Likelihood principle¶

The proposition that, for a fixed statistical model, all sample evidence about its parameters is contained in the observed-data likelihood up to proportionality.

Version: v2 · 2026-09-19 · History
Domain-specific #: 5328
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Statistical Foundations → Experimental Design & Statistics

Core Idea¶

Two experiments or outcomes with proportional likelihood functions carry the same evidential content about the parameter under the likelihood principle, irrespective of unobserved stopping or sampling possibilities. Conditioning on the observed sample produces a function of the parameter; proportional functions preserve every relative support ratio, so inference that depends on tail areas or stopping rules can conflict with the principle. The abstraction is therefore identified by a declared carrier, a transformation or constraint over that carrier, and an invariant that tells an analyst whether the named structure is genuinely present.

How would you explain it like I'm…

No faithful explanation at this level. A child-level version shrinks to 'same results, same answer'. That loses the real content: likelihood functions that are proportional over a fixed model's parameter, possibly from different experiments. It also wrongly turns a claim about evidential equivalence into a complete decision rule.

Same Story, Same Evidence

Imagine two friends testing whether a coin is fair. One planned to flip it 12 times. The other planned to keep flipping until getting 3 tails. Their results could end up giving exactly the same pattern of support for each possible amount of coin bias. The likelihood principle says that in that case their data carry the same evidence about the coin, even though their plans for stopping were different. What they might have seen, but didn't, doesn't change the evidence. It is a rule about what counts as the same evidence, not a full recipe for making decisions.

Proportional Likelihood, Equal Evidence

In statistics, a likelihood function tells you, for each possible value of an unknown parameter, how probable the observed data would be. The likelihood principle says that if two experiments or outcomes give likelihood functions that are proportional, meaning one is a constant multiple of the other with the constant not depending on the parameter, then they contain the same evidence about that parameter. This holds even if the experiments had different stopping rules or could have produced different unobserved outcomes. Proportional functions keep every ratio of support between parameter values the same. Methods that use tail areas or stopping rules, such as some p-values, can therefore conflict with the principle. It requires a fixed model and parameter, and it concerns evidential equivalence only, not how to make a complete decision.

Given a fixed statistical model and parameter θ, conditioning on the observed sample x yields the likelihood function L(θ; x). The likelihood principle holds that two experiments or outcomes whose likelihood functions are proportional—L₁(θ) = c·L₂(θ) with c independent of θ—carry identical evidential content about θ. Proportionality preserves every likelihood ratio L(θ₁)/L(θ₂), i.e., every relative support between parameter values. Hence unobserved sample-space possibilities and stopping or sampling rules are evidentially irrelevant. Inference procedures that rely on tail areas (p-values) or depend on the stopping rule can therefore violate the principle. Its scope conditions matter: the model and parameter must be fixed, the likelihoods must be based on observed data, the proportionality constant must not depend on the parameter, and the claim is about evidential equivalence, not a complete decision rule. Stretching it beyond those conditions is a misapplication.

Scope of Application¶

Likelihood principle belongs to statistical foundations and is useful where the analyst can specify the typed statistical foundations carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, then evaluate the model and parameter are fixed, likelihoods are based on the observed data, proportionality is parameter-independent, and the claim concerns evidential equivalence rather than a complete decision rule. The scope is broad within that domain but bounded by the need for the model and parameter are fixed, likelihoods are based on the observed data, proportionality is parameter-independent, and the claim concerns evidential equivalence rather than a complete decision rule. The entry records a descriptive analytical identity; practical use requires the governing domain's evidence, standards, and safety obligations.

Clarity¶

The abstraction clarifies a crowded vocabulary by making the model and parameter are fixed, likelihoods are based on the observed data, proportionality is parameter-independent, and the claim concerns evidential equivalence rather than a complete decision rule the center of the account. A claim should name the carrier, the governing operation or relation, the applicable assumptions, and the recognition test.

Manages Complexity¶

Without the abstraction, an analyst must reason directly over many local details: the carrier roles, admissibility assumptions, competing conventions, derived invariants, boundary cases, and proof or validation obligations specific to Likelihood principle. Likelihood principle compresses them into the roles in the structural signature. That compression permits comparison across instances without erasing the variables that determine validity. It also exposes which details may be varied safely and which are constitutive.

Abstract Reasoning¶

Identify the carrier. State what the elements, states, objects, or observations are: the typed statistical foundations carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets. Reject examples whose alleged carrier belongs to a different problem.
Lock the constitutive rule. Express the model and parameter are fixed, likelihoods are based on the observed data, proportionality is parameter-independent, and the claim concerns evidential equivalence rather than a complete decision rule independently of one notation or implementation.

Knowledge Transfer¶

Knowledge transfers strongly among subfields of statistical foundations because they reuse the typed statistical foundations carrier, defining objects and relations, parameters, conventions, evidence, boundary cases, and comparison targets, Conditioning on the observed sample produces a function of the parameter; proportional functions preserve every relative support ratio, so inference that depends on tail areas or stopping rules can conflict with the principle., and type the carrier, state every parameter and convention in the definition, test that the model and parameter are fixed, likelihoods are based on the observed data, proportionality is parameter-independent, and the claim concerns evidential equivalence rather than a complete decision rule, compare the nearest accepted identity, and report counterexamples, uncertainty, and limiting cases.

Relationships to Other Abstractions¶

Current abstraction Likelihood principle Domain-specific

Parents (1) — outgoing typed relations

Likelihood principle is a kind of Evidence Prime

What makes it its own entry: the domain-specific identity determined by the model and parameter are fixed, likelihoods are based on the observed data, proportionality is parameter-independent, and the claim concerns evidential equivalence rather than a complete decision rule.

Hierarchy paths (4) — routes to 4 parentless roots

Likelihood principle → is a kind of Evidence → includes Provenance → presupposes Traceability → presupposes Observability

Show alternative paths (3)

Neighborhood in Abstraction Space¶

Likelihood principle sits in a crowded region of the domain-specific corpus (8^th percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Statistical Inference & Hypothesis Testing (34 abstractions)

Nearest neighbors

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