Hello Yanchen,

I would assume that the results are not comatable with a suppression situation, as in the suppression, the distal predictor variable is not related to the outcome but when included in the equation, suddenly becomes (with a nonsensical but significant effect). In your model, this seem to be the c-effect (which is .35 and highly sign.).

If the "direct effec" when considering the mediator (c') is of opposite design as it makes sense, this could rather be a result of an unobserved confounding of the relationship between the mediator and the outcome. However, as, c' is non-significant, it is a weak sign for that. At any rate, you seem to have applied the Baron-Kenny approach. Optimally, you would have some exogeneous variables with effects on the independent and mediator variable. And yes, every mediaton model is statistically equivalent to a confounding model--that's why we should make some efforts in the planning phase of the study to rule that out.

HTH

Holger

Dear Holger Steinmetz , thanks for your respinse! Now I know for sure that it is not the suppression if the total effect was significant.

And I also found that David Kenny mentioned the "Inconsistent Mediation" in his online tutorial on SEM, "If c' were opposite in sign to ab something that MacKinnon, Fairchild, and Fritz (2007) refer to as inconsistent mediation". Would this be more in line with my results?

Thanks again!

David Kenny's online tutorial on SEM

https://davidakenny.net/cm/mediate.htm

Hello Yanchen,

after looking at Kenny's website, I assume his term "inconsistent mediation" is rather descriptively meant in that regard that you may have a positive indirect effect plus a negative (true) direct effect. If the direct and indirect effect are of equal but opposite sign, then both eliminate each other with the result that the overall relationship or total effect is zero (although an indirect effect exists). A comparable situation is when you have two mediators with opposite design.

Imagine the scenario in which you have a dummy predictor (broken leg, 0 [no] or 1 [yes]) and a single outcome "overall mood". Then imagine one mediator being "pain" and another being "getting attention". A broken leg has a positive effect on pain and pain has a negative effect on mood (=negative indirect effect). Likewise, a broken leg has a positive effect on "getting attention" which then positively affects mood (thus, a positive indirect effect). If both effects are equally strong, then the overall relationship between a broken leg and mood is zero. Beware however that this is no bias but still shows that things often have negative AND positive consequences which may counterbalance each other.

Now back to Kenny mentioning that this is like suppression. IF he simply sticks to the empirical phenomenon (without any causal explanation) of suppression--namely a zero overall effect but a negative direct effect, he is correct and this looks like inconsistent mediation. However, it is more important to consider (and THIS is the fundamental difference to inconsistent mediation) that the suppression scenerio incoporates wrong and biased effects (remember the term "true direct effect" above whereas in the incostent mediation, all effects are correct (ok we assume that for the sake of the argument). Suppression results from a causal structure in which NONE of the variables has any effect on the outcome but this strange behavior results from a sort of bias called "collider bias" in the causal inference literature. Recently, this paper solved the riddle and at least proposed two forms of causal structure that leads to the suppression scenerio (whereas all of psychology including Kenny) has failed for decades to really explain it.

Kim, Y. (2019). The causal structure of suppressor variables. Journal of Educational and Behavioral Statistics, 1076998619825679. doi:10.3102/1076998619825679

For most people this will be rather shocking, as for decades (as so often), people focused on the enhanced prediction results for the target predictor following the inclusion of the suppressor. In this regard, suppression was deemed a good thing. Enhanced prediction, yes, but no effect at all.

HTH

Holger