Indirect effects - GSEM

[Serena Menny]

Hello everybody,

I am working on a GSEM model to study the effect of emotions on investment decisions. The model includes four mediators, all estimated with linear (OLS) equations, while the final outcome (investment decision) is binary and estimated with a probit regression.

I would like to compute the indirect effects of emotions through each mediator. However, Stata does not provide a built-in command for indirect effects in GSEM when the final equation is nonlinear, and I cannot simply multiply coefficients from the OLS paths and the probit path because they are on different scales.

Is there a recommended way to correctly compute indirect effects in this mixed OLS–probit GSEM?

Thank you in advance for your help.

[Erik Ruzek]

Yes, you can do the a*b traditional mediation approach with a binary outcome, although I have mostly seen this equivalence when the outcome model is estimated using logistic rather than probit regression. It is worth investigating further. What you cannot do is the comparison of coefficients approach (c and c-prime) when you have a binary outcome due to the fact that the meaning of logistic regression coefficients change when you have different predictors in the model.

See work by Rijnhart and colleagues (2019) on this:
https://doi.org/10.1186/s12874-018-0654-z
https://doi.org/10.1080/10705511.2020.1811709 (note the supplemental material has Stata gsem code for specifying the mediation model)

[Felix Bittmann]

Are you required to use GSEM? Maybe the KHB command is an alternative option for you.

https://www.stata.com/meeting/german...g11_kohler.pdf
https://journals.sagepub.com/doi/pdf...867x1101100306