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.

Core Idea¶

The Bayes factor is the ratio of the marginal likelihoods of data \(D\) under two competing hypotheses or models, \(\text{BF}_{12} = P(D \mid H_1) / P(D \mid H_2)\), where each marginal likelihood is the likelihood of the data averaged over the model's parameter space with respect to the model's prior: \(P(D \mid H_i) = \int P(D \mid \theta, H_i) \, p(\theta \mid H_i) \, d\theta\). It quantifies how much more (or less) probable the observed data are under \(H_1\) than under \(H_2\), independently of prior model probabilities, and it updates prior model odds to posterior model odds multiplicatively: \(\frac{P(H_1 \mid D)}{P(H_2 \mid D)} = \text{BF}_{12} \times \frac{P(H_1)}{P(H_2)}\).

Three structural features distinguish the Bayes factor from the p-value and from likelihood ratios in classical statistics. First, it compares two specific hypotheses rather than testing one against an implicit null; it can therefore provide evidence for a hypothesis, not only against one. Second, the integration over parameter space introduces an automatic Occam penalty: a model with more parameters that spreads its prior mass widely over a large parameter space makes a diffuse bet on the data, and when the data fall in a small region, the model's marginal likelihood is correspondingly lower than that of a sharper competitor — complex models are penalised for their flexibility without any explicit complexity term. Third, the Bayes factor transforms naturally under sequential data accumulation: \(\text{BF}\) on the combined dataset \((D_1, D_2)\) equals \(\text{BF}(D_1) \times \text{BF}(D_2 \mid D_1)\), making it well suited to optional-stopping and adaptive designs where the p-value's type-I error guarantees collapse.

Interpretation conventionally uses Jeffreys' calibration scale (Jeffreys, Theory of Probability, 1939/1961): \(\text{BF}_{12} \in (1, 3)\) as anecdotal, \((3, 10)\) as moderate, \((10, 30)\) as strong, \((30, 100)\) as very strong, \(> 100\) as decisive evidence for \(H_1\) over \(H_2\), with the reciprocal scale for evidence in the opposite direction. These thresholds are reading aids, not licensing criteria. In forensic science the Bayes factor appears as the likelihood ratio — the weight of evidence that a forensic examiner reports to a court, leaving the prior odds on guilt to the trier of fact.

Structural Signature¶

Sig role-phrases:

the two specific models — H₁ and H₂, each a fully-specified probability model assigning likelihood to the same data (no implicit null)
the marginal likelihoods — P(D | Hᵢ) = ∫ P(D | θ, Hᵢ) p(θ | Hᵢ) dθ, each model's data-fit averaged over its parameter space with respect to its prior
the evidence ratio — BF₁₂ = P(D | H₁) / P(D | H₂), the single scalar isolating the data's weight of evidence
the multiplicative odds split — posterior odds = BF₁₂ × prior odds, the engineered identity partitioning the verdict into data-owned and analyst-owned pieces
the automatic Occam penalty — integrating fit over a diffuse prior debits a flexible model for spreading its bet, charging complexity with no explicit term
the sequential multiplicativity — BF(D₁, D₂) = BF(D₁) × BF(D₂ | D₁), the guarantee that evidence accumulates by multiplication and optional stopping does not corrupt it
the Jeffreys calibration scale — anecdotal/moderate/strong/decisive bands (with reciprocal direction) as reading aids, not licensing thresholds — its characteristic reporting convention

