|
Activity Number:
|
150
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Monday, August 7, 2006 : 10:30 AM to 12:20 PM
|
|
Sponsor:
|
Section on Bayesian Statistical Science
|
| Abstract - #307506 |
|
Title:
|
Label Switching in Finite Mixture Models
|
|
Author(s):
|
Tong Wang*+ and Steven L. Scott
|
|
Companies:
|
University of Southern California and University of Southern California
|
|
Address:
|
1138 W. 29th Street, Apt. 12, Los Angeles, CA, 90007,
|
|
Keywords:
|
finite mixture model ; identifiability ; label switching
|
|
Abstract:
|
We examine a label switching phenomenon that can occur in the Bayesian analysis of finite mixture models. Label switching occurs because finite mixture models are only identified up to a permutation of the state labels. Thus the finite mixture of $k$ distributions contains $k!$ symmetric modes. We propose a method of removing label switching by running a Markov chain that is transient with respect to all but one of the modes. The practical implementation of our method is similar to a decision theoretic relabeling scheme of Stephens (2000). Our method offers computational savings over Stephens' method, is easier to apply, and avoids imposing artificial boundaries in the multivariate parameter space.
|
