[pymvpa] decoding in an event-related design

[Stan Bullock]

Hi list,

I have an experiment with 4 scanning runs, and within each run I have trials
where I present a stimulus (S1) for 1 TR (2s), then there is 4 TR fixation,
and then the 2nd stimulus (S2) for 1 TR, then a 4 TR fixation. So the order
is S1(2s),fix(8s),S2(2s),fix(8s), repeated until the end of the scanning run
(~10 min). Thus, the trials are separated by 4 TR / 8 s. I would like to
decode an aspect of S1 by averaging over activity in the 4 volumes that
follow S1, completely ignoring S2, so really the data that I am interested
in is separately by a 6 TR / 12 s interval.

The natural way to do this in PyMVPA seems to be that you treat each
scanning run as a "chunk," use ERNiftiDataset to set up a boxcar average
over the 4 timepoints of interest for each event, and then do something like
leave-one-out with the runs for cross-validation.

1.) Would it be possible / advisable for me to do a leave-one-out
cross-validated design, where I leave out single trials instead of entire
runs, or is this problematic? My worry is that my manipulation is rather
subtle, so if I have to leave out 1/4 of my data for training on each fold,
my classifier may not learn very well.

I understand that one concern is that the training data should be
independent from the transfer data, and I suppose that if I do a
leave-one-trial-out analysis, there will be a high temporal correlation
between training data in adjacent trials and the transfer trial. However,
class 1 is no more likely to follow another class 1 trial than it is to
follow a class 2 trial. Does this alleviate the worry? What if I left out
only the adjacent trials from the training set?

2.) If this is not especially problematic on a basic level, I would
appreciate advice on the optimal way to go about setting up my splitter in
PyMVPA. Should I treat each trial as its own chunk? That has the feeling of
being stupid or kludgy, but I can't see why it would matter on the face of
it.

Thanks in advance for any advice on one or both of these problems.

Best,
Stan
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20100227/39c1d012/attachment.htm>