Likelihood-Ratio Reasoning¶
Evidence-weighting method — instantiates Bayesian Belief Updating
Updates beliefs by comparing how likely the evidence is under one possibility versus another.
Likelihood-Ratio Reasoning quantifies the weight of a single piece of evidence by asking one question: how much more (or less) likely is this observation if hypothesis A is true than if hypothesis B is true? That ratio — the likelihood ratio — is the entire contribution of the evidence; multiply the prior odds by it to get posterior odds. Its defining discipline, false of its twins, is that it is agnostic to where the prior came from and to what you will do with the answer: it owns only the weighing. And it insists that ratios may be chained across pieces of evidence only when those pieces are conditionally independent, because correlated evidence multiplied as if independent double-counts the same signal.
Example¶
A spam filter scores incoming mail. For each token it has learned the probability of seeing that token in spam versus in legitimate mail, giving a likelihood ratio per token: "invoice" ≈ 0.4 (more common in real mail), "wire-transfer" ≈ 12, "urgent" ≈ 3. The prior odds that any given message is spam are about 1:4. The filter multiplies the prior odds by the likelihood ratio of each observed token to get the posterior odds.
The independence check is the crux. "Free" and "money" tend to co-occur, so multiplying both ratios treats one theme as two independent signals and inflates the score. The filter must drop one or model the pair. The result: a message carrying a few genuinely independent strong tokens crosses the spam line, while a message whose "spamminess" is really one repeated correlated theme does not — even though a naive multiply would have flagged it.
How it works¶
- Estimate the contrast. For a hypothesis pair, estimate the probability of the evidence under each; their ratio is the weight of that evidence, meaningful only relative to the stated alternative.
- Update in odds form. Posterior odds = prior odds × likelihood ratio. On the log scale this becomes addition of weights of evidence, which makes weak evidence visibly weak.
- Check independence before chaining. For correlated evidence, combine or discount rather than multiply.
- Report the weight, not a verdict. Hand back the aggregate ratio (or its log) as the strength the evidence contributes — separate from the prior it multiplies.
Tuning parameters¶
- Likelihood estimation method — empirical frequencies vs. a fitted model, with smoothing for rare features.
- Contrast set — which alternative the ratio is taken against; a likelihood ratio is meaningless without a named alternative.
- Independence handling — assume, test, or explicitly model dependence; the naive assumption is fast but double-counts correlated signals.
- Scale — ratios vs. logs; logs turn products into sums and cap runaway multiplication visually.
- Chaining cap — a ceiling on how far accumulated evidence can move belief, to blunt cascades.
When it helps, and when it misleads¶
Its strength is that it isolates the one honest question about a piece of evidence — is it more expected under this or that? — and puts weak and strong evidence on a common, additive scale (I.J. Good's weight of evidence[n1]). Its failure mode is that the ratio only means something against the stated alternative, and chaining correlated evidence as if independent is the archetype's double-counting failure: a committee or model becomes overconfident because five correlated signals are counted as five confirmations. The classic misuse is the prosecutor's fallacy — confusing the probability of the evidence given innocence with the probability of innocence given the evidence. The guarding discipline is to always name the alternative, check independence before multiplying, and hand the prior and the action decision to other mechanisms.
How it implements the components¶
likelihood_model— it estimates how expected the evidence is under each competing hypothesis; that contrast is the mechanism.update_rule— the odds-form multiply (posterior odds = prior odds × likelihood ratio) is its combination rule.evidence_independence_check— it enforces conditional independence before chaining ratios and discounts correlated evidence.
It takes the prior odds as an input and does not establish or defend the starting rate — that is base_rate_reference in Bayesian Diagnosis, its nearest twin, which also owns the posttest posterior_estimate; Likelihood-Ratio Reasoning weighs evidence but names no hidden condition and reports no standalone probability. It sets no action cut-point (decision_threshold, Adaptive Decision Threshold) and does not vary the prior to test robustness (prior_sensitivity_check, Prior Sensitivity Analysis).
Related¶
- Instantiates: Bayesian Belief Updating — the evidence-weighting core: how strong is this observation, and don't double-count.
- Sibling mechanisms: Bayesian Diagnosis · Posterior Risk Estimation · Adaptive Decision Threshold · Prior Sensitivity Analysis · Sequential Forecast Update · Bayesian Model Update
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Likelihood-Ratio Reasoning operates as a computation, comparison, model, or analytic representation used to infer, estimate, or choose because it updates beliefs by comparing how likely the evidence is under one possibility versus another.
Independent corroboration: The frozen evidence defines Likelihood-Ratio Reasoning as 'Updates beliefs by comparing how likely the evidence is under one possibility versus another', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Universal
Rationale: Likelihood-ratio reasoning is a central statistical method for comparing how expected evidence is under rival hypotheses.
Related originating lineages:
- Mathematics — Probability theory supplies the formal ratio and odds algebra underlying Bayesian updating.
Review resolution: Both independent reviews assign primary provenance to statistics_experimental_design. The queued secondary differences (alternate_origin_disagreement) are reconciled by retaining mathematics only as formative or independently established lineage(s), not merely as application domains. origin_mode=single_lineage records the provenance relationship, while domain_reach=universal separately records applicability breadth. confidence=high preserves the more cautious assessment, and encyclopedia_synthesis=false records whether either reviewer identified a corpus-specific synthesis.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] Weight of evidence — I.J. Good's term for the logarithm of the likelihood ratio. Additive weights let independent pieces of evidence be summed, which is exactly why conditional independence must hold before you add them. ↩