What It Is Not¶

Not a p-value or a Bayesian significance threshold. A Bayes factor is a ratio of model evidences for two specific hypotheses, not a tail-area probability under a null. It answers "how much more probable are these data under H₁ than H₂?", whereas the p-value answers "how surprising would data this extreme be if the null held?" — different questions, and the two routinely disagree on the same dataset.
Not a posterior probability, and not inclusive of the prior. The Bayes factor is the data-only term in posterior odds = BF × prior odds; it deliberately excludes prior model plausibility. It is not the probability that H₁ is true — converting it to one requires supplying prior odds, which is precisely the piece the forensic examiner leaves to the court.
Not the plain likelihood ratio. Each term is a marginal likelihood, integrated over the model's parameter space against its prior, not the likelihood at a best-fit parameter value. The classical likelihood ratio is the special case with no free parameters; marginalizing is what introduces the automatic Occam penalty the bare ratio lacks.
Not a reward for better fit. A richer model that fits visibly better can still have a Bayes factor below one against a simpler rival, because spreading prior mass over a large parameter space is a diffuse bet that the integration debits. The question is not "does it fit better?" (it usually does) but "does it fit enough better to pay for its flexibility?"
Not "weak evidence" when near 1 — that is no evidence. A Bayes factor close to one means the data are uninformative between the hypotheses, which is categorically different from a Bayes factor far below one, which is positive evidence for H₂ (including for a null). The concept's signal feature is that it can support absence; conflating "uninformative" with "supports the null" discards exactly that distinction.
Not a licensing threshold. Jeffreys' bands (anecdotal, moderate, strong, decisive) are reading aids for the strength of evidence, not accept/reject cutoffs. Treating "BF > 10" as a decision rule the way 0.05 is misused reimports the dichotomous thinking the Bayes factor exists to replace; evidence here is graded and cumulative, not a one-shot verdict.

Scope of Application¶

Because the Bayes factor is a statistic — a measure summarizing the data's weight of evidence between two models — not a causal mechanism, it applies wherever its single precondition holds: two specific probability models that each assign a computable marginal likelihood to the same data. The fields below are real uses of the identical computed quantity (with its automatic Occam penalty and multiplicative accumulation), not analogies; the boundary is this precondition versus invoking "Bayes factor" loosely for any weighing of evidence, where what genuinely carries is the parent prime bayesian_updating.

Statistics — the home turf: Bayesian model comparison, replacing or complementing frequentist hypothesis tests and supplying gradable evidence for a null.
Cognitive psychology — the Bayesian-replication movement reports it precisely because null evidence is gradable rather than an inconclusive non-result.
Genetics and bioinformatics — it ranks competing causal models for variant–trait associations across the genome.
Machine learning — marginal-likelihood model selection is the same object, with the variational evidence lower bound an explicit surrogate for it.
Astrophysics — it compares cosmological models (e.g., dark-energy parameterizations) against survey data.
Forensic science — the likelihood ratio an examiner reports to a court is the Bayes factor applied to two source hypotheses, with the trier of fact supplying the prior odds.

Clarity¶

Naming the Bayes factor isolates a quantity that ordinary statistical practice keeps tangled: the evidence the data supply, held apart from how plausible the hypotheses were going in. By splitting the posterior odds cleanly into prior odds times Bayes factor, it lets an analyst point to the data's contribution alone and ask a sharp question that the p-value cannot pose — how much more probable are these observations under \(H_1\) than under \(H_2\)? — rather than the indirect "how surprising would data this extreme be if the null held?" That separation is exactly what makes the forensic division of labour coherent: the examiner reports the likelihood ratio (the data's weight) and the court owns the prior odds on guilt, each party responsible for the piece it can actually assess.

The label also dissolves two confusions that bedevil null-hypothesis testing. First, because it compares two specific models it can yield evidence for a hypothesis, not merely failure to reject one — so "we found no effect" becomes a gradable claim (anecdotal versus decisive support for the null) instead of an inconclusive non-result. Second, it makes the flexibility penalty legible without an ad hoc complexity term: a model that spreads its prior thinly across a large parameter space is seen to make a diffuse bet, and the integration over that space charges it automatically when the data land narrowly. The sharper question a practitioner can now ask of a richer model is not "does it fit better?" — it almost always does — but "does it fit enough better to overcome the diffuseness of its prediction?"

