I quite agree with Nick's "quite tricky": about the only way in which
averaging the weights over 18 the cross-validation folds will give you a
correct impression of the "important" voxels is if most of the voxels in
your ROI have no information at all, and the remaining are uniquely
informative (each distinguishes the classes, but not correlated with
each other). Needless to say, this scenario is not exactly common for
fMRI datasets. (and even more complicated if multiple people are being
analyzed.)
Searchlights can give a decent reflection of where *local* information
occurs, though there are many caveats (to cite myself, see
http://www.ncbi.nlm.nih.gov/pubmed/23558106).
I generally suggest tailoring the analysis to the hypothesis. If you're
really interested in the activity in individual voxels, some sort of
mass-univariate analysis is probably best. If you're interested in ROIs,
ROI-based MVPA can work very well. But trying to interpret *voxels* from
a *ROI-based* analysis is problematic at best.
Jo
On 1/21/2016 8:27 AM, Nick Oosterhof wrote:
On 21 Jan 2016, at 15:18, Maria Hakonen <[email protected]>
wrote:
I am working on my first fMRI data and would like to try MVPA
analysis. I have two classes that I have classified with linear
SVM. I would like to determine which voxels contribute most to the
clasifier’s successful discrimination of the classes. As far as
understand, the absolute value of the SVM weights directly reflect
the importance of a feature (voxel) in discriminating the two
classes.
Interpretation of SVM weights is quite tricky, see for example Haufe
et al 2015 Neuroimage, doi:10.1016/j.neuroimage.2013.10.067.
If you want to make inferences about the spatial location of
multivariate discrimination, you may want to consider using a
searchlight analysis instead.
I would like to average the SVM weights across all 18
cross-validation folds for each voxel and wrap the resulting map
into the standard space in order to display a map of the resulting
overlap.
Even if one would be confident that SVM weights were interpretable,
why take the absolute value? It would seem that this makes it much
more difficult to do any stats or interpret the results. In
particular, lack of signal but difference in variance of weights
across regions may then yield differences in average absolute
values.
>
_______________________________________________
Pkg-ExpPsy-PyMVPA mailing list
[email protected]
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
--
Joset A. Etzel, Ph.D.
Research Analyst
Cognitive Control & Psychopathology Lab
Washington University in St. Louis
http://mvpa.blogspot.com/
_______________________________________________
Pkg-ExpPsy-PyMVPA mailing list
[email protected]
http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
