demographic parity
**Demographic Parity** is the **fairness constraint requiring that an AI model's positive prediction rate be equal across all demographic groups** — one of the foundational fairness metrics in algorithmic decision-making, though its apparent simplicity conceals deep tensions with merit-based selection and legal frameworks.
**What Is Demographic Parity?**
- **Definition**: A model satisfies demographic parity (also called statistical parity) when P(Ŷ=1 | Group=A) = P(Ŷ=1 | Group=B) — the probability of a positive outcome is identical regardless of protected group membership.
- **Also Known As**: Statistical parity, group fairness, equal acceptance rate.
- **Example**: In a hiring model, if 40% of male applicants receive interview offers, demographic parity requires that exactly 40% of female applicants also receive offers — regardless of qualification distribution.
- **Scope**: Applies to binary and multi-class classifiers in hiring, lending, admissions, criminal risk assessment, and content recommendation.
**Why Demographic Parity Matters**
- **Discrimination Detection**: Provides a simple, auditable metric that regulators and civil rights organizations can use to detect discriminatory outcomes in automated systems.
- **Historical Redress**: In domains where historical bias has systematically excluded groups (e.g., redlining in mortgage lending), demographic parity enforces corrective equal representation.
- **Legal Context**: The "four-fifths rule" in U.S. EEOC employment law requires that selection rates for protected groups not fall below 80% of the highest-rate group — a softer version of demographic parity.
- **Auditability**: Unlike accuracy-based metrics, demographic parity can be verified from outcomes alone without knowing ground-truth labels — useful for external audits.
**Mathematical Formulation**
For a classifier with prediction Ŷ and sensitive attribute A:
Demographic Parity: P(Ŷ=1 | A=0) = P(Ŷ=1 | A=1)
Relaxed version (ε-demographic parity): |P(Ŷ=1 | A=0) - P(Ŷ=1 | A=1)| ≤ ε
Disparate Impact Ratio: P(Ŷ=1 | A=1) / P(Ŷ=1 | A=0) ≥ 0.8 (EEOC four-fifths rule)
**Critiques and Limitations**
- **Qualification Blindness**: Demographic parity ignores whether prediction errors are distributed fairly. A model could satisfy demographic parity while systematically rejecting qualified minority candidates and accepting unqualified majority candidates.
- **The Impossible Trinity**: Chouldechova (2017) and Kleinberg et al. (2017) proved that demographic parity, equalized odds, and calibration cannot all be satisfied simultaneously when base rates differ across groups — forcing a choice of which fairness notion to prioritize.
- **Data Feedback Loops**: Enforcing demographic parity on a biased dataset can entrench bias. If historical hiring data reflects discrimination, training a "fair" model on it propagates the discrimination through a mathematical proxy.
- **Legal Complexity**: In some jurisdictions, mechanically enforcing demographic parity constitutes illegal quota-setting or affirmative action beyond what law permits.
- **Intersectionality**: Demographic parity across a single protected attribute (gender) can mask severe disparities across intersecting attributes (Black women vs. White men).
**Fairness Metrics Comparison**
| Metric | What It Equalizes | Ignores | Best For |
|--------|------------------|---------|----------|
| Demographic Parity | Positive rate | Qualifications, error rates | When outcomes should reflect population |
| Equalized Odds | TPR and FPR | Acceptance rates | When accuracy parity matters |
| Calibration | Score → probability accuracy | Group outcome rates | When risk scores drive decisions |
| Individual Fairness | Similar individuals treated similarly | Group statistics | When individual justice is priority |
**Implementation Techniques**
- **Pre-processing**: Reweigh training examples or modify features to remove group information before training.
- **In-processing**: Add demographic parity constraint to the loss function during training (e.g., adversarial debiasing).
- **Post-processing**: Threshold adjustment — use different classification thresholds per group to equalize positive rates (Hardt et al. equalized odds approach).
- **Fairness-Aware Algorithms**: Frameworks like IBM AI Fairness 360, Google What-If Tool, and Microsoft Fairlearn implement demographic parity constraints with multiple mitigation strategies.
Demographic parity is **the most intuitive but mathematically contentious fairness criterion** — its simplicity makes it a powerful regulatory tool and auditing standard, while its failure to account for qualification distributions ensures that achieving demographic parity alone is neither necessary nor sufficient for genuinely fair algorithmic decision-making.