[pymvpa] effect size (in lieu of zscore)

[Mike E. Klein]

Hi all,

I'm wondering if someone could point me in the direction of calculating the
effect sizes of voxels in time series against the series' baseline
conditions. Ideally over multiple experimental chunks/runs.

For reasons that I simply can't figure out, zscore-ing my data *always* brings
down my accuracies (whether I'm using a searchlight or ROI-based analysis),
even compared to the dataset that has only been detrended. So I'm trying to
explore alternative methods to feeding the SVM that quite raw data. (To
illustrate: I'm running an auditory experiment, where my baseline condition
is silence. Without zscoring, SVMs can tell any of the sound conditions vs.
the silence condition at 98-100% accuracy...which makes sense. With
zscoring, this drops to the 80-90% level. The experiment has a good amount
of samples, is well-balanced, motion-corrected, etc., so I can't find other
obvious confounds.)

PyMVPA people on this list have been helpful in suggesting that I shouldn't
normalize by volume, so this seems like a decent alternate approach, which
should keep the voxel values in a somewhat reasonable range. Any help would
be greatly valued. (I hope some day to be someone who answers questions on
this board, as well as asks them...)

Best,
Mike Klein
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20111216/9c738407/attachment.html>