Manages Complexity¶

The question "what do these data say about which of two models is true?" otherwise drags several entangled concerns along at once: how well each model fits, how much each model was allowed to fit by virtue of its flexibility, how plausible each was before the data arrived, whether the evidence is strong enough to act on, and how to combine it with evidence already in hand. Handled informally, each of these is its own judgment call, re-argued per analysis and easily conflated — fit mistaken for evidence, a flexible model rewarded for fitting better, "no significant result" left uninterpretable. The Bayes factor compresses that whole tangle into a single scalar, the ratio of marginal likelihoods, and then makes the surrounding reasoning a matter of reading that one number through a fixed structure rather than re-deriving each concern. The organizing identity is the multiplicative split \(\text{posterior odds} = \text{BF}_{12} \times \text{prior odds}\): it factors the messy global question into three inspectable pieces — prior odds (what plausibility the analyst brings), the Bayes factor (what the data alone contribute), and posterior odds (the combined verdict) — so the data's contribution is isolated as one auditable quantity and no longer smeared together with prior belief.

Three things that would otherwise be separate adjudications collapse into properties of that single number. The fit-versus-complexity trade-off is charged automatically: because the marginal likelihood integrates the fit over the prior, a model that spreads its prior thinly across a large parameter space is debited for the diffuseness of its bet without any added complexity term, so the analyst tracks one quantity instead of trading off fit against parameter count by hand. Evidential accumulation collapses to multiplication: \(\text{BF}(D_1, D_2) = \text{BF}(D_1) \times \text{BF}(D_2 \mid D_1)\), so combining studies or stopping adaptively is reading off a product rather than reasoning afresh about error rates. And strength of evidence collapses to position on Jeffreys' calibration scale — anecdotal, moderate, strong, decisive — including graded evidence for a model, so "which model, how strongly, in which direction" is read off one scalar's magnitude and reciprocal. The high-dimensional comparison — fit, flexibility, prior plausibility, decision threshold, cross-study combination, all at once — is thereby reduced to tracking a single ratio and locating it on a fixed scale, with the qualitative verdict read off rather than re-assembled each time.

Abstract Reasoning¶

The Bayes factor licenses a set of inferential moves the p-value cannot, all flowing from its being the data-only term in the multiplicative odds split. The primary diagnostic move is evidence-direction-and-strength reading: from the magnitude and orientation of the single ratio, infer which hypothesis the data favor and by how much. A reasoner goes FROM "\(\text{BF}_{12} = 40\)" TO "the data are forty times more probable under \(H_1\) than \(H_2\), very strong evidence for \(H_1\)"; from \(\text{BF}_{12} = 0.05\), the reciprocal reading gives strong evidence for \(H_2\). Crucially this includes evidence for a null — a Bayes factor decisively below one is read as positive support for "no effect," so the reasoner can distinguish "the data argue for absence" from "the data are uninformative" (a Bayes factor near one), a separation the p-value's failure-to-reject collapses. A second diagnostic move is diagnosing a fit-versus-flexibility mismatch: when a richer model fits visibly better yet its Bayes factor against a simpler rival is below one, infer that the marginal likelihood's integration over a diffuse prior has charged the richer model for spreading its bet — so the reasoner concludes the extra fit did not pay for the extra flexibility, without computing any explicit complexity term.

The interventionist move is evidence accumulation and design adaptation under the multiplicative law. Because \(\text{BF}(D_1, D_2) = \text{BF}(D_1) \times \text{BF}(D_2 \mid D_1)\), the analyst predicts the effect of gathering more data by multiplying: a study at anecdotal strength can be pushed to decisive by a confirmatory replication, and the combined verdict is the product of the per-study factors rather than a re-derivation. This same law licenses optional stopping — the reasoner may monitor the Bayes factor as data arrive and stop when it crosses a chosen strength threshold, predicting that, unlike a p-value, the evidence summary is not corrupted by the looking, because the Bayes factor carries no type-I error rate to inflate. The interventionist can also re-target the comparison: swap \(H_2\) for a different specific rival and recompute, predicting how the same data's verdict shifts when the contrast changes — a move impossible for a test fixed against an implicit null.

