Stable Discovery of Treatment Effect Modifiers

Abstract: Identifying covariates that modify treatment effects is a critical problem in causal inference. Yet existing data-adaptive methods lack rigorous error control, risking spurious findings that fail to replicate. We propose a method combining pseudo-outcomes with a novel cross-fitted stability selection algorithm to achieve finite-sample false discovery control for effect modifiers. We prove that selection probabilities are asymptotically unbiased, converging to oracle probabilities at parametric rate under doubly robust pseudo-outcome estimation. False discovery is controlled at the nominal level while maintaining power to detect genuine heterogeneity. We demonstrate the method on simulated and real-world data.