Bayesian Model Selection

Functions for model selection/comparison.

References

[RigouxVBA3]

Rigoux L., Daunizeau J. VBA Toolbox

[RigouxBMS3]

Rigoux, L. et al. (2014) Neuroimage 84, 971–985

[GershmanMfit3]

Samuel Gershman’s mfit package (on GitHub)

Module Documentation

class fitr.model_selection.bms.BMS(model_fits, c_limit=1e-99)

Bayesian model selection [RigouxBMS].

Attributes

modelfits	(list) List of fitrfit objects from completed model fitting
nmodels	(int) Number of models to be compared
nsubjects: int	Number of subjects in the sample. (Must be equal across all fits).
c_limit	(float) Threshold at which to stop model comparison

Methods

run(self)	Runs model comparison by Bayesian Model Selection
dirichlet_exceedance(self, alpha)	Computes exceedance probabilities for a Dirichlet distribution
BOR(self, L, posterior, priors, C=None)	Computes Bayes Omnibus Risk (BOR)
FE(self, L, posterior, priors)	Derives free energy for current approximate posterior distribution
FE_null(self, L, options):	Derives the free energy of the ‘null’ hypothesis

BOR(L, posterior, priors, C=None)

Computes Bayes Omnibus Risk (BOR)

Parameters:

L

posterior

priors

C

Returns:

bor

Bayesian omnibus risk

Notes

As in [GershmanMfit].

FE(L, posterior, priors)

Derives free energy for current approximate posterior distribution [RigouxVBA].

Parameters:

L

Log model-evidence

posterior : dict

priors : dict

Returns:

F

Free energy of the current posterior

FE_null(L, options)

Derives the free energy of the ‘null’ hypothesis

Parameters:

L

Log model evidence

options : dict

Returns:

F0m

Evidence for the null (i.e. equal probabilities) over models

F0f

Evidence for the null (i.e. equal probabilities) over families

dirichlet_exceedance(alpha)

Computes exceedance probabilities for a Dirichlet distribution

Parameters:

alpha : float [0-1]

Returns:

xp

Exceedance probabilities

Notes

Implemented as in [GershmanMfit].

run()

Runs Bayesian model selection algorithm

Returns:

ModelComparisonResult :

Object representing model comparison results