The boundary-drawing moves are two. First, the Bayes factor partitions a verdict into the part the data own (itself) and the part the analyst owns (prior odds), so a reasoner can hand the data's weight to one party and the prior to another — the forensic discipline in which the examiner reports the likelihood ratio and the court supplies the prior odds on guilt, each held to what it can actually assess. Second, the Jeffreys scale draws reporting boundaries — anecdotal, moderate, strong, decisive — that let the reasoner predict whether a result is firm enough to act on, while the concept itself insists these are reading aids, not licensing thresholds, so the reasoner is warned against treating a scale crossing as a decision the way a p-value's 0.05 is mistakenly treated. The order-of-events claim the concept supports is that evidence is graded and cumulative: the same hypotheses can move from anecdotal to decisive as data accrue, so "how strong is the evidence now, and which way" is a quantity the reasoner tracks continuously rather than a one-shot accept/reject.

Knowledge Transfer¶

The Bayes factor is a statistic — a measure summarizing the data's weight of evidence between two models — not a causal mechanism, so the operative boundary (case C) is instrument-reach versus over-reading, and the construct transfers literally wherever its single precondition holds: two specific probability models that each assign a marginal likelihood to the same data. Within statistics this is its home (Bayesian model comparison, replacing or complementing frequentist tests), and it carries unchanged into every quantitative field that meets the precondition — these are not analogies but the same computed quantity on a new substrate. In cognitive psychology the Bayesian-replication movement reports it precisely because null evidence is gradable; in genetics and bioinformatics it ranks competing causal models for variant–trait associations; in machine learning marginal-likelihood model selection is the same object, and the variational evidence lower bound is an explicit surrogate for it; in astrophysics it compares cosmological models against survey data; and in forensic science the likelihood ratio an examiner reports to a court is the Bayes factor applied to two source hypotheses, with the trier of fact supplying the prior odds. In all of these the diagnostics (read evidence direction and strength off the single ratio; detect a fit-versus-flexibility mismatch via the automatic Occam penalty), the interventions (accumulate evidence and stop adaptively under the multiplicative law \(\text{BF}(D_1,D_2)=\text{BF}(D_1)\times\text{BF}(D_2\mid D_1)\)), and the prior-versus-data partition all apply directly, because the precondition is genuinely met.

The boundary to mark is therefore over-reading, not metaphor-versus-mechanism in the usual sense. Where there are not two well-defined probability models with computable marginal likelihoods — where "Bayes factor" is invoked loosely for any weighing of evidence — the construct has nothing to compute and the term is doing rhetorical, not statistical, work. In that thinner usage what genuinely carries is the parent abstraction (case B): the prior-odds × evidence = posterior-odds structure is bayesian_updating, of which the Bayes factor is the specific evidence-summary in the model-comparison step (and the classical likelihood ratio is its no-free-parameters special case; AIC/BIC are approximating cousins). So the literal cross-domain reach of the statistic is exactly co-extensive with where two probabilistic models and their marginal likelihoods exist; the broader "evidence updates belief multiplicatively" lesson, where no such models are in hand, belongs to bayesian_updating, not to the Bayes factor as named. See Structural Core vs. Domain Accent.

Examples¶

Canonical¶

A textbook worked case is testing whether a coin is fair. Let H₁ be the sharp hypothesis that the coin is fair (θ = 0.5, no free parameter) and H₂ the flexible hypothesis that the bias θ is unknown, given a uniform prior on [0, 1]. Suppose we observe 7 heads in 10 tosses. The data likelihood carries a binomial coefficient common to both models, so it cancels in the ratio. Under H₁ the marginal likelihood is proportional to 0.5¹⁰ = 1/1024 ≈ 0.000977. Under H₂ it is the integral ∫₀¹ θ⁷(1−θ)³ dθ = B(8,4) = 7!·3!/11! = 1/1320 ≈ 0.000758. The Bayes factor is therefore BF₁₂ = (1/1024)/(1/1320) = 1320/1024 ≈ 1.29 — anecdotal evidence favoring the fair coin, even though 7/10 fits a biased coin better at its best value (θ = 0.7). The flexible model's diffuse prior, spread across all θ, is debited for hedging its bet.

Mapped back: H₁ and H₂ are the two specific models; 0.5¹⁰ and B(8,4) are their marginal likelihoods, the second averaged over the whole parameter space. Their quotient 1.29 is the evidence ratio, landing in the anecdotal band of the Jeffreys calibration scale. That the better-fitting flexible model is nonetheless slightly disfavored is the automatic Occam penalty in a single computation.

