arviz.rhat

arviz.rhat(data, *, var_names=None, method='rank')[source]

Compute estimate of rank normalized splitR-hat for a set of traces.

The rank normalized R-hat diagnostic tests for lack of convergence by comparing the variance between multiple chains to the variance within each chain. If convergence has been achieved, the between-chain and within-chain variances should be identical. To be most effective in detecting evidence for nonconvergence, each chain should have been initialized to starting values that are dispersed relative to the target distribution.

Parameters
dataobj

Any object that can be converted to an az.InferenceData object. Refer to documentation of az.convert_to_dataset for details. At least 2 posterior chains are needed to compute this diagnostic of one or more stochastic parameters. For ndarray: shape = (chain, draw). For n-dimensional ndarray transform first to dataset with az.convert_to_dataset.

var_nameslist

Names of variables to include in the rhat report

methodstr

Select R-hat method. Valid methods are - “rank” # recommended by Vehtari et al. (2019) - “split” - “folded” - “z_scale” - “identity”

Returns
xarray.Dataset

Returns dataset of the potential scale reduction factors, \(\hat{R}\)

Notes

The diagnostic is computed by:

\[\hat{R} = \frac{\hat{V}}{W}\]

where \(W\) is the within-chain variance and \(\hat{V}\) is the posterior variance estimate for the pooled rank-traces. This is the potential scale reduction factor, which converges to unity when each of the traces is a sample from the target posterior. Values greater than one indicate that one or more chains have not yet converged.

Rank values are calculated over all the chains with scipy.stats.rankdata. Each chain is split in two and normalized with the z-transform following Vehtari et al. (2019).

References

Vehtari et al. (2019) see https://arxiv.org/abs/1903.08008 Gelman et al. BDA (2014) Brooks and Gelman (1998) Gelman and Rubin (1992)