Fairness 101: Understanding ML Fairness Metrics¶

A plain-language guide to fairness in machine learning for data scientists, compliance officers, and anyone building or auditing ML systems.

What is fairness in ML?¶

Fairness in machine learning means ensuring that predictions don't systematically disadvantage specific groups of people based on sensitive attributes like race, gender, age, or disability status.

The challenge: There are multiple definitions of fairness, and they often conflict. A model can be fair by one definition but unfair by another. This guide helps you choose the right metric for your use case.

The three core fairness definitions¶

GlassAlpha implements the three most widely-used fairness metrics. Here's what they mean in plain language:

graph TB
    Start[Which fairness<br/>aspect matters?]
    Start --> Concern{Primary<br/>concern?}

    Concern -->|Equal access| DP[Demographic<br/>Parity]
    Concern -->|Qualified treated equally| EO[Equal<br/>Opportunity]
    Concern -->|Consistent errors| EOdds[Equalized<br/>Odds]

    DP --> DPEx[Equal approval<br/>rates by race]
    EO --> EOEx[Qualified approved<br/>equally by gender]
    EOdds --> EOddsEx[Both qualified and<br/>unqualified equal]

    style Start fill:#e1f5ff
    style DP fill:#d4edda
    style EO fill:#d4edda
    style EOdds fill:#d4edda
    style DPEx fill:#fff3cd
    style EOEx fill:#fff3cd
    style EOddsEx fill:#fff3cd

1. Demographic Parity (Statistical Parity)¶

Definition: Groups receive positive outcomes at equal rates.

Formula: P(Ŷ=1 | Group A) ≈ P(Ŷ=1 | Group B)

Example: In a credit model, 60% of men and 60% of women are approved for loans.

When to use:

✅ You want equal access to opportunities (hiring, lending, housing)
✅ Historical discrimination has led to unequal base rates
✅ Regulators emphasize equal treatment regardless of group membership

When NOT to use:

❌ Groups have genuinely different base rates (e.g., disease prevalence)
❌ You need to optimize accuracy (demographic parity often reduces performance)
❌ The outcome variable itself is biased (e.g., arrest rates reflecting policing bias)

Real-world example:

Hiring Model (Demographic Parity)
====================================
Group           Applications    Hired   Rate
────────────────────────────────────────────
Men             1000            200     20%
Women           800             160     20%  ✓ Equal rates
Non-binary      100             20      20%  ✓ Equal rates
====================================
Status: PASS (all groups within 5% tolerance)

Regulatory context:

ECOA (US Credit): Often interpreted to require demographic parity in lending outcomes
GDPR Article 22 (EU): Emphasizes non-discrimination in automated decisions
FTC Act Section 5: Unequal rates can trigger "unfair practices" scrutiny

2. Equal Opportunity¶

Definition: Among qualified individuals (true positives), groups are selected at equal rates.

Formula: P(Ŷ=1 | Y=1, Group A) ≈ P(Ŷ=1 | Y=1, Group B)

Example: Among creditworthy borrowers, 85% of men and 85% of women are approved.

When to use:

✅ You care about true positives (qualified people getting opportunities)
✅ False positives are less harmful than false negatives
✅ The positive outcome is a benefit (job, loan, treatment)

When NOT to use:

❌ False positives are costly (fraud detection, criminal justice)
❌ You need to control both false positives AND false negatives
❌ Ground truth labels are unreliable or biased

Real-world example:

Credit Model (Equal Opportunity)
====================================
Group      Creditworthy    Approved    True Positive Rate
──────────────────────────────────────────────────────────
Men        500             425         85%
Women      400             340         85%  ✓ Equal TPR
────────────────────────────────────────────────────────
Status: PASS (TPR difference < 10%)

Key insight: Equal opportunity ignores what happens to unqualified people. If your model wrongly denies 10% of qualified men but 30% of qualified women, that violates equal opportunity even if overall approval rates are the same.

