Mediation Analysis

Ashley Naimi

Mediation analysis has become extremely popular in recent years, and there is now a vast literature on the topic. Below are a selection of articles that cover a range of topics related to defining, identifying, and estimating direct and indirect effects. They are seven in total, and represent two foundational articles, two conceptual articles, and three methods articles. The articles by Robins and Greenland (1992) and by Pearl (2001) are the first formal treatments of mediation analysis in the field of causal inference. In many ways, these set the stage for much of the research that followed. The reviews by Hafeman and Schwartz (2009) and by VanderWeele and Vansteelandt are important reviews/introductions to mediation analysis, albeit at different technical levels. Hafeman and Schwartz (2009) stands out because it highlights the distinction between natural, pure, and total direct and indirect effects. This distinction is often not considered, but can be important when there is exposure-mediator interaction.

The last three papers demonstrate how to estimate direct and indirect effects using different methods. Vansteelandt’s method is easy to use, and relies on modeling the outcome to estimate direct effects. VanderWeele’s method is also straightforward, but relies on modeling the exposure and mediator, instead of the outcome. Finally, the article by Naimi et al reviews both of these, plus two additional approaches (g estimation, TMLE), and illustrates the importance of double-robustness. While focused on health disparities, the methods in this latter paper can be applied to any substantive context where mediation is of interest.

Leave a Reply