CMStatistics 2016: Start Registration
View Submission - CMStatistics
B1041
Title: Measurement error correction through shrinkage estimation in multilevel imaging analysis Authors:  Haochang Shou - University of Pennsylvania (United States) [presenting]
Ani Eloyan - Brown University (United States)
Amanda Mejia - Johns Hopkins University (United States)
Mary Beth Nebel - Kennedy Krieger Institute (United States)
James Pekar - Johns Hopkins University (United States)
Stewart Mostofsky - Kennedy Krieger Institute (United States)
Brian Caffo - Johns Hopkins University (United States)
Martin Lindquist - Johns Hopkins University (United States)
Ciprian Crainiceanu - Johns Hopkins University (United States)
 Abstract: Imaging data as high-dimensional and complex measurements are known to be observed with errors that come from multiple sources. Such errors include both random noises and systematic errors that are spatially correlated. With relatively small sample size in most of the study and limited availability of replicates for each participant, statistical inference made based on imaging data with errors might induce bias. We extend the shrinkage estimation idea in imaging data that was previously proposed to scalar-on-function regression setting and generalize the regression calibration in scalar case to functional regression. By shrinking the individual image towards population average image on the levels of individual voxels, local neighbors and the whole brain, we are able to calibrate the spatially dependent regression coefficient via the estimated attenuation ratio. Simulation studies show that the proposed approaches are able to reduce the data noise via borrowing information from the population image, and preserve the subject-specific image features. We have evaluated our methods on seed-based connectivity maps that are calculated using resting-state functional MRI from 21 healthy volunteers (publicly known as `Kirby21' dataset). Our results have shown that we achieve substantial improvements in mean square errors for prediction, as compared to using one replicate only.