SPM/Bayesian Parameter Averaging (BPA)

From Wikibooks, open books for an open world
< SPM
Jump to navigation Jump to search

Bayesian Parameter Averaging (BPA)[edit | edit source]

Bayesian Parameter Averaging [1] is a fixed effects average of parameters over several DCMs. The models need to have the same structure, but can be of different data - e.g. you could take an average of the same model fitted to several subjects' data. Being a fixed effects average, it is appropriate to use this approach when one assumes that all subjects have the same underlying model.

Theory[edit | edit source]

Let's say we have subjects and we wish to average the parameters across one model from each subject. For each model we have a posterior covariance matrix , which contains the variance (on the diagonal) and the covariance (off-diagonal) of the parameter estimates. The inverse of this is the precision matrix , which we'll use more often than the covariance. Each model also has a vector of estimated parameter means, . The priors are assumed to be the same for all models, with prior precision matrix and vector of prior means .

Bayes law for Gaussians says that precisions are combined by addition, whereas means are combined by summing each mean weighted by its (normalized) precision:

The Bayesian Parameter Average across models is derived from this. It has posterior precision:

And posterior mean:

For a full derivation of these formulae, see the technical note by Will Penny [1].

Practice[edit | edit source]

GUI: Click "Dynamic Causal Modelling", then click "Average", then "BPA". Select the models you wish to include and a name for the output.

Scripting: Use the Matlab function

spm_dcm_average(P,name)

where 'P' is a cell array of model filenames and 'name' is the name with which to save the result.

With either method, a new model will be saved with the name "DCM_avg_name.mat". The contents of this averaged model are identical to those of the first model you selected as input, except for the following:

  • DCM.Ep (averaged posterior means, )
  • DCM.Cp (averaged posterior covariance matrix, )
  • DCM.Vp (averaged posterior variance vector of each parameter - the diagonal of DCM.Cp)
  • DCM.Pp (averaged posterior probability of each parameter).

Frequently Asked Questions[edit | edit source]

Why are my estimated parameter means larger / smaller than I expect?

First, bear in mind that this is a bayesian average, so the priors will have an impact on the outcome. If there were two parameters which both had prior mean 0 and posterior mean 0.7, you shouldn't be surprised if the average is not exactly 0.7, but rather is closer to the prior, 0. The extent to which the average is drawn towards the prior will depend on the prior's precision.

Second, some very unintuitive results can be found when parameters are correlated, as described by Kasess and colleagues [2]. The precision matrices in the formulae above reflect not only the variance of each parameter individually, but also the (off-diagonal) covariance between parameters. If the posterior covariances are non-zero, then the posterior mean of the BPA will deviate from the arithmetic mean. In fact, with certain combinations of covariances, the average of a parameter in two models could be outside the range of each parameter mean separately.

Kasess et al. [2] compared BPA (as described above) against 'Posterior Variance Weighted Averaging' (PVWA), which is the same process but with the covariance between parameters set to zero before averaging. This avoids the unintuitive result described above, but at the cost of not taking the covariance structure into account. They found that with signal-to-noise ratios of 1-2, which may be typical for fMRI data, the two approaches gave similar results. Anyone wishing to use this in SPM can call spm_dcm_average with the third parameter ('nocond') set to true:

nocond=true;
spm_dcm_average(P,name,nocond)

What's the difference between Bayesian Parameter Averaging (BPA) and Bayesian Model Averaging (BMA)?

BPA averages parameters across models without any regard for the probability of the models themselves. BMA, by contrast, weights parameter estimates by the probability of the model which contains them. As such, it can be used to average over models of different structure as well as across subjects. BMA can be used for random-effects analysis, whereas BPA cannot.

References[edit | edit source]

  1. a b http://www.fil.ion.ucl.ac.uk/~wpenny/publications/parameter_averaging.pdf
  2. a b Kasess, Christian Herbert; Stephan, Klaas Enno; Weissenbacher, Andreas; Pezawas, Lukas; Moser, Ewald; Windischberger, Christian (2010). "Multi-subject analyses with dynamic causal modeling". NeuroImage. 49 (4): 3065–3074. doi:10.1016/j.neuroimage.2009.11.037. ISSN 1053-8119.