2
$\begingroup$

I always believed that the p value was $P(H_0|Data)$ -- which to my rude shock I recently learned was not the case -- Note: I studied traditional frequentist probability.

How can one express the p-value in terms of Bayesian concepts -- assuming one can?

$\endgroup$
2
$\begingroup$

The traditional frequentist definition of a $p$-value is, roughly, the probability of obtaining results which are as inconsistent or more inconsistent with the null hypothesis as the ones you obtained. $H_0$ can be neither observed nor randomly varying, so it's appearance in a probability statement is a aberration of frequentist notation.

If $T$ is a statistic that summarizes an experiment such that $T$ takes a known value (assume WLOG this is 0) when the null hypothesis is true, and takes non-zero values otherwise, a notational expression might be $p = P_{H_0} (t > T)$, where $t$ is the sampling distribution of the test statistic, a hypothetical distribution of replications of the very experiment you conducted. The conditional, $H_0$ imposes some restrictions on $t$, so that it takes a known form under the null, and this enables comparisons without actually conducting thousands upon thousands of replications of the experiment.

To explain a frequentist "p" value in Bayesian terms we need some clarity. Bayesian probability and Bayesian testing are two separate concepts. The likelihood ratio can be roughly understood as a Bayes Factor without a prior. Bayesian probability doesn't help us understand a frequentist $p$-value because $H_0$ "fixes" the parameter to be a particular value, and Bayesians do not take data to be randomly varying (in the sense of repeated experiments).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.