We wish to model ratings (1-9) based on several predictors. We hypothesize that the effect of some predictors may vary across the ratings. That is, some predictors might distinguish higher ratings (e.g., 7-9) better than lower ratings. In this case, how reasonable is it to discretize the dependent variable into a few categories (e.g., 1-3 into 'low', 4-6 into 'middle', and 7-9 into 'high'), include the resulting categorical variable as another predictor, and examine the interaction between this variable and the other predictors in (ordinal) regression modeling?
Besides the potential issue of discretization (i.e., loss of information), the categorical variable is derived entirely from the dependent variable. Does this cause any problems? I have a feeling that it may, but cannot tell what exactly is the problem. I understand that the main effects of the predictors would not reflect their effects across the entire span of the ratings, but would the interaction terms be non-credible as well?