So to clarify, the individual subject coefficient maps were transformed to 2mm^3 resolution when they were transformed to MNI atlas space with SPM's DARTEL. This is a decision that is bigger than me and I can't change it, not that I would if I could. So what I'm asking is not whether I should change the resolution; I'm asking how to properly run monte carlo simulation given that the resolution has been changed.
As for the assertion that it is a bad idea to oversample, my experience in this case is that it works out well. First of all, it has been known for a long time that you can reconstruct a higher-resolution image from multiple lower-resolution images.
[
link.springer.com]
[
citeseerx.ist.psu.edu]
The statistical models we use assume that the individual subject maps are noisy versions of the "true" underlying population map, which is the same assumption you need for this superresolution reconstruction. And in fact I find that with 157 subjects I get a much nicer activation map with oversampling, with larger clusters that are anatomically plausible. I'm attaching a comparison between activation maps from the same data in 157 subjects, where the one on the left was done with align_epi_anat and the one on the right was done with dartel and oversampling to 2mm.
[
brainimaging.waisman.wisc.edu]
In any case, I don't want to hash that out right now, but I *am* hoping for advice on how to apply monte carlo simulation now that my maps are oversampled to 2mm. :)
-David M. Perlman