Accuracy Parity

Accuracy parity requires that a model's overall accuracy—the proportion of correct predictions—is identical across all protected demographic groups. Formally, for groups A and B: P(ŷ = y | G = A) = P(ŷ = y | G = B). This constraint targets the total error rate rather than specific error types, making it a global measure of predictive equity. Unlike demographic parity, which equalizes positive prediction rates, accuracy parity directly addresses whether the model serves all populations with equal reliability.

Accuracy parity suffers from several critical limitations:

• Accuracy Paradox: In highly imbalanced datasets where the base rate of the target variable differs across groups, enforcing accuracy parity can mask severe disparities in specific error types.
• Lack of Granularity: It provides no insight into whether errors are false positives or false negatives, which may have vastly different real-world consequences.
• Majority Group Bias: Optimizing for overall accuracy parity can inadvertently lead the model to perform well on the majority class within each group while ignoring minority class performance.

To evaluate accuracy parity, compute the accuracy score for each demographic subgroup and calculate the maximum difference or ratio between the highest and lowest values. A common threshold is a difference of less than 0.01 (1%). During training, accuracy parity can be enforced through:

• Constrained optimization: Adding the accuracy disparity as a penalty term in the loss function.
• Post-processing: Adjusting decision thresholds per group to equalize error rates after training.
• Reweighting: Assigning higher sample weights to misclassified instances from underperforming groups.

Accuracy parity is referenced in algorithmic auditing frameworks and is particularly relevant in high-stakes domains where any misclassification carries equal weight regardless of direction. Use cases include:

• Medical diagnosis: Ensuring diagnostic models do not systematically misdiagnose one demographic group more often.
• Credit underwriting: Verifying that loan default prediction models maintain consistent accuracy across protected classes.
• Educational assessment: Confirming that automated grading systems do not exhibit differential error rates. The EU AI Act and NYC Local Law 144 implicitly encourage such parity analyses through their bias audit requirements.

These two fairness criteria optimize for fundamentally different outcomes:

• Demographic Parity equalizes the decision rate (e.g., who gets a loan), which can force the model to make intentional errors to meet quotas when base rates differ.
• Accuracy Parity equalizes the correctness rate, which respects underlying statistical differences between groups but may still result in unequal decision distributions. The choice between them depends on whether the ethical priority is equality of opportunity (accuracy parity) or equality of outcome (demographic parity).

A quantitative measure used to evaluate and compare model prediction outcomes across different protected demographic groups. While Accuracy Parity focuses on equalizing error rates, other metrics like Demographic Parity (equal positive rates) or Equalized Odds (equal true/false positive rates) capture different statistical definitions of fairness. Selecting the appropriate metric requires a deep understanding of the specific societal and legal context of the deployment.

A statistical measure comparing the favorable outcome rate for a protected group against a reference group, used to detect legally actionable discrimination in automated decisions. It is calculated as the ratio of selection rates. A common rule of thumb, derived from the 80% rule in employment law, flags potential issues if the ratio falls below 0.8, indicating a substantially different rate of positive predictions between groups.

A structured transparency document detailing a machine learning model's intended use, performance metrics, evaluation data, and known limitations. For a model evaluated for Accuracy Parity, the model card would explicitly report disaggregated performance metrics, such as accuracy, false positive rate, and false negative rate, for each demographic subgroup, providing a standardized format for ethical reporting and stakeholder review.

A causal explanation method that identifies the minimal change to an input feature required to alter a model's prediction to a desired alternative outcome. In the context of fairness, it answers questions like, 'What would need to change for this loan application to be approved?' This technique is crucial for auditing models that fail Accuracy Parity tests, as it reveals the specific features driving disparate outcomes for individuals near the decision boundary.

A regulatory remedy requiring a company to delete a trained model or its associated data products when they were developed using unlawfully collected or improperly processed personal data. If a model is found to violate fairness constraints like Accuracy Parity due to biased training data that was collected without proper consent, disgorgement represents the most extreme form of enforcement, eliminating the non-compliant system entirely.

A technique for interrogating an opaque model's behavior by analyzing only its inputs and outputs to detect bias, vulnerabilities, or regulatory non-compliance without accessing internal weights. Auditors can test for Accuracy Parity by sending a controlled dataset through the model's API and statistically comparing the error rates across different demographic groups, enabling independent verification of a vendor's fairness claims.