Multi-study R-learner for Heterogeneous Treatment Effect Estimation

We propose a general class of algorithms for estimating heterogeneous treatment effects on multiple studies. Our approach, called the multi-study R-learner, generalizes the R-learner to account for between-study heterogeneity and achieves cross-study robustness of confounding adjustment. The multi-study R-learner is flexible in its ability to incorporate many machine learning techniques for estimating heterogeneous treatment effects, nuisance functions, and membership probabilities. We show that the multi-study R-learner treatment effect estimator is asymptotically normal within the series estimation framework. Moreover, we illustrate via realistic cancer data experiments that our approach results in lower estimation error than the R-learner as between-study heterogeneity increases.