Intersectional Analysis

Intersectional analysis evaluates model performance across subgroups defined by the combination of two or more protected attributes (e.g., race and gender, age and disability status). This contrasts with one-dimensional analysis that examines attributes in isolation.

• Example: Instead of evaluating fairness only for "women" or only for "Black individuals," intersectional analysis would specifically evaluate performance for "Black women," "Asian women over 65," etc.
• Technical Implementation: This requires slicing evaluation datasets and computing metrics (accuracy, FPR, FNR) for each combinatorial subgroup, often leading to a combinatorial explosion of slices that must be managed statistically.

A core tenet is that bias is not merely additive but can be multiplicative. A model may show acceptable performance for Group A and Group B in isolation but fail catastrophically for individuals at the intersection of A and B.

• Mechanism: Bias can arise from interaction effects in the data or model that are only visible when considering multiple attributes simultaneously.
• Real-World Impact: This detects the most severe equity failures, such as a facial recognition system with high accuracy for "men" and for "light-skinned individuals" but very low accuracy for "dark-skinned women," a well-documented real-world failure mode.

A primary engineering challenge is data sparsity: as you slice data by multiple attributes, the sample size for each intersectional subgroup shrinks rapidly, making statistical estimates unstable.

• Mitigation Strategies: Engineers employ techniques like:
  • Bayesian hierarchical modeling to share statistical strength across related subgroups.
  • Thresholding to only report on subgroups with sufficient sample size (n > threshold).
  • Synthetic data augmentation specifically for underrepresented intersections to improve evaluation robustness.
• Trade-off: There is a direct tension between granularity (more intersections) and reliable measurement (larger sample sizes per group).

Effective intersectional analysis interprets model disparities within the socio-structural context that creates correlated, systemic disadvantages. It recognizes that attributes like race, ZIP code, and income are not independent.

• Beyond Correlation: It asks not just "is there a disparity?" but "what systemic factors causally contribute to this compounded disparity?"
• Proxy Variable Awareness: A model may exclude protected attributes like race, but if it uses features like "neighborhood crime statistics" or "occupation," it may still enact intersectional discrimination via tightly correlated proxies. Analysis must trace these pathways.

Findings from intersectional analysis must feed directly into targeted mitigation strategies. Standard fairness interventions optimized for single attributes can inadvertently harm intersectional subgroups.

• Pre-processing: Data rebalancing or augmentation focused on the most disadvantaged intersections.
• In-processing: Using fairness constraints or adversarial debiasing that specifically penalize disparity for predefined intersectional groups.
• Post-processing: Applying different decision thresholds per intersectional subgroup (where legally permissible) to achieve equalized odds or opportunity.
• Key Limitation: Mitigation is most challenging for subgroups with very little training data.

Emerging regulations and ethical frameworks are beginning to mandate intersectional scrutiny. Analysis outputs must be documented for audits and stakeholder communication.

• Model Cards & Audits: Comprehensive model documentation now requires a section on intersectional performance, listing subgroups with significant performance gaps.
• Algorithmic Impact Assessments (AIAs): Regulators may require AIAs to specifically consider "vulnerable subgroups" defined by multiple characteristics.
• Benchmark Development: New evaluation suites (e.g., specific slices within the HELM benchmark) are being created to standardize intersectional testing across models.

Subgroup analysis is the foundational practice of evaluating a model's performance metrics separately for distinct population slices. It is the prerequisite for intersectional analysis.

• Purpose: To identify performance disparities (e.g., accuracy, F1 score, false positive rate) that are masked by aggregate metrics.
• Method: Stratify evaluation datasets by single attributes (e.g., gender or age) and calculate metrics per stratum.
• Limitation: While essential, analyzing attributes in isolation fails to capture compounded effects, which is why intersectional analysis is required.

Disparate impact is a legal and quantitative measure of algorithmic bias that occurs when a model's outputs have a disproportionately adverse effect on a protected group, regardless of intent.

• Mechanism: A facially neutral model (e.g., a credit scoring algorithm) produces outcomes that systematically disadvantage a group defined by race, gender, etc.
• Measurement: Often calculated using the 80% rule (or four-fifths rule) from employment law: if the selection rate for a protected group is less than 80% of the rate for the most favored group, disparate impact may be present.
• Intersectional Context: Disparate impact can be most severe at the intersection of multiple attributes, such as for older women of color.

Group fairness metrics are mathematical definitions used to quantify equity in model predictions across demographic groups. Intersectional analysis applies these metrics to combinatorial subgroups.

Key metrics include:

• Demographic Parity: Requires the positive prediction rate to be equal across groups. Formula: P(Ŷ=1 | A=a) = P(Ŷ=1 | A=b).
• Equal Opportunity: Requires the true positive rate (recall) to be equal across groups. Formula: P(Ŷ=1 | Y=1, A=a) = P(Ŷ=1 | Y=1, A=b).
• Equalized Odds: A stricter condition requiring both true positive rates and false positive rates to be equal across groups.

Trade-off: These metrics are often mutually exclusive and frequently in tension with overall accuracy, a phenomenon formalized as the fairness-accuracy trade-off.

Bias mitigation refers to technical interventions applied during the ML lifecycle to reduce unfair discrimination. Intersectional analysis informs where and how to apply these techniques.

Three primary stages:

1. Pre-processing: Techniques applied to the training data.
   • Reweighting: Adjusting sample weights to balance distributions across (intersectional) groups.
   • Data Augmentation: Generating synthetic samples for underrepresented intersections.
2. In-processing: Techniques applied during model training.
   • Adversarial Debiasing: Training a primary predictor alongside an adversary that tries to predict the protected attribute from the primary model's embeddings, forcing the representations to be invariant to the attribute.
   • Fairness Constraints: Adding penalty terms (e.g., for demographic parity violation) directly to the loss function.
3. Post-processing: Techniques applied to model predictions.
   • Threshold Optimization: Adjusting decision thresholds per (intersectional) subgroup to satisfy a target fairness metric without retraining the model.

A proxy variable is a feature in a dataset that is highly correlated with a protected attribute, allowing a model to discriminate indirectly even when the protected attribute is explicitly removed.

• Common Examples: ZIP/postal code (correlates with race and socioeconomic status), shopping history, university name, or even linguistic patterns in text.
• Intersectional Risk: Proxies can be complex combinations of features that uniquely identify intersectional subgroups (e.g., a specific combination of profession, location, and purchase data).
• Detection & Remediation: Identifying proxies requires techniques like residual analysis (checking if protected attributes can be predicted from other features) and causal graph modeling. Mitigation involves careful feature engineering, using in-processing techniques like adversarial debiasing, or employing causal fairness approaches.

A Model Card is a standardized documentation framework for transparent reporting of a machine learning model's performance, including its fairness characteristics across subgroups.

Proposed by Google researchers, a comprehensive Model Card should include:

• Intended Use & Limitations: Explicitly stating contexts where the model should and should not be used.
• Evaluation Data: Description of the datasets used for evaluation, including their demographic breakdown.
• Quantitative Analysis: A table or matrix of performance metrics (e.g., accuracy, precision, recall, F1) across a range of intersectional subgroups (e.g., age × gender × race).
• Ethical Considerations: Discussion of known biases, recommended mitigation strategies, and results of bias audits.
• The Role of Intersectional Analysis: It provides the granular, subgroup-specific data required to populate the quantitative analysis section meaningfully, moving beyond single-attribute reporting.