High-dimensional mediation analysis for survival data with latent variables

Understanding the effect of an exposure on a time-to-event outcome is an important task in health research. However, it remains methodologically challenging to delineate the pathways through which an exposure affects survival when key mediators are unmeasured and only numerous surrogate variables are available. This study proposes a mediation framework for survival data that accommodates latent mediators by embedding a latent-factor representation within the time-to-event regression model. An adaptive penalization is employed to select pathways, and mediation effects are estimated through an iterative expectation-maximization (EM) algorithm. We establish the large-sample properties of the proposed estimator. Simulation studies demonstrate its validity. An application to a lung cancer dataset demonstrates the practical utility of the proposed method by identifying several DNA methylation marks that may mediate the causal effect of smoking on patient survival.