LKJ Correlation Priors and Gibbs Sampling
Year: 2026
Type: article
Abstract: The separation strategy of Barnard et al. (2000) decomposes a covariance matrix parameter into its correlation matrix and its standard deviations, then places priors on these. A popular choice for the correlation matrix prior is an LKJ distribution (Lewandowski et al., 2009), and a popular choice for the standard deviation priors is a scaled half-t distribution (e.g., Gelman, 2006). Such priors are commonly used to implement Bayesian hierarchical modeling in Stan software and related packages. We propose a hierarchical representation of this joint prior, based on a scale mixture of Wishart distributions, that is amenable to Gibbs sampling. The representation enables pure Gibbs sampling for Bayesian multivariate and hierarchical normal regression models that can often be more efficient than corresponding Stan implementations.
Authors: Park, Trevor, Culpepper, Steven
Institution: University of Illinois Urbana-Champaign
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221362913
