Is it complete mediation when my independent variabel has a significant effect on the dependent variable, only after adding the mediator variable?

Question

I am currently working on my masters thesis and I am testing for mediation. For this mediation I am using the Baron and Kenny method. For path c, you need to see if the effect of the X variable (packaging material) on the Y variable (purchase intention) is significant WITHOUT the variable M (ease of use of a packagin). When I take M out of the regression, my X variable becomes insignificant. I know that complete mediation means that an effect of the X variable 'disappears' when you add the mediator M. Now in this example it actually appears when I add my mediator.

Is this still an example of mediation?

PS: I used the Sobel test to doublecheck and it says there is mediation. I just can't seem to understand why this is mediation and how I should write it down in my thesis.

Answer 1

When you add M into the model, things can go either way: the extra predictor can make a previously non-significant X become significant or vice versa. BUT those are for really very different reasons and only one relates to mediation.

Adding M can "tidy up Y" so that you can observe the X->Y relationship better -- and that can make X significant -- that sounds like what's happening for you. That can happen even if X does not have any relationship with M.

Or M can out-compete X as an explanation for Y and "steal" its credibility as the main predictor of Y. That is what you are looking for in this approach to mediation. That M has come along and been so much better as an explanation of Y than X was, that X no longer gets to be a predictor. From the limited info you've given, there isn't yet an indication of that.

BUT -- doing this based on whether it is significant or not has always sounded a bit iffy to me. Significance is as much a question about how much data you have as it is about how big an actual effect is. If you are doing this in SPSS, the PROCESS MACRO is pretty nice and gives some nice confidence intervals on the direct and indirect effects. That would give you a much clearer picture.

Though, from what you've described, it doesn't sound particularly mediation-y.