 |
 |
Dear all, We've previously used the SPM implementation of the variational free energy approximation to log model evidence for Bayesian model comparison in fMRI. We'd now like to utilize a similar approach for our EEG data, for which we have projected our sensor data onto a couple ECDs so as to obtain single-trial data on the source-level. More specifically, we intend to fit our single-trial regressors to single peri-stimulus time bin data. However, the VB / auto-regressive GLM approach (as per spm_vb_glmar) seems to utilize spatial regularization even without spatial priors. Specifically, in spm_vb_init_block the number of voxels/sensors seems to determine the initial approximate alpha and beta posteriors for all priors except the uninformative prior. In the subsequent updating procedure this leads to considerable differences int the results, which is problematic for EEG source analyses with no meaningful spatial arrangement. We'd be most thankful for any suggestions on how to implement the vb approach omitting spatial information in order to make it suitable for our (spatially-independent) data. Is it valid for us to use uninformative priors and preventing the spatial precision from being updated? (alpha/beta parameters in spm_vb_glmar) Alternatively, are there other options that'd allow us to use any priors besides an uninformative one (e.g. shrinkage prior)? Best wishes, Sam