Bayesian Econometrics Workshop — Abstracts

In the process of model formulation, specification and modification, non-Bayesians routinely employ pure significance tests. This talk, based on Geweke (2007) and subsequent work in progress, builds on three facts. (1) Bayesians correctly criticize pure significance testing. (2) Econometricians (including some sympathetic to Bayesian methods) continue to employ these tests. (3) Rational individuals, including at least some of the investigators in point (2), in fact behave as Bayesians. The resolution of these three observations is that when pure significance tests are conducted wisely – that is, with due consideration of power – investigators have in mind alternative models that are incomplete, having not yet been fully specified. The talk shows how to construct Bayes factors between complete models (those typically subject to pure significance tests) and incomplete models. The result is a procedure that accomplishes the objectives that pure significance tests try to achieve, but is wholly Bayesian. The talk illustrates these ideas in some detail with a specific example from the econometric asset return modeling literature.