Is demographic parity also called statistical parity?

Yes, that is a common alternative name.

Does it require ground-truth labels?

The binary selection-rate calculation uses predictions and group membership, not outcome labels.

Can it be used during training?

Yes. Some methods impose or approximate it as a constraint, but suitability and downstream effects still require assessment.

AI Glossary / Graphs and symbolic AI / Advanced

Demographic parity

Definition

Demographic parity is a statistical fairness criterion requiring a model's predictions to be independent of a specified sensitive attribute. For binary decisions, this means equal positive prediction rates across the compared groups. It measures one aspect of allocation; satisfying it does not establish that a system is fair in every relevant respect.

What it means

Suppose an AI system selects people for an optional training invitation. Demographic parity asks whether the proportion selected is the same in each specified group. It compares rates, not necessarily equal numbers of invitations, because group sizes may differ.
The criterion does not condition on the true outcome or a qualification label. That distinguishes it from measures such as equal opportunity and equalized odds, which compare predictions within outcome categories. Choosing among these criteria requires understanding the decision and potential harms.
A difference is evidence about selection patterns in the measured population. By itself, it does not identify why the difference arose or establish whether a particular decision was justified.

How it works

Define the positive outcome, population, time window, and group attribute. For each group, divide positive predictions by the number of evaluated people in that group. Exact binary demographic parity requires the resulting rates to match.
A common difference measure subtracts the lowest group rate from the highest. A common ratio divides the lowest by the highest, where that denominator is positive. Zero difference or a ratio of one indicates parity in the evaluated data.
Observed sample rates are estimates of population behavior. Report group counts and uncertainty where relevant. Aggregated parity can coexist with disparities within intersecting groups, so the grouping choice matters.

Three practical examples

Training invitations: An organization checks whether an invitation recommender selects comparable proportions of different age groups. The audit also examines whether staff had equal access to the application process and whether invitations led to useful participation.
Service access: A platform compares the proportions of language groups offered an optional support callback. Equal offer rates would not establish that callback quality, waiting time, or resolution rates were equal. Those require separate measurements.
Content exposure: A learning portal compares whether users from different regions receive a new-course recommendation at equal rates. The positive outcome must be defined clearly, such as receiving at least one recommendation during a week, rather than mixing users and individual impressions.

Think of it this way

Imagine comparing the fraction of tickets issued from two differently sized waiting lists. Equal fractions do not require equal ticket counts. They also do not tell you whether the waiting lists were accessible, whether selection within each list was reasonable, or whether the tickets provided the same experience.

Worked example

Consider invented training-invitation data. Group A contains 200 people, of whom 80 receive invitations. Group B contains 100 people, of whom 30 receive invitations.
The selection rates are 80/200 = 40% and 30/100 = 30%. The difference is 10 percentage points, and the minimum-to-maximum ratio is 0.30/0.40 = 0.75. The pooled rate is 110/300, approximately 36.7%, which does not reveal the group difference by itself.
Suppose a revised system invites 80 people from A and 40 from B. Both rates are now 40%, so observed demographic parity holds. The counts remain unequal because A is twice as large.
Now add an invented reference label indicating a relevant training need. In each group, exactly half the people have that label: 100 in A and 50 in B. Suppose the revised system invites 60 of those 100 people in A but only 15 of those 50 people in B. The true positive rates are then 60% and 30%, despite equal overall selection rates.
Thus, demographic parity does not imply equal opportunity. The reference label would itself need scrutiny in a real audit; here it is stipulated solely to demonstrate the mathematical distinction.

Illustrative case study

Fictional Birch Learning reviews an invitation model after staff question access to development activities. Its first dashboard shows only total invitation volume. The team adds group denominators and discovers different selection rates.
Reviewers investigate how availability, historical attendance, and manager nominations enter the process. They also examine whether past attendance reflects access barriers rather than interest. A proposed revision narrows the observed selection gap, but the team separately checks invitation relevance and participation outcomes.
The resulting assessment distinguishes allocation rates from the quality and consequences of selection. It does not treat one improved number as proof that every concern is resolved. This scenario illustrates an evaluation approach; it reports no measured results from a real organization.

Common misconceptions and mistakes

“Parity means equal numbers.” It means equal rates in the binary setting.
“Removing the sensitive attribute guarantees parity.” Other features can carry correlated information.
“Equal selection rates mean equal error rates.” The worked example shows otherwise.
“A ratio above an arbitrary threshold proves fairness.” A tolerance needs justification and does not establish overall fairness.
“A measured disparity explains its cause.” Diagnosis requires additional evidence about data, decisions, and context.

Business use

Begin with the allocation decision and the people affected. Define who enters the denominator, including whether excluded or unreachable people are missing from the audit. Document which groups are compared and why.
Use demographic parity alongside relevant quality and outcome measures. Examine intersections when data support meaningful analysis, and avoid strong conclusions from unstable small samples. Track changes over time and investigate mechanisms behind gaps. Treat any proposed correction as a design decision whose effects need evaluation, rather than assuming that equalizing one statistic resolves the whole problem.

Check your understanding

Why do 80 invitations and 40 invitations satisfy parity in the revised example?

Answer: Both represent 40% of their respective groups.

What does the initial ratio of 0.75 compare?

Answer: The lower selection rate with the higher one.

What remains unequal in the revised example?

Answer: Selection among people with the stipulated training-need label.

Related terms

Statistical parity
Sensitive attribute
Equalized odds
Fairness metric
selection rate
equal opportunity
intersectional fairness
allocation harm

Learn next

Calculate selection rates for a small invented dataset with unequal group sizes. Then add outcome labels and compare true positive rates. Explain why the two comparisons answer different questions before choosing an evaluation criterion.

True positive

Sources

Google, Machine Learning Glossary, demographic parity: https://developers.google.com/machine-learning/glossary/responsible-ai#demographic-parity Fairlearn, Common fairness metrics: https://fairlearn.org/main/user_guide/assessment/common_fairness_metrics.html Agarwal and colleagues, A Reductions Approach to Fair Classification, research abstract: https://proceedings.mlr.press/v80/agarwal18a.html

Previous: Deming regression Next: Demonic nondeterminism