Co-Supervised Tree Synthesis for Interpretable Subgroup Identification and Honest Treatment Effect Inference in Randomized Trials

Abstract

Identifying patient subgroups with heterogeneous treatment effects is central to precision medicine, yet existing approaches face a tension. Black-box methods estimate individualized effects well but yield no interpretable subgroups, while tree-based methods produce explicit partitions but are unstable and less accurate in moderate-sample trials. We propose CausalSynthTree, a co-supervised method in which black-box causal estimators guide construction of an interpretable tree. It partitions the covariate space into cells, fits cell-wise conditional average treatment effect models from both observed data and synthetic data labeled by black-box causal teachers, and synthesizes them into a tree through a treatment-effect disparity criterion. Each leaf carries a sparse linear model, so the tree path defines a subgroup and the leaf model shows which covariates modify the effect. Honest inference for leaf-wise effects remains valid despite co-supervised augmentation and data-driven subgroup selection. In simulations it closes much of the gap between the two families. From moderate sample sizes onward it is more accurate than the black-box learners whenever the effect is linear or has a single threshold, and trails them only on a finer partition than the tree it grows. It controls spurious splits and attains nominal coverage throughout. In the ACTG175 HIV trial, where standard interaction tests detect no effect modification, the method reports no subgroups while the competing tree methods report several. In an adjuvant colon cancer trial it reports a borderline covariate in part of the resamples, together with a within-region age gradient that is consistent with an independent test.