[pymvpa] Cross-participant MVPA and controlling for traits of no interest

J.A. Etzel jetzel at artsci.wustl.edu
Tue Jul 12 21:36:13 UTC 2011

I agree that it could be added as a feature (if ordinal and properly 
scaled). But that would let the classifier use the behavioral trait as 
well as the voxel values, which I think is opposite of what you want: to 
do the analysis while minimizing the effect of the behavioral trait. 
Options 2 and 3 from Cameron strike me as promising: try to take the 
trait out of the features before classification.


On 7/12/2011 1:50 PM, Cameron Craddock wrote:
> Hello John,
> I have given this alot of thought in the past, and there are a few
> things that might be worth trying:
> 1. When you say that the behavioral trait is not the same "type" as the
> features, do you mean that it is not ordinal? If it is ordinal, then I
> don't see any problem with adding it as a feature. Indeed this is how a
> lot of data fusion algorithms work. You should be careful that the
> variance of the behavioral score matches the variance of the other
> features. I.E. you should z-score it as well as the features. This will
> ensure that the resulting feature weights are comparable.
> 2. You could regress out the behavior score from your feature space. Fit
> a glm to each voxel with the behavioral score as the regressor of
> interest, and then perform the MVPA analysis on the residuals.
> 3. You could perform a MVPA regression to the behavioral score, perform
> a feature selection to find the features most predictive of the
> behavioral score, and then remove those features from the for the A vs.
> B classification.
> 4. How predictive is the behavioral trait of the group membership? Can
> you just threshold the behavioral score to ascertain group membership?
> If so then you could perform a MVPA regression to the behavioral score,
> apply a threshold to the prediction output, and see if this performs
> better to classify group membership than a classifier trained using
> class membership as the labels. I think that this is a pretty
> interesting idea.
> Just a few thoughts.
> Cheers,
> Cameron
> Hi Jo -
> Thanks for your response. The features are structural data, so yes, they are
> voxel values.
> It just so happens that a behavioral trait is also predictive of whether or
> not participants are in Group A or Group B. Since the behavioral trait is
> not of the same "type" as the features, it seems incorrect to simply add it
> to the feature space. Still, though, I would like to "control" for the
> predictability of that trait in the MVPA.
> Does that make more sense?
> Cheers,
> John
> ---
> R. Cameron Craddock, PhD
> Postdoctoral Fellow
> Virginia Tech Carilion Research Institute
> Roanoke VA
> 404-625-4973
> cameron.craddock at vt.edu
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa

More information about the Pkg-ExpPsy-PyMVPA mailing list