Assessing Mediation Using Principal Stratification in a Meta-Analysis

• Activity Number: 484
• Type: Invited
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Yun Li*+ and Jeremy Michael George Taylor and Michael R. Elliott
• Companies: University of Michigan and University of Michigan and University of Michigan
• Keywords: Bayesian Estimation ; Counterfactual Model ; Identifiability ; Multiple Trials ; Principal Stratification ; Surrogate Marker

Abstract

In this research, we aim to disentangle the causal associations among the treatment, the intermediate outcome (i.e., mediator) and the final endpoint. We focus on the use of principal stratification (PS) approach for estimating the PS direct and indirect effects in the area of surrogate marker research in a meta-analytic framework. We hypothesize that there is a latent joint distribution of the potential outcomes of surrogate markers and those of true endpoints at the population level which is invariant, and that each trial has its own distribution of potential outcomes that is linked to the population distribution. We propose a Bayesian hierarchical principal stratification model to obtain these population-level causal association measures and the trial-specific counterparts that vary from trial to trial. We examine the frequentist properties of our model estimates and the impact of the monotonicity assumption using simulations. We also illustrate challenges in evaluating surrogacy in the counterfactual framework that result from nonidentifiability. We apply the method to data from a large collection of colon cancer trials in which surrogate marker and endpoint are binary.