[pymvpa] Suspicious results

[Yaroslav Halchenko]

On Mon, 28 Feb 2011, Francisco Pereira wrote:
> Now seriously, this was just a very simple sanity check. I've been
> surprised by how much examples from trials that should have been far
> enough apart could be correlated (thanks to Michael Hanke for helping
> me realize this), and since then I've been more careful. See
> randomizing the class labels as a very simple, cheap sanity check. If
> that is not the problem, then you have an intriguing situation that is
> bound to attract curiosity and effort!

yeap!

and since "block-averaging" I guess was done per each trial just to get
a single volume out of "???" volumes following the trial onset, it is
woirth also looking at the SequenceStats (which is a part of output of
dataset.summary() in pymvpa >=0.5), e.g.:

*In [8]: print SequenceStats(normalFeatureDataset().labels)
Original sequence had 100 entries from set [0, 1]
Counter-balance table for orders up to 2:
Labels/Order O1     |  O2     |
     0:      49  1  |  48  2  |
     1:       0 49  |   0 48  |
Correlations: min=-1 max=0.96 mean=-0.01 sum(abs)=49

So what would be that on your dataset?  It should reveal order
dependencies among the two (or were there originally more than 2?)
labels.


-- 
=------------------------------------------------------------------=
Keep in touch                                     www.onerussian.com
Yaroslav Halchenko                 www.ohloh.net/accounts/yarikoptic