Bayesian inference for partially identified smooth convex models

Abstract

This paper proposes novel Bayesian procedures for partially identified models when the identified set is convex with a smooth boundary, whose support function is locally smooth with respect to the data distribution. Using the posterior of the identified set, we construct Bayesian credible sets for the identified set, the partially identified parameter and their scalar transformations. These constructions, based on the support function, benefit from several computationally attractive algorithms when the identified set is convex, and are proved to have valid large sample frequentist coverages. These results are based on a local linear expansion of the support function of the identified set. We provide primitive conditions to verify such an expansion.