Regulatory context:

SR 11-7 (Banking): Focuses on minimizing errors for creditworthy borrowers
FCRA (Adverse Actions): Requires explanations when qualified applicants are denied
EU AI Act: High-risk systems must minimize false negatives for protected groups

3. Equalized Odds¶

Definition: Both true positive rates AND false positive rates are equal across groups.

Formula:

P(Ŷ=1 | Y=1, Group A) ≈ P(Ŷ=1 | Y=1, Group B) (TPR)
P(Ŷ=1 | Y=0, Group A) ≈ P(Ŷ=1 | Y=0, Group B) (FPR)

Example: Among creditworthy borrowers, 85% approved regardless of gender. Among non-creditworthy borrowers, 5% wrongly approved regardless of gender.

When to use:

✅ Both false positives and false negatives are harmful
✅ You need consistent error rates across groups
✅ High-stakes decisions (criminal justice, medical diagnosis, fraud)

When NOT to use:

❌ False positives and false negatives have very different costs
❌ You only care about one type of error
❌ Group sizes are very imbalanced (harder to achieve)

Real-world example:

Fraud Detection (Equalized Odds)
====================================
Group          TPR    FPR    Status
────────────────────────────────────
Age 18-35      92%    3%     ✓
Age 36-55      90%    3%     ✓ Equal FPR
Age 56+        91%    4%     ⚠️ FPR too high
────────────────────────────────────
Status: WARNING (FPR disparity in Age 56+)

Key insight: Equalized odds is the most stringent fairness metric. If you pass equalized odds, you typically pass equal opportunity and are close to demographic parity.

Regulatory context:

Criminal Justice Reform: COMPAS case highlighted FPR disparities
Insurance Underwriting: NAIC Model #870 requires equal error rates
Healthcare: 21^st Century Cures Act emphasizes balanced error rates

Visual comparison: which metric detects what?¶

graph TB
    A[Credit Model:<br/>Equal 60% Approval Rates]

    A --> DP[Demographic Parity:<br/>✓ PASS]
    A --> EO[Equal Opportunity:<br/>❌ FAIL]
    A --> EOdds[Equalized Odds:<br/>❌ FAIL]

    DP --> DPWhy[Both groups<br/>60% approved]
    EO --> EOWhy[Group A: 85% TPR<br/>Group B: 70% TPR]
    EOdds --> EOddsWhy[Group A: 5% FPR<br/>Group B: 10% FPR]

    style A fill:#e1f5ff
    style DP fill:#d4edda
    style EO fill:#f8d7da
    style EOdds fill:#f8d7da

The punchline: A model can have equal approval rates (demographic parity ✓) but still treat qualified people unfairly (equal opportunity ❌) or have inconsistent error rates (equalized odds ❌).

Concrete example: hiring model¶

Let's analyze a hiring model across all three metrics:

Dataset¶

1000 applications: 600 men, 400 women
300 qualified candidates: 200 men, 100 women
100 hired: Model selects top candidates

Scenario A: perfectly fair model¶

                Men     Women   Difference
─────────────────────────────────────────────
Approval Rate   16.7%   16.7%   0.0%   ✓ Demographic Parity
TPR (Qualified) 85%     85%     0.0%   ✓ Equal Opportunity
FPR (Unqualif.) 2%      2%      0.0%   ✓ Equalized Odds
─────────────────────────────────────────────
Status: PASS all fairness metrics

Scenario B: demographic parity only¶

                Men     Women   Difference
─────────────────────────────────────────────
Approval Rate   16.7%   16.7%   0.0%   ✓ Demographic Parity
TPR (Qualified) 90%     70%     20%    ❌ Equal Opportunity (FAIL)
FPR (Unqualif.) 3%      8%      5%     ❌ Equalized Odds (FAIL)
─────────────────────────────────────────────
Status: FAIL (qualified women disadvantaged)

What happened: To maintain equal approval rates despite discriminatory scoring, the model approved more unqualified women (high FPR) while rejecting more qualified women (low TPR). This is quota-filling without fairness.

