Multivariate Bayes becomes accessible once you have loaded the results from
a first-level GLM analysis. Use the 'Results' button to do that, which will
make the three MVB-related buttons appear: multivariate Bayes, BMS, and
p-value. For details, see slide 38 in:
From: MS Al-Rawi
Sent: Thursday, February 24, 2011 12:33 PM
To: Kay H. Brodersen
Cc: [log in to unmask]
Subject: Re: [SPM] multivariate bayes analysis
I want to give MVB a try, however, I don't see anything in the SPM_GUI->MVB.
But, yet, there are mvb related functions, e.g. spm_mvb.m, etc. (SPM was
All the best,
----- Original Message ----
> From: Kay H. Brodersen <[log in to unmask]>
> To: [log in to unmask]
> Sent: Thu, February 24, 2011 8:35:18 AM
> Subject: Re: [SPM] multivariate bayes analysis
> Dear Ciara,
> Great to hear you found the SPM course useful. With regard to your first
> question: the priors in Multivariate Bayes have not been renamed.
> the code actually contains more priors than the GUI offers. You can see
> by comparing the priors listed in spm_mvb_estimate (line 78) to those
> actually supported in spm_mvb_U (starting in line 40). I would suggest
> starting off by comparing those models that are directly listed in the
> Your second question relates to fixed-effects versus random-effects model
> comparison. SPM currently does not provide a fully automated pipeline for
> comparing different MVB coding hypotheses (i.e., spatial priors).
> this is easy to do with just a few Matlab commands. You would begin by
> running MVB separately for all subjects and coding hypotheses. Given n
> subjects and m models, this results in n*m different MVB.mat files. You
> extract the log model evidences from these files (F) and arrange them in
> n*m matrix M.
> (i) For fixed-effects model comparison, you would then compute group
> factors as sum(M,1) and select the model with the highest group Bayes
> factor. This analysis relies on the assumption that the same model is
> in all subjects.
> (ii) For a random-effects model comparison, you can use spm_BMS(M, ...).
> This approach relaxes the assumption above and explicitly accounts for
> between-subjects variability.
> Let me know if this works. I'm copying this to the SPM mailing list so
> other users of MVB may benefit from our discussion.
> Very best wishes
> Kay Henning Brodersen
> Department of Computer Science
> Pattern Analysis and Machine Learning Group
> ETH Zurich
> [log in to unmask]
> From: Ciara Greene
> Sent: Monday, February 21, 2011 3:34 PM
> To: [log in to unmask]
> Subject: multivariate bayes analysis
> Hi Kay,
> Thanks for the interesting practical session on multivariate analysis at
> SPM course last week, I really enjoyed it and I'm hoping to put it to
> use. I'm trying out the multivariate bayes method with a view to
> coding models in my data, and I had a couple of questions; I hope you
> mind answering them!
> Firstly, I'm a bit confused by the terminology for the model priors. In
> Friston et al. NeuroImage paper, the various priors were listed as
> smooth, singular and support, while in the SPM GUI the options are
> sparse, smooth and support. I can't find a description of these new
> anywhere, so I'm not sure if the priors have simply been renamed (e.g.
> singular becoming compact???) or if these represent new models that
> in the old version.
> My second question related to comparing models across subjects. For
> reasons, I'm not really interested in model comparison within a single
> subject, I want to see which model seems best in my whole sample (and by
> extension, in the population). I've read a bit about the new Bayesian
> selection method for DCM implemented in SPM8 which allows a random
> comparsion of models across subjects, but I can't get that to work with
> multivariate bayes. The BMS button in the GUI allows models to be
> to one another, but presumably that uses fixed effects, and isn't
> for group analysis. As far as I understand, it also doesn't allow the
> results of the analyses to be saved out and carried forward to a second
> level. Do you know if it's possible to use the DCM model selection method
> outside DCM, or alternatively if there's another way of running a group
> level analysis here?
> Thanks in advance for any help you can provide!