We find, consistent with our modified version of the adverse selection model, that higher correlation makes joint liability borrowing more likely relative to all outside options. Surprisingly, most of the results disappear if we do not condition the sample according to the dictates of the models. When testing Ghatak, individual loans are associated with an outside option in theory, and of course group loans are those where there is joint liability.4 We also have measures of wealth, wealth distribution, technological correlation, and borrower risk. as well as multiple controls.
Our association of the borrowing forms in theory with analogues in practice deserves some explanation. To test the adverse selection models from the literature, we similarly focus on the decision whether to borrow under joint liability or any alternative (no borrowing at all or borrow under individual liability).6. Note that as W, the lender's expected income, increases, the borrower's wealth actually decreases.
In decentralizing this model, the payoff to the lender W would vary inversely with the wealth of borrowers, and the asymmetry λi would reflect the dispersion of wealth among borrowers.
Adverse Selection and Technological Correlation
It is assumed that the lender only observes whether each project succeeded or failed; thus the borrower can contract on the binary outcome of success or failure, but not on the amount of output or risk type, which would be optimal if it were observable.20 This implies that, in contrast to HM and PT, borrower repayment cannot depend on the lending pair's output other than through two binary functions, that is, depending only on whether each borrower has realized returns greater than zero. Note that the derivative of the payoff with respect to bo is [r+l(1−2p)]. As long as l ≤r, which we assume, this derivative is strictly negative forp∈[p,1). Thus, the safer an agent's type, the lower his payoff for borrowing and undertaking the project.
Since all agents have an outside option that pays u, agents will borrow if and only if the payoff from borrowing is greater than u. Given that the payoff of borrowing declines in p, there is a type of foreclosure risk, called p, such that borrowers of type p > p find it suboptimal to borrow, and borrowers of type p < p will borrow.2223 This prediction of adverse selection in the loan market with limited liability is general and can be found in the literature at least until. However, it is likely that the same qualitative results could be obtained in Ghatak with the observed output, provided certain modifications to the output distributions were made, including that the support of the output distributions does not vary by type.
This is because safe borrowers are more likely to be successful and therefore more likely to have a delinquent partner who affects their payout. For non-homogeneous groups, that is, those for ˜ for whom p = p, ˜ρ is not the correlation coefficient, but something closely related: it is the correlation, expressed as a fraction of the maximum possible correlation given individual probabilities of pandp.27 Whenp =p, the maximum possible correlation coefficient is one, so ˜ρ is equal to the correlation coefficient. For the off-diagonal cells of the matrix to be positive, we must be less than p(1−p) and p(1−p).
Clearly, for higher ˆp, ˆp will also have to be higher (since the loan payoff is still decreasing in p) in order to lower the payoff and keep u the same; the same is true for higher ˜ρ.29 As pˆ is higher, the mass of borrowers has increased, or in other words, the probability that a randomly selected agent will borrow has increased. The idea here comes from the property of the model that a positive payoff occurs only when the borrower is successful. Thus, ˜ρ has the interpretation described in the text as a correlation expressed as a fraction of the maximum possible correlation and is free to vary from zero to one.
3 Empirical Results
Empirical Strategy
In summary, the data we use are observations at the household level that show whether the household borrows and under what regime, but do not allow us to explicitly determine group membership (whether the group is a comparison group with relative performance, or an explicit lending group). group under the cooperative regime). To measure the wealth distribution within the group, we use a measure of the distance between the wealth of the borrowing household and the average wealth of the village. The assumption is that the further away the household is from the village average, the more likely it is that the household is part of a group with asymmetrical wealth.
We measure the within-group correlation with the degree of grouping of superlative years (for family income) of this family with other families in the village. Again, the assumption is that the more this family's good and bad years coincide with others in the village, the more likely it is to be in a highly correlated group. Finally, to define group wealth, we use various combinations of the household's own wealth and the average village wealth of the borrower households.
