Category: Fairness & Bias

Equalized Odds

Also known as: Conditional procedure accuracy equality, Disparate mistreatment

Simply put

Equalized odds is a way of checking whether a machine learning model treats different groups of people fairly by comparing how often the model makes correct and incorrect predictions for each group. The idea is that the model's error rates should be similar across protected groups (for example, groups defined by attributes such as race or gender), rather than being much worse for one group than another. It is one of several competing fairness measures, and satisfying it does not guarantee that a model is fair in every other sense.

Formal definition

Equalized odds, also referred to as conditional procedure accuracy equality and disparate mistreatment, is a statistical (group-level) fairness criterion for classification. As commonly defined, it requires that a model's error and success rates—typically its true positive and false positive rates, and correspondingly its false negative and true negative rates—be equal across protected groups. Because it conditions on the true outcome, it constrains the model's performance to be comparably accurate across groups, distinguishing it from criteria such as demographic parity that condition only on the prediction. It is one of multiple, often mutually incompatible, statistical fairness notions, so its applicability and appropriateness are context-dependent and it does not by itself establish that a system is non-discriminatory across all definitions of fairness.

Why it matters

Equalized odds matters because it directly addresses a concern that is often invisible in aggregate model performance metrics: a model can appear accurate overall while making its mistakes disproportionately at the expense of a particular protected group. By conditioning on the true outcome and requiring comparable error rates—typically true positive and false positive rates—across groups, equalized odds targets whether the burden of a model's errors is distributed equitably. This makes it a useful diagnostic in settings where the consequences of false positives and false negatives fall differently on different populations, such as lending, hiring, or clinical decision support.

At the same time, equalized odds is only one of several competing statistical fairness criteria, and these notions are often mutually incompatible—a model cannot generally satisfy all of them simultaneously except in special cases. As a result, choosing equalized odds is itself a normative decision about which kind of fairness a system should prioritize, and satisfying it does not establish that a system is non-discriminatory under other definitions or under applicable legal standards. Practitioners should treat it as evidence about a specific, well-defined property rather than as a certificate of overall fairness.

For governance and model risk purposes, the value of equalized odds lies in making a particular fairness trade-off explicit and measurable, so that it can be documented, reviewed, and challenged. It does not eliminate bias-related risk; it helps surface and manage one dimension of it. Its appropriateness is context-dependent, and its selection over alternative criteria such as demographic parity should be justified relative to the specific decision context and the harms at stake.

Who it's relevant to

Data scientists and ML practitioners

Those building and evaluating classification models use equalized odds as one measurable fairness diagnostic, comparing true positive, false positive, false negative, and true negative rates across protected groups. They should be aware that it is one of multiple, often incompatible, fairness notions and that satisfying it does not guarantee fairness under other definitions.

Model validators and model risk managers

In model risk contexts, equalized odds can serve as a documented, testable property when assessing whether a model's errors fall disproportionately on particular groups. Validators should scrutinize why this criterion was selected over alternatives and confirm that its use is justified for the specific decision context, since it measures one dimension of bias-related risk rather than eliminating it.

Compliance and fairness governance functions

Teams responsible for AI governance and fairness oversight may use equalized odds as part of a broader assessment framework. Because it is a statistical notion tied to a specific normative choice and is not equivalent to legal non-discrimination, it should be treated as evidence about a defined property, not as a standalone certification of fairness.

Auditors and reviewers

Those reviewing model documentation can use equalized odds results to evaluate whether disparities in error rates across groups have been measured and disclosed. Reviewers should note the contested and context-dependent nature of fairness metrics and confirm that the choice of criterion is explained rather than assumed.

Inside Equalized Odds

Equal true positive rates

A component requiring that, across the groups being compared, individuals who actually belong to the positive class are correctly identified as positive at approximately the same rate. This condition is often described as satisfying 'equal opportunity' when considered on its own.

Equal false positive rates

A component requiring that individuals who actually belong to the negative class are incorrectly classified as positive at approximately the same rate across groups. Equalized odds, as commonly defined, requires this condition in addition to equal true positive rates.

Conditioning on the true outcome

Equalized odds is typically defined as a condition on a model's error rates given the actual ground-truth label, rather than on predictions alone. This distinguishes it from group fairness criteria that only compare overall selection or prediction rates without reference to the true outcome.

Protected or sensitive attribute

The group-defining variable (for example, a demographic characteristic) across which the rate parities are assessed. The definition presumes such groups are identified for the comparison; how these attributes are chosen and whether they are legally permissible to use varies by jurisdiction and context.

Binary classification framing

Equalized odds is most commonly formulated for binary classification tasks, where true positive and false positive rates are well defined. Extensions to multi-class or continuous outputs exist but are less standardized and are generally out of scope for the core definition.

