Primary Submission Category: Mediation Analysis, Mechanisms
Causal Mediation Analysis for Survival Endpoints in Longitudinal Settings with Irregular observations and Informative Cluster Size
Authors: Jun Lu, Ming Wang, Sanjib Basu,
Presenting Author: Jun Lu*
Mediation analysis uncovers pathways through which an exposure A influences an outcome Y via intermediate factors M. While counterfactual-based approaches have advanced causal mediation analysis in non-longitudinal settings, real-world studies increasingly involve long-term follow-ups with irregular data structures, particularly in time-to-event outcomes. These studies reveal that exposures often act through dynamic mediator processes, rather than isolated mediator snapshots. For instance, the NACC study tracks MRI measures over time to explore how exposure of interest influences Alzheimer’s disease risk through irregular MRI measures. Recent methods, including functional principal component analysis and growth curve models, address irregularities but often overlook the multi-dimensional effects of mediator processes, such as mean values, variability, trends, and informative observation times tied to patient risk profiles. We propose a latent class mediation model for irregular longitudinal data with survival endpoints. Our approach captures the mediator process pattern and models the exposure-mediator-outcome relationships jointly. Estimation and inference are conducted using Bayesian methods, with sensitivity analyses addressing time-varying confounding. We demonstrate our method through simulations and an application to the NACC dataset.