[pymvpa] Found the problem (hopefully): suspicious results

[Michael Hanke]

On Wed, Mar 02, 2011 at 12:33:17PM +0100, Nynke van der Laan wrote:
> Hello all,
> 
> again, many many thanks for the responses and suggestions!
> 
> I think/hope that I've found the problem with my data. I've looked
> somewhat further into the more direct results I retrieved from the
> searchlight analysis and I think that the conversion to a nifti-map is
> causing the >0.5 accuracy bias.
> When I plot a histogram of the direct results that I get from the
> searchlight in python, more specifically, the sl_map. Then the
> distribution looks ok. The peak is at 0.5 with a kind of normal
> distribution. However, what I have been looking at earlier was a
> histogram from the data mapped back to nifti ( with:
> dataset.map2Nifti(sl_map).save('filename.nii.gz'). When I open this
> nifti in MRIcron than the distribution suddenly is totally different
> (as I reported earlier: the histogram with a peak at 0.6).
> 
> I did an additional check on it: Aside from the sl_map (the errormap
> from the searchlight analysis), I also constructed an accuracymap
> (with the following code: sl_map_accuracy = [1-x for x in sl_map]).
> When both maps are converted to nifti, they both have the >0.5 bias.
> While with a correct conversion to nifti you would expect the errormap
> to be the inverse from the accuracymap. (which is the case when I look
> at the distributions in python, so the problem seems to be in the
> conversion)
> 
> Any suggestions what could be going wrong in the conversion to nifti?
> Any known problems? Because I would really like to use the data mapped
> back to the original dataspace in nifti.

That is really strange. Reverse-mapping shouldn't touch the data at all.
Would you mind sharing the code so we can investigate this issue?

Michael

-- 
Michael Hanke
http://mih.voxindeserto.de