[pymvpa] GLM analysis questions

Nick Oosterhof nikolaas.oosterhof at unitn.it
Wed Jan 9 08:35:02 UTC 2013

On 9 January 2013 02:07, Abduljalil Sireis <kentman234 at gmail.com> wrote:
> Ok, so here you have the description. I hope it is not a problem to be so
> long! :)
> [...]
> Since the block durations are not multiples of the TR, it is needed to
> convolve the hemodynamic response function with the cognitive predictors
> (one for condition 1, and one for condition 2) at millisecond resolution,
> and then to subsample the result to get the predicted response at the TR
> resolution (one measurement each 3 s).
> There is a very unclear thing to me relating to predictors. What should I do
> exactly in multiplying the HRF with the cognitive predictors? I mean, how
> the cognitive predictors can be created using the data i have? And also the
> HRF function. Then what are the steps needed to do the subsample?

You may want to consider using the 'waver' program that is part of
AFNI [1], which runs on unix-like operating systems. Documentation [2]
and examples [3] are available. This program does the convolution for
you at high temporal resolution. In other words, no need to re-invent
the wheel (write code from scratch).

If you are happy using AFNI you may also consider using the
3dDeconvolve or 3dREMLfit programs, which can do the convolution for
all your predictors in one go, runs the full GLM and gives you beta
estimates for each predictor. 3dREMLfit also considers temporal
autocorrelation and is therefore especially suitable for fMRI data.


[1] http://afni.nimh.nih.gov/afni/
[2] http://afni.nimh.nih.gov/pub/dist/doc/program_help/waver.html
[3] http://afni.nimh.nih.gov/pub/dist/HOWTO//howto/ht02_DDmb/html/AFNI_howto.shtml

More information about the Pkg-ExpPsy-PyMVPA mailing list