Applied / In Practice¶

Forensic science operationalizes the Bayes factor as the "likelihood ratio" an expert reports to a court, and it is the backbone of evaluative reporting in DNA and trace evidence. An examiner comparing a crime-scene DNA profile to a suspect computes how probable the observed match is under two competing propositions — that the suspect is the source versus that an unrelated person is the source — using population allele frequencies for the denominator. A reported likelihood ratio of, say, one billion means the evidence is a billion times more probable if the suspect is the source than if a random unrelated person is. Crucially, the examiner reports only this ratio; standards from bodies such as the European Network of Forensic Science Institutes explicitly assign the prior odds on guilt, and their combination into a verdict, to the court rather than the scientist.

Mapped back: The source-versus-not-source propositions are the two specific models, and the reported likelihood ratio is the evidence ratio isolating the data's weight. Handing that ratio to the court while the court owns the prior is exactly the multiplicative odds split enforced as a division of institutional labor — the examiner assessing only what evidence can assess, the trier supplying prior odds.

Structural Tensions¶

T1: "Data-only" evidence versus parameter-prior dependence (the automatic Occam penalty is prior-tuned). The Bayes factor is advertised as isolating what the data alone contribute — the clean term in posterior odds = BF × prior odds — and as charging complexity automatically, with no ad hoc penalty. Both claims are only half true, because each marginal likelihood integrates the fit over the model's parameter prior, so the value of the ratio and the size of the Occam penalty depend on how diffusely that prior is spread. A wider prior debits a flexible model more; a tighter one debits it less; and for many models there is no principled prior width. The tension is that the "data-only" evidence summary is silently prior-laden, and the "automatic" penalty can be dialed up or down by a choice the data do not constrain — so two analysts with the same data and the same models can report materially different Bayes factors. What is presented as objective evidence is conditional on a within-model prior that rarely announces itself. Diagnostic: Is the Bayes factor here stable across defensible choices of parameter prior, or is its value (and the Occam penalty) being driven by an arbitrary prior width the data cannot pin down?

T2: Evidence for the null versus relativity to the specific alternative (support for absence is always contingent). The Bayes factor's celebrated advantage over the p-value is that it can support a null — "the data argue for no effect," gradable from anecdotal to decisive — rather than merely fail to reject one. But this evidence is irreducibly relative to the specific alternative chosen: a Bayes factor supports H₁ over this H₂, and swapping the alternative (a tighter effect-size prior, a different functional form) can move the verdict from support-for-null to support-for-effect on the same data. The p-value is criticized for testing against an implicit, unstated null; the Bayes factor requires a fully specified alternative and thereby imports its own modeling arbitrariness into what "no effect" means. The tension is that making null-evidence gradable requires committing to a particular rival, so "the data support absence" is never absolute but always support-relative-to-how-the-effect-was-modeled. Diagnostic: Is the evidence for the null robust across reasonable specifications of the alternative, or is "support for no effect" an artifact of the particular H₂ (and its effect-size prior) chosen?

T3: Optional-stopping freedom versus its limits (coherence is not immunity). Because the Bayes factor accumulates multiplicatively and carries no type-I error rate, an analyst may monitor it as data arrive and stop when it crosses a strength threshold without corrupting the evidence summary — a genuine and celebrated advantage over the p-value. But this coherence is easy to over-read as "anything goes." Stopping when the Bayes factor is favorable still biases the reported effect size upward, so a coherent evidence summary can accompany an inflated magnitude estimate; and the whole guarantee presumes the two models are correct, whereas under misspecification the multiplicative accumulation converges confidently to the less-wrong model, not the truth, growing more decisive with more data. The tension is that the property licensing optional stopping (evidential coherence under the multiplicative law) protects the evidence ratio while leaving effect-size bias and model-misspecification untouched, so freedom from error-rate inflation is not freedom from being confidently wrong. Diagnostic: Does the optional-stopping freedom here hold because the models are well-specified and only the evidence summary matters — or is it masking effect-size bias or convergence toward a merely-less-wrong model?

