Activity Number:
|
640
- The State-of-the-Art in Modeling and Testing of High-Dimensional Brain Images and Networks
|
Type:
|
Invited
|
Date/Time:
|
Thursday, August 3, 2017 : 10:30 AM to 12:20 PM
|
Sponsor:
|
Section on Statistics in Imaging
|
Abstract #322025
|
View Presentation
|
Title:
|
Powerful, Fast, and Robust Family-Wise Error Control for Neuroimaging
|
Author(s):
|
Simon Vandekar* and Theodore Satterthwaite and Adon Rosen and Rastko Ciric and David Roalf and Kosha Ruparel and Ruben C Gur and Raquel E Gur and Russell Shinohara
|
Companies:
|
University of Pennsylvania and Univ of Pennsylvania and University of Pennsylvania and University of Pennsylvania and University of Pennsylvania and University of Pennsylvania and University of Pennsylvania and University of Pennsylvania and University of Pennsylvania
|
Keywords:
|
Multiple testing ;
Parametric bootstrap ;
Family-wise error rate ;
Neuroimaging ;
High-dimensional testing
|
Abstract:
|
Recently, several neuroimaging studies have demonstrated that multiple testing procedures (MTPs) commonly used in neuroimaging yield incorrect false positive rates. Methods that rely on Gaussian random field theory have inflated family-wise error rates (FWERs) due to the reliance on assumptions that are invalid in neuroimaging. While classical MTPs guarantee a conservative FWER they are underpowered as they ignore the covariance structure of the test statistics. We introduce a parametric bootstrap joint (PBJ) testing procedure that leverages the covariance structure of the test statistics. We use simulations to compare the PBJ procedure to nonparametric bootstrap and permutation joint testing procedures. To generate realistic data in each simulation we draw a subsample of cerebral blood flow imaging (p=109,748) and region-wise (p=112) data from the Philadelphia Neurodevelopmental Cohort (n=1,601). The PBJ procedure maintains the FWER at the nominal level and has superior power to the nonparametric procedures. Remarkably, the PBJ procedure is 150 times faster than the nonparametric bootstrap reducing computing time from 10.5 minutes to approximately 4 seconds (for n=120, p=112).
|
Authors who are presenting talks have a * after their name.