Bayesian outliers functions for linear models

Abstract

This paper introduces two new diagnostic tools: the Bayesian outlier curve (BOC) and the Sequential Bayesian outlier curve (SEBOC). Both are built using the posterior odds for every possible number of outliers in a scale contaminated linear model. It is shown that these functions have a cross-validation interpretation, and can be useful to judge the robustness of the fitted model. The computation of these curves is carried out using ideas from stratified sampling.