Demographic Parity

Demographic Parity requires that the overall rate of receiving a positive prediction (e.g., 'approved,' 'hired,' 'high-risk') is statistically equal across different demographic groups, regardless of the actual qualifications or outcomes for individuals within those groups. It is a group fairness metric focused solely on the model's output distribution.

• Key Condition: P(Ŷ=1 | A=a) = P(Ŷ=1 | A=b) for all groups a, b.
• Focus: The selection rate or acceptance rate.
• Independence: It enforces that the predicted outcome Ŷ is statistically independent of the protected attribute A.

The condition is expressed as a probability equality. For a binary classifier with prediction Ŷ (0 or 1) and protected attribute A with groups a and b:

P(Ŷ = 1 | A = a) = P(Ŷ = 1 | A = b)

This can be measured using metrics like:

• Disparate Impact Ratio: (Selection Rate for Group a) / (Selection Rate for Group b). A value of 1.0 indicates perfect parity. The widely cited 80% rule (or four-fifths rule) in U.S. employment law considers a ratio below 0.8 as evidence of potential adverse impact.
• Statistical Parity Difference: (Selection Rate for Group a) - (Selection Rate for Group b). A value of 0.0 indicates perfect parity.

A defining—and often critiqued—characteristic of demographic parity is that it does not consider ground truth labels (Y). It mandates equal outcome rates even if the base rates of qualification differ between groups.

Example: If 60% of Group A and 40% of Group B are truly qualified for a loan, demographic parity would still demand an equal approval rate for both groups. To achieve this, the model may be forced to:

• Upsample from the qualified members of Group B.
• Downsample from the qualified members of Group A.
• Introduce randomness into decisions. This can conflict with utility and meritocratic principles, leading to justified complaints from individuals within groups.

Demographic parity is one point in a landscape of fairness definitions that often conflict.

• vs. Equal Opportunity: Equal Opportunity requires equal True Positive Rates (recall). A model can satisfy demographic parity but violate equal opportunity by having different error rates for qualified candidates across groups.
• vs. Equalized Odds: A stricter condition requiring both True Positive Rates and False Positive Rates to be equal. Demographic parity is weaker and does not guarantee Equalized Odds.
• vs. Individual Fairness: Demographic parity is a group-level criterion. It does not guarantee that similar individuals receive similar outcomes, which is the goal of individual fairness.
• The Impossibility Theorem: Barring perfect prediction or equal base rates, it is mathematically impossible to simultaneously satisfy Demographic Parity and Equalized Odds.

When is it appropriate?

• When the goal is diversification of outcomes itself (e.g., ensuring diverse candidate pools in early recruitment stages).
• When historical discrimination has severely skewed opportunity, and a corrective intervention is legally or ethically mandated.
• When ground truth labels are unreliable or biased, making metrics like equal opportunity difficult to measure fairly.

Primary Critiques:

• Ignores Merit: Can lead to reverse discrimination by disadvantaging qualified individuals from historically advantaged groups.
• Treats Groups Monolithically: Does not account for within-group heterogeneity.
• Can Harm the Intended Beneficiaries: If forced parity is achieved by approving unqualified individuals from a protected group, it can set them up for failure and reinforce negative stereotypes.

Achieving demographic parity typically requires post-processing or constrained optimization.

Common Techniques:

• Post-processing (Threshold Adjustment): Learn different classification thresholds for each demographic group to equalize selection rates. This is applied after model training.
• In-processing (Constrained Optimization): Incorporate a demographic parity constraint directly into the model's loss function during training, forcing it to optimize for accuracy subject to the parity condition.
• Pre-processing (Massaging): Re-label or reweight training data to reduce correlation between outcomes and the protected attribute before training.

Tools: Libraries like IBM AIF360, Microsoft Fairlearn, and Google's TensorFlow Fairness Indicators provide built-in functions to measure demographic parity and apply these mitigation strategies.

A bank implements demographic parity to ensure its automated loan approval model grants loans at equal rates across racial groups. This directly addresses historical disparate impact where certain groups were systematically denied credit. The bank must carefully balance this fairness goal with legitimate risk-based criteria to avoid significant profit loss or increased default rates. Post-processing techniques, like adjusting approval thresholds per group, are a common implementation method.

