A0205
Title: Decomposing causal effect heterogeneity under multiple treatment versions
Authors: Michael Knaus - University St Gallen (Switzerland) [presenting]
Abstract: A method is developed to decompose treatment effect heterogeneity when the treatment is not homogeneous and can have multiple versions. It disentangles observed aggregated treatment effect heterogeneity into true effect heterogeneity and heterogeneity due to selection into versions. This allows (i) to avoid spurious discovery of heterogeneous effects, (ii) to detect actual hidden heterogeneity in versions, and (iii) to evaluate the underlying version assignment mechanism. We propose a semiparametric method for estimation and statistical inference for the decomposition parameters. Our framework allows for the use of modern machine learning techniques in the estimation of the underlying causal effects. It can be used to conduct simple joint hypothesis tests that consider all treatment versions simultaneously. This alleviates the need for multiple testing procedures when deciding on the aggregation level of the treatment variable in empirical applications. We analyze heterogeneity due to different types of academic or vocational training in the large scale training program for the disadvantaged youth Job Corps. We find that often curricula are not better allocated than random and only specific age and income groups benefit from the actual allocation.
