Dear Jejo,

> Just a couple clarifications on your suggestions:
>
> 1. With PCA, though I haven't actually tried it yet, from what I
>understand two of the issues with PCA:
>if you do not control for physiologic noise, it's extremely difficult to
>get usable results.

That is true.  Of course,this begs the question what do you
want to 'use' the results for?.  Eigenimage analysis is just
a summary of principal modes of variation.  Your problem calls
for something slightly more focussed.

> PCA can't be used differentiate between activation and deactivation
>(without added effort) since the direction of the eigenvector is arbitrary

This may be a secondary problem.  The label 'activation' only has meaning
in relation to tasks employed.  You may be able to select an eigenvariate
that relects variance induced by task B, even if you cannot determine
the direction of the changes.

> a. Ideally we'd like to refrain from making a priori anatomical hypotheses
> and have the analysis be purely data-driven,
> b. The other issue I've run into using the time-course from the VOI as a
> covariate is that naturally the time-ourse correlates best with the region
> from which it was extracted.  So much so that activation in other regions .
> phase-locked to the original ROI are not extractable.  A couple solutions
> I've tried (please tell me if these sound reasonable) are:
>   1. denoising the time-course
>   2. somewhat 'arbitrarily' only using the 'activation' epochs from the
> time-course.  This is shady at best but I detrended the time-course and
> simply said any values < 0 correspond to 'deactivation' and hence set any
> value < 0 to 0.

First, you can slightly decorrelate the extracted time-course and the
voxel-specific repsonse by increasing the size of the VOI on extraction.

Second, you may want to think about using PPIs.  The PPI is uncorrelated
with the voxel-specific time course and often reflects the interesting
aspects of the study, namely the interaction between task A and task B.

> . In relation to point 2, I've tried using spm_peb_ppi
> (or whatever it's called) to try to deconvolve the HRF
> from the time-course from the VOI.  To not put too fine
> a point on it, the results are screwy.   The 'neuronal
> model' is corrupted by a very a low frequency cosine-type
> confound.  Basically with a period equal to the length of
> the session (400 timepoints).  I used spm_filter to remove
> this confound and essentially what I ended up with was very
> similar to the original time-course I received from VOI.
> Is it just that the time-course was too noisy or lacked
> sufficient signal for the deconvolution to be very useful.

Yes, I am sorry about that.  We have been working on spm_peb_ppi
to make it more robust.  The version in the full release (due
in a week or two) will not present these problems (it uses
a fully Bayeian deconvolution scheme as opposed to the emprical
one adopted in the beta release).

With very best wishes,

Karl