A suitable Bayesian approach in testing point null hypothesis: some examples revisited

Abstract

In the problem of testing the point null hypothesis H-0: theta = theta(0) versus H-1: theta not equal theta(0), with a previously given prior density for the parameter theta, we propose the following methodology: to fix an interval of radius epsilon around theta(0) and assign a prior mass, pi(0), to H-0 computed by the density pi(theta) over the interval (theta(0) - epsilon, theta(0) + epsilon), spreading the remainder, 1 - pi(0), over H-1 according to pi(theta). It is shown that for Lindley's paradox, the Normal model with some different priors and Darwin-Fisher's example, this procedure makes the posterior probability of H-0 and the p-value matching better than if the prior mass assigned to H-0 is 0.5.