In order to carry out a mediation analysis, we first we need to decide on which type of regression analysis that fits our outcome (y) – it could be any type (e.g. linear, logistic, ordinal, multinomial, or some other type). As we have described in this guide, the choice depends largely on the measurement scale of y.

Performing mediating analysis in the traditional way (i.e. including covariates in a stepwise fashion and comparing the estimates across models) only works satisfactorily if we perform linear regression. If we do a non-linear regression (e.g. logistic, ordinal, multinomial, Poisson, or Cox), then we should consider a different approach.