CMStatistics 2021: Start Registration
View Submission - CMStatistics
B1068
Title: On the causal interpretation of randomized interventional indirect effects Authors:  Caleb Miles - Columbia University (United States) [presenting]
Abstract: Natural indirect effects (NIEs) are mediated effects that can be identified when the exposure does not affect any confounders of the mediator-outcome relationship. To circumvent this assumption, so-called randomized interventional analog indirect effects (NIERs), which can be identified even in the presence of exposure-induced confounding, have gained popularity in the causal mediation literature. An essential property that a putative indirect effect must possess in order to have a true mediation/indirect effect interpretation is that it must be null whenever there is no one for whom both their exposure affects their mediator and their mediator affects their outcome. Without additional assumptions, the NIER does not satisfy this property. Further, examples will be provided of such additional assumptions under which this property can be recovered. Unfortunately, the NIE will also be identified under these additional assumptions, and so the NIER will provide little advantage over the NIE. Thus, while the NIER does have a meaningful interpretation pertaining to joint stochastic interventions on the exposure and intermediate variable, it cannot always be relied upon to tell stories about mediation.