Disparate-Impact Testing — the Four-Fifths Rule in One Query

Objective

Fairness testing is usually an offline project run once a quarter. Because your decisions already live in the database, it can be a query you run any day. This recipe computes approval rates by group and surfaces a disparate-impact signal against the classic four-fifths (80%) rule.

Step 1: Your decisions

(In production this is the decision record; here, a small sample with a group attribute used only for fairness monitoring.)

DROP TABLE IF EXISTS uw_dec;

CREATE TABLE uw_dec (app_id INT, grp TEXT, decision TEXT);

INSERT INTO uw_dec VALUES
 (1,'A','approve'),(2,'A','approve'),(3,'A','decline'),(4,'A','approve'),
 (5,'B','decline'),(6,'B','decline'),(7,'B','approve'),(8,'B','decline');

Step 2: Approval rate by group

SELECT grp,
       COUNT(*)                                                    AS n,
       SUM(CASE WHEN decision = 'approve' THEN 1 ELSE 0 END)       AS approvals,
       ROUND(AVG(CASE WHEN decision = 'approve' THEN 1.0 ELSE 0.0 END), 3) AS approval_rate
FROM   uw_dec
GROUP  BY grp
ORDER  BY grp;

Result:

Step 3: Read the four-fifths ratio

Group B's rate (0.25) divided by the highest rate (0.75) is 0.33 — well under the 0.80 threshold, a disparate-impact flag that warrants review. In production you would compute the ratio directly and alert when it drops below 0.8.

Why it matters

Fairness becomes a continuous query over decisions you already made, with the evidence sitting next to the decisions themselves — exactly what NAIC, Colorado, and DOI examinations ask you to demonstrate. Keep the underlying decisions in a tamper-evident record so the fairness evidence is itself defensible.