Causal Mediation Analysis for Nonlinear Models with Confounding

• Activity Number: 572
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Jeffrey M. Albert*+
• Companies: Case Western Reserve University
• Keywords: empirical distribution function ; generalized linear model ; IPTW

Abstract

Estimation of direct and indirect (or mediation) effects is conventionally performed using a product of coefficients approach based on multiple regression or structural equations models. In the case of nonlinear models, causally interpretable estimates of mediation effects may require a nonstandard expression, referred to by Pearl (2011) as the mediation formula. Implementation of the mediation formula typically uses an assumed probability distribution for the mediator based on a parametric model. We develop an approach that avoids such distributional assumptions, by using the empirical distribution function for the mediator. A challenge that arises when applying this method to observational data, is to adjust for measured confounders of the exposure-mediator relationship. We consider weighting approaches to estimate causal mediation effects while adjusting for confounding in the above context. Alternative approaches are compared via simulation studies. The methods are applied to data from a study of the effect of very low birth weight on dental caries in adolescence.