Dear Georgia,
Some additional comments.
> I have also heard this approach called a small volume correction?
It's indeed a small volume correction, as it's still voxel-by-voxel. You just restrict the volume considered for analysis. A proper ROI analysis means you extract a measure like average beta estimates, first eigenvariate, thus somehow aggregating across voxels, resulting in a single value per ROI.
> I don't know which findings are real and which are not
The FWE controls the false positive rate (in contrast to e.g. the FDR). But yes, you'll never know (except if you maybe go with simulated data).
> the most appropriate method of accounting for the multiple comparisons problem is to use FWE
I wouldn't call it the most appropriate method, it is one method among others (e.g. an uncorrected voxel threshold combined with a cluster correction, be it FDR or FWE, is another way).
> In all cases where a significant peak (FWE-corr) is found the p value for the cluster (FWE-corr) is also significant
In case you go with an FWE correction on voxel level then *any* of the resulting voxels/clusters are significant. The cluster statistics are irrelevant in that case, as you've already accounted for multiple testing on voxel level. For practical reasons (might result in a lot of very small clusters/single voxels, which would all have to be interpreted or at least mentioned) a voxel-FWE is usually combined with an arbitrary extent threshold like k > 5, 10, 20. But this doesn't change the fact that these single voxels/small clusters are significant and valid results.
> As you soon as you add an extent to the voxel-wise correction, then you have arbitrarily altered the threshold
There is a conflict indeed, as we turn from voxels to clusters, although the correction on voxel level implies to stay on voxel level. But in the end any threshold is an arbitrary decision. With an arbitrary extent threshold one can ensure that single outlier voxels are suppressed. Based on an original resolution of around 3x3x3, usually resampled to 2x2x2 and smoothed with maybe 8x8x8, a single outlier in the raw data might transform into a small cluster, therefore e.g. k > 10 or 20. But this also means loss of regions with subthreshold activation in which only one or a few voxels reach significance (which one usually observes when lowering the voxel threshold, with some exceptions like few/no smoothing, just a few subjects, in these instances the T maps are often very "pixelated").
> Many studies use Monte Carlo simulations to determine the minimum 'k'
This is another option to account for multiple testing. You would do so in case of uncorrected voxel thresholds. But as stated, you've decided to go with a voxel-FWE to account for multiple testing, thus you wouldn't determine another statistical threshold to account for multiple testing another time.
> is it appropriate to adjust this value for the size of the ROI that you are running the between groups analysis on
AlphaSim requires the spatial smoothness of the residuals, which might well be different for a search volume like the thalamus vs. the whole brain. So I would define the thresholds based on the "regional" smoothness and the mask image in case the FWHM is available (I think 3dFWHMx from AFNI does do so). In case it is not, go with the spatial smoothness based on the whole-brain data and the mask image (SVC within SPM also relies on the whole-brain estimates for FWHM, resel size, resel count and does not reestimate the parameters from the small volume as far as I remember). Very important, to be on the safe side you should resample the mask to match the spatial resolution of your data. Otherwise the output from AlphaSim might reflect k required for a resolution of 2x2x2, while your data is actually 3x3x3 (not sure about WFU right now, but most of the atlases around have a better resolution than 3x3x3).
Hope this helps a little,
Helmut
|