Dear Taylor
You have two models which differ only in the B-matrix parameters (more specifically, the priors on which parameters are non-zero). The only relevant question is whether the free energy is different between the models. Remember that the free energy is the accuracy minus the complexity, so you could have very similar predictions in DCM.y (accuracy) but a difference in complexity (parameter covariance).
If you are worried about local minima, you could specify a third model containing both modulatory inputs, estimate this full model, then test which of the nested models is better using Bayesian Model Reduction (it's like doing an f-test). This is less likely to lead to local minima, because you're only fitting one model. To do this, click Search in the DCM menu, then select DCM1 and DCM2 for a single subject.
Best
Peter
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Taylor W. Schmitz
Sent: 20 April 2016 14:17
To: [log in to unmask]
Subject: [SPM] DCM BMS model space
Dear List,
I have a question concerning DCM model specification and RFX Bayesian Model Selection.
My DCMs consist of two regions in the network (R1 and R2), and two conditions (A and B).
The question is whether it is reasonable to compare models with condition A *or* condition B modulating coupling between R1 and R2. Consider two models DCM1 and DCM2 from the model space that differ only in terms of whether condition A or B modulates coupling (B matrix). Both A and B are driving inputs (C):
drange([DCM1.DCM.xY.y]-[DCM2.DCM.xY.y]) % Data same
drange([DCM1.DCM.U.u]-[DCM2.DCM.U.u]) % Events same
DCM1.DCM.a - DCM2.DCM.a
DCM1.DCM.c - DCM2.DCM.c
DCM1.DCM.b - DCM2.DCM.b % B different (as expect)
drange(DCM1.DCM.y-DCM2.DCM.y) % Fit is (slightly) different
figure,hold on
plot(DCM1.DCM.y(:,1),'r')
%plot(DCM1.DCM.y(:,1)+DCM1.DCM.R(:,1),'r:')
plot(DCM2.DCM.y(:,1),'b')
%plot(DCM2.DCM.y(:,1)+DCM2.DCM.R(:,1),'b:')
corr(DCM1.DCM.y(:,1),DCM2.DCM.y(:,1))
drange([DCM1.DCM.y+DCM1.DCM.R] - [DCM2.DCM.y+DCM2.DCM.R]) % Reconstructed data NOT same (rounding error, or interaction between design and X0)?)
The fitted responses and F are different. And indeed, the BMS results show moderate evidence favouring models with condition B modulating coupling between R1 and R2. My question is whether DCM is capable of teasing apart which of only two conditions (with shared input) modulates connectivity, or whether this difference is more likely due to a local minimum?
Best regards
Taylor
|