Dear Statalists,
I am new to causal mediation analysis (and also statalist) but am looking into conducting a mediation analysis with multiple continuous mediators (M = 7) with a binary exposure (X) and continuous outcome (Y) where some mediators affect one another.
Does anyone know whether the command simply sums the effects of each individual mediator or whether it e.g. fits a single regression for Y with all the included mediators? Now, I have looked through STATA's documentation for the mediate command but have not figured out how it handles multiple mediators, and if there are built in ways of accounting for mediator-mediator effects.
The background:
To estimate the natural direct and indirect effects I would have to account for any mediator-mediator effects. Summing individual mediating effects of e.g. X -> M1, M2,..., Mk -> Y without accounting for effects between mediators such as X -> M1 -> M2 -> Y could end up with a sum of mediated effects over 100% as some pathways would be counted twice (in this case e.g. M2 -> Y in X -> M2 -> Y and X -> M1 -> M2 -> Y). VanderWeele and Vansteelandt (2014) has a paper on this issue.
If it fits a single regression then I am guessing separating the effect of individual mediators M1 and M2 could be problematic? The paper linked above suggests the use of an approach based on inverse probability weights of marginal natural direct and indirect effects and perhaps I could use this with the weight option. Does anyone have suggestions for how I would implement this for just those mediators that do interact with each other and therefore should be weighted? Maybe the svy command could be an option when fitting separate weighted models for only some of the mediators?
I have attached a simple DAG from dagitty.net to accompany my post with three mediators:
Thank you for any help in getting closer to understanding this!
I am new to causal mediation analysis (and also statalist) but am looking into conducting a mediation analysis with multiple continuous mediators (M = 7) with a binary exposure (X) and continuous outcome (Y) where some mediators affect one another.
Does anyone know whether the command simply sums the effects of each individual mediator or whether it e.g. fits a single regression for Y with all the included mediators? Now, I have looked through STATA's documentation for the mediate command but have not figured out how it handles multiple mediators, and if there are built in ways of accounting for mediator-mediator effects.
The background:
To estimate the natural direct and indirect effects I would have to account for any mediator-mediator effects. Summing individual mediating effects of e.g. X -> M1, M2,..., Mk -> Y without accounting for effects between mediators such as X -> M1 -> M2 -> Y could end up with a sum of mediated effects over 100% as some pathways would be counted twice (in this case e.g. M2 -> Y in X -> M2 -> Y and X -> M1 -> M2 -> Y). VanderWeele and Vansteelandt (2014) has a paper on this issue.
If it fits a single regression then I am guessing separating the effect of individual mediators M1 and M2 could be problematic? The paper linked above suggests the use of an approach based on inverse probability weights of marginal natural direct and indirect effects and perhaps I could use this with the weight option. Does anyone have suggestions for how I would implement this for just those mediators that do interact with each other and therefore should be weighted? Maybe the svy command could be an option when fitting separate weighted models for only some of the mediators?
I have attached a simple DAG from dagitty.net to accompany my post with three mediators:
Thank you for any help in getting closer to understanding this!
