[pymvpa] null classification performance in the presence of strong univariate signal??

[Nick Oosterhof]

On Sep 18, 2014, at 3:36 PM, David Soto <d.soto.b at gmail.com> wrote:

> I reiterate I did this separately for each of the 19 subjects.
> 
>  I then aimed to carry out a group analyses using the individual accuracy maps. to do this I merged the 19 nifti accuracy maps into a 4D file and run a one-sample t-test in FSL using randomise -i searchpred -o searchpredOneSampT -1 -v 5 -T. 
> 
> Weirdly the output gives a group map with all brain voxels over p<0.001 !? which cannot be right...

I am not familiar with FSL's randomise tool (and a quick google search only gave me limited information).

One possibility: did you test against the null hypothesis of a mean of zero? In your case chance level is .5 (=1/2), not 0. It could explain why you are getting all highly significant voxels.
If that is the case, you should subtract .5 from the classification accuracy maps (of individual subjects) before testing against a mean of zero.