We use the data from the National Supported Work Demonstration to study performance of non-propensity-score-matching estimators, and to compare them with propensity score matching. We find that all matching estimators we studied here are sensitive to the choice of data set. Propensity score methods are sensitive to smoothing parameters, and they usually have larger standard error. Difference-in-differences and bias-corrected matching improve the performance of the matching estimators considered here. Our results suggest that the 1974 earnings are important for Dehejia and Wahba’s PSID data but not for their CPS data in replicating experiment results. After decomposing the selection bias, we find that a sizable selection bias on unobservables is present in all data sets.