Scenario C: equal opportunity only¶

                Men     Women   Difference
─────────────────────────────────────────────
Approval Rate   20%     10%     10%    ❌ Demographic Parity (FAIL)
TPR (Qualified) 85%     85%     0.0%   ✓ Equal Opportunity
FPR (Unqualif.) 5%      1%      4%     ❌ Equalized Odds (FAIL)
─────────────────────────────────────────────
Status: PARTIAL (qualified treated equally, but unequal access)

What happened: Qualified candidates are treated equally, but men have higher overall approval rates because the model is more lenient with unqualified men (higher FPR).

How GlassAlpha measures fairness¶

GlassAlpha reports all three metrics in every audit:

fairness:
  demographic_parity_difference: 0.05 # |P(Ŷ=1|A) - P(Ŷ=1|B)|
  equal_opportunity_difference: 0.03 # |TPR_A - TPR_B|
  equalized_odds_difference: 0.04 # max(|TPR_A - TPR_B|, |FPR_A - FPR_B|)

  # Per-group breakdown
  groups:
    male:
      selection_rate: 0.45
      true_positive_rate: 0.88
      false_positive_rate: 0.03
    female:
      selection_rate: 0.40 # 5% difference → demographic parity
      true_positive_rate: 0.85 # 3% difference → equal opportunity
      false_positive_rate: 0.07 # 4% difference → equalized odds

Interpretation thresholds (industry standard):

< 5%: Minimal disparity (generally acceptable)
5-10%: Moderate disparity (review required)
> 10%: Substantial disparity (regulatory risk)

Decision matrix: which metric should you use?¶

Use this table to choose the right fairness metric for your use case:

Use Case	Primary Metric	Secondary Metric	Why
Credit/Lending	Demographic Parity	Equal Opportunity	ECOA requires equal access; SR 11-7 emphasizes creditworthy lending
Hiring	Equal Opportunity	Demographic Parity	Qualified candidates should have equal chances
Fraud Detection	Equalized Odds	-	Both false positives (customer friction) and false negatives (loss)
Criminal Justice	Equalized Odds	-	False positives (wrongful detention) and false negatives (risk)
Healthcare Diagnosis	Equalized Odds	Equal Opportunity	Balanced error rates critical for patient safety
Insurance Underwriting	Equalized Odds	Demographic Parity	NAIC Model #870 requires consistent error rates
College Admissions	Equal Opportunity	Demographic Parity	Qualified applicants should have equal chances
Content Moderation	Equalized Odds	-	Both over-moderation and under-moderation harmful
Advertising Targeting	Demographic Parity	-	Equal exposure to opportunities
Predictive Policing	⚠️ All metrics fail	-	Historical data reflects biased enforcement (avoid ML)

Rule of thumb:

Opportunities (loans, jobs) → Start with equal opportunity
Harmful outcomes (fraud, risk) → Use equalized odds
Access-based (ads, housing) → Use demographic parity

The impossibility theorem¶

Important: You usually cannot satisfy all three metrics simultaneously unless:

Base rates are equal across groups (rare in real data), OR
Your model has perfect accuracy (impossible)

Proof by example:

Base rates: Group A has 50% qualified, Group B has 30% qualified
Demographic parity requires equal approval rates (e.g., 40% each)
Equal opportunity requires equal TPR (e.g., 80% of qualified approved in both groups)
Math: Group A approves 80% × 50% = 40% (matches parity) ✓
Math: Group B approves 80% × 30% = 24% (doesn't match parity) ❌

Takeaway: You must choose which fairness definition matters most for your use case. Document this choice in your audit policy.

