The key to estimating the impact of a programme is constructing the counterfactual outcome representing what would have happened in its absence. This problem becomes more complicated when agents, such as individuals, firms or local governments, self-select into the programme rather than being exogenously assigned to it. This paper uses data from a major social experiment to identify what would have happened to the earnings of self-selected participants in a job training programme had they not participated in it. We investigate the implications of these earnings patterns for the validity of widely-used before-after and difference-in-differences estimators.