mediation analysis with binary outcome and survey weights (gsem in Stata 14)

[Laura Gibson]

Hello -

I'm trying to interpret the indirect and direct effects of a weighted mediation model with a binary DV, a 4-category mediator that I'm treating as continuous, and a set of two dummy variables for the categorical IV. It looks like I can only run a logit model with gsem. Gsem works with svy, but unfortunately estat teffects does not.

svy, subpop(if age2cat==1): gsem(dv<-iv1 iv2 med) (iv1<-med) (iv2<-med), logit

How might I go about calculating the indirect and direct effects from the output of this model?

This is the output:

(running gsem on estimation sample)

Survey: Generalized structural equation model

Number of strata = 1 Number of obs = 3,181
Number of PSUs = 3,181 Population size = 3,186.0806
Subpop. no. obs = 1,099
Subpop. size = 815.754285
Design df = 3,180

Response : dv Number of obs = 1,099
Family : Bernoulli
Link : logit

Response : iv1 Number of obs = 1,099
Family : Bernoulli
Link : logit

Response : iv2 Number of obs = 1,099
Family : Bernoulli
Link : logit

-------------------------------------------------------------------------------
| Linearized
| Coef. Std. Err. t P>|t| [95% Conf. Interval]
--------------+----------------------------------------------------------------
dv <- |
iv1 | .6390778 .4056427 1.58 0.115 -.1562701 1.434426
iv2 | 1.305039 .3572413 3.65 0.000 .6045924 2.005486
med | 1.242894 .1640752 7.58 0.000 .9211896 1.564598
_cons | -5.682145 .4373785 -12.99 0.000 -6.539717 -4.824573
--------------+----------------------------------------------------------------
iv1 <- |
med | -.0006646 .1235467 -0.01 0.996 -.2429038 .2415746
_cons | -.8246543 .2127554 -3.88 0.000 -1.241806 -.4075025
--------------+----------------------------------------------------------------
iv2 <- |
med | .3443353 .1153281 2.99 0.003 .1182104 .5704602
_cons | -1.935966 .2165425 -8.94 0.000 -2.360544 -1.511389
-------------------------------------------------------------------------------

Thanks so much,
-Laura

[Laura Gibson]

Hello again,
I realize that the direct effects are what I'm seeing in the output now. So I'm just looking for how to calculate the indirect effect.
Thanks,
-Laura

[Alejandro Sevilla]

Have you tried nlcom?
Something like

gsem (dv <- med iv, probit) (med <- iv, reg)

//Indirect effect
nlcom _b[dv:med]*_b[med:iv]

//Total effect
nlcom _b[dv:iv]+_b[dv:med]*_b[med:iv]

where dv=binary dependent variable; med=continuous mediator; iv=continuous independent variable