Statistics Seminar
Date/Time: | Monday, 07 Oct 2013 from 4:10 pm to 5:00 pm |
---|---|
Location: | Snedecor 3105 |
Cost: | Free |
URL: | www.stat.iastate.edu |
Contact: | J LaGrange |
Phone: | 515-294-3440 |
Channel: | College of Liberal Arts and Sciences |
Categories: | Lectures |
Actions: | Download iCal/vCal | Email Reminder |
In Bayesian nonparametrics, unknown distribution functions are viewed as parameters, and priors are placed on them. Mixtures of Dirichlet processes are the most commonly used priors, and these are currently used to deal with many problems, for example mixed effects models, meta-analysis, and very recently, some problems in machine learning. Their use always involves choosing certain hyperparameters, including the so-called precision parameter. These hyperparameters have a big impact on subsequent inference, so it is important to choose them well. Almost always, these are selected in an ad-hoc way. In this talk, I will discuss a principled approach for selecting the hyperparameters. Implementation of the approach relies on a likelihood ratio formula for Dirichlet process models, which can be used to compute Bayes factors. Because we may view parametric models as limiting cases where the precision hyperparameter is infinity, the method also enables us to decide whether or not to use a semiparametric or an entirely parametric model. I will illustrate the methodology through two detailed examples involving meta-analysis.