Estimating the indirect effect

[Dai Tran]

Dear Community,

I am having a problem with obtaining the standard errors of indirect effect. My research topic is to examining the impact of locus of control on life satisfaction and using physical activities as a mediator. I want to obtain the coefficient and standard errors of the indirect effect. Here are my equations with random effect estimation:

(1) Physical activities = a1 locus of control + a2 control variables + error term
(2) Life satisfaction = b1 locus of control +b2 physical activities + b3 control variables + error term

The indirect effect is a1*b2. I have checked the previous posts and follow this way: estimate both equations, use estimates store and use suest, then nlcom. But xtreg is not supported by suest. I don’t know whether this is the right way. If not, it would be great if you can suggest another way. Can anyone support me to solve this problem?

Thank you and best regards.

[Erik Ruzek]

You could switch over to gsem and then use nlcom to calculate the indirect effect. See example 42g in the SEM manual. Note that nlcom will give you confidence intervals based on a normal distribution. Indirect effects are known to have non-symmetric confidence intervals, so bootstrapping is often used for these.

Depending on your field and its recognition of the problems of classic mediation analysis, you may want to look into more recent work on causal mediation. This work illuminates the assumptions necessary for mediation to be causal and identifies the modeling additions necessary to meet some of these assumptions. A brief introduction to this work with an eye towards available programs in Stata can be found here. Currently none of the available causal mediation packages in Stata deal with nested data. You can calculate the causal quantities (direct and indirect effects) yourself, of course. The only package that can do it automatically for nested data is Kosuke Imai's R package, called mediation.

[Dai Tran]

Thank you very much, Mr. Erik Ruzek.