Dear Xingbao,
The BMS file which is written to disc contains all the results from the Bayesian model selection procedure. For the random effects analysis, have a look at the structure BMS.DCM.rfx. In this structure, you will find all results, e.g. “xp” is the exceedance probability for each model or “exp_r” is the expected posterior probability for each model. If you search the list for BMS, you will find earlier comments by Maria who detailed all fields and their meaning in this file.
Concerning your problem that SPM appears to freeze when you run the BMS procedure, is it possible that you are comparing a very large number of models or are using a very large number of subjects? Because a sampling procedure is used for computing exceedance probabilities in the case of more than two models, this may take a long time. You could speed this up by reducing the number of samples (currently we’re using a very generous number of one million samples). You would have to change line 30 in the function spm_BMS.
Best wishes,
Klaas
________________________________
Von: "Li MD, Xingbao" <[log in to unmask]>
An: Klaas Enno Stephan <[log in to unmask]>
Gesendet: Montag, den 22. Juni 2009, 15:02:01 Uhr
Betreff: RE: DCM model selection
Thank you so much, Klaas. I re-read your 2009 paper which
helps me lots in understanding BMS. I still have a basic questions on SPM 8.
Within Batch Editor, select Stats--Bayesian Model Selection--BMS:DCM, select
multiple subjects, each subject with 3 models, then select RFX for Inference
Method and default for Long-evidence. I successfully created BMS.mat and
had a graph for Model Comparsion. How can I get the BMS results (a
number in Matlab Command Window, not just a graph)? When I tried to
run Bayesian Model Selection---BMS:DCM(Result), the estimation procedure
took much long time for running (it may freeze in process). Each time I have to
click "Bye" to stop Results Window. No error message came out. Could you please
point out a possible incorrect?
All the best,
Xingbao
________________________________
From: Klaas Enno Stephan
[mailto:[log in to unmask]]
Sent: Thursday, June 18, 2009 3:32
PM
To: Li MD, Xingbao
Cc: [log in to unmask]
Subject: RE: DCM model
selection
Dear Xingbao,
I do not think that you did anything incorrect in your
analyses. Results from Bayesian model selection can differ between SPM5
and SPM8 because they use different approximations for the log evidence.
BMS in SPM5 is based on AIC/BIC whereas in SPM8, BMS uses the negative free
energy (F). The critical difference between these approximations is how
they define the complexity of a model. For example, AIC/BIC only take into
account the number of model parameters, regardless of the degree of
interdependence amongst these parameters. In contrast, F does consider
such interdependencies (and has some other properties), which, in most
instances, renders it a more appropriate measure of the complexity of a
model. For this reason, all model comparisons in SPM8, whether for DCM or
other applications, now use F. (For a full discussion of the relative
advantages of F and AIC/BIC see our recent random effects BMS
paper).
Because of these differences in defining complexity, BMS results
can differ when using SPM8 versus SPM5. This is not necessarily the case
(for my interhemispheric model, for example, the model rankings are identical
for F and AIC/BIC), but it is possible. As long as you clearly describe
which approximation you have used, you can use either F or AIC/BIC; the informed
reader will then know what assumptions your BMS procedure is based on. And
concerning your final question: yes, of course, you can still cite the older DCM
papers that used AIC/BIC.
All the best,
Klaas
________________________________
Von: "Li MD, Xingbao"
<[log in to unmask]>
An: "[log in to unmask]" <[log in to unmask]>
Gesendet: Dienstag, den 16. Juni 2009,
21:20:04 Uhr
Betreff: DCM model
selection
Dear Dr. Stephan,
You wrote an email on Feb 16,
2009 regarding the Bayesian model selection tool in SPM8. " All model
comparisons now use the negative free energy (F) as an approximation to the log
evidence; the use of AIC/BIC has been abandoned. The results of the model
comparison are described in a "BMS.mat" file that is written to disk." Do you
mean that SPM5 is not available to do BMS? Can I still cite these paper from
SPM5 (Stephan 2007, J of Neuroscience).
One data set with 3 models was run
with SPM 5 (compare Model) and SPM8 (Batch) respectively. The results puzzled
me. Could you please help me to understand the results? Why they are
different? If I am wrong, can you advise where I did incorrectly. (see attached
results)
Best regards,
Xingbao Li, MD
Assistant
Professor
Brain Stimulation Laboratory
Institute of Psychiatry
Medical
University of South Carolina
67 President Street, room 504 North
PO Box
250861
Charleston, SC 29425 USA
Phone: 1 (843) 792 5729
Fax: 1 (843)
792 5702
email: [log in to unmask]
www.musc.edu/fnrd/tms.htm
