Permutation testing has been applied in the context of PET [Holmes 1996; Arndt & al. 1996; Heckel & al. 1998] and fMRI [Brammer & al. 1997; Locascio & al. 1997]; in general, though, it has not been a widely used technique in functional neuroimaging. A major reason for permutation testing's limited application thus far, it seems, is that this technique has not been integrated into a self-contained, widely distributed software package tailored for fMRI analysis. The present report, along with the software that it describes, aims to fill this need.
Our objectives are limited to the study of permutation testing, and limited to within-voxel analysis. We do not attempt to implement preprocessing filters (e.g. for the removal of autocorrelation [Locascio & al. 1997]), nor do we apply supra-voxel techniques such as cluster analysis. (As Locascio et al. [1997] observe, cluster analysis is less important in the context of a method such as permutation testing that already takes into account the spatial correlational structure of the data.) The software that we detail is coded in a modular manner, so that such adjuncts can be implemented as pre- and post-processing steps.
Our focus is on algorithmic optimisations and data structures that speed up the permutation test, making it feasible as an interactive procedure. We include sufficient detail to allow others to re-implement our optimisations as part of their own software systems, should they so choose.
Finally, we aim to compare quantitatively the results of the permutation test in this implementation with those of a Bonferroni-corrected parametric test, applied to data from a cognitive activation paradigm. We choose a higher-order cognitive task in order to supply an appropriate challenge. Although intellectual processes are often the focus of neurobehavioural studies, the activation in a cognitive paradigm is less robust than simple sensory or motor activations. It is for these research applications that sensitive methods such as the permutation test are truly needed.