Primary Submission Category: Heterogeneous Treatment Effects
Generalized Causal Rule Ensembles: Interpretable Heterogeneous Exposure–Response Functions for Continuous Exposures
Authors: Suhwan Bong, Heejun Shin, Francesca Dominici,
Presenting Author: Suhwan Bong*
In many areas of scientific research, it is often of interest to estimate an exposure response function that has a causal interpretation, and most importantly to also assess whether the shape of that ERF is heterogeneous across subgroups of the population. However, most of the recent work on heterogeneous causal effects is limited to the case where the exposure is binary. In this paper, we introduce the Generalized Causal Rule Ensemble (GCRE), a data-driven, rule-based framework that uncovers and summarizes heterogeneity in continuous exposures. The GCRE can be summarized in two steps. First, in a subgroup discovery step, we rely on targeted-smoothing Bayesian Additive Regression Trees (tsBART) to identify covariate-based decision rules whose effects are smooth in the exposure, and then applies group LASSO on spline expansions of exposure to select a sparse, stable subset of subgroups. Second, in a inference step, we estimate subgroup-specific deviation functions and ERFs for the selected rules with favorable theoretical guaranties. In simulations, GCRE recovers heterogeneous ERF that are invisible in dichotomized analyses and improves estimation accuracy relative to standard ERF estimators. We illustrate the approach in a ZCTA-level study of the health effects of PM 2.5 among U.S. Medicare beneficiaries in 2021, and provide summaries of how the ERF varies between sociodemographic subgroups.