Bayes Factor

The ratio of the marginal likelihoods of the data under two specific models, isolating the data's weight of evidence between them and updating prior model odds to posterior odds multiplicatively — with an automatic Occam penalty on flexible models.

Version

v4 · 2026-09-19 · History · 1 correction

Domain-specific #

87

Domain group

Formal Sciences

Origin domain

Mathematics

Subdomain

Statistics → Mathematics

Core Idea

The Bayes factor is the ratio of the marginal likelihoods of data under two competing models, BF12 = P(D|H1)/P(D|H2), each marginal likelihood being the fit averaged over the model's parameter space against its prior. It quantifies how much more probable the data are under one hypothesis than the other, and updates prior model odds to posterior odds multiplicatively. Three features distinguish it from the p-value: it compares two specific hypotheses (so it can give evidence for a hypothesis), the integration imposes an automatic Occam penalty, and it accumulates multiplicatively over data.

How would you explain it like I'm…

Whose Guess Called It Better?

Two friends each make a guess about what will happen before a game. Then the game happens. The Bayes factor tells you how many times better one friend's guess matched what happened than the other's. A friend who guessed "anything could happen" spreads their guess thin, so they get less credit if something specific happens.

Which Explanation Predicted Best

Scientists often compare two ideas, or models, to see which explains the data better. The Bayes factor compares how likely the data were under each idea, as a ratio, for example "the data were 10 times more likely under idea 1 than idea 2." Each idea's chance of the data is averaged over all the settings it considers possible. That means an idea that allows too many possibilities gets penalized if the data land in a small spot. Unlike some tests, a Bayes factor can give evidence for an idea, not just against one. You multiply it by how much you believed each idea before to get how much you should believe them after.

Ratio of Marginal Likelihoods

The Bayes factor compares two hypotheses, H1 and H2, by the ratio of how probable the observed data D are under each: BF12 = P(D | H1) / P(D | H2). Each of these marginal likelihoods averages the likelihood of the data over the model's parameters, weighted by that model's prior. It does not depend on how likely you thought the hypotheses were beforehand; instead, it multiplies the prior odds to give the posterior odds. It differs from a p-value in three ways. It compares two specific hypotheses and can support either one. The averaging over parameters automatically penalizes models that spread their prior over many possibilities, an Occam penalty. And it combines neatly as data arrive in stages, which helps in adaptive designs where p-values break down. People often read it with rough bands, such as 3 to 10 as moderate and above 100 as decisive or extreme, but these are only reading aids, and published versions of the scale differ. In forensic science it appears as the likelihood ratio reported to a court.

 

The Bayes factor is the ratio of marginal likelihoods of data D under two competing hypotheses or models, BF12 = P(D | H1) / P(D | H2), where P(D | Hi) = integral of P(D | theta, Hi) p(theta | Hi) d theta, the likelihood averaged over the model's prior. It measures how much more probable the data are under H1 than H2 without reference to prior model probabilities, and it updates prior odds to posterior odds multiplicatively: posterior odds = BF12 times prior odds. Unlike p-values, it compares two specified hypotheses and so can provide evidence for a hypothesis, not only against one. Integration over parameters imposes an automatic Occam penalty, since a flexible model that spreads prior mass diffusely has lower marginal likelihood when data fall in a narrow region. It composes sequentially, BF(D1, D2) = BF(D1) times BF(D2 | D1), which suits optional stopping and adaptive designs where p-value error guarantees fail. Conventional interpretive bands (roughly 1-3 anecdotal, 3-10 moderate, 10-30 strong, 30-100 very strong, above 100 decisive or extreme) derive from Jeffreys' grades, but Jeffreys used half-powers of ten and different labels, and the rounded version comes from Lee and Wagenmakers, so the thresholds are reading aids, not decision criteria. In forensic science the same quantity is the likelihood ratio an examiner reports, leaving prior odds to the trier of fact.

Scope of Application

Being a statistic, the Bayes factor applies wherever its precondition holds: two specific probability models each assigning a computable marginal likelihood to the same data.

• Statistics — the home turf: Bayesian model comparison, gradable evidence for a null.
• Cognitive psychology — the Bayesian-replication movement, because null evidence is gradable.
• Genetics and bioinformatics — ranking causal models for variant-trait associations.
• Machine learning — marginal-likelihood model selection; the variational ELBO as surrogate.
• Astrophysics — comparing cosmological models against survey data.
• Forensic science — the reported likelihood ratio, the court supplying prior odds.

Clarity

The Bayes factor isolates the evidence the data supply from how plausible the hypotheses were going in, splitting posterior odds into prior odds times Bayes factor. It poses the sharp question the p-value cannot — how much more probable are these data under H1 than H2 — yields evidence for a hypothesis (including a null), and makes the flexibility penalty legible without an ad hoc complexity term.

Manages Complexity

An entangled tangle — fit, flexibility, prior plausibility, decision threshold, cross-study combination — compresses into a single scalar read through the multiplicative split posterior odds = BF × prior odds. The fit-versus-complexity trade-off is charged automatically, evidential accumulation collapses to multiplication, and strength collapses to a position on Jeffreys' calibration scale.

Abstract Reasoning

The Bayes factor licenses a diagnostic evidence-direction-and-strength reading (including support for a null versus uninformativeness); a diagnostic detecting a fit-versus-flexibility mismatch via the Occam penalty; an interventionist move accumulating evidence and stopping adaptively under the multiplicative law; and boundary-drawing moves partitioning data-owned from analyst-owned and treating Jeffreys' bands as reading aids, not thresholds.

Knowledge Transfer

As a statistic the Bayes factor transfers literally wherever two specific probability models assign marginal likelihoods to the same data — psychology, genetics, ML, astrophysics, and forensics are the same computed quantity, not analogies. The boundary is over-reading: where no such models exist, "Bayes factor" does rhetorical work. There the portable content is the parent Bayesian Updating (prior odds × evidence = posterior odds), of which the Bayes factor is the model-comparison evidence-summary.

Relationships to Other Abstractions

Current abstraction Bayes Factor Domain-specific

Parents (1) — outgoing typed relations

• Bayes Factor includes Bayesian Updating Prime

  A Bayes factor is the evidence multiplier that converts prior model odds into posterior model odds within Bayesian updating.

Hierarchy paths (3) — routes to 2 parentless roots

• Bayes Factor → includes Bayesian Updating → is a kind of Inductive Reasoning

Show alternative paths (2)

Neighborhood in Abstraction Space

Bayes Factor sits in a moderately populated region (60^th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Unclustered & Miscellaneous (2769 abstractions)

Nearest neighbors

• Jeffreys-Lindley Paradox — 0.86
• Variational Bayesian Methods — 0.86
• Hannan–Quinn information criterion — 0.85
• Monotone Likelihood Ratio Property — 0.85
• Machine-Learning Learning Curve — 0.84

Computed from structural-signature embeddings · 2026-10-25