T4: Jeffreys scale as reading aid versus inevitable threshold (a calibration that becomes a cutoff). The concept insists that the Jeffreys bands — anecdotal, moderate, strong, decisive — are reading aids for graded, cumulative evidence, not accept/reject cutoffs, and that treating "BF > 10" as a decision rule reimports the dichotomous thinking the Bayes factor exists to replace. Yet supplying a labeled scale to make evidence interpretable all but guarantees the labels harden into bright lines: "BF > 3" and "BF > 10" are already used in practice exactly as p < 0.05 is, as publication and decision thresholds. The tension is that the very feature meant to escape dichotomous testing — a human-readable calibration of strength — hands practitioners a fresh set of thresholds they will treat as cutoffs, so the gradable-evidence tool manufactures new dichotomies at its reporting bands. The remedy (the scale) and the disease (thresholding) share a form. Diagnostic: Is the Bayes factor here being read as a continuous, accumulating weight of evidence, or has a Jeffreys band been turned into a de facto accept/reject cutoff?

T5: Autonomy versus reduction (the Bayes factor or the Bayesian-updating it instantiates). The Bayes factor is a named statistic with specific cargo — marginal likelihoods, the automatic Occam penalty, sequential multiplicativity, the Jeffreys scale, the forensic likelihood-ratio division of labor. As a statistic it transfers literally wherever two specific probability models assign computable marginal likelihoods to the same data, across psychology, genetics, machine learning, astrophysics, and forensics — these are the same computed quantity, not analogies. But where there are not two well-defined models with marginal likelihoods, "Bayes factor" is doing rhetorical work, and what genuinely carries is the parent bayesian_updating — the prior-odds × evidence = posterior-odds structure of which the Bayes factor is the model-comparison evidence-summary (the classical likelihood ratio its no-free-parameter special case, AIC/BIC approximating cousins). The tension is between a construct that transfers literally exactly where two probabilistic models exist and the broader "evidence updates belief multiplicatively" lesson that reaches further under the parent. Diagnostic: Resolve toward bayesian_updating when no two models with marginal likelihoods are in hand; toward the Bayes factor when comparing two specific probability models' marginal likelihoods on the same data in situ.

Structural–Framed Character¶

The Bayes factor sits toward the structural side — best read as mixed-structural, and among the most structural entries in the corpus, because it is not a causal mechanism but a statistic, a piece of mathematics that transfers literally rather than by analogy. Like available-to-promise and Battle of the Sexes, it is a substrate-indifferent object recognised as the same computed quantity across fields. The five criteria lean strongly structural, with the framed pull isolated in its reporting conventions. On evaluative weight it reads structural: the Bayes factor is a ratio — it renders no verdict on the world, isolating what the data alone contribute, and the concept explicitly insists its Jeffreys bands are "reading aids, not licensing thresholds." On human-practice-bound it reads mostly structural: the core is a mathematical relationship (a ratio of marginal likelihoods, the multiplicative odds split) derivable from Bayes' theorem, which does not dissolve when the statisticians leave, though it presupposes the analytic construction of two probability models. On institutional origin it is mostly structural with one framed thread: the object itself is mathematics, not an artifact of a survey or agency, but the Jeffreys calibration scale (anecdotal/moderate/strong/decisive) is a reporting convention and the forensic likelihood-ratio division of labour is an institutional practice. On vocab-travels it reads structural: the operative vocabulary — marginal likelihood, evidence ratio, automatic Occam penalty, multiplicative accumulation — is pure mathematics that travels intact to psychology, genetics, machine learning, astrophysics, and forensics, with only the Jeffreys bands and forensic framing as domain dress. On import-vs-recognize it reads strongly structural, its signature mark: the entry states the statistic "transfers literally wherever its single precondition holds... these are not analogies but the same computed quantity on a new substrate."

