Dear Christian For the RDFX or FFX model comparison, this may be a good explanation. And in fact, we see identical posterior evidence here. But what I'm most puzzled by is the fact, that for each individual model the evidence as stored in DCM.F is exactly identical even though the models are clearly different and the parameters look reasonable. Best Simon ________________________________ Von: Grefkes, Dr. Christian [[log in to unmask]] Gesendet: Donnerstag, 4. Februar 2010 14:29 An: Eickhoff, Simon; [log in to unmask] Betreff: Re: [SPM] Strange behaviour of model evidence Dear Simon! We encountered a similar problem some weeks ago when comparing 10 DCMs with different connectivity patterns. After extensive testing we found out that this strange result was caused by the inclusion of one particular model. When we removed this model from our sample of DCMs, we got reasonable results. Interstingly, this problem only occured for a random-effects approach but not for a fixed-effects analysis. But I still don't have an explanation for this... Best wishes, Christian -------------------------------------------------- Dr. med. Christian Grefkes Klinik und Poliklinik für Neurologie Universität zu Köln Kerpener Straße 62, 50924 Köln phone: +49-221-4726-310 fax: +49-221-478-7005 http://cms.uk-koeln.de/neurologie/content/mitarbeiter/index_ger.html Max Planck Institute for Neurological Research Research Group Leader Neuromodulation & Neurorehabilitation Gleueler Str. 50, 50931 Köln phone: +49-221-4726-310 fax: +49-221-4726-298 http://www.nf.mpg.de/index.php?id=255 "Eickhoff, Simon" <[log in to unmask]> hat am 4. Februar 2010 um 14:05 geschrieben: > Dear Klaas and DCM-experts > > > Recently, we encountered what we considered a rather strange behaviour of model evidence in DCM for fMRI (SPM8): > > We tried to explain an (positive) interaction observed in the GLM analysis of a 2x2 design in region Z by different models of feed-forward from the two regions [X, Y] showing the main effects for the two factors [x and y]. > > So we set up and estimated several model. > > In all models, the values obtained for the different parameters (driving input, intrinsic connections and modulations) were well plausible. > > > Strangely, however, the model evidence DCM.F as well as AIC/BIC is exactly (!) identical across models. > > Closer analysis showed, that in spite of the different connection patterns (e.g., different patterns of modulation such as X->Z modulated by y or Y->Z modulated by x) and the plausible parameters (in the range of 0.01 to 0.1), the residuals in Z were basically identical. > > > Any thoughts on this? Seems not logical to me! > > Thanks in advance > Simon > > ------------------------------------------------------------------------------------------------ > ------------------------------------------------------------------------------------------------ > Forschungszentrum Juelich GmbH > 52425 Juelich > Sitz der Gesellschaft: Juelich > Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498 > Vorsitzende des Aufsichtsrats: MinDir'in Baerbel Brumme-Bothe > Geschaeftsfuehrung: Prof. Dr. Achim Bachem (Vorsitzender), > Dr. Ulrich Krafft (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt, > Prof. Dr. Sebastian M. Schmidt > ------------------------------------------------------------------------------------------------ > ------------------------------------------------------------------------------------------------