This article evaluates a residential alcohol/drug treatment program for the homeless. The evaluation documents numerous deviations from the program as designed and other implementation problems. Foremost among these was that the project was designed as a randomized experiment, but randomization was sabotaged by the treatment staff. Non-random research attrition constituted another potential source of bias. The authors employ econometric modeling techniques to correct for these selection biases. Consistent with the drug treatment literature, evidence suggests that people retained in treatment for more than a few months exhibit more positive outcomes than those staying for shorter periods (authors).