The former corresponds to small volume correction on cluster level = adjusting cluster p values. This is valid, but it's going to rely on the global estimate for smoothness, which might be a good or bad estimate. In any case it's accepted to go this way.
The latter is no true small volume correction from a technical point of view, as SVC implies setting up a GLM for some (larger) volume and then restricting it to some smaller volume. If you start with a restricted brain volume right from the beginning there's nothing you correct for. But indeed, the hyperparameters, statistics and also the smoothness are based on that restricted brain volume then.
If you're really interested in just some particular brain region then the latter is valid, but one would usually also just acquire data for this brain region, as it allows to shorten the TR, increase the spatial resolution. E.g. if you want to conduct a retinotopy you might cover occipital and parietal lobe only, if you're interested in fusiform face area just the ventral occipito-temporal regions, if you're interested in amygdala possibly a ZOOM EPI for that region. Thus there would usually be no need to restrict the volume when setting up the models, as one would have done so well before during data acquisition.
Usually people go with the former, as at some point they look at the whole-brain results. Which also points to a major problem with practical application of SVCs. For a proper SVC, we should take into account *all* the a-priori regions, also those that have already reached significance on whole-brain level, not just those that did not.
Leaving this aside, at least in theory it should be possible to combine SVC on cluster level with non-stationary cluster extent correction, which should provide a good estimate for the smoothness within the volume of interest. Non-stat. correction takes into account the local smoothness, which results in longer computation times for statistics (when pressing the "Results" button). For smoother regions clusters have to be larger to reach significance then, for less smooth regions smaller clusters are already sufficiently large. As a consequence, you can no longer display significant data based on an extent threshold of k voxels, as 100 voxels might be sig. for one region, with a larger cluster failing sig. in another. You can enable non-stat. correction by setting defaults.stats.rft.nonstat = 0; in som_defaults.m to 1. It works well for whole-brain statistics, but I have no idea whether it works properly when combined with SVC.
Best
Helmut
|