Common questions

Answers to the questions practitioners most commonly ask about Equalized Odds.

Does satisfying equalized odds mean a model is fair overall?

No. Equalized odds is one specific group-fairness criterion that concerns whether error rates are equalized across groups; it does not capture all conceptions of fairness. Equalized odds is a technical parity condition, whereas fairness is a broader normative concept that depends on context, legal standards, and stakeholder values. A model can satisfy equalized odds and still be considered unfair under other criteria, and vice versa. Professionals should avoid treating any single mathematical criterion as equivalent to fairness.

Is equalized odds the same as demographic parity or other group-fairness metrics?

No. These are distinct criteria that are often mutually incompatible. As commonly defined, equalized odds requires that true positive rates and false positive rates be equal across groups (conditioning on the actual outcome), whereas demographic parity requires that positive prediction rates be equal across groups regardless of the actual outcome. Research has shown that different group-fairness definitions typically cannot all be satisfied simultaneously except in special cases, so selecting one criterion generally involves trade-offs against others. Treating them as interchangeable is a frequent error.

What information do I need to evaluate equalized odds, and when can it be measured?

Evaluating equalized odds typically requires access to the protected group attribute, the model's predictions, and the ground-truth outcomes for each group, because the criterion conditions on the actual outcome. This means it is generally assessed where reliable outcome labels are available, often in validation or post-deployment monitoring settings. Where the protected attribute is unavailable, unreliable, or restricted by law or policy, measuring equalized odds directly may not be feasible, and this limitation should be documented.

How can equalized odds be enforced or improved in practice?

In many treatments, interventions are grouped into pre-processing (adjusting training data), in-processing (adding constraints or penalties during model training), and post-processing (adjusting decision thresholds or outputs by group after training). Each approach has trade-offs affecting accuracy, other fairness criteria, and operational complexity. The appropriate choice depends on context, and any group-specific adjustment may itself raise legal and policy questions in some jurisdictions, so such decisions should be reviewed with legal and compliance stakeholders.

How does equalized odds relate to model performance and validation?

Equalized odds concerns the distribution of specific error types across groups, which is distinct from overall model performance. A model with strong aggregate performance can still exhibit disparate true positive or false positive rates across groups. In many model risk management practices, disaggregated error-rate analysis across relevant groups can be incorporated into validation and ongoing monitoring, but it should be treated as one component alongside broader performance, stability, and control assessments rather than a substitute for them.

Should equalized odds be documented in model governance artifacts, and what should be recorded?

Where equalized odds is used as an evaluation criterion, it is generally useful to document the definition applied, the groups analyzed, the data and labels used, the measured disparities, any trade-offs against other criteria, and any interventions taken. Documenting why a particular criterion was chosen over alternatives, and its limitations, supports transparency and review. The specific documentation expectations depend on the applicable framework, sector, and jurisdiction, and vary across organizations.

Common misconceptions

Equalized odds is the single correct or universal definition of fairness.

Equalized odds is one of several group fairness criteria, and it can conflict mathematically with others such as demographic parity or calibration. As commonly noted in the fairness literature, satisfying multiple such criteria simultaneously is generally not possible except in restricted cases, so the appropriate criterion depends on context and cannot be treated as authoritative across all settings.

Achieving equalized odds means the model is unbiased or fair overall.

Equalized odds addresses a specific statistical parity in error rates conditioned on the true outcome; it does not by itself guarantee the absence of bias in the underlying data, the appropriateness of the ground-truth labels, or fairness in a broader legal or ethical sense. Bias and fairness are distinct concepts, and this metric speaks to only one narrow operationalization.

Equalized odds is the same as equal opportunity.

Equal opportunity, as commonly defined, requires only equal true positive rates across groups, whereas equalized odds additionally requires equal false positive rates. The two are related but not interchangeable, and conflating them can lead practitioners to under-specify the parity condition they intend to enforce.

Best practices

State explicitly which fairness criterion you are applying and why, since equalized odds is one option among several that can conflict with each other and is not universally required.

Report both true positive and false positive rate parities across groups, rather than only one, to avoid mistakenly satisfying the weaker equal-opportunity condition while claiming equalized odds.

Document the choice of protected or sensitive attributes and confirm, in consultation with legal or compliance functions, whether their use is permissible in the relevant jurisdiction and context.

Assess the quality and provenance of the ground-truth labels, because equalized odds is defined conditional on the true outcome and inherits any bias present in those labels.

Treat equalized odds as one measure that helps evaluate and manage a specific form of disparity, not as evidence that the model is free of bias or that broader fairness obligations are satisfied.

Note the scope limits of your assessment, including whether the metric was computed for a binary task and whether any extension to multi-class or continuous outputs was used, since such extensions are less standardized.