How to do mediation analysis with GSEM

[julian maron]

Hi, I seek to do a mediation analysis with the generalized structural equation modeling (GSEM) in Stata 14 (dependent variable, independent variable and mediator are not continuous). Unfortunately, the "estat teffects" command (which is the proper command for doing mediation analysis with SEM) is not allowed for the GSEM (why?). So: Does anyone know an adequate method to do a mediation analysis on the basis of the GSEM? Is Stata the proper software for this? Can I do this in Mplus? Many thanks in advance! Best, Julian

[wbuchanan]

It's possible to do in Mplus, but the Mplus user manual explains why it can be difficult/problematic. In short it is an issue with scaling. For example, if x is a variable containing measurements in inches and is mediated by a variable indicating whether or not the person's mother smoked during pregnancy, how would you interpret the indirect effect? This is obviously a non-sensical model, but helps to illustrate the issue with scaling between variables measured on different measurement scales.

[Dick Campbell]

I'm not quite sure what you are trying to do here, that is, I can't tell exactly what your GSEM model looks like. In any case, mediation analysis is a very active research area right now. If you do a search on "mediation" in Stata you will be led to at least three different packages that will handle the kind of problem that you describe. You can find an extremely thorough discussion in Tyler Vanderweele's book Explanation in Causal Inference, Oxford University Press, 2015. The Stata package paramed implements Vaderweele's methods.

[Roman Mostazir]

The indirect effect is a product term (path a x path b), therefore, needs a non-linear test. Fitting after gsem, use 'nlcom'. Type

help nlcom

. But the standard errors will need bootstrapping to be equivalent of Sobel's test which can be done saving the standard erros from 'rclass' in scalar after gsem and then bootstrapping. This is just to tell you how to test an indirect effect test outside the model, but as others mentioned, it is difficult to grasp what you are after and may require caution if you have both mediator and moderator.