With evoked sets, we distinguish between good and bad alternatives, where bad alternatives are those the respondent says they would not choose. This means we can avoid showing the bad alternatives (all or subset). But this in turn requires informing the model that those bad missing alternatives are not missing at random, but because they are bad. We review a few of the approaches to this problem, and describe a new approach modeling ~min(good utilities) > ~max(bad utilities), where ~min and ~max are continuous approximations to min and max functions.
