Primary Submission Category: Generalizability/Transportability
Causal machine learning for generalizing heterogeneous treatment effects
Authors: Vanessa Rodriguez, Karla Diaz Ordaz, Brieuc Lehmann,
Presenting Author: Vanessa Rodriguez*
Many methods exist to generalize inferences from randomized trials to target populations, with most approaches focusing on the average treatment effect (ATE). However, the relative performance of methods for generalising conditional average treatment effects (CATEs) using machine learning (ML) meta-learners remains under explored, particularly under varying degrees of sampling bias, CATE complexity, and runtime confounding.
In this talk, we compare two approaches: the generalized T-Learner, which uses inverse probability of sampling weights (IPSW) but lacks rate robustness when using ML models, and the generalized DR-Learner, a debiased estimator that addresses these limitations. Following an extensive simulation study, we observe that the generalised DR-Learner consistently exhibits lower median mean squared error (MSE) than both the standard and generalized T-Learner in almost all cases, but especially in settings with complex sampling mechanisms and smaller sample sizes.
We also explore the use of conformal prediction to ensure valid inference, which has traditionally been a limitation of using ML methods for causal inference. We present the coverage and interval lengths obtained by using weighted conformal inference, allowing us to obtain prediction intervals for causal effects under covariate shift.
We anticipate these findings will provide practical guidance to practitioners wanting to incorporate ML methods in their analysis.