【正文】 een advantage is 55% versus 40%. These simple parisons appear to be driven entirely by selection biases, however. Once appropriate parisons with control groups are made, access to the three microfinance programs does not yield meaningful increases in per capita consumption, the education of sons, nor the education of daughters. If anything, the levels are slightly lower than for control groups. The results are surprising and contradict frequent claims made about the programs in international discussions of microfinance. Access to the programs does, however, appear to aid the diversification of labor supply across seasons. In turn, access is associated with a reduction in the variability of consumption across seasons. Thus, while the programs may not increase consumption on average, they may offer households ways to smooth consumption through smoothing ine. In pointing to impacts on vulnerability, the results highlight an advantage that is seldom considered in the emerging microfinance literature (an exception is Pitt and Khandker, 1998a). These benefits should be judged against the tens of millions of dollars that have supported the programs. The results also demonstrate how misleading simple performance indicators can be, and they hold lessons for evaluations of similar public health and other social programs in lowine As here, such programs are often limited to particular regions and particular target groups, typically poor households. Unlike in wealthier countries, inebased means tests are almost never used. Instead, for example, the microfinance programs in rural Bangladesh focus on the “functionally landless” implemented as a rule barring lending to households owning over a half acre of cultivable land. The program rule can be the basis of a plausible econometric strategy if the eligibility requirement is strictly enforced and built around a feature that is exogenous to the , clean impacts can be gauged by paring the status of households clustered just below the arbitrary dividing line to households clustered just above. This approach is a form of regression discontinuity design (Campbell, 1969), and the insights provide the basis of Pitt and Khandker’ evaluations are subject to multiple selection biases: selfselection into the programs by the most able, nonrandom program placement, and endogenous determination of the intensity of participation (., the size of loans in microfinance)