In conclusion, caching of substitute values renders the step-down permutation test computationally feasible.

The permutation test activates more voxels, at higher levels, with greater anatomical clustering than standard parametric methods.

Permutation testing may be especially useful in studies of cognitive activations with low signal-to-noise.

Our immediate focus in future work is going to be implementing methods of removing autocorrelation. Shuffling `whitens' the noise in the time series, and while we can ignore this confound at 1.5 Tesla and a two- or three-second TR, it becomes a consideration when we go to high fields or low TR's.

Thank you.