= Jeffreys-Lindley paradox for a Gaussian point null
{c}
Suppose $X_i\sim N(\mu,1)$ independently, and compare $H_0:\mu=0$ with $H_1:\mu\sim N(0,\tau^2)$ for fixed $\tau>0$, using equal prior model probabilities. Put $z=\sqrt n\,\overline X$. The <posterior probability of a Gaussian point null> is
$$
P(H_0\mid\overline X)=\left[1+\frac{\exp\{z^2n\tau^2/[2(1+n\tau^2)]\}}{\sqrt{1+n\tau^2}}\right]^{-1}.
$$
Holding $z$ fixed while $n\to\infty$ leaves the two-sided <p-value> $2\Phi(-|z|)$ unchanged but makes the null <posterior probability> tend to one. For $n=100$, $\tau=1$ and $z=2$, the <p-value> is about $0.0455$, while the null posterior is about $0.581$.
The two procedures answer different model-based questions: a tail probability under the point null and a comparison of that null with a specified alternative <prior distribution>. The alternative's fixed prior scale spreads mass over parameter values much wider than the increasingly precise likelihood. This is the Jeffreys-Lindley phenomenon, not a conversion formula from <p-values> to <posterior probabilities>.<Bayes Factors by Robert E. Kass and Adrian E. Raftery>[Section 5.2, discussion of sample size and point-null testing]{ref}
Back to article page