Handbook of Statistics, Vol. 25
ISSN: 0169-7161
© 2005 Elsevier B.V. All rights reserved.
DOI 10.1016/S0169-7161(05)25005-8
5
Role of P-values and other Measures of Evidence in
Bayesian Analysis
Jayanta Ghosh
a
, Sumitra Purkayastha and Tapas Samanta
Abstract
We review the relation between P-values and posterior probabilities, and also the
Bayesian P-values for composite null hypothesis when there is no alternative.
Keywords: Bayes factors; Bayesian P-values; Pitman alternative; posterior proba-
bility
1. Introduction
A P-value is “the smallest level at which the observations are significant in a partic-
ular direction” (Gibbons and Pratt, 1975). Equivalently, it is the probability, under a
simple null or the supremum of probabilities under a composite null, of deviations of
the test statistic being greater than that currently observed. As a data based measure
of evidence against the null hypothesis H
0
, it contains more information than just a
statement of significance or lack of it at a predetermined level of significance, or the
Neyman–Pearsonian probability of Type I error. Therein lies its attraction as a measure
of evidence in classical statistical methodology for testing.
Gibbons and Pratt (1975) argue that in “any single experiment the P-value measures
the agreement or disagreement in a specific direction” and that, therefore, it serves a
useful purpose. On the other hand they point out that one cannot compare P-values
across sample sizes or across experiments – a point also made very effectively in Lindley
(1957). It is also clear and has been noted many times that the frequentist interpretation
of a P-value under a simple null is somewhat odd in that it depends on hypothetical
repetition leading to new samples which have more deviant values of a test statistic
than the value actually observed. Among other things it implies an apparent use of
unobserved more deviant data as evidence. A fact which is both an advantage and a
disadvantage is that a P-value only depends on the distribution under the null hypothesis,
it ignores the alternative. This is an advantage because the null distribution is usually
a
Corresponding author.
151