Not all policy questions can be addressed by social experiments. Nonexperimental evaluation methods provide an alternative to experimental designs but their results depend on untestable assumptions. This paper presents evidence on the reliability of propensity score matching (PSM), which estimates treatment effects under the assumption of selection on observables, using a social experiment designed to evaluate the PROGRESA program in Mexico. We find that PSM performs well for outcomes that are measured comparably across survey instruments and when a rich set of control variables is available. However, even small differences in the way outcomes are measured can lead to bias in the technique.