Dear Sam
> While playing around with DCM/PEB, I discovered a few things that I'm not quite sure how to interpret. For context: I ran `spm_dcm_fit` to fit DCMs for different subjects, then ran `spm_dcm_peb` with different covariates and finally used `BMA=spm_dcm_peb_bmc` to 'prune' parameters.
OK
> Example 1) [Using fields={'B'} for the PEB]
> BMA.Pp of B(1,2,3) for the commonalities is (very very close to) 0, while BMA.Pp of B(1,2,3) of a covariate is >0.95. Would this mean that across subjects, there was no evidence for modulation from input 3 of that connection, but the model that best explains individual differences does include such a modulation?
As you've mean-centred your covariates, the first regressor (the commonalities) is the mean connectivity across subjects. Then it's perfectly possible that you'll have no modulation on average across subjects, but you'll have a positive or negative difference in modulation across subjects associated with your covariates. E.g. perhaps you had positive modulation in group 1 and negative modulation in group 2, if you had two groups of subjects indicated by your covariates.
> Example 2) [Using fields={'A', 'B'}]
> BMA.Pp of a connection A(1,2) is pretty much 0 for both commonalities and all covariates, but BMA.Pp of the modulation of that connection (e.g. B(1,2,3)) is non-zero for one of the covariates.
This is a similar situation. If you mean-centred your DCM regressors (DCM.options.centre), then your average connectivity across experimental conditions (A) can be zero with non-zero modulations. Alternatively, like above, the A matrix is the baseline connectivity, akin to the intercept of a linear model.
Best
Peter
