[pymvpa] data scaling and accounting for nuisance factors
David V. Smith
david.v.smith at duke.edu
Wed Sep 14 17:42:11 UTC 2011
I've poked around a little more, and I'm wondering if L2 normalization would ameliorate my second issue regarding lesion size being a confounding factor in my classification? http://www.pymvpa.org/generated/mvpa.misc.transformers.l2_normed.html
At the end of the day, all I mainly want to ensure that (a) 0/1 scaling is correct and (b) determine how much of the CV is due to lesion size alone. If it were a simpler univariate regression, I could include lesion size as a covariate, but I'm not sure how to do something analogous in PyMVPA.
On Sep 6, 2011, at 12:52 AM, David V. Smith wrote:
> I have lesion data, and I am trying to test whether particular patterns of lesions distinguish two classes of patients. I have two questions:
> 1) What is the best way to scale the lesion data? Traditionally, these data are represented with 1s (lesion) and 0s (no lesion). I've played around with different scalings, and I've gotten different (but replicable) results using the SMLR classifier in PyMVPA 0.4. See below: first column is the leave-one-out CV; second column the value for the spared voxels; third column is the value for the damaged voxels.
> CV NoLesion Lesion
> 83.571 000 001
> 75.000 001 002
> 77.143 002 004
> 81.429 100 200
> 81.429 200 400
> 2.) What is the best way to control for a nuisance factor? I know there is an additional variable (i.e., lesion volume) that can distinguish between my two patient groups, so I would like the resulting CV and heavily weighted voxels to be uncontaminated by this nuisance factor. Ideally, I would like to know how much additional predictive power is gained over and above this nuisance factor.
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Pkg-ExpPsy-PyMVPA