Predictive Counterfactuals for Treatment Effect Heterogeneity in Event Studies with Staggered Adoption

I propose a method for estimating treatment effect heterogeneity in event studies with staggered adoption. This method serves as an alternative to two-way fixed effects regressions, which have been shown to be near-term biased in these settings. I build on insights from the machine learning (ML) literature. The first step is to use ML algorithms to predict counterfactual outcomes in the absence of treatment, which can then be used to estimate a distribution of treatment effects. From this distribution, it is possible to estimate different causal parameters, depending on the researcher's objective. With simulations, I show that the ML approach is unbiased and recovers treatment effects more efficiently than traditional methods. I conclude with an application of the method to a large energy efficiency program in the US, revealing patterns of heterogeneity in energy savings that had not yet been documented