Targeted Maximum Likelihood Estimation of a Natural Direct Effect

• 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): Mark van der Laan*+ and Wenjing Zheng and Alan Hubbard
• Companies: University of California at Berkeley School of Public Health and University of California at Berkeley and University of California at Berkeley School of Public Health

Abstract

Our talk focuses on recent advances in locally efficient estimation of natural direct effects rather than identifiability issues, about which a rich literature exists (e.g., Robins and Greenland, 1992; Pearl, 2000; Robins 2003; van der Laan, Petersen 2004; Hafeman and VanderWeele 2010; Imai et al 2010; Pearl 2011). We discuss estimation of the NDE based on Targeted Maximum Likelihood Estimation (TMLE) in three different settings. First, we consider estimation of NDE in the context of randomized treatment such as gender (Chapter 8, van der Laan, Rose, 2011). Next, we allow treatment to be confounded by baseline covariates, allow for continuous multivariate intermediate mediators, and target the natural direct effect among the untreated. The TMLE for the latter parameter avoids inverse weighting by the conditional probability of the mediator, and is thereby well adapted to handle continuous multivariate mediators. Finally, we present a locally efficient TMLE of the NDE which is shown to be triple robust. We demonstrate the virtue of the general approaches with simulations and an analysis of data targeting the natural direct effect of gender on salary at academic institutions.