Equalized odds¶

A classifier fairness criterion requiring protected groups to have equal true-positive and false-positive rates conditional on the actual outcome.

Version: v2 · 2026-09-06 · History
Domain-specific #: 1775
Origin domain: machine learning
Subdomain: group fairness for supervised prediction
Aliases: Equalized odds criterion

Core Idea¶

Equalized odds is a classifier fairness criterion requiring protected groups to have equal true-positive and false-positive rates conditional on the actual outcome. ^[1]

A predictor satisfies equalized odds with respect to protected attribute A and outcome Y when its prediction is conditionally independent of A given Y. In binary classification, this requires equal true-positive rates and equal false-positive rates across protected groups, or equivalently equal prediction distributions within each actual-outcome stratum.

Its operative boundary is not supplied by the name alone. Preserve this identity: A classifier fairness criterion requiring protected groups to have equal true-positive and false-positive rates conditional on the actual outcome. Validity boundary: Both true-positive and false-positive rates must be equal across groups conditional on ground truth; demographic parity alone does not satisfy the criterion. The entry therefore captures a reusable specialist role structure rather than a topic label, a single historical instance, or a loose analogy.

Structural Signature¶

Sig role-phrases:

the protected attribute — the group membership across which rates are compared
the ground-truth outcome — the actual label on which comparisons are conditioned
the predictor — a score or decision derived from available features
the group-conditioned true-positive rate — positive predictions among actual positives in each group
the group-conditioned false-positive rate — positive predictions among actual negatives in each group
the conditional-independence constraint — prediction independent of group given outcome
the performance frontier — accuracy–fairness tradeoff under feasible post-processing
the evaluation sample — data supporting uncertainty-aware estimates of each conditional rate

Recognition test. A case qualifies only when the analyst can map the declared the protected attribute, the ground-truth outcome, the predictor, the group-conditioned true-positive rate, the group-conditioned false-positive rate and preserve the specialist validity conditions. Shared vocabulary, a similar output, or a generic instance of one parent relation is insufficient.

What It Is Not¶

Not demographic parity. Demographic parity equalizes positive prediction rates without conditioning on Y.
Not equal opportunity alone. Equal opportunity commonly requires only equal true-positive rates.
Not individual fairness. Equalized odds is a group-conditional rate criterion.
Not proof of justice in all respects. It formalizes one contested fairness condition and can conflict with calibration.
Not valid measurement without trustworthy labels. Biased or noisy Y changes the meaning of conditional rates.

Scope of Application¶

The abstraction recurs literally within binary and multiclass supervised decisions where protected-group error rates are audited or constrained. The following habitats preserve the same recognition machinery; they are not invitations to extend the name metaphorically.

Lending. approval errors are compared conditional on repayment outcome.
Medical prediction. sensitivity and false-positive rates are compared across patient groups.
Risk assessment. error-rate disparities are measured by outcome strata.
Post-processing. randomized thresholds move a score onto an equalized-odds frontier.
Model selection. accuracy and group-conditional error constraints are optimized jointly.

Clarity¶

Declare A, Y, the prediction threshold, positive class, population, and whether equality is exact or tolerance-bounded. Report uncertainty for every group–outcome cell. Equalized odds can be technically satisfied against labels that encode unequal treatment, so label validity and decision consequences remain separate normative questions.

A practical identification audit begins with the typed roles rather than the title: establish the protected attribute, verify the ground-truth outcome, then test the remaining conditions and exclusions. If the case retains only the portable skeleton described below, it should be named through a parent abstraction rather than as Equalized odds.

Manages Complexity¶

The criterion decomposes disparity into false-negative and false-positive treatment conditional on actual status. ROC geometry makes feasible post-processing explicit and reveals the performance cost of equalizing rates.

The compression remains accountable because each simplification has a named failure condition. Disagreement can be localized to a missing role, an invalid assumption, an ambiguous measurement, or a neighboring abstraction instead of being hidden inside an unanalyzed label.

Abstract Reasoning¶

R1. Define protected groups, outcome labels, prediction, and decision population. R2. Estimate true- and false-positive rates within every adequately sampled group. R3. Test both rate equalities with uncertainty or a declared tolerance. R4. If post-processing, derive group-specific randomized thresholds from the validation ROC curves. R5. Audit label quality, calibration, intersectional groups, and downstream consequences beyond the criterion.

These moves separate definition, derivation, measurement, and interpretation. A formal consequence does not by itself prove that an observed case instantiates the abstraction, while an observed resemblance does not relax the formal or institutional recognition conditions.

Knowledge Transfer¶

The criterion transfers literally across supervised predictions with protected group, outcome, and both conditional error rates. Fairness and constraint are parents; equal acceptance rates or equal accuracy alone do not satisfy equalized odds.

The transfer boundary is explicit: DOMAIN-SPECIFIC PASS / PRIME FAIL: The criterion is applied across classifiers, protected attributes, and binary outcomes by comparing group-conditioned error rates. Literal recognition retains the specialist vocabulary and validity conditions of machine-learning fairness; outside that setting only broader parent operations transfer. The safe move beyond the home habitat is to carry the applicable parent relation and leave the specialist name behind unless every defining role remains literal.

