PROXY DISCRIMINATION DETECTION GAP (PDDG)

GMI-IT-050IT OperationsResearch-Forward

Definition

Proxy Discrimination Detection Gap (PDDG) measures the share of model features with significant correlation to protected-class attributes that are not in the governance platform's fairness monitoring scope.

The Measurement Gap

A regional bank's lending team operates a credit decisioning model that was certified compliant with fair lending regulations 14 months ago. The model does not use protected-class attributes as inputs: no race, no gender, no national origin. The fairness monitoring platform checks the model's predictions against four protected-class slices monthly and reports zero significant disparities. The model card is current. The compliance team treats the model as a stable, governed system.

Then the bank's internal audit team runs an independent fair lending review. They look at the model's inputs not against the protected-class attributes directly, but against features known to correlate with them: ZIP code, employer name, college attended, prior account types held. They find that the model places significant weight on prior account types, which historically correlate strongly with race in this region's customer base. The model is not using race as input. It is using a feature whose information content is largely race. Predictions are different by race despite the model never seeing race directly. The fairness monitoring platform did not detect this because it monitors prediction parity across protected slices, not feature-to-protected-attribute correlation.

The bank is now in a difficult position. The compliance program reports that the model is fair. The audit report shows the model produces racially disparate outcomes through a proxy feature. Both are correct. The compliance program measures the wrong thing: output disparity at slice level rather than input correlation with protected attributes. The audit caught what the monitoring missed. Without the audit, the proxy discrimination would have continued.

The legacy compliance metrics on the bank's dashboard report a healthy state.

Models with fairness monitoring: 100%.

Slice-level disparity tests passing: 100%.

Model card completion: 100%.

Compliance audit completion: 100%.

Models with documented fair lending review: 100%.

Predictive parity by protected slice: within tolerance for all slices monitored.

None of these report on whether the features used by the model are themselves proxies for protected attributes. The fairness monitoring platform applies a post-hoc check on predictions against protected slices. It does not analyze whether the model's input features encode the same information as the protected attributes through correlation. The proxy variables become the discrimination pathway, and the monitoring is structurally blind to that pathway.

These metrics were designed for compliance programs that included independent fair lending review as a regular practice. The proxy-variable analysis was done by human auditors using their judgment about which features were likely correlated with protected attributes. The check existed; it lived outside the automated monitoring. With AI governance platforms automating the monitoring, the proxy-variable check has not been reproduced as an automated metric. Organizations assume the platform's monitoring is comprehensive when it is structurally incomplete.

There is no metric that tells you whether a model that passes standard fairness checks is discriminating through proxy variables, which means governance platforms certify models as fair while discriminatory outcomes persist through correlated features.

Formula

Proxy Discrimination Detection Gap (PDDG) measures the share of model features with significant correlation to protected-class attributes that are not in the governance platform's fairness monitoring scope.

In plain language: for each model, identify all features that correlate strongly with any protected-class attribute (computed independently, not from the model's own outputs). Of those features, count how many are not being monitored by the fairness platform. PDDG is the unmonitored fraction. A low value means the monitoring scope covers the proxy pathways. A high value means proxy variables are operating outside the monitoring.

The formula:

PDDG = Features_correlated_unmonitored / Features_correlated_total

Where:

  • Features_correlated_total = the count of model input features with |r| > threshold to any protected-class attribute
  • Features_correlated_unmonitored = the subset that are not included in the governance platform's fairness monitoring scope

The correlation threshold recommendation is |r| > 0.3. Lower thresholds catch weaker correlations but produce more noise; higher thresholds miss meaningful proxies. The threshold is a calibration parameter that should be set deliberately based on the model's risk class.

Units are a ratio from 0 to 1. Lower is better.

A worked example. The regional bank from the opening scenario runs PDDG on their credit decisioning model after the audit finding. The model uses 47 input features. Independent correlation analysis against the four protected-class attributes (race, gender, age, national origin) finds:

  • 12 features with |r| > 0.3 to at least one protected attribute (some features correlate to multiple)
  • Of those 12, the governance platform monitors 3 (because those 3 are explicitly tagged as sensitive variables in the platform configuration)
  • 9 features correlate with protected attributes but are not in the platform's fairness monitoring scope

PDDG = 9 / 12 = 0.75

Three edge cases need explicit treatment. First, PDDG identifies monitoring gaps, not confirmed discrimination. A feature correlated with race that is being monitored is no guarantee of fair outcomes; the monitoring still has to be tuned to detect actual disparity. PDDG is necessary but not sufficient for fair outcomes. Second, multi-collinearity complicates the analysis. When several features correlate with each other and with the protected attribute, attributing the correlation cleanly requires partial correlation analysis. Track direct correlation and indirect correlation (via mediating features) separately. Third, protected-class data access is sensitive and access-controlled in most jurisdictions; the correlation analysis must be done by people with appropriate access and the results stored in audited systems.

Cross-references. Downstream of PDDG is Compliance Documentation Currency Index (CDCI), a model with high PDDG should have its compliance documentation flagged because the fairness statement in the documentation is incomplete.

Proposed formula, pending validation pilots.

At a Glance

Category
AI-Era IT Operations
← Provisioning Entitlement Utilization (PEU)RATE LIMIT EQUITY INDEX (RLEI) →

Assess Your Organization

The AI-GPS Assessment evaluates your AI governance maturity and recommends KPIs like this one for your implementation roadmap.

Take the Free AI-GPS Assessment