On Optimal Imputation Strategies for Natural Effect Models Probing Mediation

• Activity Number: 184
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Johan Steen*+ and Tom Loeys and Beatrijs Moerkerke and Stijn Vansteelandt
• Companies: Ghent University and Ghent University and Ghent University and Ghent University
• Keywords: mediation analysis ; natural effect models ; imputation ; causal inference ; indirect effect

Abstract

Natural effect models parameterize the natural direct and indirect effects of an exposure on an outcome in function of baseline covariates. Vansteelandt and colleagues introduced a regression mean imputation strategy for fitting these models. Compared to direct application of the mediation formula, this framework allows for (i) more easily interpretable effect estimates and (ii) more convenient hypothesis testing. So far, the statistical properties (robustness to model misspecification, and efficiency) of the considered imputation strategy are not well understood. For instance, in non-linear settings where traditional product-of-coefficients estimators for the indirect effect are applicable for assessing mediation under the null of no effect of exposure on mediator, these estimators outperform the imputation estimators in terms of efficiency when the imputation model is fitted using MLE. In this talk, we will discuss advanced imputation strategies to improve efficiency and robustness. We will discuss their implementation using the R package flexmed.