[pymvpa] Group analyses of 2-class classification based on individual V1

[Nick Oosterhof]

On 11 Dec 2014, at 17:51, David Soto <d.soto.b at gmail.com> wrote:

> My concern is that because each individual V1 is different to the other individuals (no one to one mapping), the inputs fed onto the group analyses will contain voxels with empty values for some subjects but not for others and this may perhaps reduce the sensitivity of my analyses. Any advise/commment would be useful? is it just fine to do it like that?

Indeed you may loose some sensitivity. 
The 3dtest++ program in the AFNI suite has a “-zskip” option which skips values of zero. It adjusts the degrees of freedom across voxels and then converts the t-value to a z-score, so this may be more optimal than using df=N-1 (N the number of subjects) across all voxels.

Or you could run a surface-based searchlight of course. It will most likely give better alignment, is arguably anatomically more appropriate, and reduces the issue of correction of multiple comparisons because it only considers the grey matter. See http://www.pymvpa.org/examples/searchlight_surf.html