1
$\begingroup$

I've run a parallel mediation analysis with bootstrapped confidence intervals and found that one of my three mediators had a significant indirect effect, one mediator was very close to significance, and one mediator was not significant at all. The total indirect effect was significant, but the direct effect and total effect were not significant. The coefficient between X and Y was reduced from 0.05 to -0.035 in the full model. Can I say that I have partial mediation in this case?

Thanks!

$\endgroup$
1
$\begingroup$

Since you are using a bootstrap approach, I assume that you do not adhere to Baron & Kenny's (1986) mediation procedure. As you most likely know, anyone following Baron & Kenny's procedure would have stopped their analysis upon finding that there is no significant total effect. In such a case we would conclude that there is simply nothing there to be explained by mediation.

More modern approaches to mediation (e.g., Shrout & Bolger 2002, who advocated bootstrapping) do not require a significant total effect in order to speak of mediation. A significant indirect effect would be considered sufficient. What is more, the distinction between partial or full mediation makes much less sense in this paradigm than it did for Baron & Kenny. If we no longer require a significant total effect, then what becomes of the idea that we have obtained partial mediation if the direct effect is still substantial when controlling for the mediator? It seems to make no sense anymore - we never asked for the total effect to be substantial prior to introducing the mediator, then why ask for a substantial direct effect after the mediator has been added to the model?

And what is a substantial remainder anyhow? A significant one? That does not seem right: our remainder can be statistically non-significant (our Null is that there is no effect, after all) and still be substantial. We are increasingly told not to rely on p-values alone to judge and describe effects. Accordingly, relying on p-values to arrive at a binary classification as "partial" or "full" mediation seems neither necessary nor advisable. In the modern view, mediation is essentially a set of SEM scenarios. I therefore suggest you report your mediation as such, if possible, stating the coefficients and confidence intervals for all relevant paths and avoiding the notion of partial or full mediation altogether.

If you have not read it already, Rucker et al. (2011) demonstrate how significant indirect effect may come about in the absence of a significant direct or total effect (e.g., because of differences in statistical power available to detect effects on different paths).

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Stopped their analysis for lack of direct effect... do you mean lack of total effect? Full mediation removes direct effects. $\endgroup$ – AdamO Jul 6 '17 at 18:15
  • $\begingroup$ Yes, total effect. Sorry about the confusion. $\endgroup$ – Benedict M.J.G. Jul 15 '17 at 21:17
0
$\begingroup$

One possibility is that there is a possible quadratic total effect which flatlines in marginal associations between the "treatment" and outcome. That is to say that the treatment is positively associated with the mediator and has a direct effect positively associated with the outcome, but the mediator is negatively associated with the outcome. To say the mediation effects exist, well that relies firstly on having specified a correct causal model. If we accept this, then lack of statistical significance cannot be taken as evidence against associations, it may merely be insufficient sample size. To better explain your results, make use of a coplot demonstrating the association between the mediator and the outcome over a range of possible treatment values.

| cite | improve this answer | |
$\endgroup$

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.