Reconciling Bayesian and Frequentist Tests: the Imprecise Counterpart
Proceedings of the Tenth International Symposium on Imprecise Probability: Theories and Applications, PMLR 62:97-108, 2017.
Imprecise Dirichlet Process-based tests (IDP-tests, for short) have been recently introduced in the literature. They overcome the problem of deciding how to select a single prior in Bayesian hypothesis testing, in the absence of prior information. They make use of a “near-ignorance” model, that behaves a priori as a vacuous model for some basic inferences, but it provides non-vacuous posterior inferences. We perform empirical studies regarding the behavior of IDP-tests for the particular case of Wilcoxon rank sum test. We show that the upper and lower posterior probabilities can be expressed as tail probabilities based on the value of the $U$ statistic. We construct an imprecise frequentist-based test that reproduces the same decision rule as the the IDP test. It considers a neighbourhood around the $U$-statistic value. If all the values in the neighbourhood belong to the rejection zone (resp. to the acceptance region), the null hypothesis is rejected (resp. accepted). Otherwise, the judgement is suspended.