Easy estimation of normalizing constants and Bayes factors by posterior simulation: Stabilizing the harmonic mean estimator.

Easy estimation of normalizing constants and Bayes factors by posterior simulation: Stabilizing the harmonic mean estimator.

J.M. Satagopan, M.A. Newton , and A.E. Raftery.

Technical Report 1028, Department of Statistics, University of Wisconsin, Madison.

First issued November 2000. Under review at JRSSb.

(Issued simultaneously as TR 382 by the Department of Statistics, UW Seattle.)

Abstract:


The Bayes factor is a useful summary for model selection. Calculation of this measure involves evaluating the integrated likelihood (or prior predictive density), which can be estimated from the output of MCMC and other posterior simulation methods using the harmonic mean estimator. While this is a simulation-consistent estimator, it can have infinite variance. In this article we describe a method to stabilize the harmonic mean estimator. Under this approach the parameter space is reduced such that the modified estimator involves the harmonic mean of heavier tailed densities, thus resulting in a finite variance estimator. We discuss general conditions under which this reduction is applicable and illustrate the proposed method through several examples.

Postscript copy
NBA Data from which free throw example arises, and some SPlus code to read it.