Hybrideseminar: Estimating causal forests in a difference-in-differences research design

Causal forests (Athey et al., 2019; Wager and Athey, 2018; Athey and Imbens, 2016) are a popular machine learning tool to estimate heterogeneous treatment effects. However, application of this estimator to common difference-in-differerences settings is computationally infeasible or leads to inconsistent estimates. We therefore present a computationally feasible algorithm to estimate causal forests in the presence of many fixed effects. Our modification identifies treatment effects by partialling out fixed effect using group averages. Simulation results suggest that our algorithm provides consistent estimates of the Conditional Average Treatment Effect in a (staggered) difference-in-differences research design. Finally, we use our method to document heterogeneity in the treatment effect of payrolling on worker outcomes following Goos et al. (2022). We find robust evidence that outcomes for some workers improve after the treatment. Such evidence was not found when we performed an elaborate heterogeneity analysis using manually formed subgroups.