Results from gsem are not consistent?

[Max Coleman]

I'm trying to test mediation effects with gsem, but my total effects aren't adding up correctly. If I have a binary outcome and treat it as linear, things are okay:

sysuse auto
gsem (trunk -> mpg)(trunk mpg -> foreign)
nlcom _b[mpg:trunk]*_b[foreign:mpg] + _b[foreign:trunk] = -.0386701
// Indirect + direct = total

This total effect is the same as:

gsem (trunk -> foreign) = -.0386701 √

However, if I use logit, the total effect from the mediation model is different from the total effect from the simple model.

gsem (trunk -> mpg)(trunk mpg -> foreign)
nlcom _b[mpg:trunk]*_b[foreign:mpg] + _b[foreign:trunk] = -.2149484

vs.

gsem (trunk -> foreign, logit)
lincom _b[foreign:trunk] = -.2092823

This makes me unsure which to use for the denominator when I calculate % mediated as indirect effect over total effect.

sysuse auto
gsem (trunk -> mpg)(trunk mpg -> foreign, logit)
nlcom (_b[mpg:trunk]*_b[foreign:mpg]) / -.2149484  = 39.6%

or

nlcom (_b[mpg:trunk]*_b[foreign:mpg]) / -.2092823 = 40.7%

These differences might not seem like much, but in the actual data I'm working with, the difference can be large. Any advice?

[Genping Ye]

Hi Max, how do you solve this issue? I have the same problem. Looking forward to your advice

[Erik Ruzek]

Unless you are dealing with random intercepts or slopes, I would personally not use gsem for these analyses. Instead, I would suggest Hicks & Tingley's (2011) mediation program and the associated medeff and medsense commands. medeff can handle mediators and outcomes that are binary and medsens produces sensitivity analyses that tell you how large a common cause of M and Y would have to be to eliminate the observed mediation effect.

net sj 12-2 st0243_1

[Felix Bittmann]

If you have a binary outcome variable, I would suggest the KHB approach.

https://www.stata.com/meeting/german...g11_kohler.pdf

https://gitup.uni-potsdam.de/ukohler/khb