Fairness vs accuracy tradeoffs¶

Enforcing fairness constraints typically reduces overall accuracy:

graph TB
    A[Unconstrained<br/>Accuracy: 85%]
    A --> B{Apply Fairness<br/>Constraint}

    B -->|Demographic Parity| C[Accuracy: 82%<br/>Disparity: 2%]
    B -->|Equal Opportunity| D[Accuracy: 83%<br/>Disparity: 3%]
    B -->|Equalized Odds| E[Accuracy: 81%<br/>Disparity: 4%]

    style A fill:#e1f5ff
    style C fill:#d4edda
    style D fill:#d4edda
    style E fill:#fff3cd

Why this happens:

Optimal decision boundaries differ by group when base rates differ
Fairness constraints force the model to use a common boundary
Result: Some groups get less-than-optimal predictions

Practical guidance:

2-3% accuracy drop: Typical and acceptable
5-10% accuracy drop: Review if fairness constraint is appropriate
>10% accuracy drop: Investigate data quality or model architecture issues

Exception: If accuracy improves with fairness constraints, your model was likely overfitting to spurious correlations with protected attributes.

Beyond group fairness: individual fairness¶

The metrics above measure group fairness (comparing groups). GlassAlpha also supports individual fairness:

Definition: Similar individuals should receive similar predictions.

Example: Two loan applicants with identical income, debt, and credit history should get similar scores, regardless of race or gender.

Metrics:

Consistency Score: How often similar individuals get similar predictions
Individual Treatment Disparity: Average prediction difference for matched pairs

When to use:

✅ You have matched pairs or can construct them
✅ You want to catch one-off discriminatory decisions
✅ Regulators ask about "similarly situated" applicants (ECOA, FCRA)

Learn more: Individual Fairness Metrics

Intersectional fairness¶

Warning: Analyzing single attributes (gender alone, race alone) can hide discrimination at intersections.

Example:

Single-attribute analysis (PASS):
  Men:     60% approval
  Women:   60% approval  ✓ Equal rates

  White:   60% approval
  Black:   60% approval  ✓ Equal rates

Intersectional analysis (FAIL):
  White Men:    70% approval
  White Women:  50% approval
  Black Men:    50% approval
  Black Women:  30% approval  ❌ 40% disparity

GlassAlpha supports intersectional analysis:

fairness:
  intersectional:
    enabled: true
    combinations:
      - [gender, race]
      - [age_group, income_bracket]

Learn more: Intersectional Fairness Guide

Common pitfalls¶

1. "We removed protected attributes, so we're fair"¶

Wrong: Correlated features (zip code, name, education) can serve as proxies.

Solution: Test fairness metrics even when protected attributes aren't model inputs.

2. "Our overall accuracy is high, so we're good"¶

Wrong: High accuracy can coexist with large group disparities.

Example: 95% accuracy but 30% gap in false negative rates for minorities.

Solution: Always report fairness metrics alongside accuracy.

3. "We balanced our training data by group"¶

Wrong: Balancing data helps but doesn't guarantee fair predictions.

Solution: Measure fairness on test set predictions, not training set composition.

4. "We have equal error rates within each group"¶

Wrong: This checks calibration, not fairness.

Solution: Compare error rates ACROSS groups (equalized odds).

5. "We used a neutral dataset without bias"¶

Wrong: No dataset is truly neutral; historical data reflects societal biases.

Solution: Acknowledge data limitations and test for bias regardless.

Regulatory expectations by industry¶

Different industries emphasize different fairness metrics:

Banking (SR 11-7, ECOA, FCRA)¶

Primary: Demographic parity (equal access to credit)
Secondary: Equal opportunity (creditworthy borrowers approved)
Threshold: <10% disparity in approval rates or adverse action rates
Evidence: Disparate impact analysis, counterfactual explanations

Insurance (NAIC Model #870, California SB 221)¶

Primary: Equalized odds (consistent error rates)
Secondary: Calibration (accurate risk prediction within groups)
Threshold: <5% difference in FPR/FNR by protected class
Evidence: Actuarial justification for any disparities

Healthcare (HIPAA, 21^st Century Cures Act)¶

Primary: Equalized odds (equal diagnostic accuracy)
Secondary: Calibration (accurate risk scores for treatment decisions)
Threshold: <5% disparity in sensitivity/specificity
Evidence: Clinical validation studies by demographic subgroup

