Let's say I'm using Baron & Kenny's method for mediation, which has the following requirements: (a path) IV has effect on mediator; (b path) mediator has effect on DV; (c path) IV has effect on DV; (c' path) when IV and mediator considered jointly, the effect of the IV is eliminated. Only the mediator has an effect on DV.
In an experiment I've run I exogenously manipulated someone's appearance in a photo. I theorized it would affect people's disgust, and subsequently their desire to be friends. There are very large effects for paths a-c (p all <.0001). However, when I run c', the effect of the mediator disappears (p=.2989) BUT the effect of the IV does not (p<.0001).
Now, if I were looking at the regression results without any knowledge of what the variables are, I would say "oh, I guess you made a mistake. Clearly what you have marked as the IV is the real mediator. Since the effect of the other variable disappeared when both were added, your 'IV' is soaking up the other variable's effect."
Which might be fair, except for one thing: it's theoretically impossible, since the IV in my experiment was exogenously manipulated. That interpretation would say "the level of disgust you have for that person causes them to have a different appearance in the photo, which changes how much you want to be friends."
How do I interpret these results?
Thank you so much for your help!