[pymvpa] question about cross-subject analysis

J.A. Etzel jetzel at artsci.wustl.edu
Wed Jan 18 17:16:41 UTC 2012

My instinct in that situation is not to look for MVPA-specific solutions 
but rather general fMRI ones: how to best align your specific people to 
a common space. For example, if you are interested in a particular part 
of the brain or have a patient population, use methods tweaked for those 
situations; there's been quite a bit of work for techniques (though no 
perfect solution).

This assumes that you want to do a "standard" MVPA (e.g. linear svm) 
using voxels from many people (something like a leave-one-subject-out 
cross-validation); methods like Raj's similarity comparisons are quite 


On 1/18/2012 11:10 AM, Yaroslav Halchenko wrote:
>> This sometimes works surprisingly well and is fairly straightforward.
> yeap -- but I thought that the main question behind was on how could we
> account for areas where nice (linear) alignment is difficult so
> that union (pretty much your scenario) of voxels across subjects would
> not be ideal either.
> On Wed, 18 Jan 2012, J.A. Etzel wrote:
>> To run multiple-subjects tests I've usually converted everyone's
>> functional images to a standard space first (MNI or whatever), then
>> subsetted to only have voxels with non-zero variance in all
>> subjects.
>> This sometimes works surprisingly well and is fairly straightforward.
>> Jo

More information about the Pkg-ExpPsy-PyMVPA mailing list