Ceiling effect in mediation AND/OR moderation

Question

I have a variable (Y) with a mean that is probably too high. (and large prop of participants scoring the maximum score)

My hypothesis was that X-Y is stronger at high (vs. low) levels of the moderator M, but now I see that the interaction comes out to be significant but in a way that X-Y assocation is stronger at lower, rather than higher, levels of M. I think this is because of the ceiling effect (i.e., it was not possible that X-Y at high levels of M can be higher than one at low levels of M because there was no way Y could go further up).

But I was wondering if it makes sense to use this variable as a mediator.
I have another hypothesis using this variable as a mediator and the indirect effect does come out to be significant, and in a predicted way. So I was wondering if it is more of a problem when such a variable is used as a outcome variable in a moderation than as a mediator...

On the one hand, this does not make sense because after all mediation IS a regression analysis and uses this variable as a DV in some way, but on the other hand, looking at how its association with another variable varies by the level of the third variable (DV in a moderation) does look different from just looking at it as a mediator...

I would really appreciate any help!!!

Answer 1

I don't think that having the results in hand and rewiring a post-hoc solution to the problem should drive how you construct models. You were already testing a model, now you have to deal with that information rather than reconstruct that narrative. I would simply ask yourself why it was that this turned out the way it was, since you have not provided a plausible theoretical rationale for this difference in results.

The ceiling effects probably have little to do with this issue unless you have some defensible reason for why ceiling effects address the theory you were testing. But without more information about what specifically you were testing in your model, it is difficult to say for certain.