0
$\begingroup$

I understand that it is a convention for statisticians to control for mediators to observe the direct effects of an exposure. However, is it appropriate to do so?

Let's look at an example below: Causal diagram of the relationship between variables X, M, and Y

In Scenario A, X is the exposure, and M is the mediator between X and Y. In Scenario B, the diagram is flipped horizontally, making M the exposure and X the confounder between M and Y.

We can then build a regression model Y = Intercept + mM + xX, with m = effect size of M and x = effect size of X.

The model depicts the relationship in both scenarios. However, the interpretation is different. In Scenario A, we would say that the effect of X on Y is x adjusted for M. In Scenario B, we would say that the effect of M on Y is m adjusted for X. But is it okay to treat a mediator the same way we would treat a confounder (i.e., adjustment)?

Improve this question
$\endgroup$
5
  • $\begingroup$ If you think M is a mediator, you need to run more than one regression. Approaches vary beyond that. If you browse the mediation tag you will find a lot of threads. $\endgroup$ Commented Jul 26 at 16:47
  • $\begingroup$ Yes, but a lot of researchers just want to see the direct effect of X and Y, so they control for mediator. $\endgroup$ Commented Jul 26 at 17:05
  • $\begingroup$ The real question is isn't controlling for mediator essentially "stratifying" it, just like for confounders? $\endgroup$ Commented Jul 26 at 17:16
  • $\begingroup$ For a discussion of the issue that you're raising, see MacKinnon, D. P., Krull, J. L., & Lockwood, C. M. (2000). Equivalence of the mediation, confounding and suppression effect. Prevention Science, 1(4), 173-181. $\endgroup$ Commented Jul 26 at 21:57
  • 1
    $\begingroup$ If you control for the mediator, you will definitely NOT get the full causal effect of $X$ on $Y.$ $\endgroup$ Commented Jul 28 at 13:29

0

Reset to default

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.