A partially collapsed Gibbs sampler for Bayesian quantile regression

Abstract

We introduce a set of new Gibbs sampler for Bayesian analysis of quantile re-gression model. The new algorithm, which partially collapsing an ordinary Gibbs sampler, is called Partially Collapsed Gibbs (PCG) sampler. Although the Metropolis-Hastings algorithm has been employed in Bayesian quantile regression, including median regression, PCG has superior convergence properties to an ordinary Gibbs sampler. Moreover, Our PCG sampler algorithm, which is based on a theoretic derivation of an asymmetric Laplace as scale mixtures of normal distributions, requires less computation than the ordinary Gibbs sampler and can significantly reduce the computation involved in approximating the Bayes Factor and marginal likelihood. Like the ordinary Gibbs sampler, the PCG sample can also be used to calculate any associated marginal and predictive distributions. The quantile regression PCG sampler is illustrated by analysing simulated data and the data of length of stay in hospital. The latter provides new insight into hospital perfor-mance. C-code along with an R interface for our algorithms is publicly available on request from the first author.

JEL classification: C11, C14, C21, C31, C52, C53.