Employment (EEOC, OFCCP)¶

Primary: Equal opportunity (qualified candidates advance equally)
Secondary: Demographic parity (equal access to interviews)
Threshold: <20% relative difference (80% rule)
Evidence: Adverse impact analysis, job-relatedness studies

Learn more: Compliance Guides

How to report fairness in audits¶

When presenting fairness results to stakeholders:

For technical teams¶

Fairness Analysis (threshold=0.5, n=1000)
==========================================
Metric                    Value    Status    Threshold
──────────────────────────────────────────────────────
Demographic Parity Diff   0.08     WARN      <0.10
Equal Opportunity Diff    0.12     FAIL      <0.10
Equalized Odds Diff       0.14     FAIL      <0.10

Group-level breakdown:
  Male   (n=600): TPR=0.88, FPR=0.03, Selection=0.45
  Female (n=400): TPR=0.76, FPR=0.05, Selection=0.37

Action: Investigate equal opportunity violation (12% gap in TPR)

For compliance officers¶

This model exhibits moderate to substantial fairness violations:

1. Demographic Parity: 8% gap in approval rates (ACCEPTABLE)
   - Male applicants: 45% approved
   - Female applicants: 37% approved

2. Equal Opportunity: 12% gap in true positive rates (UNACCEPTABLE)
   - Among qualified applicants, 88% of men approved
   - Among qualified applicants, 76% of women approved
   - Regulatory risk: Violates ECOA requirement for equal treatment

3. Recommendation: Recalibrate threshold or add fairness constraints

For regulators¶

Disparate Impact Analysis (ECOA Compliance)
===========================================
Protected Class: Gender (Male vs Female)
Sample Size: 600 male, 400 female applicants
Time Period: Q4 2024

Finding: Substantial disparate impact detected
  - Adverse action rate (Male): 55%
  - Adverse action rate (Female): 63%
  - Relative difference: 14.5% (exceeds 10% tolerance)
  - Statistical significance: p < 0.001

Business necessity justification: Under review
Alternative methods with less disparity: Being evaluated

Evidence pack: SHA256 checksum attached

Next steps¶

Learn more¶

Understanding Protected Attributes - What makes an attribute "protected"
Debug Fairness Violations - Step-by-step troubleshooting
Fairness Metrics Reference - Technical definitions and individual fairness

Try it yourself¶

Quickstart Guide - Generate your first fairness audit
German Credit Example - Walkthrough with real data
Configuration Guide - Complete configuration reference

Regulatory compliance¶

Questions? Join our community or contact us.

Fairness 101: Understanding ML Fairness Metrics¶

What is fairness in ML?¶

The three core fairness definitions¶

1. Demographic Parity (Statistical Parity)¶

2. Equal Opportunity¶

3. Equalized Odds¶

Visual comparison: which metric detects what?¶

Concrete example: hiring model¶

Dataset¶

Scenario A: perfectly fair model¶

Scenario B: demographic parity only¶

Scenario C: equal opportunity only¶

How GlassAlpha measures fairness¶

Decision matrix: which metric should you use?¶

The impossibility theorem¶

Fairness vs accuracy tradeoffs¶

Beyond group fairness: individual fairness¶

Intersectional fairness¶

Common pitfalls¶

1. "We removed protected attributes, so we're fair"¶

2. "Our overall accuracy is high, so we're good"¶

3. "We balanced our training data by group"¶

4. "We have equal error rates within each group"¶

5. "We used a neutral dataset without bias"¶

Regulatory expectations by industry¶

Banking (SR 11-7, ECOA, FCRA)¶

Insurance (NAIC Model #870, California SB 221)¶

Healthcare (HIPAA, 21st Century Cures Act)¶

Employment (EEOC, OFCCP)¶

How to report fairness in audits¶

For technical teams¶

For compliance officers¶

For regulators¶

Next steps¶

Learn more¶

Try it yourself¶

Regulatory compliance¶

Healthcare (HIPAA, 21^st Century Cures Act)¶