Dear Darren,
>It seems that spm_peb_ppi as implemented in spm2b is not quite right
>because it doesn't implement the decorrelating matrix.
That is correct (this is only an issue if you have specified
non-sphericity). However, I think there are convergence and
robustness issues for this over-determined problem that are not
related to decorrelation.
>spm_peb_ppi as implemented in spm_devel using empirical bayes seemed to
>assign most of the variance to the xu part of the design matrix (original
>design). This produces an odd looking result if we assume we're getting the
>neuronal activity (as you recall it looked like a boxcar). spm_peb_ppi in
>spm_devel using full bayes appeared to produce a time shifted version of
>the BOLD response, which appeared to be at least intuitively more in the
>right direction.
>So if I want to do PPI what should I use?
I would use the Full Bayes option in spm_devel. I have thought about it and
I think that should be the default. It retains the beauty of a Bayesian
Weiner filtering, as described in your paper, but will be robust and quick.
In fact, I will change spm_peb_ppi to always use Full bayes and see how
it works on Beta testing here.
All the very best - Karl
|