انتخاب در سراسر قراردادهای اعتباری: نظریه و تحقیق میدانی

Lenders may choose to encourage borrower side contracting using group, or co-signed, loans or discourage it using individual loans, so as to make relative performance comparisons. In this context wealth of the agents relative to outsiders, and wealth inequality among potential joint liability partners, are important factors determining the choice among loan contracts. In a related model of whether to borrow, higher covariance of returns mitigates an adverse selection effect. We test these models using relatively rich data gathered in field research in Thailand. The prevalence of joint liability contracts relative to individual contracts exhibits a U-shaped relationship with the wealth of the borrowing household and increases with the wealth dispersion. The likelihood of joint-liability borrowing increases the lower is the probability of project success, a direct affirmation of adverse selection. Higher correlation across projects makes joint liability borrowing more likely relative to all other alternatives. Strikingly, most of the results disappear if we do not condition the sample according to the dictates of the models, with selection into and across credit contracts.

Microcredit is viewed as a major tool to alleviate world poverty, but practitioners are polarized in a debate concerning the virtues of individual versus group, joint liability lending. For the most part, only one contract per lender is observed, e.g., group-based loans under the Grameen Bank in Bangladesh and individual loans under BRI in Indonesia. It is thus difficult to make progress in the debate. We do not see what would happen if the alternative contract were available, not to mention all the difficulties inherent in cross country comparisons. Here we take advantage of the menu of contracts offered by a dominant rural lender in a single country and some unusual data gathered in field research to make some headway on the issue. Potential borrowers from the BAAC1 in Thailand decide whether to borrow at all, and if so, whether to borrow as individuals or in a group. That is, potential borrowers select into and across loan contracts. There is in fact much variation in these choices in the data. Our use of the data is naturally enough dictated by various well-known models from the contract theory, mechanism design literature that debates the virtues of individual versus group lending. We exposit and test two related models that compare an individualistic, relative performance regime with a group, cooperative regime in the context of moral hazard in effort provision. The defining feature of each regime is whether or not the borrowers are able to side-contract. The basic advantage of the individualistic regime, where there is no side-contracting, is the opportunity for the lender to gage if one borrower has been diligent by looking to see if other individual borrowers are doing well. The group regime, where borrowers can side-contract, does not allow this kind of information extraction since borrowers can collude against the lender, coordinating in the choice of low mutual effort or reallocating internal resources (e.g., tunneling). Its potential advantage, however, is the enhanced ability to monitor and enforce intra-group agreements on actions and transfers (i.e. risk sharing). It is thus not obvious the kind of regime for which moral hazard considerations argue; the optimal regime may vary with the parameters of the model. Indeed, Holmström and Milgrom (1990) show that one crucial determinant of the optimal regime is the degree of correlation between the random component of borrowers’ returns. In the individualistic regime, the principal, or lender, can use the correlation to mitigate its imperfect information. It does so by rewarding or punishing based on borrower performance comparisons, which diminishes the risk cost of high-powered incentives. The higher the correlation, the more effectively the information problem is dealt with in this regime. In the group regime, however, rewards based on performance comparisons can be manipulated, so higher correlation does not carry the same benefits. In fact, this regime turns out to be more effective the lower is the correlation, since low correlation makes the ability of the borrowers to coordinate and share risk among themselves more valuable. As intuition suggests and Holmström and Milgrom (1990) show formally, it would be Pareto optimal for the borrowers to be acting non-cooperatively if technological correlation is high enough, and cooperatively if it is low enough. This is a testable implication. Prescott and Townsend (2002) address similar questions but focus instead on wealth levels of borrowers relative to the lender and wealth dispersion across borrowers. They offer an extended version of the above-mentioned unobserved effort model, generalized in several directions, but without closed form solutions available. In simulations, they find that for sufficiently asymmetric Pareto weights on the two borrowers, interpretable as high wealth dispersion, the cooperative regime dominates. When the weights are more similar, the non-cooperative regime dominates. Apparently the cooperative regime is better at extracting effort from the low-weight borrower. Further, the cooperative regime dominates for high enough or low enough reservation payoff of the borrowers, interpretable as the borrowers’ wealth, while an intermediate value makes the non-cooperative regime optimal. Again, both of these predictions are in principle testable. These two models focus on conditions under which one borrowing regime dominates another. A third model, Ghatak (1999), makes predictions about whether a given borrower will choose to borrow in a group or take the best alternative, which may mean not borrowing at all. Some borrowers are inherently more risky than others. In a setting of limited, individual liability and lender ignorance of risk-type, risky borrowers are