We argue that non-experimental impact estimators will continue to be needed for evaluations of interventions in developing countries as social experiments, for various reasons, will never be the most preferred approach. In a survey of four studies that empirically compare the performance of experimental and non-experimental impact estimates using data from development interventions, we show that the preferred non-experimental estimators are unbiased. We try to explain the reasons why the non-experimental estimators perform better in the context of development interventions than American job-market interventions. We also use the survey as a source for suggestions for implementation and assessment of non-experimental impact evaluations. Our main suggestion is to be more careful and precise in the formulation of the statistical model for the assignment into the program and also to use the assignment information for model-based systematic sampling.