BIostatistics Seminar : Doubly Robust Causal Interaction Analysis for Multilevel Exposures in Observational Survival Studies Using Restricted Mean Survival Time

About this event

Dr. Arnab Aich, University of Tennessee Health Science Center

Doubly Robust Causal Interaction Analysis for Multilevel Exposures in Observational Survival Studies Using Restricted Mean Survival Time

Whether two exposures interact in their effect on survival is a common question in observational epidemiology. The conventional product term in a proportional hazards model measures interaction on the noncollapsible hazard ratio scale and restricts its form for multilevel exposures. I develop a causal framework for interaction between two multilevel exposures on the restricted mean survival time scale, which is collapsible and interpretable in units of time. The target estimands are population-standardized restricted mean survival times for each joint exposure level and contrasts that measure departures from additivity of the single-exposure effects. These are estimated by an augmented inverse probability weighted estimator that combines a generalized propensity score, an outcome regression, and a correction for censoring. The estimator is doubly robust with respect to both the exposure and censoring models, and its influence function supports design-based inference under complex survey sampling. A simulation study shows that the estimator has small bias and close to nominal coverage when a single working model is misspecified. We illustrate the method with NHANES data on smoking, alcohol consumption, and all-cause mortality, in which smoking is associated with a graded loss of survival time but there is little evidence of interaction.