We assess selection bias in estimated returns to workplace training by exploiting a field experiment with random assignment of workers to a one-week training program. We compare experimental estimates of this program with non-experimental estimates that are estimated by using a sample of agents who were selected by management not to participate in the experiment. Our results show that non-experimental estimates are biased, yielding returns about twice as large as the causal effect. When controlling for pre-treatment performance or individual fixed effects, only about one tenth of this bias remains and is even further reduced when applying common support restrictions.