Bayesian Statistics Seminar
North Carolina State University
presents
Dr. Sujit Ghosh
North Carolina State University
"Prior Distributions for the Variable Selection Problem"
ABSTRACT
Given a prior distribution on the parameters and models, the Bayesian approach provides a comprehensive solution to the variable selection problem in terms of the (marginal) posterior distribution of the models by integrating out the other parameters. However the issue of prior specification is non-trivial, especially when the number of variables is large and subjective considerations are almost formidable. The use of semi-automatic methods that avoids the explicit elicitation of prior distributions becomes almost necessary for a variable selection problem. For the canonical normal linear models Casella and Moreno (2006) obtained explicit expressions for the intrinsic prior for the variable selection problem and based the inference on model posterior probabilities and Bayes factors. In addition to the use of such default prior distributions I will discuss the construction of prior distributions that correspond to some of the popular penalized sum of squares (e.g., LASSO, ElasticNet etc.).
Slides of the talk
Tuesday, October, 3, 2006
4:00 - 5:00 pm
208 Patterson Hall