The portable structural skeleton is Bayesian updating — the prior-odds × evidence = posterior-odds structure in which evidence updates belief multiplicatively — of which the Bayes factor is the specific evidence-summary at the model-comparison step. That skeleton is fully substrate-spanning, but it is exactly what the Bayes factor instantiates from its umbrella prime bayesian_updating (with the classical likelihood ratio as its no-free-parameter special case, and AIC/BIC as approximating cousins) — not a force the named statistic carries beyond its precondition: the entry is explicit that where two well-defined models with marginal likelihoods are not in hand, "what genuinely carries is the parent bayesian_updating," not the Bayes factor. So the cross-domain reach belongs to the updating prime, while the domain-accented specifics — the marginal-likelihood integration, the automatic Occam penalty, the sequential-multiplicativity guarantee, the Jeffreys calibration scale, the forensic reporting division — are what make it this named statistic rather than the general prime. Its character: an evaluatively neutral, substrate-indifferent mathematical statistic genuinely recognised as the same computed quantity across quantitative fields, structural in skeleton yet a specific instance of its umbrella bayesian_updating and dressed in statistical reporting conventions — mixed-structural, and short of a prime only because it is the sharp model-comparison instance, not the general updating force.

Structural Core vs. Domain Accent¶

This section settles why the Bayes factor is a domain-specific abstraction and not a prime — and, since the entry is a piece of mathematics that transfers literally, the argument turns on the exact scope of that literal transfer versus the more general updating lesson beneath it.

What is skeletal (could lift toward a cross-domain prime). Strip the statistical apparatus and a thin relational structure survives: prior odds times a data-borne evidence term equals posterior odds — belief is revised multiplicatively, with the evidence term isolating what the observations alone contribute, held apart from how plausible the hypotheses were going in. That is Bayesian updating, and the multiplicative split (posterior odds = evidence × prior odds) is its bookkeeping identity. This skeleton is fully substrate-spanning — it recurs anywhere belief is revised by evidence — which is exactly why the Bayes factor is the model-comparison evidence-summary within the parent bayesian_updating. But it is the core the entry shares, not what makes the Bayes factor the distinct named statistic it is.

What is domain-bound. What makes the entry the Bayes factor in particular is statistical machinery that presupposes a specific precondition and does not travel without it. The evidence term is a ratio of marginal likelihoods — each model's fit integrated over its own parameter space against its own prior, \(P(D\mid H_i)=\int P(D\mid\theta,H_i)\,p(\theta\mid H_i)\,d\theta\) — not a bare likelihood at a best-fit value. From that integration flows the automatic Occam penalty (a diffuse prior debits a flexible model for spreading its bet, with no explicit complexity term); the sequential-multiplicativity guarantee that licenses optional stopping without inflating an error rate; the Jeffreys calibration scale (anecdotal/moderate/strong/decisive) as reporting convention; and the forensic likelihood-ratio division of institutional labor, examiner reporting the evidence while the court supplies the prior odds. The load-bearing precondition is two specific probability models each assigning a computable marginal likelihood to the same data. The decisive test: where there are not two well-defined models with marginal likelihoods — where "Bayes factor" is invoked loosely for any weighing of evidence — there is literally nothing to compute, the automatic Occam penalty and the multiplicative accumulation have no referent, and the term is doing rhetorical rather than statistical work.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose cross-domain transfer is recognition of the same mechanism, not analogy. The Bayes factor's transfer is bimodal, and the twist is that its literal reach is genuine but sharply bounded by its precondition. Wherever two probability models with marginal likelihoods exist — Bayesian model comparison in statistics, the gradable null-evidence of the psychology replication movement, variant–trait model ranking in genetics, marginal-likelihood model selection in machine learning (with the variational evidence lower bound as an explicit surrogate), cosmological-model comparison in astrophysics, the forensic likelihood ratio — it is the same computed quantity on a new substrate, recognition rather than metaphor. But that reach is exactly co-extensive with the precondition; it is the portability of an instrument to every field that meets its input requirement, not substrate-spanning structure. When the broader lesson is wanted — evidence updates belief multiplicatively — where no two models with marginal likelihoods are in hand, what carries is the parent bayesian_updating (of which the classical likelihood ratio is the no-free-parameter special case and AIC/BIC are approximating cousins), not the Bayes factor as named. So the genuinely cross-domain content is already carried in more general form by the updating prime, while everything distinctive to this entry — the marginal-likelihood integration, the automatic Occam penalty, the sequential-multiplicativity guarantee, the Jeffreys bands, the forensic reporting split — is domain accent that stays home. That is what keeps the Bayes factor below the prime bar even as it clears the domain-specific one decisively.

