Using Gibbs Sampling for Bayesian Inference in Multidimensional Contingency Tables
Abstract
This paper discusses a method suggested by Epstein and Fienberg (1991) for the Bayesian analysis of multidimensional contingency tables in connection with the Gibbs sampler to calculate posterior densities. The method consists of a two-stage hierarchical prior. The first stage is a Dirichlet distribution with a loglinear reparametrization for its means. The second stage is a multivariate normal distribution on the loglinear parameters. However, other distributions can be used if the Dirichlet-normal combination is not flexible enough to accommodate one's prior beliefs. These prior distributions are useful when one believes, with uncertainty, in a given loglinear structure for the cell probabilities. Contingency tables; Bayesian estimation; Dirichlet prior distribution; Gibbs sampler; Loglinear model; Maximum likelihood estimation of Dirichlet distributions.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jan 01, 1992
- Accession Number
- ADP007137
Entities
People
- Leonardo D. Epstein
- Stephen E. Fienberg
Organizations
- Johns Hopkins University