mediation in panel data - mediator and IV appear interchangeable?

[Isabelle Solal]

Hi,
I am struggling with a mediation model I am doing on panel data (using the sureg command and instrumented variable estimation approach, followed by bootstrapping.)
I get significant mediation effects, in the direction expected, using my intended mediator. But I am also getting significant mediation effects when I switch the mediator and the IV. Why might this be happening?

I have tried two IV variables (one continuous, one categorical), several mediator variables (all categorical), as well as instruments for the mediator (also categorical) to address any endogeneity problem. I also have controls in the sureg command. But everything I try comes out the same way: significant mediation for both IV->Med->DV and Med->IV->DV.

I feel like there must be an obvious reason why this is happening, but I can't figure it out!

Thanks for your help,
Isabelle

[ben earnhart]

It comes down to the basics: correlation (or covariance) is not causation. There is no way for the model to detect what direction the relationships go. So swapping layers of causality won't make a difference. Fall back on theory to defend which is IV and which is mediator.

[Phil Bromiley]

If you were doing interactions, then the empirical evidence cannot differentiate between x1 moderating the influence of x2 on y and x2 moderating the influence of x1 on y. They are empirically identical.

The association of IV and Med will be the same regardless of which appears as the iv and which as the dv. However, the association between Med and DV and between IV and DV will generally not be the same.

I don't do a lot of mediation models, so take what I suggest with that caveat. Traditional mediation analysis interprets mediation as meaning that IV ->DV has a significant coefficient on IV, IV->Med has a significant coefficient on IV, and Med->DV has a significant coefficient on Med. Then, when you add IV to the equation where Med->DV, you test whether the influence of IV on DV is fully mediated by Med or only partially mediated by Med. In full moderation, IV is insignificant in the last estimate, and in partial mediation both IV and Med are significant.

So, there is no problem that IV->DV is significant. I suspect some would claim without IV->DV it is not correct to talk about Med mediating the influence of IV on DV.

However, there has been a lot of methods work recently on mediation estimation that casts doubt on the traditional independent regressions approach. I know there are other relevant publications, but the paper I am familiar with is Myles Shaver, "Testing for Mediating Variables..." J of Management 2005. This is available at https://faculty.fuqua.duke.edu/~char...ions_Paper.pdf

Phil