Block Hyper-g Priors in Bayesian Regression

Abstract: The development of prior distributions for Bayesian regression has traditionally been driven by the goal of achieving sensible model selection and parameter estimation. The formalization of properties that characterize good performance has led to the development and popularization of thick tailed mixtures of g priors such as the Zellner--Siow and hyper-g priors. The properties of a particular prior are typically illuminated under limits on the likelihood or the prior. In this paper we introduce a new, conditional information asymptotic that is motivated by the common data analysis setting where at least one regression coefficient is much larger than others. We analyze existing mixtures of g priors under this limit and reveal two new behaviors, Essentially Least Squares (ELS) estimation and the Conditional Lindley's Paradox (CLP), and argue that these behaviors are, in general, undesirable. As the driver behind both of these behaviors is the use of a single, latent scale parameter that is common to all coefficients, we propose a block hyper-g prior, defined by first partitioning the covariates into groups and then placing independent hyper-g priors on the corresponding blocks of coefficients. We provide conditions under which ELS and the CLP are avoided by the new class of priors, and provide consistency results under traditional sample size asymptotics.