In this paper we estimate, for the 1989-93 period in Belgium, the effect of vocational classroom training on the rate of transition from unemployment. We propose a “control function” estimator accounting for variable treatment effects. In the absence of interaction effects between explanatory variables this estimator identifies treatment effects free from selection bias. A natural experiment induces exogenous sub-regional variation in the training supply. This provides over-identifying restrictions that cannot be rejected. During participation, the transition rate decreases by 23% to 30%. Afterwards it increases by 47% to 73%. Making training available for a broader population would, however, reduce the