Mediation

[Muhammad Yahya Khan]

Hello. Hope everyone’s good. Need some help. So my Independent Variable (X) is binary. My Mediator (M) is a continuous variable. My Dependent Variable (Y) is also binary. Due to the nature of Y being binary, I have to do binary logistic regression.

The problem begins here. There are 4 steps of mediation:
1) X to Y (Direct effect)
2) X to M
3) M to Y
4) X to M to Y (Indirect effect)

As my M is continuous, I can’t do step 2. I’m doing this for my thesis. Do I have to do my regression in binary logistic, but then do OLS regression again for the 4 steps for meditation?

Help would be appreciated. Thanks.

[Rich Goldstein]

the issue is with the calculation of the indirect effect as you want comparable coefficients for your multiplication; I recently had a related issue and I did the following: estimate twice, once using OLS for each step and once using some form of logistic regression for each step (i.e., use ordinal logistic regression in step 2; if the results are qualitatively similar, fine (mine was)

[Muhammad Yahya Khan]

the issue is with the calculation of the indirect effect as you want comparable coefficients for your multiplication; I recently had a related issue and I did the following: estimate twice, once using OLS for each step and once using some form of logistic regression for each step (i.e., use ordinal logistic regression in step 2; if the results are qualitatively similar, fine (mine was)

hello Rich. Thanks for the reply. You talked about multiplication? I don’t get it. Isn’t the multiplication for moderation? Where the moderator and IV are multiplied? Whereas I have mediation.

but overall, I get it. I should do binary regression for the direct effect and indirect effect. But for the step number 2 specially, I have to do the OLS regression. I tried this, and the results are qualitively similar yes. The direct effect is significant. The indirect effect (mediation) is insignificant). This is what you mean, right? Thanks.

[Rich Goldstein]

for indirect effects, you multiply the effect of X on M by the effect of M on Y

[Joerg Luedicke (StataCorp)]

mediate fits generalized linear mediation models. For a simple model without covariates and no mediator-treatment interactions (as implied in OP), it could look like:

mediate (Y, probit) (M) (X), nointeraction

For further info, see:
https://www.stata.com/new-in-stata/c...tion-analysis/
https://www.stata.com/manuals/causalmediate.pdf