The wealth of the household itself is definitely part of the group wealth, as we know with certainty that this household belongs to the borrower group in question. However, the wealth of the rest of the group is a more significant component of total group wealth, since groups contain more households; but using the village average wealth of borrowers measures the wealth of the rest of the group with error. Ultimately, we use these measures of group wealth, wealth dispersion and correlation, all at the household level, to explain whether the household has a group loan.
We use the same household correlation measure described in the preceding paragraphs, which reflects the coincidence of the household's good and bad years with other households in the town. The risk-type measure applies the theory directly to subjective evaluations that the household makes about the distribution of its future income prospects. This would argue for their inclusion with non-borrowing households in the regression, as full liability renders risk type unimportant.
Description of Variables
WEALTH Household wealth, in million baht 1.8 (5.8) WEALTH SQ Wealth in square meters, in million square baht. LAND Average land holdings (in rai) of the household EDYEARH Years of education achieved by the head of the household 4.1 (2.6). Two analogous variables measure whether the household has an individual loan contract from a lending institution.
The part of this wealth attributable to the land property, to which the family has legal title, is called TITLE and its square is TITLESQ. To measure the distribution of wealth WLTHDSPR, which corresponds to the λi's of PT, we use the following function of household wealth and average village wealth, we call it W EALT H:. Each family was asked which of the last five years was the best and which was the worst, in terms of family income.
Loans that are actually secured by assets between contractual links, such as future wages, if the loan is made by an employer, were not categorized as either group or individual. .. 38 In other words, the WLTHDSPR does not change if the wealth of everyone in the village increases or decreases by the same multiplier factor. Since power utility is used in the calculation results, and since the ratio of marginal utilities under power is scale-free, WLTHDSPR must also be scale-free. . income.). Specifically, each household was asked what their income would be in the coming year if it were a good year (Hi), what their income would be if it were a bad year (Lo) and what their income would be (eg ) .
We assume the income distribution is binomial across the high and low states, as in the Ghatak model. The household head's gender and years of education are controlled in MALEH (male coded as one) and EDYEARH, respectively. We also include a dummy variable, NRTHEAST, equal to one for the 50% of households in the poorer Northeast region.
Logit Results
It is crucial to control for occupation, as BAAC and several other institutional lenders exclusively targeted agricultural workers at the time of the study. Made by the U-shaped ratio is about 6.4 million baht for BAGPLOAN and 8.4 million baht for GRUPLOAN. The U-shape remains significant when excluding households with wealth over 20 million (1.4% of the data).43 We check for a cubic relationship for robustness and find a very similar picture: coefficients on WEALTH and OPPORTUNITY with the same sign as in Table II and significant at 1%, and the coefficient on WEALTHCB negative and significant at 5% or 10% depending on specification.
We use a bandwidth to include 80% of the data and a three-cube weighting function (see Cleveland 1979 and Fan 1992) to reduce the weight of observations that are, say, further away than the current one. This is not necessarily due to a flaw in the model, but rather an imperfect approximation of group wealth. This result is even stronger than the U-shape result: it remains significant in every specification reported in the previous paragraphs.
This allows us to remain agnostic about the functional form of the wealth distribution variable. With regard to technological correlation, none of the measures show significance, except SAMEBEST at the 90% level in the specification with GRUPLOOAN as dependent variable. We experimented with excluding EXINCOME, with and without including a measure of the variance,49 and the results did not change.
The basic results are that loan size is a negative and significant predictor of group regime, always at the 1% level, except under BAGPLOAN in the third sample, when it is at the 10% level. In short, the somewhat weakened significance of the U-shaped relationship with wealth is the only change in the results when controlling for loans. This is evidence that households that are highly correlated with the rest of the village are more likely to borrow with a collective contract.
The result is all the more remarkable because of the unique, literal interpretation we use to measure the type of risk.51. 54 Note that moral hazard here in the sense of Stiglitz/Weiss and Ghatak refers to effects on risk, not the mean of a distribution.
4 Conclusion
