[pymvpa] Accounting for the haemodynamic lag

[Shane Hoversten]

On Jan 20, 2014, at 10:55 AM, David Watson wrote:

<<<<
As it happens I have a block design so perhaps I could get away with just
offsetting the timeseries, although convolving an HRF might still be
preferable. But if I ever wanted to do an event-related design where
measuring timings precisely is more important then I'm not sure if simply
offsetting the timeseries would still be considered acceptable. Also, are
there any other commonly used methods of accounting for the lag that I have
missed?
>>>>

Your block design will allow you a lot of flexibility with your analysis;
I'm doing a rapid event-related design, and so using a TR lag proved
non-viable, since regardless of how you lag the volumes from a rapid
event-related experiment each volume is polluted with the hemo response
from a bunch of different event presentations.

Then I stumbled onto this paper [1] by Turner et al (2012).  The idea is
basically to do a univariate-style regression using an HRF convolved with
your event schedule, except each event gets its own regressor and the other
event types get their own nuisance regressor.  You might take this idea and
implement it a number of ways; I've done it a ridiculously computationally
intensive way, in which each event in the experiment gets its own design
matrix.  For instance, if the experiment consisted of 10 presentations each
of stimuli from four stimulus classes (say, distinguishing between pix of
faces, houses, chickens, and pizzas) then there would be 10*4 design
matrices, each encompassing the totality of stimulus presentations.  One of
the design matrices for a "face" event would have the following five
regressors (not counting motion / drift / etc.):

current_face, other_faces, houses, chickens, pizzas

the idea is to soak up the variance in the HR owing to the event
presentations at each of the trials, so that the variance left over is
parcelled into the 'current' event regressor.  I use a volume of these
regression coefficients as input to the classifier.

That's a super terse and crappy description, sorry.  The paper has better
details, although the paper itself is not as comprehensible as might be
wished for.  The idea in the paper is based on this [2] somewhat more
comprehensible Mumford et al (2012) paper, which is a similar idea, except
it jams _all_ the other nuisance events into a single regressor; so you'd
only have two regressors in each design matrix:

current_face, other_stimulus_presentations

AFNI provides some support for doing a variation of this techniqe (called
"least squares separate") but it does not imo do it the right way, which is
to build a whole design matrix for each event presentation, as described
above.  The bad thing is that it takes _forever_ to run; luckily, I have
access to a beast of a 16-core Xeon server I can destroy with this
analysis, or it wouldn't be feasible to use.

Shane


[1] Turner, B. O., Mumford, J. A., Poldrack, R. A., & Ashby, F. G. (2012).
Spatiotemporal activity estimation for multivoxel pattern analysis with
rapid event-related designs. NeuroImage, 62(3), 1429–1438.
doi:10.1016/j.neuroimage.2012.05.057

[2] Mumford, J. A., Turner, B. O., Ashby, F. G., & Poldrack, R. A. (2012).
Deconvolving BOLD activation in event-related designs for multivoxel
pattern classification analyses. NeuroImage, 59(3), 2636–2643.
doi:10.1016/j.neuroimage.2011.08.076
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20140122/4ce430b0/attachment.html>