Nonidentifiable Issues in a Bayesian Cox Cluster Process Model for Functional Neuroimaging Data Analysis

• Activity Number: 337
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Jian Kang*+ and Timothy D. Johnson
• Companies: University of Michigan and University of Michigan
• Keywords: Nonidentifiability ; Bayesian Cox cluster process ; Hierarchical model ; Functional neuroimaging ; Meta analysis

Abstract

A Cox cluster process is a spatial point process that models the clustering of points about a latent parent process. To account for more complicated point patterns, the parent points may also cluster about a grandparent process, resulting in a hierarchical Cox cluster process model. This model shares the same nonidentifiability issues as finite mixture and hierarchical models. First, the intensity function is invariant to component label permutations, rendering the locations of the latent points nonidentifiable at both levels. Second, the grandparent process is nonidentifiable in the sense that its posterior conditional distribution given the parent process does not depend on the data. In this talk, we discuss these issues in general and show how they can be addressed in an analysis of functional neuroimaging meta-analysis by leveraging anatomical/functional knowledge of the brain.