Hi Ian,

An easy solution to your problem is DCM12 (see Karl’s email for more details). You can adjust the precision of the observation noise in the model structure DCM.M. For instance, in DCM12, the precision of the observation noise is set at 4 in DCM.M.hE; see spm_dcm_estimate. As Karl explained in his email, this would tell the inversion scheme that a reasonable signal to noise ratio is expected in the fMRI data.

Regarding the stochastic DCM, DCM.options.s is not enabled. The only way to reduce the smoothness of fluctuations form 1/2 to 1/4 is to edit spm_dcm_estimate and change the value in DEM.M(1).E.s.

I hope this helps,
Mohamed




On 22/05/2012 18:50, Ian Ballard wrote:
[log in to unmask]" type="cite"> Hello,
I have some questions relating to the thread below about the development of DCM over the past several years. I have also noticed with my data that often models that were inverted relatively well (at least some of the parameters were substantially different than their prior expectations) in DCM8 now flatline in DCM 10 (I am not sure if the 'DCM12' Dr. Friston refers to is just shorthand, but I have the most recent updates and my DCM still says it is version 10). 

I understand that DCM has changed as the priors and inversion method have evolved, but I wonder whether this corresponds to any principled reason why DCM should more often conclude that the data are purely observation noise. For example, has the variance on the priors changed in a way to make the "shrinkage" priors more conservative? As I have a strong prior belief that my data are not purely observation noise (after all, ROIs needed to pass some activation threshold to be included), is it appropriate for me to manually adjust the priors on the precision of the observation noise? If so, could you please offer some details on how to do this (I can't seem to find this in the spm_dcm_fmri_priors script)?

I am not sure if this next point is related, but I have been playing around with stochastic DCMs on the same data (this gives an inversion that is a little better. It's not a flatline but it has far fewer significant posterior estimates than with the DCM8 inversion). It gives me some concerning output that I wonder may indicate some problem with my data:
LAP: 1 (1)      F:0.0000e+00    dF:0.00e+00   (3.07e+02 sec)  
LAP: 1 (1)      F:0.0000e+00    dF:0.00e+00   (3.05e+02 sec)  
LAP: 1 (1)      F:0.0000e+00    dF:0.00e+00   (3.09e+02 sec)  
LAP: 1 (1)      F:0.0000e+00    dF:0.00e+00   (3.10e+02 sec)  
LAP: 1 (1)      F:0.0000e+00    dF:0.00e+00   (3.06e+02 sec)  
LAP: 1 (1)      F:0.0000e+00    dF:0.00e+00   (3.09e+02 sec)  
This inversion may have been unstable;
try reducing DCM.options.s to 1/4

I don't see DCM.options.s anywhere. Is this shorthand for DCM.options.stochastic?

Thanks for any help or input!
-Ian