Simulation Comparison of Generalized and Block Generalized Pseudo-Likelihood for Autologistic Models

• Activity Number: 78
• Type: Contributed
• Date/Time: Sunday, July 31, 2011 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Jordan Earl Purdy*+ and Jon Graham
• Companies: University of Montana and University of Montana
• Keywords: Autologistic ; Gibbs Sampler ; Markov Chain Monte Carlo ; Pseudolikelihood ; Generalized Pseudolikelihood ; Block Generalized Pseudolikelihood

Abstract

A regular lattice of spatially dependent binary observations is often modeled using the autologistic model. It is well known that likelihood-based inference methods cannot be employed in the usual way to estimate the parameters of the autologistic model due to the intractability of the normalizing constant for the corresponding joint likelihood. Two popular and vastly contrasting approaches to parameter estimation for the autologistic model are maximum pseudolikelihood and Markov Chain Monte Carlo Maximum Likelihood (MCMCML). Two newer and less understood approaches are maximum generalized pseudolikelihood and maximum block generalized pseudolikelihood. Both of these newer methods represent varying degrees of compromise between maximum pseudolikelihood and MCMCML. We will present simulation results comparing the four aforementioned estimation methods under varying lattice sizes, degrees of spatial correlation, neighborhood structures, and covariate dependencies.