DINA: Estimating Heterogenous Treatment Effects in Exponential Family and Cox Models

We propose to use the difference in natural parameters (DINA) to quantify the heterogeneous treatment effect for exponential family models, in contrast to the difference in means. Similarly we model the hazard ratios for the Cox model. For binary outcomes and survival times, DINA is both convenient and perhaps more practical for modeling the covariates' influences on the treatment effect. We introduce a DINA estimator that is insensitive to confounding and non-collapsibility issues, and allows practitioners to use powerful off-the-shelf machine learning tools for nuisance estimation. We use extensive simulations to demonstrate the efficacy of the proposed method with various response distributions and censoring mechanisms. We also apply the proposed method to the SPRINT dataset to estimate the heterogeneous treatment effect, demonstrate the method's robustness to nuisance estimation, and conduct a placebo evaluation.