Something I've been wondering about for years: every explanation of AlphaSim or 3dClustSim talks about the tradeoff between p-value and cluster size in very vague terms. For example Gang Chen's web site says "Basically you need to compromise among overall significance level (alpha), minimum cluster size, and individual voxel significance level, and may have to vary different p value to find the compromise you are interested. Keep in mind about the compromise between p and minimum cluster size: smaller the p value, smaller the minimum cluster size when other parameters are fixed, but it is not true that smaller the p value the better. Any p from 0.05 and less is appropriate, and the appropriate the compromise depends on each circumstance."
I'm wondering if there is a detailed explanation somewhere of *how* to actually determine this compromise? I mean, in general there's the never-fails technique of "mess with the parameters until you get results you are happy with", but I'm wondering if there is any documentation of a more principled approach.
Another question: Gang's page says: "As data are acquired in original voxel size, do make sure that Monte Carlo simulations are done in original voxel size (but not original brain shape - you will see the difference later) instead of higher resolution such as tlrc space. This is to make sure cluster formation is properly simulated." But if I'm doing my statistics (e.g. t-tests) in group resolution, which is, say, 2mm cube, then isn't that the resolution that clusters will be forming in? So shouldn't I do the monte carlo simulation in the resolution that I will be doing the stats in?
Thanks very much!
-David Perlman
Edited 1 time(s). Last edit at 05/19/2013 06:08PM by dperlman.