Dear Isabel,

is see not rational, why it is necessary to dichotomize any variables, it has been shown quite the opposite, that it is detrimental in most cases. In fact, in his 2015 paper Hayes demonstrates the index of moderated mediation for continous variables and shows additionally that it can be also done with (natural) dichotomous variables (which is just a special case in the linear model). Therefore, your arguments seem not right to me.

BUT you are right that model 58 does not give you the index if the variable is continous (contrary to model 14, where it DOES prints it for continous W) and only if dichotomous. Therefore, I would assume that this is up to the programming of the PROCESS makro and does not indicate statistical constraints. Maybe I am wrong about it, but I suggest to contact Dr. Hayes directly. In my experience he will answer quite quickly to requests. It would be helpful if you could post a possible answer by Mr Hayes here on RG. THX

Hi Rainer, thank you for your suggestion. I have contacted Prof. Hayes. I will post his response to the group. On the Hayes website, he recommends the approach followed by Edwards and Lambert which is to split M based on what makes clinical sense. This may not always be clear, as in my case.

Apart from this Hayes recommends his new book published this year named The Cambridge Handbook of Research Methods in Clinical Psychology, edited by Aiden G.C. Wright and Michael N. Hallquist. I couldn't find it in Google Scholar yet or in my library yet, but it is available on eBay.

Adam and Lambert Edwards, J. R., & Lambert, L. S. (2007). Methods for integrating moderation and mediation: a general analytical framework using moderated path analysis. Psychological methods, 12(1), 1.

Dear Isabel,

I found something in his 2020 paper (not the book chapter):

"Observe that Equation 13 is a nonlinear function of W. When W is a continuum, there is no index of moderated mediation for the first- and second-stage conditional process model, and so a test of moderation of the indirect effect of X by W must be conducted differently than in models with only a single path defining an indirect effect specified as moderated. Hayes (2015, 2018a) recommends choosing two values of the moderator W, w1 and w2, and constructing a bootstrap confidence interval for the difference between the conditional indirect effects of X at those two values. If the confidence interval does not include zero, this supports moderation of the indirect effect by W. But a confidence interval that includes zero does not disconfirm moderation of mediation, because it is possible that alternative choices for w1 and w2 would produce a bootstrap confidence interval for the difference that does not include zero. So, this approach can be used to confirm but not disconfirm moderation of mediation."

So, in the special case of two W paths there seems to be another complication (I didnt think about it enough, before answering, to be honest), but his recommendation is NOT to dichotomize the variabel, but to chose 2 (hopefully meaningful) values for W for which you then calculate the indirect effects and test them. This is completely different to dichotomization. It is similar to "probing" the moderation, to use Hayes' terms and then comparing the indirect effects with each other.

Hayes, A. F., & Rockwood (2020). Conditional process analysis: Concepts, computation, and advances in modeling the contingencies of mechanisms. American Behavioral Scientist, 64, 19-54.

Good point Rainer

Thank you, Haroon, this is good background information that others may find useful.