Similarities in resting state and feature-driven activity: Non-parametric evaluation of human fMRI

Introduction and Motivation: fMRI is a natural source of high-dimensional time series data. Recordings are typically acquired in several hour sessions, with processing and analysis done offline. Due to the short amount of recording time available from any session, there is an imbalance between the comparatively small number of time slices and the high dimensionality of the data at each slice. Additionally, low spatial variation in the activation across voxels may imply the need for a spatial regularizer. Classical statistical tests, such as the Kolmogorov-Smirnov test of independence work on univariate data samples [5]. Recent extensions to high-dimensional non-parametric testing typically do not assume spatial or temporal dependence [3]. While these statistical tests are nevertheless commonly applied to fMRI recordings, the underlying generating process clearly violates the assumptions inherent in the design of the statistical tests.