Relationships to Other Abstractions¶

Current abstraction Bayes Factor Domain-specific

Parents (1) — more general patterns this builds on

Bayes Factor is part of Bayesian Updating Prime

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

Not to Be Confused With¶

AIC / BIC (information criteria). Point-estimate model-selection scores that reward fit at the maximum-likelihood value and subtract an explicit penalty for parameter count (2k for AIC, k·ln n for BIC). They are approximating cousins, not the thing: BIC is in fact a large-sample Laplace approximation to −2 ln BF, but it fixes the flexibility penalty by a coarse formula in k rather than by the actual integration over the prior. The Bayes factor's Occam penalty is automatic and prior-dependent — it emerges from averaging fit over the whole parameter space, so it can differ from any fixed function of parameter count. Tell: is the complexity charge a fixed term in the number of parameters (AIC/BIC), or does it arise from integrating the likelihood over the model's own prior, so a wider prior debits the model more (Bayes factor)?
Frequentist likelihood-ratio test (Neyman–Pearson / Wilks). A classical procedure that forms a ratio of maximized likelihoods, then refers −2 ln Λ to a chi-square sampling distribution to produce a p-value and an accept/reject decision at a fixed level. It evaluates one hypothesis's fit at its best-fit parameters against a null's, with type-I error control as the guarantee. The Bayes factor instead integrates over parameters (never maximizing), carries no sampling distribution or error rate, and yields graded evidence that can favor either model. Tell: is the ratio taken at best-fit parameter values and calibrated against a null sampling distribution to control error (LRT), or is it a ratio of prior-averaged marginal likelihoods read directly as weight of evidence (Bayes factor)?
Marginal likelihood (model evidence). The ingredient, not the ratio: P(D | H_i), a single model's fit averaged over its parameter space, is one of the two quantities the Bayes factor divides. Part-versus-whole — a marginal likelihood alone is an unnormalized number with no interpretation until set against a rival's. Tell: are you looking at one model's averaged data-fit in isolation (marginal likelihood), or the quotient of two such quantities that isolates the evidence between them (Bayes factor)?
Cross-validation / predictive model comparison. An alternative model-selection paradigm that scores models by out-of-sample predictive accuracy (held-out log-loss, LOO, WAIC), estimating how well a fitted model predicts new data. It targets predictive performance and is comparatively insensitive to the parameter prior that the Bayes factor is acutely sensitive to (T1). Tell: is the model being judged by how it predicts held-out data under repeated splitting (cross-validation), or by the prior-weighted probability it assigned to the data actually observed (Bayes factor)? The two can disagree, and the disagreement is the prior's fingerprint.
Bayesian parameter estimation / credible intervals. A different Bayesian task: given one accepted model, infer the posterior distribution of its parameters and summarize it with a credible interval. This assumes the model and asks "what are its parameter values?"; the Bayes factor stands one level up and asks "which of two models?" — comparison, not within-model estimation. Tell: is the question the value (and uncertainty) of a parameter inside a fixed model (estimation), or the relative evidence for one whole model over another (Bayes factor)?
bayesian_updating (the parent prime). The substrate-spanning umbrella — prior odds × evidence = posterior odds, belief revised multiplicatively — that the Bayes factor instantiates at the model-comparison step, not a confusable peer. Where no two probability models with computable marginal likelihoods are in hand, "Bayes factor" is doing rhetorical work and the genuine content is the parent's. Tell: is there literally a ratio of two models' marginal likelihoods to compute (Bayes factor), or only the general lesson that evidence updates belief multiplicatively (bayesian_updating)? The parent — treated fully in the sections above — carries the cross-domain reach; this named statistic carries only where its precondition holds.

Neighborhood in Abstraction Space¶

Bayes Factor sits in a sparse region of the domain-specific corpus (62^nd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Statistical Inference & Model Failure Modes (16 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12