Print

Print


Dear Mina,

The Bayes factor you should take (in the procedure as it currently stands in SPM5) is the last line of the output in the MATLAB command window.  In your example below, this is BF = 0.01237.

In the current implementation BFs are computed on the basis of both the AIC and BIC approximations to the log evidence and and the more conservative of the resulting BFs is provided as result of the model comparison.  See Penny et al. 2004, this paper describes all issues of model comparison (as curently implemented for fMRI) in full detail:
http://www.ncbi.nlm.nih.gov/pubmed/15219588

Best wishes,
Klaas


----- Ursprüngliche Mail ----
Von: Mina Khoshnejad <[log in to unmask]>
An: Klaas Enno Stephan <[log in to unmask]>
Gesendet: Mittwoch, den 23. Juli 2008, 19:49:12 Uhr
Betreff: Re: AW: [SPM] Group Bays Factor - DCM Analysis

Dear Klass,

 

I was refering to DCM.BIC.mat file that is stored. these BIC values are kind of similar for all models.

 

the command matlab output that you mentioned looks like something like this:

-------------------------------------------------------------------------------------------------------------------------------

Model 9: C:\Users\mina\Documents\DNIC\Pain only- Control group\subject1\analyzed\GLM_High\DCM_high\DCM_Model9_suje1_h.mat
          versus
Model 1: C:\Users\mina\Documents\DNIC\Pain only- Control group\subject1\analyzed\GLM_High\DCM_high\DCM_Model1_suje1_h.mat
 
All costs are in units of binary bits
 
Region S1L: relative cost  = -0.2625, BF= 1.2
Region S2L: relative cost  = -1.554, BF= 2.935
Region S1R: relative cost  = 0.6212, BF= 0.6501
Region S2R: relative cost  = -4.01, BF= 16.11
AIC Penalty = 11.5416, BF = 0.0003355
BIC Penalty = 28.0895, BF = 3.501e-009
AIC Overall = 6.3365, BF = 0.01237
BIC Overall = 22.8844, BF = 1.292e-007
 
Strong evidence in favour of model 1
Bayes factor >= 0.01237

-------------------------------------------------------------------------

 

-- Lets say if I am interested in BIC, which of the above BF should I take?? there is 2 of them above.

 

-- I have 12 models per subject, shall I look at all comparison of each of these models VS all the rest?

ex: Group BF (model 1 Vs All)

     Group  BF (model 2 VS All)

     ... etc

 

-- Is there a measure to look at the residuals (estimated and real time courses) for the models, to tell how good DCM can predict the BOLD time series?

 

 

Thanks,

 

Mina

--- On Wed, 7/23/08, Klaas Enno Stephan <[log in to unmask]> wrote:

From: Klaas Enno Stephan <[log in to unmask]>
Subject: AW: [SPM] Group Bays Factor - DCM Analysis
To: "Mina Khoshnejad" <[log in to unmask]>, [log in to unmask]
Date: Wednesday, July 23, 2008, 7:52 AM

Dear Mina,

I am confused:  where do you get a BIC.mat file from?  The model comparison in DCM should only display values for both the AIC and BIC approximation in the MATLAB command window, but not save them as files.  The Bayes factor that DCM currently returns is based on the more conservative of the two approximations; it is displayed at the bottom of the statements in the MATLAB command window.  If you store the BF for each subject (for the same model comparison) and then multiply all subject-specific BFs, you get the group Bayes factor.  At the moment, you can only do this by hand in SPM.

All the best
Klaas


----- Ursprüngliche Mail ----
Von: Mina Khoshnejad <[log in to unmask]>
An: [log in to unmask]
Gesendet: Mittwoch, den 23. Juli 2008, 06:23:09 Uhr
Betreff: [SPM] Group Bays Factor - DCM Analysis

Hello All,

 

I have a question about calculating the Group Bays Factor for Group DCM analysis.

 

when I run the 'compare option' in DCM for my estimated models for each subject, I clearly get an evidence in favor one of my models, however when I look at the BIC.mat file for all the models in every subject, they have more or less the same BIC value.

 

I thought plotting the BIC values for each model should give me what I can see from compare option. Am I right?

 

Is this BIC the Bays factor?

so for the Group Bays factor, shall I multiply these BICs for each model in all subjects?

 

Thanks,

Mina




Gesendet von Yahoo! Mail.
Dem pfiffigeren Posteingang.



Gesendet von Yahoo! Mail.
Dem pfiffigeren Posteingang.