Dear Michaël
Bayesian model comparison is based on an approximation of the log model evidence - i.e. the log of the probability of having observed your data, given your model: log p(y|m). The approximation used is the free energy, which can be decomposed into two terms: accuracy - complexity. The explained variance corresponds to the accuracy. The complexity is based on how far the parameters have moved from their priors (a distance measure called the KL divergence).
Bayesian Model Selection, therefore, does not try to find the model with the best explained variance. Instead, it is used to find the model with the optimal tradeoff between accuracy and complexity. It is perfectly possible, therefore, that models with higher explained variance do not do well in the BMS, because the complexity cost was not worth the extra explained variance.
As for whether 10% explained variance is too low... You could have a useful model that only explains a little variance. E.g. with a sparse, event-related design where much of the timeseries is noise. However, personally, I prefer to have a model which explains most of the variance - especially given how much cleaning up is done on the fMRI timeseries before entering DCM (PCA across voxels, removal of nuisance regressors).
Best
Peter
-----Original Message-----
From: Mouthon Michaël [mailto:[log in to unmask]]
Sent: 05 June 2017 09:28
To: [log in to unmask]; Zeidman, Peter <[log in to unmask]>
Cc: Mouthon Michaël <[log in to unmask]>
Subject: {SPAM?} DCM : opposite result between variance explained and BMS
Dear Peter,
I would like to understand something.
I specify several DCM models (on a population of 19 subjects). Then I tested the best model with a Bayesian Model selection RFX inference method.
For data quality check, I have also look at the variance explained by the model using the code 'spm_dcm_fmri_check(DCM)' across my models and subjects.
I am surprise because both methods give opposite result. The best model for BMS method has usually a very low variance explained (below 10%) while several other models have a much better explained variance (above 10%).
Can you explain me why it happen ?
I would like to make comparison between connection weight of the A matrix between experimental conditions. Does it make sense to do it with my best BMS model although the variance explain is very low ?
Thank you very much in advance.
Ps. models which modulates connectivity had usually a better variance explain as modulation on the VOIs according 'spm_dcm_fmri_check(DCM)'. I don't know if this result is related to this fact.
|