This paper analyzes the returns to training-occupation combinations. I use administrative panel data on apprenticeships and employment for German workers and identify the returns using data on occupation-specific vacancies. For the estimation, I set up a Roy model and extend existing control function approaches to deal with selection in a two-stage, high-dimensional setting. I find sizable returns to training in one's occupation, and substantial bias when not controlling for selection. Returns are decreasing in the task distance between training and occupation. I argue that imperfect information leads to ex-post suboptimal training choices, and that retraining could address this ex-ante friction.