Dear Klaas, dear Volkmar,
thanks a lot for your answers. I've tried around a bit with scaling the VOI data; however, my confusion has got worse.
With unscaled VOI (i.e. with peaks at the very beginning and end, and with eigenvalues around 1xx), DCM takes about 40 steps and estimates coupling parameters ranging up to 1.0. As a first step, I tried several models in the model space, and DCM came up with a most probable model that is anatomically plausible, too.
When I tried this model with EoI-scaled VOI (i.e. no peaks, eigenvalues around median 0), only seven steps are conducted, only one connection is found, with a value of .4. In tests with synthetic data It showed to me that scaling didn't make any difference regarding DCM results, unscaled data (mean around 1xx) and scaled data resulted in identical values.
The peaks are indeed a side effect of whitening. In SPM.xX.W, only the main and first secondary diagonal are filled, and thus with AR(1), for the first and last line of the matrix one summand is missing. Maybe the 'strong' connections in the unscaled model are just an artifact of DCM estimating a relationship between the peak values.
So I wonder: Is there any best practice when picking VOI, and beforehand specifying a model? I've read several papers so far, and many things (theory behind DCM, model space, interpretation) have gotten much clearer. Setting of these technical parameters, however, remain a bit of a mystery, as these 'swtches' have a strong impact on the results.
I've read some posts by Darren Gitelman, in general he seems to recommend concatenation -- although it seems to be no necessity in case a session has 'enough trials', so it's hard for me to judge if concatenation would be worth the effort. I'm still not sure how to set up such an unified model, regarding the transistions between sessions: The temporal order not necessary for the GLM should be preserved for the DCM; but on the other hand I can't just assume ceteris paribus and treat the data as if it was acquired in one run.
Any further hints would be appreciated -- I think (hope) a successful and valid DCM analysis comes into reach, and the list of obstacles grows shorter.
Thanks for your assistence, best regards, Marius
|