Asymptotic Properties of HPD Regions in the Discrete Case

This paper obtains asymptotic expansions of the frequentist distributions of modified likelihood ratio statistics when the observations are discrete. An upper bound of the uncertainty due to the discrete nature of the observations is obtained, which is slightly larger than Yarnold's result (obtained in the case of elliptic confidence regions). Higher order results are also derived from continuity corrections. They are instrumental in the determination of matching priors for HPD regions to the orders o(n-1) and O(n-3/2).