[pymvpa] Justification for trial averaging?

[Shane Hoversten]

Thanks Brian and Francisco.

Francisco, you said:

<<<<

You' want to average if the analysis you want to do is to look at
similarity patterns, rather than train classifiers. It might have to
be combined with permutation tests (e.g. you average within things
labelled as being in the same class within each label permutation).

>>>>

would you mind elaborating on these points?  I think I know what you're
getting it with the classification vs. pattern similarity, but you lost me
on the permutation tests.

Here's a related question: it's my perception that most of the people who
use a convolution solution (as opposed to lagging the volumes under
consideration and using the activity directly) use the regression weights
for the column of interest at the event of interest for their
classifications.  What would this mean wrt averaging and pattern analysis?
 It seems weird to average the regressor weights, but maybe it shouldn't.
 Is that something that's done, or is the averaging process only used with
raw voxel activity?

Thanks,

Shane
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20140123/8bee0f8e/attachment.html>