Hierarchical Mixture Modeling for Multiple Testing and Effect Size Estimation in Voxel-Level Inference of Neuroimaging Data

• Activity Number: 173 - Recent Advances on Neuroimaging Analysis
• Type: Contributed
• Date/Time: Monday, July 30, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Ryo Emoto* and Atsushi Kawaguchi and Hisako Yoshida and Shigeyuki Matsui
• Companies: Nagoya University Graduate School of Medicine and Saga University and Saga University and Nagoya University
• Keywords: Effect size; False discovery rate; Hierarchical mixture models; Local significance index; Multiple testing

Abstract

In association analysis of neuroimaging data with disease status, it is important to consider the spatial structure in neuroimaging data. Conventional multiple testing methods for voxel-level inference, however, often ignore the spatial dependency and thus can induce substantial loss of efficiency. In this paper, we consider a model-based framework for identifying disease-related voxels, while allowing for estimation of voxel-specific effect sizes. We employ hierarchical mixture models with a hidden Markov random field structure to incorporate the spatial dependency among voxels. A non-parametric effect size distribution is assumed to flexibly estimate voxel-specific effect sizes. Simulation studies demonstrated less estimation bias compared to when parametric, finite mixture normal distributions are specified for the effect size distribution. An application to neuroimaging data from an Alzheimer's disease study is provided.