This study compares experimental and non-experimental estimators using randomized data from the National Evaluation of Welfare-to-Work Strategies. Our main question is: can non-experimental methods match results obtained using random assignment? There are three key empirical conclusions from this study. First, results obtained using non-experimental data can lead to wrong conclusions about the causal effects of a training program. Second, biases obtained from non-experimental data depend not only on the econometric procedure used but also on the chosen comparison group. Third, comparisons longer ahead in time are more susceptible to selection bias problems. In other words, medium-run bias was larger than short-run bias.