Non-Parametric Regulatory Disparate Impact and Parity Audit Script

Role: Lead Forensic Statistician and Regulatory Compliance Data Scientist.

Context

• Protected demographic variable: {{protected_attribute}}
• Regulated outcome metric: {{outcome_variable}}
• Statutory equity threshold: {{benchmark_threshold}}
• Input table structure: {{dataset_schema}}
• Control confounding factors: {{confounding_covariates}}
• Bootstrap sample size: {{resample_iterations}}

Task

Construct a comprehensive Python forensic audit script that calculates adverse impact ratios, performs non-parametric permutation tests, and computes bootstrap confidence intervals on {{outcome_variable}} across groups defined by {{protected_attribute}}.

Method

1. Ingest input data conforming to {{dataset_schema}} using pandas with schema validation and type coercion.
2. Compute empirical group summary distributions, cross-tabulations, and adverse impact ratios against {{protected_attribute}}.
3. Evaluate the 4/5ths (80%) rule and custom {{benchmark_threshold}} statutory compliance metrics.
4. Build a non-parametric bootstrapping engine with {{resample_iterations}} iterations to calculate 95% bias-corrected accelerated (BCa) confidence intervals.
5. Execute two-sample permutation tests to determine exact non-parametric p-values without assuming normality.
6. Fit a propensity-score or regression-stratified model controlling for {{confounding_covariates}}.
7. Evaluate whether raw disparities remain statistically significant after conditioning on {{confounding_covariates}}.
8. Generate a structured JSON audit artifact containing test statistics, odds ratios, and compliance flags.

Constraints

• Code MUST execute deterministically using configurable random seeds across all {{resample_iterations}}.
• The script MUST NOT rely on unstated asymptotic normality assumptions for small subgroup slices.
• Implement strict vectorized computations in numpy/scipy to optimize resampling loops.
• Output variables and summary keys MUST match exact regulatory terminology (e.g., adverse_impact_ratio, p_value, ci_lower, ci_upper).

Output format

Provide a production Python script containing:

1. Input Data Contract & Validation Function (min 20 lines).
2. Bootstrap & Permutation Testing Core Engine (min 45 lines).
3. Covariate-Adjusted Disparity Analyzer incorporating {{confounding_covariates}} (min 30 lines).
4. Report Dispatcher printing both formatted console tables and returning a structured dictionary.

Self-review

• Confirm that both raw and adjusted disparity measures are calculated and logged.
• Ensure {{resample_iterations}} is parameterized with a reasonable fallback for fast dry-runs.
• Verify vectorization prevents memory exhaustion during heavy bootstrap matrix expansion.