 |
 |
Hi Tallie I hope you don't mind me CC'ing the SPM mailing list so others can benefit. > After I have created my design matrix with columns for the various group comparisons, and then run the PEB on the full DCM for all my subjects, I have been putting the result from the PEB through spm_dcm_peb_bmc. This gives me a set of P values that match the posteriors (i.e. Pp's for my Ep's). If you run spm_dcm_peb_bmc without specifying specific reduced models, then it performs a series of steps: 1. It will do an automatic search of potentially many hundreds of reduced models, with different combinations of parameters switched off. 2. It will then take the set of models from the final iteration of the search (the best 256 models overall) and perform Bayesian Model Averaging (BMA). This means averaging the parameters over models, weighted by the model probabilities, so that better models contribute more to the average than the worse models. 3. To help with determining which parameters are important, it does a separate model comparison for every parameter. So for parameter 1, it compares the evidence for those models (within the 256) where parameter 1 was switched on, vs the evidence for those where it was switched off. It then repeats for parameter 2, 3 etc. This is what's stored in BMA.Pp. If you use the review GUI - spm_dcm_peb_review(BMA,GCM) - this is what is used if you select thresholding by 'Free energy (with vs without)'. > When I met with Karl in UCL he suggested that this BMC step was not necessary, implying that the parameters are already tested for significance in the PEB. However, I can find no P values in the output of the PEB. But that would imply that ALL values (i.e. all Ep's) are significant, which doesn't seem right. Any parameters not contributing to the free energy should have been pruned during the automatic search. Any which remain are likely to be doing something useful. So Karl is correct in saying that further thresholding is not required. However, you may want to focus the discussion in your paper on just the most important parameters, e.g. Pp > 0.95, which is why the thresholding is useful. But it's important to bear in mind that all the parameters retained by the search are probably doing something useful. Best Peter