§Method

Every number, and how we got it.

An audit is only as defensible as its method. Here is exactly what Isonomy computes, with the choices and thresholds behind each finding. Nothing is a black box.

1. Selection and scoring rates

For a tool that makes a decision, the selection rate of category g is the share of people in g who were selected. For a tool that produces a score, Local Law 144 uses the scoring rate: the share of g scoring above the median of the whole sample.

rate(g) = selected(g) / assessed(g)    scoring_rate(g) = #{ score > median } in g / assessed(g)

Records with an unknown or declined category are excluded from that table and their count is reported.

2. Impact ratio

impact_ratio(g) = rate(g) / maxh rate(h)

The reference is the category with the highest rate. Categories smaller than 2% of the sample are reported but excluded from the ratio, as the Local Law 144 rules permit; the threshold is configurable. The four-fifths figure (0.80) is a long-standing rule of thumb, not a legal safe harbour, which is why Isonomy pairs it with significance testing.

3. Uncertainty and significance

Each rate carries a Wilson score 95% interval. Each impact ratio carries a Katz log interval:

CI = exp( ln(RR) ± 1.96 · √( 1/k₁ − 1/n₁ + 1/k₂ − 1/n₂ ) )

with a 0.5 continuity correction when a count is zero. Each group is compared with the reference by Fisher’s exact test, and p-values are Holm-adjusted within each table so that testing many groups doesn’t manufacture findings.

4. Status rules

StatusRule (defaults: threshold 0.80, α = 0.05)
Adverse impactImpact ratio below the threshold and adjusted p < α
Below four-fifthsImpact ratio below the threshold, not statistically significant: collect more data or monitor
MonitorStatistically significant difference with impact ratio between the threshold and 0.90
PassEverything else
ExcludedCategory under the minimum share of the sample

5. Intersectional analysis

Every combination of the selected attributes (all pairs, and the full combination when there are more than two) gets its own table under the same rules. Two findings are specific to intersections:

  • Hidden disparity: an intersectional group is below the threshold while every constituent single-attribute ratio is at or above it.
  • Amplification: the lowest intersectional ratio is more than 0.10 below the lowest single-attribute ratio.

6. Proxies and leakage

Each input is binned (numeric inputs into deciles) and compared with each protected attribute and with the outcome using bias-corrected Cramér’s V (Bergsma, 2013). An input is a strong proxy at V ≥ 0.30 and moderate at V ≥ 0.15; proxy risk is the geometric mean of the two associations.

Leakage asks whether the inputs jointly encode an attribute: a gradient-boosted classifier predicts the attribute from all inputs, and its 3-fold cross-validated macro AUC is reported. At 0.70 or above, removing the protected column has not removed the information.

7. Explanation fairness

Given per-decision attributions (e.g. SHAP values), each person’s absolute attributions are normalized to sum to one. For each group:

  • Explanation disparity: Jensen–Shannon distance (base 2) between the group’s mean importance profile and the population’s.
  • Stability deviation: the group’s mean total-variation distance from its own centroid minus the population’s, in the spirit of MESD, our published metric for explanation fairness across intersectional subgroups.
  • Score penalty: the feature whose mean signed attribution is most below the population mean for that group, in score units.

8. Drift

With a date column, impact ratios are recomputed for each month (groups need at least 30 people in a period and must clear the minimum share overall). The reference group is re-chosen in each period. A fall of 0.10 or more between the first and last period is flagged; if the latest ratio is also below the threshold, it is a fairness regression.

9. Less discriminatory alternatives

Candidates are logistic models retrained on the supplied inputs with 3-fold cross-validation: the inputs as they are; each flagged proxy dropped; all proxies dropped; reweighing (Kamiran & Calders) so group and outcome are independent in training; and iterative nullspace projection, which removes directions in feature space that linearly predict a protected attribute until it is no longer recoverable on held-out data (AUC < 0.55), capped at a third of the dimensions.

Every candidate selects the same share of people as the current tool. Validity is precision against a job-related outcome when one is supplied, otherwise agreement with the current tool. A candidate is eligible if it loses no more than 2 points of precision, has no more significant adverse groups, and improves the worst impact ratio (over groups of 100 or more). Among eligible candidates, Isonomy prefers the fewest adverse groups, then, within 0.02 of the fairest, one that never uses protected data, then the fairest, then the most valid.

No candidate uses protected attributes or group-specific cut-offs at decision time. Adjusting employment-test scores or cut-offs by race, sex, religion or national origin is prohibited under 42 U.S.C. § 2000e-2(l).

10. Limitations we state in every report

  • Statistical association is not proof of discrimination, and passing the four-fifths rule is not proof of fairness.
  • Job-related outcomes are usually observed only for people who were selected (selective labels); validity estimates are directional and should be confirmed by a validation study.
  • Candidate models show what is achievable on the supplied inputs. They are not a drop-in replacement for a vendor’s model.
  • Results describe the data supplied, and are only as complete as its demographic fields.

The research behind these methods