Does it ever make sense for one to split their sample by values of the dependent variable and then run a separate regression for each group?