Speaker: Michael Knaus, University of Tübingen

His research interests are at the intersection of causal inference and machine learning to answer questions in empirical, mostly labor, economics. In particular he is currently working on the estimation of average and heterogeneous treatment effects as well as policy learning. 

Title:  Implied Weights of Estimators for (Conditional) (Local) Average Treatment Effects

Abstract: Estimators that can be represented as linear combination of observed outcomes are very useful in practice. Their implied weights allow, e.g. to check covariate balancing or to investigate complier characteristics. This paper provides a unified framework for a range of common estimators and shows under which conditions their implied weights are defined in closed form. It covers parametric estimators like OLS and 2SLS, semi-parametric estimators like (augmented) inverse probability weighting (IPW) and partially linear estimators for (local) average treatment effects, as well as causal forest, R-learner and DR-learner for conditional average treatment effects. We analyze basic properties of the weights in the binary treatment case and find, e.g. that weights of partially linear estimators do not sum up to one in the treatment and control groups, while they do so for augmented IPW even without normalized IPW weights.

Host: Phillip Heiler

Organisers: Luke Taylor and Mikkel Bennedsen