4

I am testing a mediation model for a research question using PROCESS in SPSS. While I realize Baron and Kenny (1986) would not test this model, I have read quite a bit about it not being necessary for there to be a total effect between x and y.

Where I am confused: How can I make sense of the two positive paths, with a significant negative direct effect?enter image description here

Robert Long
  • 53,316
  • 10
  • 84
  • 148
psychdoctoralstudent
  • 41
  • 2

1 Answers1

4

This seems entirely plausible.

If the direct effect of $X$ on $Y$ is negative, while the mediated effect is positive, these will oppose each other and, if they are of similar magnitude, result in a weakly positive or negative total effect.

This is easy to demonstrate with an example based on your causal diagram/DAG with R (hopefully this is sufficiently self-explanatory to make sense to non R users):

set.seed(1)
N <- 100
X <- rnorm(N)
M <- X + rnorm(N)

So $X \rightarrow M$ is positive with magnitude 1.

Now let us simulate $Y$ such that $M \rightarrow Y$ is positive, while the direct effect $X \rightarrow Y$ is negative (with magnitude negative 1):

Y <- M - X + rnorm(N)

then we find the direct effect, by regressing Y on X and M, since conditioning on M will block the path $X \rightarrow M \rightarrow Y$:

summary(lm(Y ~ X + M))

## X           -0.92542

So we have a strong negative direct effect as expected. But when we compute the total effect we obtain:

summary(lm(Y ~ X))

## X            0.02011

which is consistent with your scenario.

Robert Long
  • 53,316
  • 10
  • 84
  • 148
  • Thank you! I think my main problem now is how to explain this relationship theoretically. I originally hypothesized that X would predict Y (like in a traditional mediation according to Baron & Kenny) but what really confuses me is explaining why there is a direct negative effect. Is this essentially controlling for the other variables? (I actually did Model 21, with moderators significantly moderating the a and b paths, which all panned out, but the mediation part of this is where things are getting confusing which is why my question just addressed the mediation aspect – psychdoctoralstudent Feb 12 '21 at 14:47
  • Worth noting: the same thing happens with the mediation relationship in Model 4 and Model 21. Insignificant total effect, negative direct effect – psychdoctoralstudent Feb 12 '21 at 14:49
  • I can't help explain the theoretical relationship. Statistically it all makes sense. I would suggest drawing a causal diagram / DAG with ALL the relevant variables (measured or unmeasured) and see where that takes you. This may help: https://stats.stackexchange.com/questions/445578/how-do-dags-help-to-reduce-bias-in-causal-inference/445606#445606 – Robert Long Feb 12 '21 at 15:00
  • Thank you for your help! – psychdoctoralstudent Feb 12 '21 at 15:32
  • @psychdoctoralstudent you're welcome. If this answers the question posed in the OP then please consider marking it as the accepted answer and feel free to ask a seperate question about the issue raised in your comments above :) – Robert Long Feb 13 '21 at 09:57
  • Do you know how this type of mediation would be described? I can't find any examples of this occurring. I essentially have the same question as this person who did not receive any answers: https://stats.stackexchange.com/questions/473984/inconsistent-competitive-mediation-vs-suppression-positive-indirect-negative – psychdoctoralstudent Feb 13 '21 at 17:10
  • I would describe it as opposing indirect and direct effects. It's very common. – Robert Long Feb 13 '21 at 17:18