bayespecon.diagnostics.ModelComparison¶

class bayespecon.diagnostics.ModelComparison(models, labels=None, method='bridge', **kwargs)[source]¶

Compare a collection of fitted Bayesian models.

Object-oriented wrapper around bayes_factor_compare_models(), post_prob(), and arviz.compare(). Marginal likelihoods are computed once on first access and cached, so repeated calls to bayes_factors(), posterior_probabilities(), etc. reuse the same estimates.

Parameters:¶

models : list, tuple, or dict¶: Fitted model objects (each with inference_data and, for bridge sampling, pymc_model). A dict is interpreted as {label: model}; otherwise pass labels explicitly.
labels : list of str, optional¶: Display labels. Required when models is a list/tuple and no __name__-style label is available; auto-generated when omitted.
method : {"bridge", "bic"}, default "bridge"¶: Marginal-likelihood estimator. See bayes_factor_compare_models().
**kwargs¶: Forwarded to bayes_factor_compare_models() (e.g. repetitions, maxiter for bridge sampling).

Examples

>>> cmp = ModelComparison({"OLS": ols, "SAR": sar, "SDM": sdm})
>>> cmp.log_marginal_likelihoods
OLS   -123.4
SAR   -120.1
SDM   -119.8
Name: log_ml, dtype: float64
>>> cmp.bayes_factors        # pairwise BF matrix
>>> cmp.posterior_probabilities()
>>> cmp.loo_comparison()     # arviz.compare on LOO

__init__(models, labels=None, method='bridge', **kwargs)[source]¶

Methods

`__init__`(models[, labels, method])
`loo_comparison`(**kwargs)	Compare models by LOO via `arviz.compare()`.
`posterior_probabilities`([prior_prob])	Posterior model probabilities given the estimated marginal likelihoods.

Attributes

`bayes_factors`	Pairwise Bayes factors \(BF_{ij} = ML_i / ML_j\).
`diagnostics`	Per-model diagnostics from the marginal-likelihood estimator.
`labels`	Model labels in the order supplied.
`log_marginal_likelihoods`	Estimated \(\log p(y \mid M_k)\) for each model.

property bayes_factors : pandas.DataFrame[source]¶: Pairwise Bayes factors \(BF_{ij} = ML_i / ML_j\).

property diagnostics : dict | None[source]¶: Per-model diagnostics from the marginal-likelihood estimator.

property labels : list[str][source]¶: Model labels in the order supplied.

property log_marginal_likelihoods : pandas.Series[source]¶: Estimated \(\log p(y \mid M_k)\) for each model.

loo_comparison(**kwargs)[source]¶

Compare models by LOO via arviz.compare().

Each model must have a log_likelihood group in its inference_data. Extra keyword arguments are forwarded to arviz.compare().

Returns:¶: Output of arviz.compare (one row per model).
Return type:¶: pandas.DataFrame

posterior_probabilities(prior_prob=None)[source]¶

Posterior model probabilities given the estimated marginal likelihoods.

Parameters:¶

prior_prob : array-like, optional¶: Prior model probabilities. Defaults to uniform.

Returns:¶

Posterior probabilities, summing to 1, indexed by label.

Return type:¶

pandas.Series