# Mediation via regression analysis - b path non-signifcant - how do I interpret?

I'm doing mediation analysis using hierarchical regression, using Baron and Kenny's 4 steps. I have one IV (mental health stereotype activation), one mediator (rejection expectation), and one DV (comfort with disclosure). Path c (IV -> DV) is significant, Path a (IV -> M) is significant, and Path b (M -> DV) is significant

In the second step, when I add the M to the IV, Path c' (IV -> DV) is significant, but the link between M and DV is non-significant.

Does this mean that M is not a mediator? Does it also mean that IV -> DV is significant even when M is controlled for?

A second related question, I have a variable that I would like to control for (well-being). Can I do mediation analysis using hierarchical regression, whilst controlling for another variable? To put this in context, I hypothesis that stereotypes cause discomfort with disclosure partly because people expect rejection. However, I believe this will be affected by how unwell one is (this factor does correlate with IV, M and DV). Is this a moderator?

I'm a novice in statistics, so this is all quite new to me and many of the tutorials I'm reading have significant outcomes.

• If you were my +1 for this answer you may want to look it over again. I posted the answer without finishing initially and have added to it to address your other questions. Commented Feb 12, 2018 at 14:44

If you go back to Baron and Kenny's 1986 article, you will see that they outline the following conditions for mediation.

1. The IV is related to the DV (seems to be met in your case)
2. The IV is related to the mediator (M).
3. M is related to the DV, controlling for the IV.
4. The relation between the IV and the DV is appreciably weakened (may or may not be reduced to non-significance, but the effect is not as strong when M is in the equation).

If any one of those 4 conditions does not hold, according to Baron and Kenny, you don't have mediation.

Regarding your second question, you can include a covariate like well-being. It just needs to be in every equation required of the mediation analysis.

However, you should note that you seem to have asked about including a covariate (i.e., controlling for well-being), which is separate from testing the hypothesis that you have presented - that the relations between your variables are influenced by well-being. That sounds a lot like a moderation model to me.

If that is where you want to take this analysis then you may want to look into moderated-mediation models. The Wikipedia page may be a good place to start.

As Matt said, you do not have mediation because the third requirement of the causal steps approach is not met.

I would encourage you to explore alternative options to the causal steps approach. It's quite possible you're getting a Type II error (since the M -> DV relationship was significant before it was entered into the mediation model). You should look into a bootstrapping method; this will increase your power. The attached reference below may be helpful to you.

Hayes, A. F. (2009). Beyond Baron and Kenny: Statistical mediation analysis in the new millennium. Communication Monographs, 76(4), 408-420.