The main concern for many evaluation studies is that controlling for individuals' observed characteristics may not be enough to obtain valid treatment effects. We exploit a unique dataset that contains a rich set of administrative information on individuals newly entering unemployment in Germany, as well as several usually unobserved characteristics like personality traits, attitudes, expectations, social networks and intergenerational information. This allows us to empirically assess the effect of including these usually unobserved variables on the propensity score distribution, the matching quality, and the treatment effects obtained using unconfoundedness-based estimators. Our findings indicate that these variables play a significant role for selection into active labor market programs (ALMP), but do not make a significant difference in estimating treatment effects on wages and employment prospects. This suggests that the usually unobserved variables we analyze are not a threat to the validity of the estimated treatment effects, if comprehensive control variables of the type usually used in modern ALMP evaluations (which include labor market histories) are available. Our results also suggest that rich administrative data may be good enough to draw policy conclusions on the effectiveness of ALMPs.