We estimate the short‐, medium‐, and long‐term effects of different types of government‐sponsored training in West Germany using particularly rich data that allows us to control for selectivity by matching methods and to measure interesting outcome variables over eight years after a program's start. We use distance‐weighted radius matching together with a bias removal procedure based on weighted regressions in order to increase the precision and robustness of standard matching estimators. We find negative employment effects in the short term for all program types, effects whose magnitude and persistence is directly related to program duration. In the longer term, training seems to increase employment rates by 10–20 percentage points. For most programs the longer‐term positive effects seem to be sustainable over the eight‐year observation period.

We analyse the effects of government-sponsored training for the unemployed conducted during East German transition. For the microeconometric analysis, we use a new, large and informative administrative database that allows us to use matching methods to address potential selection bias, to study different types of programmes and to observe labour market outcomes over eight years. We find strong evidence that, on average, the training programmes under investigation increase long-term employment prospects and earnings. However, as an important exception, the longer training programmes are not helpful for their male participants. At least part of the explanation for this negative result is that caseworkers severely misjudged the structure of the future demand for skills. Copyright Verein für Socialpolitik and Blackwell Publishing Ltd. 2007.

Labor market effects of training programs, Dynamic treatment regimes, Nonparametric identification, Causal effects, Sequential randomization, Program evaluation, Treatment effects, Dynamic matching, Panel data, C21, C31, J68,

Recent examples are papers by Imbens and Angrist (1994) for the identifying power of instrumental variables, Heckman and Vytlacil (2001b) and Vytlacil (1999) for nonparametric selection models, and by Rubin (1977) and Lechner (2001b) for the conditional independence assumption. Lemma 1-II (connection of treatment effects defined for different lengths of treatments)