Based on new, exceptionally informative and large German linked employer-employee administrative data, we investigate the question whether the omission of important control variables in matching estimation leads to biased impact estimates of typical active labour market programs for the unemployed. Such biases would lead to false policy conclusions about the cost-effectiveness of these expensive policies. Using newly developed Empirical Monte Carlo Study methods, we find that besides standard personal characteristics, information on individual health and firm characteristics of the last employer are particularly important for selection correction. Moreover, it is important to account for past performance on the labour market in a very detailed and flexible way. Information on job search behaviour, timing of unemployment and program start, as well as detailed regional characteristics are also relevant.