in effect subsidized by safe ones. This can cut out socially productive loans to safe borrowers and leave only the risky types to borrow.2 In this setting, loans that make use of joint liability within an endogenously formed group can mitigate the adverse selection problem. They do so by inducing homogeneous matching among risk types and reducing the subsidy from safe to risky borrowers, drawing back into the market relatively safe borrowers. While group loans always lessen adverse selection relative to individual loans, they may not eliminate it. A testable implication of Ghatak (1999) is thus the existence of adverse selection, which some argue is negligible in practice due to lenders’ ability to collect information and their offering of other incentives.3 We modify this model to introduce correlation in borrower returns and find that borrowers with higher correlation of returns are more likely to self-select into group contracts (a testable proposition). They do this because correlation raises their payoff of borrowing by lowering the chances of facing liability for their partner's loan. We attempt to test these predictions from the various models using relatively rich data gathered in field research in Thailand. The BAAC is the primary institutional lender in rural Thailand: for example 64% of the institutional loans in our sample are from the BAAC. But it is not the only lender. In our data we have a complete enumeration of loans outstanding or repaid in the last twelve months, and for each loan the household respondent is asked what if any collateral was used (land title, cosigner, other). Among these are loans from village funds, money lenders, commercial banks, and the informal sector. The BAAC offers both individual loans and group, joint liability loans; other institutional lenders may offer only one or the other. We test the models on the smaller sample of BAAC loans and the larger sample of all individual and co-signed loans from institutional lenders, including but not restricted to the BAAC. In addition to loan data, we also have measures for wealth, wealth dispersion, technological correlation, and the risk of the borrowers, as well as numerous controls. We link the non-cooperative regime in the moral hazard theories with individual loans in the data, and the cooperative regime in these theories with joint liability (i.e. co-signed), group loans in the data. Since these models focus on comparing two regimes against each other but not against outside alternatives, we restrict the sample to those who have had a group or individual loan outstanding in the last twelve months and exclude those who did not borrow at all. By contrast, when testing the adverse selection models of the literature, we include the whole sample and focus on the decision whether to borrow under joint liability or to choose any alternative (not borrow at all or borrow under individual liability).4 Our association of the loan types in the theory with analogs in actual practice deserves some elaboration. BAAC policy dictates that to receive a group loan, one must form or join an official BAAC-registered borrowing group and enter into a joint liability agreement. A box denoting that the group is the collateral is checked off on the loan form, and sometimes a particular member in the group is named as a cosigner. In contrast, individual loans must be guaranteed by some form of collateral, usually land. Thus the link of group loans in the data to group loans in Ghatak (1999) is relatively clear. Liability of one borrower for another within the group is written into the borrowing contract explicitly. Cases do exist in which the BAAC required group members to pay for a delinquent member of the group. Further, liability does indeed appear to be limited since no explicit collateral is required. The key element in Holmström and Milgrom (1990) and Prescott and Townsend (2002) is the re-allocation of risk. The BAAC has in place a risk-contingency system, as documented in Townsend and Yaron (2001), for example. Farmers experiencing force majeur events report their difficulties to the local branch. The loan can be rescheduled, and in some instances interest and even part of principal forgiven, as if an insurance indemnity had kicked in. Thus group and individual credit contracts are state contingent loan repayment agreements. The key distinction in Holmström and Milgrom (1990) and Prescott and Townsend (2002) between the group and individual regimes is whether the borrowers can side-contract. Of course in the data, group loans are not accompanied with a full legal apparatus that enables borrowers to enforce side contracts; neither do individual loan contracts stipulate that borrowers not coordinate on financial or production-related decisions.5 However, the group loan format does appear to be a natural way of facilitating, even encouraging, side-contracting between borrowers. Group loans are accompanied by required group meetings and explicit group-related contingencies. Conversely, the individual loan format builds in an institutional neglect of borrower side-contracting capabilities. Further, while not explicitly in the loan contracts, performance comparisons seem to be made across individual borrowers as theory would predict. A credit officer visits the village (the farmer and neighboring households) to verify the adverse event. More generally the BAAC is well positioned to make relative performance comparisons in deciding what to do about repayment problems. Thus, while undoubtedly imperfect, there is a highly plausible link between group and individual loans in the data and the group and individual regimes in Holmström and Milgrom (1990) and Prescott and Townsend (2002). It should be noted that some key assumptions are different across the models as written. Whether or not the models are irreconcilable at a deeper level is beyond the scope of this paper. A priori, however, each seems to have a reasonable chance at explaining the data