Controlling for group-level heterogeneity in causal forest

The use of causal forests to recover heterogeneous treatment effects for randomized interventions is growing. As with any estimator, controlling for group- or strata-level unobservables is necessary to recover unbiased estimates. However, there exists no proven method to control for group-level unobservables in causal forests. We provide a novel solution. We use Monte Carlo simulations and two applications to demonstrate the effectiveness of this solution and the shortcomings of a broad set of alternatives sourced from the classical econometrics and machine learning literatures. Our method greatly increases the number of settings in which unbiased, heterogeneous treatment effects are recoverable.