UCLA

Recent work has considerably advanced the definition, identification and estimation of different types of effects in causal mediation analysis. It extended the traditional approaches for mediation analysis by using the counterfactual or potential-outcome framework to allow for nonlinearities and exposure-mediator interaction. Despite the various estimation methods and statistical routines being developed, a unified approach is lacking, which incorporates recently introduced causal decompositions. Also, relatively few studies explored scenarios with more than one mediator. In this work, we used causal diagrams and potential-outcome framework to contribute to the literature on causal mediation analysis. We first provided a unified framework for estimating targeted effect(s) from the most recent 4-way decomposition in the single-mediator setting. We demonstrated that g-computation, implemented via Monte Carlo simulations in standard statistical software, can offer such unification and is flexible in accommodating different types of exposure, mediator and outcome variables. We also extended some of the existing estimation techniques to more complicated mediation settings that involve contextual exposure, intermediate confounding, multiple causally ordered mediators, time-varying mediators, and time-to-event outcomes. We applied regression-based techniques, g-computation, and inverse-probability-weighted (IPW) fitting of marginal structural model (MSM) to investigate mechanisms underlying the effects of human development on individual health, the health disparity in education, and the effect of different physical activity domains on acute myocardial infarction. The flexibility of g-computation comes at a large cost: it becomes computationally intensive as the number of variables and sample size grow. Alternatively, mediator and outcome regression-based methods and IPW fitting of MSM can be applied in general linear systems and survival context respectively. The use of causal inference techniques did not preclude the possibility of model misspecification and the presence of uncontrolled confounding, which may bias our results. Future work should explore the properties of different estimation techniques and their use in estimating targeted quantities, and incorporate sensitivity analysis for uncontrolled confounding in causal mediation analysis.