This is a more conceptual question, rather than mathematical per se, so I will talk through an example.

To test the proportion of an effect (outcome ~ exposure) mediated by a variable I have performed 2 linear regression analyses;

  1. Outcome ~ exposure
  2. Outcome ~ exposure + mediator

By comparing the coefficients I have determined the proportion of the effect (outcome ~ exposure) mediated by my proposed mediating factor.

If the effect size gets smaller I interpret this as some of the effect of the exposure actually going via the mediator.

My question relates to the occurrences where the effect size gets larger after inclusion of the mediator. I have heard the term "negative mediation", but am having trouble conceptualizing this - how can one have a mediator that does not account for any of the variance between the exposure and the outcome?

Thanks for your time. Ideal answers would keep this simple, as I am not a stats guru - a real world example would be great.


Since you ask this conceptually....

If the results change when you add the mediator, the mediator is doing something, but what? Since the traditional use of the term "mediation" is one that reduces the effect, negative mediation would be one that increases the effect.

As for a real world example.... well, I don't have one that's from actual data, but Google found this article.

  • $\begingroup$ Hi Peter, thanks for the answer. I realize that "negative mediation" increases the effect size, but I am just unsure how this is still "mediation"? Sure this is closer to confounding? $\endgroup$ – Luke Jun 25 '13 at 11:26
  • 1
    $\begingroup$ Well, the vocabulary isn't completely clear. According to (this paper)[ncbi.nlm.nih.gov/pmc/articles/PMC2819361/] confounding and mediation are the same thing. MacKinnon is one of the gurus of mediation, so his word is probably good. $\endgroup$ – Peter Flom Jun 25 '13 at 11:36
  • $\begingroup$ Thanks for that great link. Having had a skim, it appears that "negative mediation" is actually known as "suppression" (a 3rd kind of effect that may be observed when adding a 3rd variable to a model - the other 2 being mediation and confounding). Definition of suppression: "a variable which increases the predictive validity of another variable (or set of variables) by its inclusion in a regression equation" $\endgroup$ – Luke Jun 25 '13 at 12:26

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.