0
$\begingroup$

When I use Baron & Kenny's approach to test mediation, I receive the following results (IV: independent variable, M: mediator, DV: dependent variable):

DV ~ IV : IV is significant

M ~ IV : IV is significant

DV ~ IV + M : M is significant, but IV not

Now, the curious thing is that these data are based on an experiment where I manipulated the independent variable IV, but not the mediator M (M measures perceptions of a person who acted differently depending on condition IV). Hence, theoretically, it seems implausible that in the multiple regression M is significant, but IV not. My question then is: what is going on here?

I believe that this question is different from related questions in that I provide information that IV is manipulated and M is not. I know from reading related questions (e.g., here) that two highly correlated variables may compete for variance. Yet it seems strange that a variable that has no causal relation (as it was not manipulated) "wins" this competition for variance over a variable that might have causal influence (as it was manipulated; the two variables IV and M are correlated .58).

Update: I typically use Preacher and Hayes' approach (in which case the mediation analysis is significant). Does this mean that Baron & Kenny's approach is not suited in this situation?

$\endgroup$
1
$\begingroup$

This is precisely what you are looking for: After controlling for the mediator, the IV is no longer significant. This, according to the theory of mediation analysis, is because M is what accounts for the effect of the IV on the DV. Remember that manipulation is a methodological aspect; the statistical model does not know whether or not you manipulated something.

Let's say I assign people to drink (a) a rum and Coke or (b) water (IV), measure their blood alcohol content (M), and then see how well they do on a reaction task (DV).

If you ran a mediation model with this, drinking alcohol (IV) would no longer have an effect on the reaction task (DV) after controlling for blood alcohol content (M). This is exactly what we are looking for: The major shrinkage of the effect of the IV on the DV after controlling for the M shows that the M accounts for a lot of the effect of the IV on the DV.

Now, there are a lot of really good criticisms of mediation analysis and how this might not be justified, but according to the canon of mediation analysis, this is what you would expect; I don't see a problem in the IV no longer having an effect—to the contrary, it is what most researchers look for.

As for B&K vs. P&H, never use B&K. The Baron and Kenny approach assumes that the indirect effect (a path times the b path) is normally distributed, but products of coefficients are not normally distributed. This is why we use bootstrapping instead (e.g., P&H).

$\endgroup$
  • 1
    $\begingroup$ I think I was staring at my screen for too long, of course this is what mediation is about. Thanks for your friendly clarification! $\endgroup$ – Flo Jun 7 '17 at 18:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.