B0423
Title: Ironing out the wrinkles in the cluster failure
Authors: Anders Eklund - Linköping university (Sweden) [presenting]
Abstract: Functional magnetic resonance imaging (fMRI) results in nice brain activity maps, but many of the used statistical methods have not been validated using real data. We recently showed that parametric statistical methods for cluster inference in fMRI can lead to inflated false positive rates, due to the fact that two assumptions for Gaussian random field theory are violated in real data (a Gaussian spatial autocorrelation function, and a constant spatial smoothness in the brain). A permutation test was for two-sample t-tests also shown to result in nominal false positive rates, as the permutation test is not based on these two assumptions. However, for one-sample t-tests, the permutation approach (based on sign flipping) failed for some parameter combinations. We will therefore focus on how to improve the sign flipping test for one-sample t-tests.
