Dear Shoichi
> ================================
I found previous my description was wrong.
However, can I use the fully connected model (5 regions,
DCM.a = [1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1],
DCM.b = [1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1],
DCM.c = [1 1 1 1 1]'), in which all the regions have direct inputs ( DCM.c =
[1 1 1 1 1]').
> ================================
Generally it is best to drive the network through only 1 (or a small subset
of) regions. Providing input to every region turns the DCM into more of an
SPM-type model. I believe this was a comment made in the Friston et al., DCM
paper.
> Is this mathematically correct?
I'm not sure about mathematically per se, but it is likely that by driving
all regions you will end up with a model that is either difficult to
estimate or has all influences attributed to the driving connections.
> Instead, if I have a hypothesis, is it better to use another model such
as:
DCM.a = [1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1],
DCM.b = [1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1; 1 1 1 1 1],
DCM.c = [1 1 0 0 0]'?
Yes- as Klaas and others have commented it is always best to test specific
hypotheses with DCM. If you have hypotheses about the architecture then test
those hypotheses. If you are unsure then it is reasonable to choose a
driving region, and then use a fully connected model for the A and B
matrices.
Best regards,
Darren
|