Robust Bayesian model selection for heavy-tailed linear regression using finite mixtures (Q2180257)

The distribution of r.v. \(Y\) belongs to the normal independent (NI) class of distributions if \(Y=\mu+U^{-1/2}W,\) where \(\mu\) is a location parameter, \(W\sim N(0, \sigma^2), U\) is a mixing positive r.v. which is either degenerate or continuous with pdf \(h(\cdot | \nu)\) and \(\nu\) is a scalar or vector indexing the distribution of \(U.\) Apart from the normal model, where \(U=1,\) two types of heavy-tailed densities based on the choice of \(h(\cdot | \nu)\) are explored: (a) the Student-\(t\) distribution, and (b) the Slash distribution. For each of the 3 classes of NI distributions, a novel methodology is presented to perform Bayesian model in linear regression models, based on a finite mixture of distributions to model the latent variable \(U\), where each component of the mixture corresponds to one possible model within the underlying class. A simultaneous analysis is made using the posterior probability of each model. Inference is performed via MCMC. An extension to censored linear regression is introduced. Simulations show the advantages of the proposed approach to a segregated analysis based on a fixed model selection criterion. Examples with real data are presented.