What is Demographic Parity?
Demographic Parity is a group fairness criterion requiring a model prediction or downstream decision to be statistically independent of protected-group membership, commonly assessed by comparing favorable-selection rates across groups.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Measure selection on a declared eligible population
For each group, divide favorable decisions by eligible cases, not by whichever records happen to have labels. Compare rates using a difference such as max(rate)-min(rate) and, where the reference rate is nonzero, a ratio. The Fairlearn metric guide distinguishes demographic parity difference and ratio and makes the grouping function part of the metric contract.
Use the criterion when allocation itself is the concern
Parity can be relevant when access to interviews, outreach, exposure, or another scarce opportunity is itself the object of review, or when outcome labels are structurally unavailable. Dwork and coauthors distinguish statistical parity from individual-fairness reasoning. The metric still requires a normative decision about which population is eligible and which outcome counts as beneficial.
Do not infer equal error or equal welfare
Equal selection rates can coexist with unequal false rejections, false approvals, utility, or burden when labels, base rates, measurement, and intervention effects differ. Conversely, a justified program targeted at a disadvantaged group may intentionally produce unequal rates. Review label quality, error-conditioned criteria, calibration, benefit and harm magnitude, uncertainty, and the downstream process before changing a threshold to force parity.
Key Characteristics
- Compares favorable-selection rates across declared protected groups
- Imposes statistical independence between the decision and group attribute
- Does not condition on observed outcome labels
- Can be summarized with absolute gaps and reference-relative ratios
- Requires explicit eligibility, decision unit, threshold, window, and missingness rules
- Does not by itself establish equal errors, equal utility, or legal compliance
Common Use Cases
- Auditing access to interviews among otherwise eligible applicants
- Comparing exposure allocation in recommendation or advertising systems
- Monitoring approval rates before delayed outcome labels are complete
- Checking intersectional selection patterns after a policy threshold change
- Evaluating whether a limited outreach opportunity is distributed as intended
Example
Loading code...Frequently Asked Questions
How is Demographic Parity calculated?
Calculate the favorable-selection rate for every declared group on the same eligible population and time window. Report numerators, denominators, absolute rate differences, ratios when their denominator is nonzero, and uncertainty. Do not compare rates built from different eligibility or missing-data rules.
Is Demographic Parity the same as equal accuracy?
No. Demographic Parity ignores the outcome label and compares how often a favorable decision is issued. Equal accuracy compares correctness, while Equalized Odds compares true-positive and false-positive rates conditional on labels. Selection rates can match even when errors differ substantially.
Does the 80 percent rule define fairness?
No. A four-fifths ratio may be used as a screening convention in particular legal or organizational contexts, but it is not a universal fairness threshold or proof of compliance. Sample size, uncertainty, job or product context, applicable law, and other evidence still require review.
When is Demographic Parity a poor fit?
It can be misleading when the primary harm concerns false positives or false negatives, when groups have different valid eligibility definitions, when labels are trustworthy and error balance matters, or when a remedial program intentionally targets unequal access. The decision context should select the criterion.
Can matching selection rates make outcomes worse?
Yes. Threshold changes can shift errors, withhold beneficial interventions, encourage gaming, or mask poor data and process design. Evaluate who gains and loses, label quality, error rates, utility, uncertainty, and downstream effects rather than optimizing the parity statistic in isolation.