Peer Monitoring and Microcredit: Field Experimental Evidence from Paraguay

Abstract

Given the substantial amount of resources currently invested in microcredit programmes, it is more important than ever to accurately assess the extent to which peer monitoring actually reduces moral hazard among borrowers faced with group liability. We conduct a field experiment with women about to enter a group loan programme in Paraguay and then gather administrative data on their repayment behaviour in the 6-month period after the experiment. In addition to the experiment, which is designed to measure individual propensities to monitor one's peers, we collect a variety of other potential correlates of behaviour and repayment. Controlling for other factors, we find a very strong causal relationship between the average monitoring propensity of a person's loan group and repayment. Our most conservative estimate suggests that borrowers in highly monitored groups are 36% less likely to have problems repaying their portions of the loan. In addition, confirming previous results, we also find some evidence that risk preferences, social preferences and cognitive skills affect repayment.

Keywords::

• C93
• D03
• D14
• G21
• O16
• O54

Notes

¹ Other valuable contributions to the survey-based literature include Hermes et al. (2005), Kritikos & Vigenina (2005), Barboza & Barreto (2006), Simtowe & Zeller (2007) and Feigenberg et al. (2009).

² The Paraguayan currency is the guaraní; the exchange rate at the time was about 1 PGY = $0.00017. All conversions in the paper use this rate.

³ Clearly, 58 participants cannot be evenly divided into groups of four. Instead of turning away people from our limited subject pool, we relied on the fact that participants could not know who the other members of their group were and formed groups with “shadow members”. These randomly chosen shadow members contributed to their own group, but their behaviour was also counted in another group to get the total up to four persons.

⁴ The first session lasted 10 rounds. Because this took longer than our time allocation, in all subsequent sessions eight rounds were played. Our analysis uses all the available data.

⁵ Average earnings in the experiment were 15 400 PGY or about $2.60 at the time of the study. Considering that 18% of Paraguayans live on less than $2 per day and participants' median daily earnings were 50 000 PGY ($8.50), the incentives appeared to be salient.

⁶ To save time and to eliminate end-game effects, the monitoring and punishment ended after round seven. This was not announced until contributions had been made for the eighth round near the end of the experiment.

⁷ Originally, we also wanted to examine the propensity of our participants to punish free riders, in addition to the propensity to monitor them. However, as one can see in Table 2, there was not a lot of punishment and so there is not enough variation in the punishment choices to be useful.

⁸ Manski (1993) outlines this common shocks problem in peer effects estimation.

⁹ Of course we had to deal with the fact that some people always monitor and therefore the residual is zero. We assumed that there was some small error (0.01) to perturb our estimates. Experimenting with different small error values does not affect the predicted probabilities of not being monitored appreciably and prevents a few zeros from removing much of the variation in this measure.

¹⁰ Because or sample is limited, the incidental parameters problem does not affect our results appreciably (i.e. our results are very similar to linear probability estimates).

¹¹ In earlier versions of the analysis, we explored including other measures from the experiment, including one's own monitoring propensity and average contribution. However, neither of these had much predictive power nor did they affect the other point estimates substantially.

¹² Although we realize that linear approximations in probit models are often a problem, we decided to report standardized marginal effects because the effects were otherwise very hard to compare.

¹³ Recall that long-term cohabitation is also common in Paraguay and, therefore, marriage may also proxy for other characteristics, such as religiosity.

¹⁴ Our “family member” effect is similar to that described in Ahlin & Townsend (2007).

¹⁵ Indeed, neither mean frequencies of peer monitoring nor mean predicted monitoring rates differ significantly from one loan group to the next.

¹⁶ In addition, and as suggested by a referee, our earlier note that controlling for one's own monitoring propensity did little to affect our estimates also suggests that the correlation between nosiness and “responsibility” may not be strong.

¹⁷ Formally, where M_− i is our measure of the propensity to monitor of the other group members, β is our estimate of the effect of peer monitoring on repayment, and u_i is some (un)observed trait of the borrower then . Notice that only if (i) δ, the direct effect of on repayment is positive and or (ii) and .

We thank the editor, our referees, Mary Burke, Matthieu Chemin, Ben Feigenberg, Erick Gong, Julian Jamison and Elizabeth Murry for valuable comments on earlier drafts.