Dear Siri,
my apologies for the delay of getting back to you, but life is
currently very, very busy here. Hope this is soon enough to be
still helpful. I have copied my reply to the SPM list so that others
can benefit from the exchange.
At 15:18 20/04/2006, you wrote:
>Hi, Klaas!
>
>Thanks for getting back to me about the convergence issue. I'm
>afraid I'm bothering you yet another time to get your thoughts on
>the following approach (hope you don't mind):
>
>I have a question about the usage of the Bayesian Model averaging utility
>for DCMs. I have two sets of indentically set-up DCMs that differ by one
>experimental factor, and I would like to contrast the two sets of DCMs at
>every connection.
I assume that by "connections" you are referring to the values of the
A matrix, correct? What exactly do you mean by "DCMs that differ by one
experimental factor"?
>For this purpose, would it be valid to multiply all of
>the parameters for one set of DCMs, and then submit all DCMs into a fixed
>effects average, to then see which connections are significantly greater
>than zero for the average, and thus which connections stand
>as "significantly different" between the two groups of DCMs?
I don't think this is a valid approach, I'm afraid. If the
difference in "experimental factor" that you refer to above means
that the DCMs differ by the presence of a particular input, the only
way to compare the two groups is by using a model selection
approach. One should never compare parameter estimates between
non-identical models: any change in the model structure will lead to
differences in the modelled dynamics and thus to parameter changes
everywhere in the model.
If the models *are* structurally identical (i.e. all regions,
connections, inputs are the same), you could compare each pair of
corresponding connections by entering the subject-specific values
into two-sample t-tests.
Hope all is well - very best wishes
Klaas
|