Conditional Generative Adversarial Network for Individualized Causal Mediation Analysis with Survival Outcome

Causal mediation analysis aims to investigate the underlying mechanism of how an exposure exerts its effects on the outcome mediated by intermediate variables.However, existing methods for causal mediation analysis in the context of survival models are primarily focused on estimating average causal effects and are difficult to apply to precision medicine.Recently, machine learning has emerged as a promising tool for precisely estimating individualized causal effects without assuming specific model forms.This study proposes a novel method, conditional generative adversarial network (CGAN)-based individualized causal mediation analysis with survival outcomes (CGAN-ICMA-SO), to infer individualized causal effects with survival outcomes based on the CGAN framework.We show that the estimated distribution of the proposed inferential conditional generator converges to the true conditional distribution under mild conditions.Our numerical experiments indicate that CGAN-ICMA-SO surpasses five other state-of-the-art methods.Applying the proposed method to an Alzheimer's disease (AD) Neuroimaging Initiative dataset reveals the individualized direct and indirect effects of the APOE-4 allele on time to AD onset.