[pymvpa] Questions about mvpa preprocessing...

Michael Hanke michael.hanke at gmail.com
Fri Apr 15 18:19:10 UTC 2011

On Fri, Apr 15, 2011 at 07:45:31PM +0200, Roberto Guidotti wrote:
> Thank you, for the quick response and for the advice.
> For the first question: in the review papers I've read this question is
> never treated in depth, but only if the study is across-subject, thank you
> for the paper you link to me. My study is preliminary, I want to check that
> MVPA technique tells me something else... My analysis is firstly
> within-subject and then I could "upgrade" to cross-subject!!
> The second question: I know it were incomprehensible ;) now I'll explain
> simply :D I have a task with two conditions plus resting, without resting
> state (fds = fds[fds.sa.targets != 'Rest']) my accuracy is good, with the
> full dataset I have bad results (no. of training examples for each class are
> balanced) I don't know if I the independence conditions, with this dataset,
> could be true. My question is the fact that my resting state is not
> classified as resting state, or my conditions are misclassified, is it an
> alarm bell?

That depends on how your design looks like. If "rest" is the time
between two stimulations (and hence adjacent in time to one or more
stimulation conditions) a classifier might get confused by the remaining
effects of a stimulation that "leak" into samples that you have labeled
as "rest". Make sure that any "rest" data is at least as far away from
the stimulation data than the stimulation conditions from each other.


Michael Hanke

More information about the Pkg-ExpPsy-PyMVPA mailing list