Mediation Analysis of Randomized Experiments

• Activity Number: 115
• Type: Topic Contributed
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Stijn Vansteelandt*+ and Sjouke Vandenberghe
• Companies: Ghent University and Ghent University
• Keywords: Mediation analysis ; Double robustness ; Efficiency ; Semiparametric inference ; Bias

Abstract

Motivated by the analysis of the EORTC 10994 /BIG 1-00 randomised phase 3 trial on the effect of taxane and anthracycline based chemotherapies on the overall survival of advanced breast cancer patients, we focus on the general problem of mediation analysis of randomized experiments. Traditional product and difference of coefficient methods for mediation analysis lack theoretical validity and, moreover, ignore knowledge of the randomization probabilities. In view of this, building on work by Tchetgen Tchetgen (2011) and Tchetgen Tchetgen and Schpitser (2012), we develop semi-parametric efficient strategies to decompose the intention-to-treat effect into a (natural) direct and indirect effect. By exploiting the known randomization probabilities, the considered strategies are double-robust: valid when either a working model for the mediator or a working model for the outcome is correct, but not necessarily both. We focus on optimal strategies for fitting these working models, with a view on bias under misspecification of both models, and give results from the analysis of the EORTC 10994 /BIG 1-00 phase 3 trial.