Dear Jonathan,
I think I would use B.
C won't reduce the search volume (in a statistical sense) but just mask
in or out the voxels to display. That's why volume_C = volume_D, so we
can ignore that option from now on here.
A and B will give you the same search volume (being the intersection of
the brain mask (mask.img) and the mask you provided), however the
estimated smoothness will be performed using data from the whole brain
in B and just the reduced mask in A - hence the difference in FWHM. You
could also expect (probably very small) differences in the
hyperparameters for the non-sphericity (SPM.xVi.h) as different sets of
voxels might have been picked up to estimate the sample covariance.
I would however have expected same corrected p-values (both peak and
cluster) for C and D. Are you sure they were different?
Best regards,
Guillaume.
On 10/08/10 19:57, Jonathan Peelle wrote:
> I would like to constrain my 2nd-level MRI analysis using an explicit
> mask to restrict the search volume. In SPM8, there are at least three
> ways to do this:
>
> A. When estimating the model, specify an explicit mask;
>
> B. After running whole-brain analysis and results, do a small-volume
> correction and specify the mask image;
>
> C. When running results, during masking options specify an image to
> inclusively mask by.
>
> And of course, there is (D), the unmasked whole-brain analysis on the
> same images.
>
> When running these 4 analyses, approaches A-C give different results,
> although I would expect them to be the same (because it's the same
> search volume, and the same voxelwise uncorrected threshold I'm
> using).
>
> One difference is the smoothness (as reported at the bottom of the
> results table):
> B==C==D, A is different.
>
> The volume (i.e. just below the FWHM in the results table) also
> differs, with A==B and C==D.
>
> The cluster-level corrections (both FWE and FDR) also differ:
> A, B, C, D all differ (D I would of course expect to be different, but
> I also expected A==B==C)
>
> The FDR p value for the peak voxel changes as well:
> A, B, C, D all differ.
>
> (Needless to say, the t/Z stats for the peaks are identical across all
> of these approaches.)
>
> Any insight into why these 3 methods of constraining the search volume
> give differing results would be appreciated...as well as which method
> might be most sensible.
>
> Thanks in advance,
> Jonathan
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|