A0331
Title: Robust FWER control in neuroimaging using random field theory
Authors: Fabian Telschow - Humboldt University zu Berlin (Germany) [presenting]
Samuel Davenport - University of California, San Diego (United States)
Thomas Nichols - University of Oxford (United Kingdom)
Armin Schwartzman - University of California, San Diego (United States)
Abstract: The Gaussian kinematic formula (GKF) is a computationally efficient tool for performing statistical inference on random fields over large and complex domains. It is also a well-established tool in the analysis of neuroimaging data. The validity of methods based on the GKF, however, has come under scrutiny following a prior seminal work, which utilized error models derived from resting state data collected in the 1000 functional connectomes project. They revealed that while voxelwise inference yields conservative control of the familywise error rate (FWER), cluster-size inference tends to inflate false positive rates. The purpose of this talk is to review the primary factors leading to these findings, notably the unrealistic assumptions regarding "sufficient" smoothness, stationarity, and Gaussianity. Subsequently, a novel method based on the GKF is introduced, which accurately controls the FWER in voxelwise inference. Furthermore, the outcomes of the validation efforts under realistic error models are presented. A big data Eklund style approach is employed, based on resting-state data of 7000 subjects from the UK BioBank.
