A New Multiple Imputation Method for High-Dimensional Neuroimaging Data

Abstract

Missing data are a prevalent challenge in neuroimaging, with significant implications for downstream statistical analysis. Neglecting this issue can introduce bias and lead to erroneous inferential conclusions, making it crucial to employ appropriate statistical methods for handling missing data. Although the multiple imputation is a widely used technique, its application in neuroimaging is severely hindered by the high dimensionality of neuroimaging data, and the substantial computational demands. To tackle the critical computational challenges, we propose a novel approach, High dimensional Multiple Imputation (HIMA), based on Bayesian models specifically designed for large-scale neuroimaging datasets. HIMA introduces a new computational strategy to sample large covariance matrices based on a robustly estimated posterior mode, significantly improving both computational efficiency and numerical stability. To assess the effectiveness of HIMA, we conducted extensive simulation studies and real-data analysis from a Schizophrenia brain imaging dataset with around 1000 voxels. HIMA showcases a remarkable reduction of computational burden, for example, 1 hour by HIMA versus 800 hours by classic multiple imputation packages. HIMA also demonstrates improved precision and stability of imputed data.