Mediation with binary mediator and binary outcome

[Anna Warren]

Hello,

I am trying to run a mediation using the following variables:

PTSD (binary variable - outcome)
Abuse (binary variable - mediator)
PSAI (3 level categorical variable - predictor) For this variable the important comparison is between level 1 and 3 of the variable, so this could be considered binary

I am aware of the paramed programme which allows the use of binary mediators. However, I am hoping to deal with missing data in my variables. Paramed will not work with multiple imputation nor will it work with other forms of handling missing data (e.g. method(mlmv)). The following is the model I ran:

paramed ptsd, avar(psai) mvar(abuse) a0(1) a1(3) m(1) yreg(logistic) mreg(logistic) bootstrap nointer

I am also aware of gsem and can run the following model:

gsem (ptsd <- psai abuse, family (binomial) link (logit)) (abuse <- psai, family (binomial) link (logit))

And can handle missing data with this by using

mi estimate, cdok

However I am not sure whether using gsem is an accurate way of handling binary variables in a mediator, and how I would get the breakdown of direct/indirect/total effects from this, as 'estat teffects' will only run with sem, not with gsem.

If anyone has advice about running a mediation using binary predictors/mediators/outcomes on stata I would really appreciate your input.

Thanks,
Anna