Methodology

How OnHirely detects hiring bias and assesses compliance.

1. The 4/5ths Rule (Adverse Impact Ratio)

For each hiring stage and protected category, we calculate the selection rate of each group. The impact ratio is the lowest group's rate divided by the highest group's rate. If this ratio falls below 0.8 (80%), adverse impact is indicated per EEOC guidelines.

impact_ratio = min(selection_rates) / max(selection_rates)
< 0.8 = Adverse Impact | 0.8-0.9 = Warning | ≥ 0.9 = Pass

2. Statistical Significance Testing

We apply chi-squared tests (when sample size > 40) or Fisher's exact test (when sample size ≤ 40) to determine whether observed differences are statistically significant (p < 0.05) rather than due to random chance.

3. AI Score Distribution Analysis

When AI scores are provided, we analyze the distribution across groups using the Kolmogorov-Smirnov test to detect whether scores are systematically different between demographic groups.

4. Intersectional Analysis

Bias can compound at intersections. We cross-reference protected categories (e.g., Black Female, Asian Male) to detect compound discrimination that single-category analysis might miss. Minimum 10 candidates per combined group.

5. Overall Compliance Score

A weighted average of all impact ratios across stages and groups. Screening stage is weighted 1.5x (where AI bias is strongest), interview 1.2x, and offer 1.0x.

0-59: HIGH RISK — Non-compliant

60-79: MODERATE RISK — Action needed

80-100: LOW RISK — Likely compliant

6. Confidence Levels

Results are tagged with confidence levels based on sample size per group: Insufficient (<5), Low (5-29), Medium (30-99), High (100+). Small samples produce unreliable statistics — we flag them accordingly.

References

EEOC Uniform Guidelines on Employee Selection Procedures (29 CFR 1607)
NYC Local Law 144 of 2021 — Automated Employment Decision Tools
California Civil Rights Council — Automated-Decision Systems Regulations
EU AI Act — Regulation (EU) 2024/1689, Annex III