interpreting moderation interaction versus main effects

Question

Moderation interpretation when one predictor is insignificant? My moderation analyses is testing whether self-compassion moderates the relationship between pain severity and emotion (negative and positive, separately). In this moderation, pain severity contributed to significant variance in positive emotion until self-compassion was added to the model (and then it became insignificant). Pain and self-compassion are both unique predictors of negative emotion in the same model. So self-compassion significantly moderates the relationship with negative emotion, but not with positive - which means that at high levels of pain, negative emotion is not going up when self compassion is high (the slope of the line gets flatter). However, the lines are all equally parallel with positive. Is this because in the positive emotion moderation, pain severity's effect gets wiped out by self-compassion when its added in, and what does that mean?

I have another moderation with quality of life as the DV, where there is no significant main effects from pain severity or self-compassion, but only the interaction is significant. Not sure how that works either.

Answer 1

The coefficient for the predictor corresponds to the slope of the predictor when the moderator is at 0. So the coefficient for self-compassion corresponds to the effect of self-compassion when pain is zero. Is this a meaningful value? If not, you can center the pain variable at different levels of pain (e.g., low, medium, and high), run separate regressions, and see what the coefficient of self-compassion is at these levels.

Note that the concept of "main effect" is very different in regression from what it is in ANOVA. In regression, there is no main effect; there is only the effect at the reference level (i.e., zero value) of the mediator. There is a vast but accessible literature on moderation probing, which will help you understand your data.