Dear Tammar
> Is there a reason to get the best reduced PEB model (A or C matrices), and then recreate and rerun a new DCM model based on these results?
> This could results in better free energy, though I'm not sure if it has an added value over our original reduced model.
Re-estimating the DCMs, with the starting values or priors for the connectivity parameters set to the group average (as estimated from PEB), can help with local optima problems. In general, if you have a non-linear model, then the estimation algorithm is not guaranteed to converge to the optimal solution. This means that individual subjects' estimates could get 'stuck' in sub-optimal solutions. This is less of a problem for fMRI DCMs, which are only weakly non-linear, but is a greater issue for EEG / MEG DCMs, which are highly non-linear. For this reason, I don't usually bother for fMRI.
If you wish to re-estimate DCMs from the group mean, this should be done using a PEB model that does not know about any covariates - or you could bias the results. The recipe is, 1) estimate the DCMs, 2) re-estimate the DCMs based on group-level priors from PEB, and 3) specify a PEB model (with covariates if you have any) and then do the Bayesian model reduction on the PEB model. The function for re-estimating is called spm_dcm_peb_fit and there's a batch option for it.
> 2. We explored effective connectivity in our data, using PEB of PEBs method.
> For each matrix (A,B,C) separately, we executed BMC to find the connections that best explain the data.
> As for C matrix, we enabled all VOIs to be pinged by the task, i.e. driving inputs were set to all VOIs.
> The percent of explained variance was mostly > 10% per each DCM and each VOI.
> Nevertheless, none of the driving input survived at the group level (peb of pebs), even before model reduction.
> The model seems ok, but we are trying to understand:
> a. Could it be possible? How is the model informed the data if not through the driving input?
> b. If possible, is it ok to leave the model that way? or should I change the driving input?
This means that the driving input parameters couldn't be explained by the PEB model - either the form of the model or the specification of the priors wasn't right. One possibility is that the parameters weren't normally distributed across subjects - e.g., half the subjects had positive driving inputs and half had negative driving inputs. This can happen when the baseline isn't well sampled or varies across subjects. You could investigate this further - e.g., by plotting the C-matrix parameters of all subjects, but I don't think this is important. The C-matrix parameters aren't very interesting - they depend on the scale of the data, so for example, subjects with brain regions closer to the head coil will have larger C-matrix parameters. I'd focus on the B-matrix parameters, and use the A-matrix parameters to contextualize them.
All the best
Peter
|