Wither Random Field Theory?
A fundamental goal in "brain mapping" with functional Magnetic Resonance Imaging (fMRI) is localising the parts of the brain activated by a task. The standard tool for making this inference has been Random Field Theory (RFT), a collection of results for Gaussian Processes of the null statistic image (implemented in the two most widely used packages, SPM & FSL). RFT provides inference on individual voxels (voxel-wise) and sets of contiguous suprathreshold voxels (cluster-wise) while controlling the familywise error rate, the chance of one or more false positives over the brain. I will discuss how RFT methods have been used for the past 25 years, show some small-scale evaluations that pointed to problems with RFT when the degrees-of-freedom are low. I will then show results from a recent study based on the wealth of (1000's of) publicly available resting-state fMRI datasets; these massive evaluations show that, even with n=20 or 40 subjects, RFT suffers from slightly conservative voxel-wise inferences and sometimes catastrophically liberal cluster-wise inferences. I will discuss the reasons for these failures of RFT and practical solutions going forward.
Joint work with Anders Eklund & Hans Knutsson.