Abstract

To mitigate the Look Elsewhere Effect in multiple hypothesis testing using _p_-values, the paper suggests an “entropic correction” of the significance level at which the null hypothesis is rejected. The proposed correction uses the entropic uncertainty associated with the probability measure that expresses the prior-to-test probabilities expressing how likely the confirming evidence may occur at values of the parameter. When the prior-to-test probability is uniform (embodying maximal uncertainty) the entropic correction coincides with the Bonferroni correction. When the prior-to-test probability embodies maximal certainty (is concentrated on a single value of the parameter at which the evidence is obtained), the entropic correction overrides the Look Elsewhere Effect completely by not requiring any correction of significance. The intermediate situation is illustrated by a simple hypothetical example. Interpreting the prior-to-test probability subjectively allows a Bayesian spirit enter the frequentist multiple hypothesis testing in a disciplined manner. If the prior-to-test probability is determined objectively, the entropic correction makes possible to take into account in a technically explicit way the background theoretical knowledge relevant for the test.