence of a significant correlation between cortisol response and self-reported stress measures appears to hold more generally 
. conducts a literature review of the correlation between salivary cortisol and self-reported mental stress measures and conclude, “the evaluation of the studies in this paper showed insufficient evidence for an association between self-reported mental stress and the cortisol response in field studies.” The authors also detail some of the difficulties in collecting saliva swabs in a reliable way for cortisol testing.
Because the contracts were randomly assigned to clients, a comparison between treatment arms has a causal interpretation. Ordinary Least Squares (OLS) regression analysis allows us to compare monthly and weekly clients controlling for variables such as day of the week, whether the survey took place in the morning, and cohort effects.
For all outcome variables we estimate simple ordinary least squares regressions of the following form:
is the outcome of interest for client i
in group g
on day d
is an indicator variable that equals one if the group was assigned to the five-week repayment schedule. All regressions include dummies for stratification batch (Bg
), day of the week (Pdig
), whether the survey was taken in the morning (Mdig
), number of weeks since disbursement (Wdig
), and the calendar week (Cdig
). The vector Xig
, which is present only in the specification labeled as including controls, consists of age, marital status, household size, Muslim, literate, has savings, negative shock in last month, has household business, total asset value, and loan size. Regressions including Xig
also control for loan officer fixed effects. In all regressions, standard errors are corrected for clustering within loan groups using Huber-White standard errors.
Before turning to the results, we discuss our primary hypotheses regarding the possible channels of influence. Increased flexibility in repayment can influence mental stress among clients through several channels:
- Income: Greater flexibility in repayment may allow clients to invest their microfinance loans in less liquid but more profitable business assets and inventory by providing more time between repayments to earn a return. As a hypothetical example, consider a small hardware store owner who knows that buying and selling higher-quality light fixtures yields higher profits. Yet, he cannot use his microfinance loan to buy this inventory because higher-quality light fixtures do not sell as quickly as lower-quality ones; hence, he will not have the money in time to make his first repayment if he invests in this more illiquid (but more profitable) inventory. If he repays every month instead of every week, he may be able to invest in the higher-quality inventory and increase profits. Since wealth is negatively correlated with financial stress, we should see higher income and reduced financial stress among clients with more flexible repayment schedules.
- Self-control: Recent research has suggested that the poor may be more susceptible to temptation . In such cases, less frequent repayment may increase stress and default. Specifically, less frequent repayment should increase both overall household expenditure and spending on “temptation goods” (for example, tobacco, alcohol, and ready-made foods).
- Consumption smoothing: Much evidence indicates that avoiding large fluctuations in consumption is costly for the poor in developing countries . Less frequent repayment can reduce the cost of smoothing consumption in the event of negative health and business shocks. This reduced cost to smoothing consumption may reduce variance of household expenditure and make it easier for the poor to meet their repayment obligations. Alternatively, if households would have used costly methods such as liquidating business assets or removing children from school to smooth consumption, it could manifest itself in higher income. In either case, we would expect a reduction in both stress and default.
- Time burden: By reducing the number of repayment meetings clients are required to attend, a monthly repayment schedule relaxes the time constraint on clients by approximately 1.5 hours per month, which could reduce their overall stress levels.
We expect that these channels of influence will interact. For instance, less frequent repayment can increase client income while at the same time increasing default through its impact on fiscal discipline. Hence, the net effect of less frequent repayment on mental stress remains an empirical question.