by **mcfanda@gmail.com** » Wed Mar 17, 2021 10:42 pm

Hi

Ordered variables can be used in GLM mediation only if you can assume that they are continuous variables, otherwise, the GLM underlying the mediation analysis is not correct. If you really have ordered categories, you cannot use glm mediation. At the moment, there are not many options available in terms of software, because I'm not aware of any software (among the main statistical software) that allows doing categorical ordered dependent variables mediation.

Nonetheless, you can run a "logical" mediation analysis: First, establish that the IV has an effect on the DV, then on the Med, and then run IV+Med predicting the DV. Each of these models is run with different types of analyses (ordinal logistic or GLM) depending on the scale of the dependent and med variables. You have evidence of mediation if IV has an effect on Med, and Med has an effect on DV keeping constant IV. The limitation of this approach is that you cannot compute the mediated effect, because the coefficients (a and b in the slides) come from very different analyses, so they have different scales and the interpretation. Nonetheless, the logic of mediation holds.