MCube78 Wrote:
-------------------------------------------------------
> I'm trying to use 3dClustSim to perform a cluster
> correction on group data. However, rather than
> being derived from the standard 1st level GLM and
> then group analysis, my group data is a set of
> 0-centered classification accuracies from an
> SVM-MVPA analysis.
>
> As things stand my process does the following main
> steps:
> 1. Run a group test (e.g., 3dttest++ or a sign
> test with FSL's randomise)
> 2. Create, within my mask of interest, a mean
> across all subjects' data, and feed that to
> 3dFWHMx
> 3. run 3dClustSim with the 3dFWHMx-y-z values I
> got in #2
>
> --> My question is about #2. SInce the typical
> explanation for the null data that should be fed
> to 3dFWHMx is in the fMRI 4D context, is it
> correct feed to it the average across subjects
> (within the mask of interest) -- so in my case,
> the average classification accuracy across my
> subjects (within the mask of interest)? Is there
> any better alternative?
Your questions is similar to
an earlier post, although that post concerned data on the surface, not the volume. Yet the same principles apply; using group accuracy maps seems an invalid approach to estimate the smoothness of the noise.
As in the previous post, I would suggest to use either:
- PyMVPA's
afni_surface_alphasim.py - unlike what its name suggests, it should work for both surface and volume data. The approach to estimate smoothness is similar to what SPM does, but note that his approach has not been properly validated.
- CoSMoMVPA's
cosmo_montecarlo_cluster_stat, which supports both 'standard' fixed-uncorrected cluster thresholding or Threshold-Free Cluster Enhancement.
Alternatively, PyMVPA provides
group_clusterthr.py for fixed-uncorrected cluster thresholding; using this method will require probably require that your MVPA pipeline (searchlight and classification) runs in PyMVPA , in order to generate null maps.