Is statistical parity another name for demographic parity?

In common binary-classification usage, yes.

Does it require equal accuracy across groups?

No. It concerns selection rates rather than accuracy or conditional error rates.

Does achieving parity prove fairness?

No. Fairness involves context, competing concerns, and consequences that one metric cannot fully represent.

AI Glossary / Statistics and probability / Intermediate

Statistical parity

Definition

Statistical parity is a group-level criterion requiring a model's predictions or decisions to be independent of membership in a specified group. For binary decisions, it means equal positive-selection rates across the groups being compared. It is commonly called demographic parity. Statistical parity measures one aspect of how outcomes are allocated; it does not by itself establish that a system is fair, accurate, lawful, or appropriate for its purpose. Its relevance must be considered in the social and operational context.

What it means

If a model selects 40% of one group and 40% of another, its observed selection rates satisfy parity for that comparison. Equal rates do not require equal numbers selected when the groups have different sizes.

The criterion concerns the overall decision rate, without conditioning on an outcome label. It therefore answers a different question from measures comparing error rates among people with the same observed outcome. Neither approach is universally sufficient. The choice of groups, decision, population, and denominator determines what the metric actually describes, while sampling uncertainty affects conclusions drawn from finite data.

How it works

Define the decision being assessed and which value counts as positive. Specify the population, period, group attributes, and treatment of missing group information. For each group, calculate the number receiving a positive decision divided by the number eligible for that decision.

Compare rates using a clearly defined difference or ratio. A maximum-minus-minimum difference of zero indicates observed parity across the included groups. A minimum-to-maximum ratio of one does likewise when the denominator is positive. Signed differences against a reference group use a different convention and should be labeled accordingly. Ratios can be unstable or undefined when rates are very small or zero.

Assess uncertainty and examine relevant intersections of groups and stages of the decision process. A coarse aggregate can hide disparities within subgroups. Review errors, data quality, access to the process, and real-world consequences alongside allocation rates. If parity is used as a design constraint, evaluate the resulting tradeoffs rather than assuming a numerical target automatically resolves the underlying fairness question.

Three practical examples

1. A training program compares the proportions of eligible applicants offered places across specified demographic groups.

2. A digital service examines whether a recommendation system distributes access to a beneficial feature at different rates across user groups.

3. An organization reviews an automated screening stage and then checks whether later human decisions change the group-level allocation pattern.

Think of it this way

Think of comparing how often different queues receive invitations. Equal invitation rates answer a question about allocation. They do not tell you whether the invitations went to the right people, whether some people could not join the queues, or whether the event itself meets their needs.

Worked example

Consider an original fictional selection dataset. Group A contains 100 eligible people, of whom 40 are selected. Group B contains 200 eligible people, of whom 80 are selected. Both selection rates are 0.40, so the observed maximum-minus-minimum parity difference is zero and the minimum-to-maximum ratio is one.

Now suppose an independently assessed binary outcome label is available for illustration. In A, 60 people have a positive label and 30 of them are selected. In B, 100 have a positive label and 70 of them are selected. The true-positive rates are therefore 30/60 = 0.50 and 70/100 = 0.70.

The remaining selected people number 10 in each group. There are 40 negative-label people in A and 100 in B, giving false-positive rates of 10/40 = 0.25 and 10/100 = 0.10. Thus equal selection rates coexist with unequal true-positive and false-positive rates.

This constructed example shows that statistical parity does not imply equalized odds. It does not establish which metric should govern a real decision, and the outcome labels themselves would need scrutiny. A label may reflect historical access or measurement problems rather than an unquestionable definition of merit or need.

Illustrative case study

In a fictional skills-development service, an allocation dashboard shows equal offer rates across broad groups. The team initially describes this as a complete fairness assessment.

Further review finds differences in who completes the application process and in which subgroups receive offers within each broad category. Participants also report that some offered courses are inaccessible because of scheduling. The service keeps parity as one useful allocation measure while adding analysis of access, subgroup outcomes, and practical usefulness. It avoids treating a single balanced percentage as evidence that every relevant harm has been addressed.

Common misconceptions and mistakes

Equal rates are not equal counts, and parity at one stage does not imply parity throughout a pipeline. Removing group attributes from model inputs does not guarantee parity because other variables can act as proxies.

Do not interpret observed equality in a small sample as proof of population equality. Conversely, a measured difference needs contextual and uncertainty analysis rather than an automatic conclusion about its cause. Parity also differs from calibration and conditional error-rate criteria, which can conflict in some settings. A numerical threshold borrowed from a particular policy context is not a universal fairness or legal-compliance certificate.

Business use

Statistical parity can help businesses investigate allocation patterns in automated or human-supported decisions. Its value lies in making one question measurable and prompting deeper review. Define the metric with affected stakeholders and domain experts, document the reasons for choosing it, and monitor changes after deployment. Report counts, rates, uncertainty, and complementary measures so readers can understand both the finding and its limits.

Check your understanding

Group A has 20 selections among 50 people, and B has 60 among 150. Are their observed selection rates equal?

Answer: Yes. Both are 40%, despite different selection counts. This establishes observed parity for that defined comparison only.

Related terms

Demographic parity
Group fairness
Equalized odds
Calibration
selection rate
equal opportunity
intersectional analysis
allocation harm

Learn next

Learn confusion matrices and group-specific rates, then compare statistical parity with conditional error-rate criteria. Study how metric choice relates to the actual decision and affected people.

Error rate

Sources

Fairlearn, Common fairness metrics, Demographic parity and Equalized odds: https://fairlearn.org/main/user_guide/assessment/common_fairness_metrics.html

The primary project documentation was read. The entry uses the statistical criterion rather than offering a legal assessment. All selection counts, calculations, and the fictional case are original teaching examples.

Previous: Statistical MT Next: Statistical process control