 |
 |
Dear Alex (CC’d to the mailing list), Sorry for the delay in replying. Although very large free energies can be a sign of a problem with estimation, the absolute values aren’t meaningful. As the for the difference between models – the log bayes factor –it’s perfectly possible to be highly confident in one model rather than another. As only one connection has changed, this suggests that adding this connection has either greatly improved the accuracy of the model, or reduced the complexity by allowing various other connections to become closer to their prior values (zero). To dig deeper, I would first check each of the two models in your example pass certain sanity checks. Run spm_dcm_fmri_check(DCM) on each and ensure the explained variance is non-trivial (>10%). Next, have a look at the parameter estimates – is it clear what’s changed between each model? You can plot them by using: figure,spm_plot_ci(DCM.Ep,DCM.Cp) All the best, Peter. From: Alex [mailto:[log in to unmask]] Sent: 12 August 2015 11:09 To: Zeidman, Peter <[log in to unmask]> Subject: Fwd: DCM Bayes Factor comparison Dear Peter! First of all, thanks for your explanation last time and sorry for bothering you! I thought probably you could help with my confusion? Thanks a lot, Alex -------- Пересылаемое сообщение -------- От кого: Alex <[log in to unmask]<mailto:[log in to unmask]>> Кому: [log in to unmask]<mailto:[log in to unmask]> Дата: Четверг, 16 июля 2015, 23:12 +03:00 Тема: DCM Bayes Factor comparison Dear SPM team! After doing some DCM analysis I'm a bit confused by results: for each subject there are three competing models A,B,C and these models are alike (with or wothout ONE particular connection). What seems strange to me is that we choose a winning model if the difference in log(BF)>3, but for a particular subject the dispersion in model evidence could be as big as 200 or 300 (F(A)=1700 F(B)=1500 for example). How could it happen so, that we are 99.999999% confident in model A, though it differs from model B only by one connection? I studied the article by M. Garrido "Evoked brain responses are generated by feedback loops" - there is a graph (Fig 3A) where my situation is illustrated - differences in BF are always rather big and are practically always outside the gray zone (99% confidence that one model is more likely). Models in this article are also alike. Thanks, Alex ________________________________ Alex