Susie,

Could somebody advise on the relative merits of using the existing SVC
tool, versus making your small search region explicit during the
statistical modelling part of SPM?

There are two differences between using a mask at the Estimation stage vs. using a mask at the Results stage with SVC. 

The first, and most important difference, is that the smoothness is estimated over all voxels in the analysis during estimation.  With a whole-brain mask, this is no problem.  But if you have a very small ROI used as an explicit analysis mask, say just a few 1000 voxels, can have a very poor/unstable estimate of the smoothness.  (You can view an image of the estimated *roughness* by looking at the RPV image; better is to view an image of average voxel FWHM=(RPV).^(1/3) computed ImCalc).  The topology of the ROI also matters; if you have a checker-board-like ROI, there are no contiguous voxels, and hence no way to estimate the smoothness. 

The second difference is a very minor one.  The SVC button uses an correction factor for the topology of the search region (see line 93 in spm_resels.m).  Very 'crinkly' search regions sample more 3D space than a compact spherical search region of the same volume, and so require a more severe correction.  For the standard results (no SVC), the search region is assumed to be spherical and no crinkly correction is applied.


I did a quick experiment with firstly a whole-brain threshold of 0.99,
and then applying a small volume correction, vs. specifying the small
volume as an explicit mask in my model.  The former method included
121424 voxels (93 resels), the latter 122673 voxels (56.6 resels).
Peaks were similar but not the same, and there were more peaks using
the SVC tool than by specifying the small volume explicitly.

You didn't list them, but I assume that the FWHM estimates were different, as mentioned above.

I would have expected that you have identical voxel counts; I can only assume that in the 2nd instance, ROI as explicit mask, you're including some voxels with values less than 0.99 (and hence 122673 > 121424).


Why does the SVC button only look at voxels which already
exceed an (arbitrary) threshold?  Is there a recommended whole-brain
threshold to apply before using the SVC button, and/or is it ok to set
this as near to 1 as possible?

There are two different issues wrapped up here:  One is how the analysis mask is determined... either with an analysis threshold (e.g. .99) or via an explicit mask constructed manually (somehow) (you can also do the both, too, you know); the other issue is whether to constrain inference to an ROI as part of the estimation or part of the Results via the SVC button.  As for the former question, I find that I trust my analysis mask most when I manually construct and review it, and then use it as an explicit mask.


As for Estimate vs Results/SVC for ROI's, I'd generally stick with the Results/SVC option, unless you've got a large, regular ROI and you worry that the smoothness elsewhere in the brain is very different from the ROI.


Hope this helps.

-Tom


PS: Whenever the next release of SPM occurs, the crinkly correction will be uniformly applied, both in estimation and in Results/SVC.

____________________________________________
Thomas Nichols, PhD
Director, Modelling & Genetics
GlaxoSmithKline Clinical Imaging Centre

Senior Research Fellow
Oxford University FMRIB Centre