Dear Mike,
Your FWHM values sound reasonable. For fMRI I usually go with a smoothing kernel of ~ 8 mm and observed FWHM values somewhere above 10 mm, similar to your values. More complex contrasts (e.g. a three-way within-subject contrast) tend to have somewhat lower FWHM than simpler contrasts (just an observation).
> which is usually less than 100 voxels for an alpha level of 0.05
As already stated, this well depends on the voxel resolution/no. of voxels. If you don't resample to the default 2x2x2 but e.g. 3x3x3 there are less voxels of course. Often this is not reported though. In addition, if you only covered half of the brain there are also less voxels to account for.
However, there are also several sources of error. E.g. people might accidentally take the voxel resolution of the raw data instead of the resolution after preprocessing. When switching between programs people might mistake FWHM values in mm for FWHM values in voxel units or vice versa (as programs might have different default settings). They might just go with the FWHM from smoothing. There might also be some confusion on the meaning of "clustering neighbourhood"/"cluster connectivity criterion" as pointed out by Donald at https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;a4568e15.1601 . Plus over time people have relied on different files for obtaining the FWHM values, e.g. estimating from the group T map, estimating from the group ResMS image, from the individual ResMS images, possibly also the 4D residuals, then there's the issue with taking the square root or not (see https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;4403627.1310 ). Thus make sure you report everything properly, did you rely on e.g. AFNI or did you go with another implementation (e.g. REST offers some Alphasim correction as well), did you take the estimated FWHM from SPM results table or did you determine them with AFNI, ... .And of course, you should report the initial threshold as well. This is quite similar for topological FDR in which people often forget to report the initial thresholds.
Best
Helmut
|