Hi Stefan,
Ah, I see.
Yes - you will have to go to the matlab function here.
The relevant function is spm_BMS_gibbs.m if you're doing model comparison or
spm_compare_families.m if you're doing family comparison. This second function then calls the first function. Look at the help for each to see how to enter the number of samples.
The default number of samples is 1e4 for both functions (although the help of spm_BMS_gibbs says its 1e6 - but its not - the code says 1e4 - I'll change the help here).
If its not sampling variability there's another possibility.
Do you have families with different numbers of models ? eg. 10 in family 1, and 2 in family 2. If so, this could be the source of the discrepancy, because spm_compare_families sets up a uniform prior over families which does not equate to a uniform prior over models if you have families with different numbers of models (see the families paper for discussion of this). So spm_BMS_gibbs is called with a different prior (second argument).
If the discrepancies are substantial - which they don't seem to be in your case - in your paper I would report the model inferences from spm_BMS_gibbs.m, and family inferences from spm_compare_families.m.
Best,
Will
> -----Original Message-----
> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
> On Behalf Of Stefan Frässle
> Sent: 19 June 2013 14:50
> To: [log in to unmask]
> Subject: Re: [SPM] DCM: BMS
>
> Dear Dr. Penny,
>
> thank you very much for your help.
>
> I already tried to run the algorithm a couple of times using the same
> configurations and I indeed noticed the variability due to the Gibbs
> sampling. However there is still a clear difference between exceedance
> probabilities when defining families as compared to when not.
>
> I tried to increase the number of samples (as you suggested). However I
> only find that option in the BMS:Maps GUI and not in the BMS:DCM GUI,
> which I am using for model comparison (is that wrong?).
>
> So do I have to increase the number of samples directly within the
> Matlab-functions?
>
> Thanks again for your help. Its great to get feedback when you're stuck
> with a question concerning SPM/DCM.
>
> Best wishes
> Stefan
>
>
> Quoting "Penny, William" <[log in to unmask]>:
>
> > Dear Stefan,
> >
> > The computation of the exceedance probability is based on a sampling
> > approach - which can potentially produce small differences each time
> > you run it.
> >
> > To see if the differences are due to this 'sampling variability'
> > perhaps you could just run
> > the algorithm twice with exactly the same configuration (eg. both
> with
> > no families) to see what the variability is.
> >
> > To reduce this variability you can increase the number of samples
> used
> > - there's an option in the GUI for this.
> >
> > Best,
> >
> > Will.
> >
> >> -----Original Message-----
> >> From: SPM (Statistical Parametric Mapping)
> >> [mailto:[log in to unmask]] On Behalf Of Stefan Frässle
> >> Sent: 13 June 2013 14:17
> >> To: [log in to unmask]
> >> Subject: [SPM] DCM: BMS
> >>
> >> Dear DCM experts,
> >>
> >> I have a question regarding BMS RFX for DCM. I compared 16 DCM
> models
> >> for fMRI data and found one model to be the best with an exceedance
> >> probability of around 0.6.
> >>
> >> In a subsequent analysis step I intended to compared the 16 models
> >> according to a specific property, thus creating 4 different families
> >> (each including 4 models) in order to compare those families.
> >>
> >> When running the batch (including the definition of the families)
> SPM
> >> still shows me information about the comparison of the individual
> >> models and I noticed that the exceedance probability for my winning
> >> model is now around 0.7!
> >>
> >> I'm confused why there are differences in the exceedance
> >> probabilities when defining families as compared to when I just
> >> compare the individual models?
> >>
> >> I would be very grateful, if anyone could give his opinion on why
> >> this occurs.
> >>
> >> Thanks a lot in advance
> >> Stefan
|