Dear Guillaume,
Yes, I didn't expect this to be a new issue, but there's no particular reference to the implementation within SPM (as far as I could see, didn't check all the help sections in the m files though). Different calculations depending on spheres/boxes/images is mentioned somewhere as far as I remember, but some of the methods papers imply a precise calculation depending on the exact shape of the VOI (e.g. the Worsley et al., 1996, "A Unified Statistical Approach for Determining Significant Signals in Images of Cerebral Activation" paper, page 63, 68 considering different brain regions with different shapes). I would have assumed any of the implementations to work on non-NaN voxels only. Worsley (2003, "Developments in Random Field Theory") also goes:
The approximation (2) is accurate for search regions of any size or shape, even a single
point, but it is best for search regions that are not too concave. Sometimes it is better
to surround a highly convoluted search region, such as grey matter, by a convex hull with
slightly higher volume but less surface area, to get a lower and more accurate P-value.
But this does not stress it also takes into account NaN voxels. Well, it's no manual of course.
Concerning peak p and FWE-R and FWE-Bonferroni, this makes sense, didn't think about it.
Best
Helmut
|