The presented paper contributes to a growing literature on machine learning applications for causal inference. The author focuses on estimation of treatment effect heterogeneity, specifically for event studies with time-varying treatment/exposure dates (i.e. staggered adoption). Recent econometric literature shows that traditional impact evaluation methods (such as fixed effects regressions) can be near-term biased in those settings. He introduces an approach that does not suffer from such bias, and that can recover heterogeneity more efficiently than standard fixed effects models. The first step of the approach is to accurately predict a distribution of counterfactuals, by applying flexible machine learning algorithms to pre-treatment data only. Then, by comparing counterfactual and true outcomes, it is possible to estimate a distribution of treatment effects. Those effects can be summarized for different portions of the sample, thus recovering heterogeneity. With simulations, the author demonstrates how that approach can be accurate and efficient, even in the presence of dynamic (time-varying) treatment effects. He concludes with an application to real data from a large residential energy efficiency program in the US. While prior literature is restricted to estimating average program effects, he identifies which types of upgrades are associated with higher energy savings. Those results provide insight about which upgrades to target to improve the program's cost-effectiveness.
Please contact Lisa Vorbeck if you wish to participate in the online seminar.