Pay Equity Model Adequacy Score (PEMAS)

GMI-HCM-017Human Capital ManagementResearch-ForwardPattern O-7: Inspection Coverage

Definition

PEMAS evaluates whether an organization's AI pay equity model has an appropriate feature set: neither under-controlling (missing legitimate explanatory variables, which inflates apparent gaps) nor over-controlling (including tainted variables that encode historical discrimination, which masks real gaps).

The Measurement Gap

Measurement Gap and Decision Impact

It produces a composite score from three sub-scores: Completeness (are all legally recognized explanatory variables present?), Contamination (do any features encode historical bias?), and Stability (how sensitive are gap estimates to adding or removing individual variables?).

The Decision It Supports:

Can the organization trust its AI equity analysis for compliance reporting and remediation decisions? If the model is under-controlled, the organization may remediate gaps that are actually explained by legitimate factors. If over-controlled, the organization may miss gaps that constitute real inequity. Both errors have material financial and legal consequences.

Why It Requires a New Metric:

Legacy pay equity analysis was typically conducted by compensation consultants who made explicit, documented decisions about model specification. AI-driven pay equity platforms often use opaque feature sets: the organization may not know exactly which variables the model controls for, whether those variables are legally appropriate, or how sensitive the results are to model specification choices. No legacy metric evaluates the model itself rather than the model's output.

The measurement gap PEMAS addresses is the absence of any quality assessment for the AI pay equity model itself: organizations are treating equity analysis outputs as authoritative without evaluating whether the model that produced them is methodologically adequate.

Pay equity analysis has become a compliance and risk management function. Organizations use regression-based models to identify unexplained pay gaps by gender, race, and other protected categories. Increasingly, these models are embedded in AI-driven compensation analytics platforms that produce automated equity reports, flag outliers for remediation, and track gap closure over time. The outputs look precise and authoritative. But the quality of every output depends entirely on the quality of the model specification.

Model specification in pay equity analysis is genuinely difficult because the choice of control variables is both a statistical and a legal question. Controlling for job level is generally appropriate: people at higher levels are paid more, and that is legitimate. Controlling for department may be appropriate if compensation structures vary by function. But controlling for prior salary history may encode the very discrimination the analysis is supposed to detect. And failing to control for performance ratings may inflate apparent gaps if performance genuinely varies across groups.

AI pay equity platforms make these specification decisions algorithmically or through default configurations that the organization may never review. The platform says "your adjusted gender pay gap is 2.3%" and the organization reports that number to the board, to regulators, and to employees. But nobody asked: is the model over-controlling by including variables that mask real gaps? Is it under-controlling by omitting variables that explain legitimate differences? How much does the gap estimate change if a single variable is added or removed?

PEMAS forces these questions by evaluating the model on three dimensions. Completeness checks whether all legally recognized explanatory variables are present. Contamination assesses whether any features may encode historical discrimination. Stability measures the sensitivity of gap estimates to specification changes through leave-one-out analysis. The composite tells the organization whether its equity model is a reliable instrument or a false confidence generator.

Without PEMAS, organizations report pay equity numbers with no quality assurance on the model that produced them, a risk that compounds with every compliance cycle.

Formula

Formula:

PEMAS = 0.35 × Completeness + 0.35 × (1 − Contamination) + 0.30 × Stability

Variable Definitions:

Variable Definition Unit Source
Completeness Proportion of legally recognized explanatory variables that are present in the model specification Ratio (0 to 1) Model documentation + legal adequacy rubric
Contamination Proportion of model features that may encode historical discrimination (salary history, biased performance ratings, tainted job level assignments) Ratio (0 to 1) Model documentation + legal counsel review
Stability 1 minus the maximum absolute change in gap estimate from leave-one-out sensitivity analysis (removing each variable individually and recomputing the gap) Ratio (0 to 1) Computed from model re-runs

Edge Cases and Exclusions:

  • The distinction between legitimate and tainted variables is legally contested and jurisdiction-dependent. This assessment requires employment counsel, not data science alone
  • If the model uses opaque vendor-provided features whose definitions are proprietary, Contamination cannot be assessed: flag these features as "opaque" and note the risk
  • Stability analysis via leave-one-out may be computationally expensive for models with many features: use bootstrap sampling as an approximation

Data Requirements:

Data Element Source System Classification
Equity model specification (full variable list) Compensation analytics platform or vendor NATIVE / NOT CAPTURED
Variable definitions and data sources Model documentation MANUAL
Legal adequacy rubric (appropriate vs. inappropriate controls) Legal counsel / compliance MANUAL
Leave-one-out sensitivity results Computed from model re-runs DERIVED

Proposed formula, pending validation pilots.

At a Glance

Unit
Score
Trend Direction
Higher
Reporting Frequency
Annually
Category
AI Compensation Integrity
← Offer Anchor Bias Index (OABI)Learning Competence Conversion Rate (LCCR) →

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