Dear Shoichi,
in reply to your questions:
1) The Bayesian model comparison tells you that DCM_control is a lot
better than DCM_test. If, as you say, that is a consistent finding
across subjects, then there is no reason to stick to DCM_test.
2) The overall Bayes factor is mentioned at the bottom of the MATLAB
output, in your example 414.24. See this paper for detailed
explanations on model accuracy and model complexity:
Penny WD, Stephan KE, Mechelli A & Friston KJ (2004)
Comparing dynamic causal models.
NeuroImage 22: 1157-1172.
3) Given a flat prior p(m) on the models, the posterior probability
of the model is identical to the model evidence, i.e. p(m|y) =
p(y|m). You can derive this mathematically from Bayes theorem.
Best wishes,
Klaas
At 06:52 08/12/2006, you wrote:
>Dear DCM experts,
>
>I analyzed fMRI data with DCM, but I have trouble with the interpretation
>of the results. Now I want to check whether my model (DCM_test) is
>reasonable or not. Based on the previous message, I compared the model to
>the other model (DCM_control) which have no intrinsic connections (i.e.,
>DCM.a = [0 0 0;0 0 0;0 0 0], DCM.b = [0 0 0;0 0 0;0 0 0], DCM.c = [1 1
>1]'). We got following result with the subject 1:
> -------------------------------------------------
> Model 1: D:\DATA\sub01\DCM_control.mat
> versus
> Model 2: D:\DATA\sub01\DCM_test.mat
>
> All costs are in units of binary bits
> Region ROI_1: relative cost = -0.00, BF= 1.00
> Region ROI_2: relative cost = -0.11, BF= 1.08
> Region ROI_3: relative cost = 0.07, BF= 0.95
> AIC Penalty = -8.66, BF = 403.43
> BIC Penalty = -24.60, BF = 25412184.00
> AIC Overall = -8.69, BF = 414.24
> BIC Overall = -24.64, BF = 26093096.24
>
> Consistent evidence in favour of model 1
> Bayes factor >= 414.24
> -------------------------------------------------
>The parameter of the DCM_control and DCM_test are as follows:
>DCM_control
>DCM.A = [-1 0 0; 0 -1 0; 0 0 -1];
>DCM.B = [0 0 0; 0 0 0; 0 0 0];
>DCM.C = [0.0167, 0.0060, -0.0018]
>
>DCM_test
>DCM.A = [-1.0000 -0.0000 0.0000; 0.0098 -1.0000 0.0000; -0.0508 -
>0.0000 -1.0000]
>DCM.B = 1.0e-005 *[0 0 -0.2646; 0 0 -0.5407; 0 0 0]
>DCM.C = [0.0171 0 0]
> -------------------------------------------------
>
>(1) Should I give up DCM_test model? I calculated the data of ten
>subjects, and found that all the results were similar. Should I not compare
>DCM_test model to DCM_control which is the special case of the DCM?
>
>(2) In recent DCM papers, they compared two models and they reported their
>Bayes factors (for example, Smith et al., Neuron 49, 631-638, 2006, page
>637, Figure 3 legend; Each subject's BF was more than 7). In the upper
>case, there are a lot of BFs. Which is my Bayes factor.
>
>(3)Could you tell me how to extract the posterior pobability value.
>
>I am sorry for my basic questions, but I hope someone helps me.
>
>
>Sincerely,
>
>
>Shoichi Ugai
>[log in to unmask]
>Department of system engineering
>Tokyo Metropolitan University
>Tokyo, JAPAN
|