The Multiset Sampler

Abstract

We introduce the multiset sampler (MSS), a new Metropolis–Hastings algorithm for drawing samples from a posterior distribution. The MSS is designed to be effective when the posterior has the feature that the parameters can be divided into two sets, X, the parameters of interest and Y, the nuisance parameters. We contemplate a sampler that iterates between X moves and Y moves. We consider the case where either (a) Y is discrete and lives on a finite set or (b) Y is continuous and lives on a bounded set. After presenting some background, we define a multiset and show how to construct a distribution on one. The construction may seem artificial and pointless at first, but several small examples illustrate its value. Finally, we demonstrate the MSS in several realistic examples and compare it with alternatives.

Keywords:  :

• Data augmentation
• Gibbs sampler
• Markov chain Monte Carlo
• Metropolis–Hastings algorithm
• Multimodal
• Proposal distribution