Statistical Image Analysis with Missing Data

• Activity Number: 159
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Jian Kang*+
• Companies: Emory University
• Keywords: Missing data problem ; Brain imaging ; Gaussian process ; Multiple imputation

Abstract

It is common that neuroimaging data such as positron emission tomography (PET) contain incomplete data in some voxels, either due to subject motion, variable patient positioning, or acquisition artifacts. However, most neuroimaging tools (such as SPM) do not analyze voxels with missing data. Thus, the information concerning certain regions of the brain is usually discarded. To reduce the information loss, in this talk, we mainly discuss two approaches to handling missing values in neuroimaging data. Assuming the data are completely missing at random, a simple dummy variable approach is proposed to perform complete-case analysis of the voxel-wise general linear regression. This method can be valid under some other missing data mechanisms of neuroimaging. Also, we propose a Gaussian process guided imputation method taking into account for the spatial correlation between voxels. Some fast computational algorithms are developed to mitigate the extensive computational burden. Both simulation studies and a real data analysis are conducted to evaluate and illustrate the proposed methods.