CMStatistics 2022: Start Registration
View Submission - CMStatistics
B1567
Title: The central role of the mediator process in mediation analysis Authors:  Caleb Miles - Columbia University (United States) [presenting]
Abstract: Traditionally, mediation analysis involves analysis of exposure, mediator, and outcome, each observed at sequential discrete points in time. The natural direct and indirect effects are then defined based on these three-time points. Identification relies on the assumption that no effect of the exposure can cause both the mediator and the outcome. However, the mediator of interest will often be a stochastic process varying from baseline to follow-up, and its value observed at an individual point in time but a coarse measurement of this process. When the intermediate variable is a mediator, we will argue that earlier instances of the intermediate variable will often be exposure-induced confounders of the mediator at its observed time. Thus, the mediated effects defined in terms of the coarsened mediator process will not be identified. Further, we will argue that the mediated effects of greatest substantive interest are those involving the full mediator process, and that the coarsened mediator process effects can have nonsensical interpretations. To make progress, one must instead rely on strong exclusion restriction assumptions or account for the full mediator process. Lastly, we will discuss an effect decomposition relating the full mediator process's indirect effect to the coarsened mediator process's indirect effect.