Identifiable Preference Types and Optimal Menu-Choice Experiments

Abstract

A planner wants to elicit partial information about a subject's preference relation. Preferences are grouped into ``types" relevant to the planner's purpose. Can the planner identify the subject's type through menu-choice data or experiment, without distinguishing preferences within the same type? We provide a complete answer by characterizing identifiable type spaces. The characterization uses a novel geometric condition, called face consistency, which governs distinctions within and across square and hexagonal faces of the permutohedron. For every identifiable type space, we explicitly construct the unique irreducible identifying experiment from its boundary. This experiment is optimal among all identifying experiments under any cost criterion for which adding menus cannot reduce cost. Using the equivalence between identifiability and elicitation with acts, as shown in Azrieli et al. (2021), we study symmetric policies requesting unranked shortlists that contain only acceptable products, include every product preferred to a selected one, and are nonempty whenever an acceptable product exists. Only two such policies are elicitable with acts: requesting the favorite acceptable product or the entire acceptable set.