Decomposing a Matching Estimation - Correlated Unobserved non-random Selection, Substitution, and Causal Effects in a Large Scale ALMP Program in Germany

We analyse the treatment effects of a large scale German active labour market program. Modellprojekte Bürgerarbeit is a modified public employment scheme, with an activation period as a first step. Using data from individual employment records we perform matching estimations to analyse the treatment effect on the treated. In the cross-section estimation, we find reemployment rates to about 10%-points higher in the treatment group. Additionally, we consider that this effect might be upward biased due to correlated time-invariant unobserved heterogeneity and negative spill-over effects on non-participants. Therefore, we combine matching with difference-in-difference estimation (to control for unobserved heterogeneity), Doing so, we estimate a slightly lower treatment effect (about 8%-points). Finally, we use untreated individuals from regions with treated individuals (pseudo-treatment), and compare them to untreated individuals from regions without treatment. The negative pseudo-treatment effect lies around 2%-points, indicating negative spill-over effects. We conclude that the 10%-point difference from the matching estimation can be decomposed into a causal effect (6%-points), correlated unobserved heterogeneity and negative spill-over effects (2%-points each).