A Bayesian look at diagnostics in the univariate linear model

Abstract

This paper develops diagnostics for data thought to be generated in accordance with the general univariate linear model. A first set of diagnostics is developed by considering posterior probabilities of models that dictate which of k observations form a sample of n observations (k < n/2) are spuriously generated, giving rise to the possible outlyingness of the k observations considered. This in turn gives rise to diagnostics to help assess (estimate) the value of k. A second set of diagnostics is found by using the Kullback-Leibler symmetric divergence, which is found to generate measures of outlyingness and influence. Both sets of diagnostics are compared and related to each other and to other diagnostic statistics suggested in the literature. An example to illustrate to the use of these diagnostic procedures is included.