Activity Number:
|
403
|
Type:
|
Contributed
|
Date/Time:
|
Thursday, August 15, 2002 : 10:30 AM to 12:20 PM
|
Sponsor:
|
Section on Bayesian Stat. Sciences*
|
Abstract - #300476 |
Title:
|
Bayesian Analysis of Robust Mixture Modeling Using the Multivariate t Distribution
|
Author(s):
|
Tsung-I Lin*+ and Jack Lee
|
Affiliation(s):
|
National Chiao Tung University and National Chiao Tung University
|
Address:
|
1001 TA Hsueh Road, Hsinchu, International, 30050, Taiwan
|
Keywords:
|
Bayesian prediction ; Gibbs sampling ; Maximum likelihood estimation ; MCMC ; Posterior mode ; Proper prior
|
Abstract:
|
Finite mixture models using the multivariate t distribution provides a useful extension of the normal distribution for the modeling of data containing groups of observations with longer than normal tails or atypical observations. In this paper, we consider a Bayesian approach to t mixture models, while assuming a fixed number of components. Our specification of the prior distributions are proper but weakly informative. For parameters estimation, efficient EM-type algorithms, the ECM and ECME algorithm are derived based on the observed data and partially observed future values. Markov Chain Monte Carlo (MCMC) schemes are also developed to obtain more accurate Bayesian inference for parameters. Issues related to the assessment of MCMC convergence are also considered. Numerical results are illustrated with real data.
|