March 9, 2018 – 9:30. Dipartimento di Fisica, Edificio 8, Aula Rossa, Piazza Leonardo da Vinci 32

The most common way of testing statistical hypotheses is to conduct null hypothesis significance tests (NHST) and to use a p-value to describe evidence against the null hypothesis. In this talk, I would like to highlight a fundamental conceptual problem with this approach: the impossibility to express support for the null hypothesis. Since null hypotheses are often simple and precise idealizations of complex models with substantial theoretical importance, a good method for scientific hypothesis tests has to be able to express support for them. Also, the one-sided nature of p-values arguably aggravates the replication crisis in various scientific disciplines. My approach is twofold: first I explain why classical NHST and classical Bayesian inference fail to evaluate a null hypothesis in an appropriate way; then I develop a measure of corroboration, taking inspiration from both Bayesian and frequentist procedures. I argue that degrees of corroboration achieve a more nuanced judgment on the evidence in favor of a null hypothesis and that they can be used in a variety of cases in statistical inference.