An admissions office uses an AI tool to pre-screen applicants. To promote diversity, they constrain the model with demographic parity, requiring an equal selection rate for applicants from different geographic (urban/rural) or socioeconomic backgrounds. This application highlights the tension between group fairness and meritocratic principles, as it may necessitate accepting some lower-scoring applicants from underrepresented groups to meet the parity target.

A company employs an AI resume screener to handle high-volume applications. To prevent the model from perpetuating historical hiring biases (e.g., against women in tech roles), they enforce demographic parity on gender for the "pass to interview" stage. This requires the tool to shortlist an equal proportion of male and female applicants from the pool. A key challenge is ensuring the model doesn't use proxy variables like hobbies or club memberships to infer and circumvent the protected attribute.

A law enforcement agency considers using a model to predict areas at high risk for crime. Analysts apply demographic parity as an audit metric, discovering the model disproportionately flags neighborhoods with higher minority populations. This disparity often stems from historical bias in arrest data used for training. Enforcing parity in this context is highly controversial, as it may involve artificially lowering scores for some areas or raising them for others, potentially misallocating resources.

A hospital system uses a model to identify patients for a high-risk care management program. To ensure equitable access across racial groups, they mandate demographic parity in referral rates. This aims to correct for underdiagnosis and unequal treatment historically experienced by minority groups. The implementation must be carefully monitored to ensure medical need remains the primary driver, avoiding the allocation of limited resources to healthier patients simply to meet a quota.

The core critique of demographic parity is the qualification paradox: it mandates equal outcomes regardless of actual qualification rates between groups. In practice, qualification rates often differ due to complex socioeconomic factors.

• Example: If Group A has 60% qualified candidates and Group B has 40%, strict parity forces equal selection rates, unfairly disadvantating qualified candidates in Group A and/or advantaging unqualified candidates in Group B.
• This makes it a blunt instrument, often unsuitable for contexts where meritocratic selection is paramount. It is most justifiable when differences in qualification rates are themselves believed to be the result of past discrimination.

A group fairness metric requiring the true positive rate (recall) to be equal across demographic groups. It ensures qualified individuals from each group have an equal chance of receiving a positive outcome (e.g., loan approval).

• Key Difference from Demographic Parity: Equal opportunity conditions on actual qualification (the true label), whereas demographic parity does not.
• Use Case: Appropriate when the cost of a false negative is high, such as in hiring or medical screening, as it focuses on not missing qualified candidates from any group.

A stricter group fairness criterion requiring both the true positive rate and the false positive rate to be equal across groups. It ensures parity in both beneficial and harmful error rates.

• Comparison: A superset of equal opportunity. A model satisfying equalized odds automatically satisfies equal opportunity.
• Implication: This is often harder to achieve than demographic parity, as it tightly couples fairness with the model's accuracy on the ground truth for all groups.

A legal doctrine and form of algorithmic bias that occurs when a facially neutral model produces outcomes that disproportionately harm a protected group. It is typically measured using the 80% rule (or four-fifths rule).

• Relation to Demographic Parity: A finding of disparate impact is evidence that demographic parity has been violated. Demographic parity is the technical formulation of avoiding disparate impact.
• Example: If the loan approval rate for Group A is 50% and for Group B is 10%, the ratio is 20% (<80%), indicating illegal disparate impact.

A systematic, documented evaluation of an AI system to detect, measure, and report on potential discriminatory biases against defined protected groups. It is a core governance activity.

• Process: Involves subgroup analysis using fairness metrics (like demographic parity), examining training data for representation bias, and testing for proxy variables.
• Output: Often results in a model card or Algorithmic Impact Assessment (AIA) that transparently documents performance disparities.

A technique applied to a trained model's predictions to achieve a target fairness metric without retraining. The most common method is to apply different decision thresholds to different demographic groups.

• Mechanism: To achieve demographic parity, the approval threshold might be lowered for a group with a lower base rate of positive outcomes.
• Advantage: Simple and fast to implement post-deployment.
• Disadvantage: Can violate individual fairness by treating two identical individuals differently based solely on group membership.

The observed phenomenon where enforcing a strict fairness constraint (like demographic parity) often requires a reduction in the model's overall predictive accuracy. This occurs because the optimal accuracy-maximizing model may rely on statistical patterns correlated with protected attributes.

• Key Insight: Perfect demographic parity and perfect accuracy are frequently incompatible when base rates of outcomes differ between groups.
• Engineering Implication: Requires stakeholders to explicitly define the acceptable trade-off frontier, making fairness an explicit design requirement rather than an afterthought.