https://www.na-mic.org/w/index.php?action=history&feed=atom&title=FBIRN%3AMarch2006AHMStatsNichols 2026-06-05T19:53:33Z Revision history for this page on the wiki MediaWiki 1.33.0 https://www.na-mic.org/w/index.php?title=FBIRN:March2006AHMStatsNichols&diff=4160&oldid=prev 2006-12-18T13:37:05Z

<p>Update from Wiki</p> <p><b>New page</b></p><div>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.<br /> <br /> 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.<br /> <br /> 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.<br /> <br /> 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.</div>

Andy