Sequential G-Estimation and SEM: Viable Alternatives to Inverse Probability Weighting in Structural Nested Direct Effect Models

• Activity Number: 416
• Type: Topic Contributed
• Date/Time: Tuesday, August 6, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Tom Loeys*+ and Stijn Vansteelandt and Beatrijs Moerkerke
• Companies: Ghent University (Belgium) and Ghent University and Ghent University
• Keywords: G-estimation ; mediation ; SEM ; latent variables

Abstract

Traditional regression approaches for direct effects focus on the residual association between exposure X and outcome Y after adjusting for the given mediator M and baseline confounders C. However, they deliver a biased estimator for the controlled direct effect of X on Y other than through M, whenever there are common causes of M and Y that are themselves affected by X. Robins (1999) accommodated this by introducing linear structural nested direct effect models with direct effect parameters that can be estimated using inverse probability weighting (IPW) under specific no unmeasured confounding assumptions. Goetgeluk et al. (2009) proposed an alternative G-estimator for this controlled direct effect, which avoids IPW. Their estimator typically has less finite-sample bias and more precision by avoiding IPW, and is computationally straightforward to obtain. We extend their estimation strategy to deal with latent variables on the mediator and/or outcome when one or several measured constructs exist. The proposed estimator is easy to obtain using SEM-software. We conclude that IPW-based estimators remain useful, but are primarily indicated to handle certain forms of non-linearity.