Dear Shoichi,
in response to your questions:
(1) No. A value in DCM.pB reflects the
conditional probability that this parameter,
given the posterior mean and posterior variance
of its estimate, exceeds a chosen
threshold. Your result means that you have only
25% confidence that your parameter exceeds this threshold.
Furthermore, from your comments, I understand
that you are using SPM2 where the threshold was
set to be log(2)/4. This was a choice that was
motivated specifically for the "attention to
motion" model shown in Friston et al. 2003. We
subsequently changed it to be zero, and this is
what I would recommend for your case as well. It
means you are testing how confident you can be
that the modelled effect exists. Generally, I
would recommend to use SPM5 for DCM analyses.
(2) You are correct that the numbers in DCM.B are
arbitrary in the sense that they depend on the
scaling of the input functions. However, this
scaling is linear and, because the PDF is
Gaussian, it does not change any inference,
neither at the single-subject nor at the group
level. I would say that you should always report
the DCM.B values (together with a description of
your input functions). If it is a
subject-by-subject analysis, the DCM.pB values
are essential as well because this is what your
inference is about. If it is a group analyses,
your inference is across subjects and whether or
not DCM.pB is relevant depends on how this
inference is obtained. If you go for classical
stats (e.g. t-test applied to the DCM.B values),
the DCM.pB values are not relevant. If you go
for the Bayesian approach, you need the "average"
function in DCM. This will give you an "average"
DCM.B value to report along with the associated
DCM.pB for inference (see previous postings on
this and on the caveats concerning
interpretation). Generally, never ever do any
stats on the DCM.pB values themselves.
Best wishes
Klaas
At 02:05 17/01/2007, you wrote:
>Dear DCM experts,
>
>I studied DCM analysis, but I do not have confidence in my understanding.
>My comments (q1-q2) are correct?
>Please help me.
>
>I analyzed fMRI data with 5 ROIs and got the results:
>¥mean DCM.B (ROI1 to ROI2) = 0.0122
>¥mean DCM.pB(ROI1 to ROI2) = 0.2518
>¥DCM.T = 0.1733
>¥The mean DCM.B (ROI1 -> ROI2) was statistically significant using t-test
>(p<0.05).
>
>
>(q1) Because 0.2518 is larger than 0.17 (=log(2)/4), the modulation from
>ROI1 to ROI2 was significant with the probability of 95 %. Because DCM
>uses Bayesian inference in each subject analysis (fix effect model), I
>cannot say that the significant level is P < 0.05.
>Is my comment correct?
>
>(q2) Because the DCM.B values are arbitrary, I should report the DCM.pB
>values not DCM.B value in my paper.
>Is my comment correct?
>
>Thank you in advance,
>
>Shoichi Ugai
>[log in to unmask]
>Department of system engineering
>Tokyo Metropolitan University
>Tokyo, JAPAN
|