In this paper, we investigate the impact of classroom training programmes on individual unemployment rates in Denmark. In 1994 a social experiment was conducted, where unemployed applicants for labour market training were randomised into treatment and control groups. We formulate and estimate experimental impact estimators of the effect of treatment on the treated. The experimental data is polluted by the presence of no-shows and crossovers, which implies that traditional experimental estimators are biased. Therefore we formulate and estimate an endogenous variables model (using the randomisation indicator as a perfect exclusion restriction) and implement various matching estimators. We find – surprisingly – that classroom training significantly increases individual unemployment rates. We discuss some possible reasons for this surprising finding and some related policy issues.