View source for FBIRN:March2006AHMStatsNichols

Permutation tests: assumes exchangeability of the data. FMRI timeseries have temporal autocorrelation, which means permutation tests on a subject by subject basis not a good thing.

ASL as a perfusion response: That reduces temporal autocorrelation--differenced ASL data are white, which makes the permutation test easy on those data. His opinion: Don't difference the ASL data. Use standard BOLD fMRI modeling tool, code for tag and control frames in the model, and keep the power.

Tons of non-parametric tests out there--all control the false-positive rate but may not have the same sensitivity to the alternative hypothesis. His example in the slides shows how using the median rather than the t-test can be more sensitive in the presence of heterogeneity/outliers.

Examples: a cluster of voxels around t = 5.4 not at all unusual when searching 26,000 voxels. That alone is not a measure of robustness. There must be correction for family-wise error, either using FWER or FDR, Bonferroni or permutation tests etc.