7.11 Takeaway Points: Interactions and Cross-Model Comparisons
The identification assumptions of logit/probit (distribution of the errors) and the functional form restrictions (compression) cause the coefficients to be inappropriate as tests of interaction, mediation, and other cross-model comparisons
Taking the methods you learned for linear models and applying them in the same way to nonlinear models can lead to incorrect conclusions
Everything discussed in this section of the course is applicable to nonlinear linear regression models and all models for categorical DVs (e.g., binary logit/probit, ordinal logit/probit, multinomial logit, count models, etc.)
Predicted probabilities and average marginal effects are unaffected by the identification assumptions in logit/probit
MEs should generally be the focus of tests of interaction, mediation, and other cross-model comparisons