I have a question about the interpretation of mediation models.

So i have a situation where I meet all but one of requirements for a significant mediation model.

This is my modelenter image description here

It meets the following requirements:

X significantly predicts M

X predicts Y not controlling for M

The relationship between X and Y is reduced to non-significance when controlling for M

And the indirect effect through M is significant (due to the 95% CI not including zero)

So the only thing it doesnt satisfy is that the relationship between M and Y is not significant when controlling for X.

However, when I look at the simple relationship between M and Y (without X) it is significant.

Therefore I am confused how I would interpret these results. Clearly M is important in explaining the relationship between X and Y but without M to Y being significant, I can't claim that it is mediated by it.

Could this circumstance result from a lack of power? Is there not enough variance when controlling for X to find a significant relationship between M and Y? (I have a sample size of 27).

I really appreciate your help

  • 1
    $\begingroup$ "So the only thing it doesnt satisfy is that the relationship between M and Y is not significant when controlling for Y." — Do you mean "when controlling for X"? (If so, edit your question appropriately rather than replying with a comment.) $\endgroup$ – Kodiologist Aug 18 '17 at 16:03
  • $\begingroup$ A couple thoughts: 1) You might find my answer to a similar question (though it's focused on the c' path) helpful, because I see your question rife with causal steps logic for testing mediation, and there are many reasons why this approach is overly conservative: stats.stackexchange.com/questions/185626/… and 2) How are estimating the CI for the indirect effect? It is so borderline, that alongside the non-significant b path, I would be tempted to check its robustness if you are bootstrapping $\endgroup$ – jsakaluk Aug 18 '17 at 16:58
  • $\begingroup$ Hi jsakaluk, thanks for your response. Your answer on the other thread was very helpful. In terms of the causal steps logic, as soon as I find that the b path is not significant I should not proceed to the next step - so really does that mean that I should ignore all the results that come after that and not try to interpret them? the problem is with the method I used it gives me all the paths in one go so its tempting to try to make sense of what is going on. $\endgroup$ – neij Aug 19 '17 at 9:42
  • 1
    $\begingroup$ In regards to 2, i'm using the bootstrap method using Hayes Process model in SPSS. Are you suggesting Bootstrapping doesnt help with robustness? I was under the impression that bootrapping gives you more confidence in the robustness of your results (but i'm not that familiar with it)? But yes I agree that the confidence intervals are borderline. I used 5000 bootstrap samples, but perhaps if I used 10,000 the confidence intervals would include zero? $\endgroup$ – neij Aug 19 '17 at 9:42
  • $\begingroup$ The point of the other post is that the causal steps process has some noteworthy flaws, so while your mediation "fails" according to it, I would put greater stock in the bootstrapped estimates. But yeah, I would boost your resamples to 10,000 and see if the indirect effect still holds up. You may have just gotten an (un)lucky group of resampling "rolls" $\endgroup$ – jsakaluk Aug 23 '17 at 2:01

Your Answer

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

Browse other questions tagged or ask your own question.