Empirical Null Distribution Modeling for Signal Detection in fMRI

• Activity Number: 518
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Shuzhen Li*+ and Lynn Eberly
• Companies: University of Minnesota and University of Minnesota
• Keywords: fMRI ; FDR ; Empirical null distribution

Abstract

In signal detection of fMRI data, the inherent neural and spatial correlation among voxels could change the theoretical distribution of test statistics across voxels and thus compromise the accuracy and efficiency of region of interest (ROI) detection. In this article, a statistical parametric map (SPM)with moderated t statistics (Smyth, 2004) is created by considering a stabilized variance estimate instead of the usual sample variance estimate for typical fMRI data with small sample size compared to the large number of voxels. The empirical null distribution of the moderated t statistics, an assumed t distribution with scale parameter s and non-central parameter ?, is estimated through a maximum likelihood approach based on truncated data. A Poisson regression model is also implemented to estimate the nonparametric marginal distribution of all voxels. False discovery rate (FDR) thresholding can then be applied. We consider methods of Benjamini and Hochberg (1995) and computed local FDR (fdr, Efron 2007b) based on both regular and moderated t statistics. The criteria of true FDR and sensitivity are compared to demonstrate the performance of our methods in large simulations.