Dear Klaas
Thanks for you quick response.
The problem is that we are getting different results using the two approaches:
At the individual level, using the Bayesian model comparison (in the GUI) we
get strong evidence for one of the models, while the DCM.F of these two
models is identical.
What could be the reason for this peculiarity?
Thanks again
Tali Bitan
On Fri, 28 Nov 2008 15:12:19 +0000, Klaas Enno Stephan
<[log in to unmask]> wrote:
>Dear Tali,
These two approaches are identical: multiplying the BFs gives you the same
model selection as adding the differences in log evidence. Mathematically, the
latter is just the log transform of the first:
log GBF = log [product [p(y_n|m1)/p(y_n|m2)]] = sum[log p(y_n|m1) - p
(y_n|m2)]
Here, n is an index subjects; product and sum is over n; m1 and m2 are the
two models compared.
See also Equations 19 and 20 in the Stephan et al. 2007 paper you cite.
Best wishes,
Klaas
________________________________
Von: Tali Bitan <[log in to unmask]>
An: [log in to unmask]
Gesendet: Freitag, den 28. November 2008, 14:24:03 Uhr
Betreff: [SPM] Different ways for group model comparison with DCM?
Dear DCM experts
I realize there are (at least) 2 ways for doing the DCM model comparison at
the group level, and I would like to ask what is the difference between them,
and in which conditions should each one be applied (for fMRI):
1- Do the Baysian model comparison at the individual level, multiply the BFs of
all individuals for each comparison to get GBF, and choose the model with the
largest GBF given that the PER is reasonalbe (??). (e.g. as in Stephan
Weiskopf et al. 2007).
2- Get the model evidence from DCM.F, and sum up this value across
individuals for each model. Choose the model with largest sum (suggested by
Stefan Kiebel in SNDS workshop (Dead Sea 2007), and based on Kiebel Garrido
et al. 2007).
I'd appreciate your advice
Tali Bitan
Haifa University, Israel
|