Jeffreys–Lindley Paradox

The Jeffreys–Lindley paradox is a fundamental conflict between frequentist and Bayesian hypothesis testing, particularly in the context of large sample sizes. While p-values from frequentist tests may indicate strong evidence against a null hypothesis, Bayesian methods often yield high posterior probabilities in favor of the same null—leading to sharply contradictory conclusions.

Originally identified by Harold Jeffreys and formalized by Dennis Lindley, the paradox stems from testing point null hypotheses within Gaussian models. It reveals that for fixed significance levels and increasing sample sizes, the Bayes factor and posterior probability can increasingly favor the null—even when the p-value is highly significant. This is due to the influence of priors concentrated on the null and the properties of the "spike-and-smear" model.

The article presents Lindley’s mathematical formulation of the paradox, including his now-famous expression for posterior probability involving prior weights. Bartlett’s correction, later known as Bartlett’s anomaly, exposes how prior variance can be manipulated to make posterior probability arbitrarily high—raising concerns about misuse of Bayesian inference.

The paradox is visualized using comparative charts of p-values, Bayes factors, and posterior probabilities, showing at which sample sizes each begins to support the null. It is shown that Bayes factors can grow unbounded with increasing sample size—even when the data under the null are improbable—highlighting a troubling inconsistency.

Numerous proposed resolutions are reviewed: altering priors, redefining significance, using Bayesian reference criteria, and even dissolving the paradox by testing interval null hypotheses instead. The article introduces the idea of a "Jeffreys event horizon"—a critical point where posterior belief becomes fixed despite contradictory data.

This work also connects the paradox to the Rao–Lovric Theorem and explores simulation results showing that the paradox arises frequently in the era of massive datasets. It concludes with a set of eight possible responses, from abandoning p-values and Bayes factors to redefining null hypothesis testing altogether.

For complete mathematical derivations, visualizations, and critical reflections, see the full article in the International Encyclopedia of Statistical Science.