Our methodology is published because transparency is not optional in independent auditing.
Paritas collects employment decision data through two methods:
All candidate personally identifiable information (names, emails, direct identifiers) is hashed with SHA-256 + organization-specific salt before storage. Paritas never stores raw candidate PII. Only demographic categories, scores, and outcomes are retained for analysis.
For AEDTs that produce a binary outcome (hired/rejected, or pass/fail based on a threshold):
Selection Rate = (# selected in group) / (# total in group)
Calculated for each demographic group within sex, race/ethnicity, and intersectional (sex × race/ethnicity) categories.
For AEDTs that produce a continuous score rather than a binary decision:
Scoring Rate = (# in group scoring at or above threshold) / (# total in group)
Default threshold is the overall median score unless the customer specifies one (e.g., score of 60).
Impact Ratio = (selection/scoring rate of group) / (selection/scoring rate of most-selected group)
The four-fifths (80%) rule is the regulatory benchmark for identifying potential adverse impact. Paritas classifies results using a three-tier system:
Impact Ratio ≥ 0.90
Impact Ratio 0.80–0.89
Impact Ratio < 0.80
All combinations of Sex × Race/Ethnicity (e.g., "Asian Female", "Black or African American Male").
Per DCWP rules, categories comprising less than 2% of total applicants are excluded from impact ratio calculations. These groups are still reported for transparency with their raw counts and selection rates.
Beyond the four-fifths rule, Paritas applies additional statistical rigor:
Applied for all pairwise comparisons, especially valuable for small sample sizes where large-sample approximations may not hold.
Applied when both comparison groups have n > 200. Uses significance threshold p < 0.05.
Calculated for each impact ratio to quantify the range of uncertainty around the point estimate.
Flagged when statistical significance exists but the impact ratio is in the MONITOR zone (0.80–0.89). This indicates a result that warrants attention even if it technically passes the four-fifths threshold.
Simpson's Paradox occurs when aggregate data shows one pattern, but the pattern reverses when data is stratified by a confounding variable (e.g., department or job category).
Paritas performs stratified analysis by:
This analysis helps identify cases where overall "passing" results may mask localized discrimination patterns.
Per LL144 requirements, Paritas reports the count and percentage of applicants with unknown or missing demographic data. These applicants are excluded from the analysis.
High unknown rates are flagged prominently in the report as they reduce confidence in findings. If more than 20% of records have missing demographic data, a warning is included in the executive summary.
Every Paritas audit report includes:
This methodology document is versioned. The specific methodology version used for each audit is recorded in the published report.
Updated methodologies do not retroactively change published reports. All methodology versions remain accessible at paritas.ai/methodology.
Current version: 1.0 (February 2026)
Our methodology is published. Our results are defensible.