TY - CHAP
TI - Model Choice and Variable Selection for Discrete Valued Data Using an Auxiliary
Mixture Sampler
AB - In this talk we will be concerned with model and variable selection for discrete-valued data, which are modelled based on generalized linear models, like Poisson regression models, state space models for count data, or binary random effect models. First we will discuss MCMC estimation for these types of models, which is based on an approximate, but very accuarte new Gibbs sampler, which introduces two sequences of artificial latent variables. This Gibbs sampler leads to a conditionally linear Gaussian model. Then we will show that this Gaussian structure is useful with respect to model selection. We will show how to obtain Bayes factors using the candiate's formula introduced by Chib (1995) and how to perform variable selection through Bayesian indicators. Any of these procudures is straightforward and does not involve Metropolis-Hasting components or other tuning constants.
AF - Workshop on Model Choice and Validation
PP - München
UR - http://www.stat.uni-muenchen.de/sfb386/workshop/mcv2005/
PY - 2005-01-01
AU - Frühwirth-Schnatter, Sylvia
ER -