This paper proposes a new approach to identify the wage effects of training. The idea is to narrow down the comparison group by only taking into consideration the workers who wanted to participate in training but did not do so because of some random event. The point estimate of the return to training consistently drops when restricting the comparison group this way. While the OLS estimate of the return to training participation is significantly positive, this is no longer the case when we use the new comparison group. This outcome suggests that a large share of what is usually interpreted as returns to training is actually the return to some unobservable characteristic.