 |
 |
Hi Matteo, Happy to look at this but might not get a chance until next week - my guess is that you're getting results which surprise you due to your family definitions. If you switch off families, are the results more consistent with what you expected? Best, Peter. -----Original Message----- From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Matteo Tonietto Sent: 03 September 2014 13:26 To: [log in to unmask] Subject: [SPM] BMS with rfx: strange results Dear all, I am testing 27 DCM on 136 subjects but I am having very strange results when performing BMS with random effect. Looking at the free energy there are 2 models which are predominant (around 30% of times one is the best and 30% of times the other one is the best), but when calculating the BMS I got a third model with more than 50% probability in my population while the other two are around 5%. I tried to recalculate the BMS with 100000 samples using spm_BMS_gibbs, but the results didn’t change much, so I guessed is not a problem of model inversion. I am not interest in defining the best model, as I am using the results of BMS to perform BMA (over a family of 8 models). However, with the results I am obtaining, my average model is too polarized by one single model which is almost never the best in my population. When then I compare the parameter estimates of this group of subjects with another group which underwent the same analysis, I detect significant differences in all the connections, but probably these are due to the different weights used in the model average and not to differences in the connection strengths. My only guess is that the Dirichlet distribution may be inadequate for describing the posterior distribution of the model frequencies in my particular case but I am open to other suggestions. Thank you very much for your help Matteo