Decomposition after GSEM

[Joost de Moor]

Dear Stata Users,

I am testing some indirect effects on a binary coded variable using the Bernoulli logit link in Stata's GSEM. The mediator variables are measured on a 7-point scale, and I use OLS to estimate those. Now, I want to decompose the effect of the indpendent variable to obtain the direct and indirect effect, but since this effect is built up of a logistic and a linear regreassion, decomposition is not that straigthforward. In fact, GSEM does not support the post estimation command for decomposing effects from SEM (estat teffects). When doing a 'traditional' mediation analysis built up of seperate regressions, I often use the 'khb' program for this problem, but after a GSEM, this seems inappropriate because the 'khb' program does not allow for several functions I use in the GSEM (like correlating the standard errors of the mediators) - it would therefore decompose a differently estimated effect.

Could anybody suggest any sollutions that would still allow decomposition after GSEM?

Thanks a lot.

Best,
Joost

[Rebecca Bucci]

Was this ever solved? I have a similar dilemma where I want to estimate indirect and direct effects of a categorical IV via two mediators (one dichotomous, one continuous) on a continuous outcome. There are benefits to KHB which I am familiar for dichotomous outcomes, but does the same logic (regarding rescaling of the DV) apply when the dichotomous measure is the mediator (and the outcome is continuous)? Furthermore, I wanted to use GSEM as it is more straightforward for moderation analyses and complex mediation/moderation models, but I cannot find a clear answer for how to estimate the indirect paths when one mediator is a dichotomous measure. I am assuming using the postestimation nlcom command is inappropriate.