Dealing with Missing fMRI Data in Whole-Brain Analysis

• Activity Number: 541
• Type: Invited
• Date/Time: Wednesday, August 1, 2012 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Mulugeta Gebregziabher*+ and Kenneth Vaden and Stefanie E Kuchinsky and Mark A. Eckert
• Companies: MUSC and MUSC and MUSC and MUSC
• Keywords: fMRI ; group analysis ; missing data ; multiple imputation ; spatial coverage ; neuroimaging methods

Abstract

Functional magnetic resonance imaging(fMRI) analyses rarely include the entire brain because of missing data due to image-acquisition space limitations and susceptibility-artifact. This missing data problem is typically addressed by excluding voxels from analysis. As a result, brain regions are excluded from analysis that may be of theoretical or clinical interest and increases risk for Type-II and Type-I error. In this study, we evaluate multiple imputation(MI) that exploits the spatial nature of fMRI data assuming missing at random. Comparisons are made with simple-neighbor-imputation(SNI), regression imputation that accounts for covariates and neighborhood dependencies(RMI) and available case analysis in a general-linear-model framework. In both real and simulated data analysis, we study the extent to which these methods quantitatively(effect-size) and qualitatively(spatial-coverage) increased sensitivity of group analyses. Variance and mean estimates from RMI show less bias and it does not appear to increase false positive errors. Compared to omitting of voxels, RMI shows increased brain-coverage, increased size and number of significant clusters.