Examples¶

Canonical: post-processing a risk score¶

For each protected group, choose a possibly randomized threshold so the resulting classifier has a common TPR and FPR on validation data. The selected point lies in the intersection of feasible ROC regions and trades some accuracy for the two rate equalities. ^[1]

Mapped back: the protected attribute; the ground-truth outcome; the predictor; the group-conditioned true-positive rate; the group-conditioned false-positive rate; the performance frontier.

Applied / In Practice: auditing a diagnostic model¶

Estimate sensitivity and false-positive rate separately for each patient group, with confidence intervals and sample counts. Equal sensitivity alone establishes equal opportunity, not equalized odds; the negative-outcome stratum must also match. ^[2]

Mapped back: the evaluation sample; the group-conditioned true-positive rate; the group-conditioned false-positive rate; the conditional-independence constraint.

Structural Tensions¶

T1: Equalized errors vs calibration. When base rates differ, nonperfect predictors generally cannot satisfy both criteria simultaneously. Diagnostic: Which property serves the decision context?

T2: Observed label vs morally relevant outcome. Y may encode historical access or enforcement disparities. Diagnostic: Is the conditioning label valid?

T3: Group equality vs individual treatment. Rates can match while similar individuals receive different randomized outcomes. Diagnostic: Are individual-level constraints also required?

T4: Exact equality vs sampling noise. Finite groups make exact empirical matching unstable. Diagnostic: What tolerance and uncertainty rule is used?

T5: Single attributes vs intersections. Parity on broad groups can conceal disparities at intersections. Diagnostic: Which subgroup resolution is adequately powered?

T6: Domain autonomy vs prime reduction. Fairness and Constraint omit the specialist objects, constraints, and validity tests named above. Diagnostic: Would retaining only the portable parent pattern still satisfy the recognition test?

Structural–Framed Character¶

The five-criterion aggregate is 0.90 (framed). The judgment is criterion-specific:

Vocabulary travels — high (0.75). The complete vocabulary remains tied to the typed roles in the Structural Signature.
Evaluative weight — high (1.00). Application carries the stated degree of normative or interpretive judgment beyond structural recognition.
Institutional origin — high (1.00). The abstraction depends to this degree on a scholarly, technical, legal, or social convention.
Human-practice bound — high (1.00). Recognition depends to this degree on organized practice, language, measurement, or institutional action.
Import versus recognize — high (0.75). Beyond its home habitat, use of the full name increasingly becomes analogy rather than literal recognition.

The portable skeleton is a decision rule is constrained so its error behavior is independent of group membership within each true-outcome stratum. The named abstraction remains framed because that skeleton alone does not supply its specialist objects, constraints, or tests.

Structural Core vs. Domain Accent¶

Structural core: A decision rule is constrained so its error behavior is independent of group membership within each true-outcome stratum.

Domain accent: Protected attributes, ground-truth labels, true-positive and false-positive rates, roc curves, randomized thresholds, and fairness tradeoffs.

Why it does not clear the prime bar: Fairness and constraints travel; equalized odds is the exact outcome-conditioned error-rate criterion. Generalization therefore routes through parent abstractions; preserving the specialist name requires the full accent.

Fairness (prime:fairness). The criterion expresses one group-comparative allocation of classification errors.
Constraint (prime:constraint). Two conditional rate equalities restrict the feasible predictor set.

These are prose placement proposals only. They create no dag_edges; endpoint, redundancy, and cycle checks are recorded separately in the bundle's placement memo.

Relationships to Other Abstractions¶

Current abstraction Equalized odds Domain-specific

Parents (2) — more general patterns this builds on

Equalized odds is a kind of Constraint Prime

Constraint (prime:constraint).
Equalized odds is a kind of Fairness Prime

Fairness (prime:fairness).

Hierarchy paths (2) — routes to 2 parentless roots

Equalized odds → Constraint

Show alternative path (1)

Neighborhood in Abstraction Space¶

Equalized odds sits in a sparse region of the domain-specific corpus (76^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Statistical Paradoxes & Model Reliability (20 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Demographic parity. equal positive prediction rates across groups. Tell: Is comparison conditioned on actual outcome?
Equal opportunity. equal true-positive rates only. Tell: Are false-positive rates also equal?
Predictive parity. equal positive predictive value. Tell: Is conditioning on prediction or outcome?
Calibration. outcome frequency matches score within groups. Tell: Are scores or binary error rates constrained?
Overall accuracy parity. equal total error or accuracy. Tell: Can opposite error types offset each other?

References¶

[1] Moritz Hardt, Eric Price, and Nati Srebro, “Equality of Opportunity in Supervised Learning”, NeurIPS 29 (2016). registry ↩a ↩b

[2] Solon Barocas, Moritz Hardt, and Arvind Narayanan, Fairness and Machine Learning, 2023. registry ↩