at hand. Adverse selection is an obstacle thought to be pervasive in many contexts, and quite plausible in the Thai rural context. Moral hazard and the use of risk-contingencies in loan contracts, consistent with the practice of the BAAC, are also not unlikely here. Indeed, we find some support for both types of model. As predicted by Prescott and Townsend (2002), the prevalence of joint liability contracts relative to individual contracts exhibits a U-shaped relationship with the wealth of the borrowing pair and increases with the wealth dispersion. The latter result is especially robust. The results surface only when we limit the sample to that suggested by theory: households with either a group or an individual loan, but not both. The results do not appear to be due to a conventional collateral effect (in which higher wealth enables households to borrow without cosignors); we separate out and control for the subset of wealth that is most commonly used as collateral. Contrary to the theory of Holmström and Milgrom (1990) (and presumably also Prescott and Townsend, 2002), we find no evidence that joint liability borrowing becomes less likely as covariance of output increases. This is true when restricting the sample only to those with group or individual loans. When we use the full sample, however, as Ghatak (1999) would dictate, we find some evidence for the opposite. That is, we find that higher correlation makes joint liability borrowing more likely relative to all outside options, just as our modification of Ghatak (1999) suggests. Finally, we find direct evidence consistent with the prediction of adverse selection in the credit market, in that the likelihood of joint-liability borrowing increases the lower is the probability of project success. This result only appears when the full sample is used. Thus the more risky households are borrowing. Ausubel (1999) tests this proposition with data from credit card companies: those willing to select into higher interest rates are more likely to default. A parallel insurance literature also tests whether households paying for more complete coverage are more likely to experience the adverse, insured event. (See the literature review in Chiappori and Salanie, 2003). Here we take advantage of measurements of risk type also of those who choose not to borrow and offer a direct test of this prediction of the adverse selection model. As Chiappori et al. (2002) emphasize, it is difficult to determine whether accident probabilities increase for those purchasing more complete coverage because of adverse selection (ex ante selection) or moral hazard (ex post shirking). Here we further restrict our sample to those who have borrowed in the last year but are not currently doing so, and find that the adverse selection remains a force in the data. Strikingly, as alluded to above, most of these results confirming the models’ predictions disappear if we do not condition the sample according to the dictates of the models.

The results on wealth and contract selection are interesting in several respects. Indirectly, they give plausibility to the interpretation of group lending as a way for the lender to promote side-contracting between borrowers. This has not always been the primary view of group lending, but it does appear to have some empirical validity. More directly, these results suggest that wealth distribution, as emphasized by Prescott and Townsend (2002), can be a key determinant of the optimal contract. Specifically here, a wealth level further away from the village average, or further away from the population cutoff of 6–8 million baht, makes choice of a group loan over an individual loan more likely. From a theoretical point of view, moving away from a transferable utility framework seems warranted. From a policy point of view, one might be led to wonder whether group lending accomplishes its goal of reducing inequality, or whether, at least locally, it thrives on inequality and perhaps perpetuates it through unequal intra-group allocation of consumption and labor. The absence of corroboration in the data for the correlation result of Holmström and Milgrom (1990) is puzzling. There is a range of possible explanations. First, the restricted sample size used to test Holmström and Milgrom (1990) is less than half of the full sample size used to test Ghatak (1999). Second, the result is proved under very specific functional form assumptions, which may not be satisfied. On the other hand, the intuition appears likely to be valid more generally. Third, it may be that correlation is unobservable or unused by the lender. But as discussed in Section 2.3, one might still expect the patterns in the data that we look for even if only borrowers observe correlation. Finally, it is possible that the correlation measure is too noisy. This is less likely since the measure does fairly well in the Ghatak (1999) specification. The results confirming Ghatak (1999) on adverse selection and correlation indirectly support the more common interpretation of group lending as joint liability. Though we do not view this interpretation as necessarily contradictory with the other, we know of no attempt to analyze their interaction. This attempt could lead to fruitful research. At any rate, adverse selection of risky borrowers seems to be an issue not only in the ivory tower and formal insurance companies, but also in the rice patties. In terms of policy, at least partial screening by risk information that is available is an obvious first step in improving the extension of credit and the allocation of funds. Since the survey, the BAAC has in fact begun offering different interest rates depending on credit history. Still, there may remain room for more efficient screening by offering a wider array of contracts with different combinations of interest rate and joint liability (as in Ghatak, 2000). On the theoretical front, introducing loan size into such an adverse selection model could also prove fruitful.