Definition

Fairness parity is an umbrella term describing approximate equality across demographic groups or cohorts according to a specified fairness metric.

In plain words:

The system produces similar group-level outcomes or performance under a chosen definition of fairness, with measured gaps kept within a documented tolerance.

Meeting one parity criterion does not establish that a system is fair in every respect. Different fairness metrics capture different harms and may conflict with one another.

---

Types of Parity (Common Fairness Metrics)

Depending on what is measured, parity can refer to different criteria:

1. Demographic Parity (Statistical Parity)
   • Each group receives positive outcomes at the same or approximately the same rate.
   • Example: Group A has a 60% loan approval rate and Group B has a 58% approval rate, producing a 2-percentage-point gap.
   • Whether this satisfies a guardrail depends on the predefined tolerance.
2. Equal Opportunity
   • The True Positive Rate (TPR) is equal or approximately equal across groups.
   • Example: The model correctly identifies 80% of qualified applicants in each group.
3. Equalized Odds
   • Both the True Positive Rate and False Positive Rate are equal or approximately equal across groups.
4. Predictive Parity
   • Positive Predictive Value, also called precision, is equal or approximately equal across groups.
   • Among individuals predicted positive, the proportion who are actually positive is similar across groups.
5. Calibration Within Groups
   • A given predicted probability has approximately the same observed meaning within each group.
   • For example, among people assigned a 70% risk score, the observed outcome rate should be close to 70% in each evaluated group.

Predictive parity and calibration are related but not interchangeable. Predictive parity usually compares precision at a decision threshold, whereas calibration evaluates predicted probabilities across score levels.

---

“Fairness parity within Xpp”

• This usually means that the absolute difference in a specified group-level metric is no greater than X percentage points.
• The metric, evaluated groups, reference group, statistical method, and tolerance should be stated explicitly.
• Example using churn-model accuracy:
  • Group A accuracy = 85%
  • Group B accuracy = 83%
  • Absolute gap = 2 percentage points
  • The model satisfies an accuracy-parity guardrail of no more than 3 percentage points.

Passing this guardrail does not demonstrate parity for selection rates, TPR, FPR, calibration, or other fairness metrics.

---

Why It Matters

• Helps identify whether a model produces meaningfully different outcomes or error rates across evaluated groups.
• Supports ethical risk management, model governance, and compliance analysis.
• Builds trust with customers, employees, regulators, and other stakeholders.
• Encourages teams to evaluate subgroup performance rather than relying only on aggregate metrics.

In U.S. employment selection, the four-fifths rule compares a group’s selection rate with the highest group’s selection rate. It is a rule of thumb for indicating possible adverse impact, not a legal definition of discrimination or a universal fairness threshold.

The GDPR requires lawful, fair, and transparent processing of personal data, but it does not prescribe one universal statistical-parity metric.

---

Practical Implementation

1. Choose the fairness metric: selection rate, TPR, FPR, precision, calibration, or another context-appropriate measure.
2. Define the evaluated groups: use legally protected and fairness-relevant attributes appropriate to the application.
3. Specify the comparison: compare each group with a documented reference group or evaluate the maximum pairwise gap.
4. Quantify uncertainty: report sample sizes and, where appropriate, confidence intervals or statistical tests.
5. Set a documented tolerance: define the threshold based on the use case, potential harm, and applicable requirements.
6. Investigate violations: determine whether a gap reflects data quality, sampling variation, label bias, model behavior, or decision policy.
7. Monitor continuously: group metrics and data distributions can change after deployment.

---

Summary

• Fairness parity describes approximate equality across groups under a specified fairness metric.
• Common criteria include demographic parity, equal opportunity, equalized odds, predictive parity, and group calibration.
• A tolerance such as 3 percentage points is an operational threshold, not proof that a difference is immaterial or legally acceptable.
• Every parity claim should identify the metric, evaluated groups, comparison method, tolerance, sample size, and uncertainty.