Bayesian variable selection in linear regression models with non-normal errors

Multiple linear regression is a prime statistical tool used to discover potential relationships between an outcome and some explanatory variables of interest. One of the common required assumptions is for the error terms in the model to be Gaussian. Instead of assuming normality, an alternative is to use a finite mixture of normal distributions, allowing for a more flexible definition of the heterogeneity structure of the data. We use this approach to develop a Bayesian linear regression model with non-normal errors, and through variable selection we focus on finding active predictors effectively contributing to explaining patterns in the observations.