Department of Statistical Science
Duke University

presents:

Mario Peruggia
mperuggia@virginia.edu
Department of Health Evaluation Sciences
University of Virginia

"A Hybrid MCMC Algorithm for Fitting Outlier Accommodation Models"

Abstract:

In Bayesian outlier detection, the usual extended models include a set of 0-1 variables that indicate whether or not specific observations are outliers, and continuous components that quantify the amount by which outlying observations deviate from ``typical'' observations. In principle, the Gibbs sampler provides a computational tool to analyze the required posterior distributions.

However, it has been recognized by various investigators that, in this and similar problems with mixed discrete-continuous priors (e.g., Bayesian variable selection), specific combinations of modeling assumptions and observed data can give rise to a posterior distribution that assigns high probability to several nearly disjoint regions. In such instances, the Gibbs sampler does not mix well over these regions and its output yields poor estimates.

As an alternative, a hybrid MCMC algorithm is proposed that allows more rapid movement among the nearly disjoint regions of the posterior. The algorithm uses a clustering initialization followed by a mixture of pure Gibbs and Metropolis-Hastings cycles. The implementation of the algorithm will be illustrated with some simple examples.

This is joint work with Thomas Santner of the Ohio State University.

January 30, 1998

3:30 pm - 4:30 pm

116 Old Chem Building