Group Fairness

Explanation

Group fairness — also called demographic or statistical parity — checks whether the model produces a favourable outcome at roughly the same rate for every protected group (e.g. gender, age band, region, or, in Indian lending, caste and religion). It matters because a wide gap signals disparate impact and possible unlawful discrimination. It is measured by the Statistical Parity Gap, calculated as P(Ŷ=1|A=0) − P(Ŷ=1|A=1): the difference in the positive-decision rate between two groups. A gap near zero means the groups are treated alike; for example a loan model that approves 60% of one group but only 45% of another shows a wide gap signalling hidden bias. To implement it, compute the parity gap on validation data and in production by logging predictions alongside protected-attribute labels, slice decision rates per group, and surface the gap on a fairness dashboard with drill-downs and evidence stored for audit. Run it both pre-deployment and on a recurring monitoring cadence. The threshold requires a gap below 0.05 after controlling for legitimate credit risk; breaching it triggers an alert and reviewer investigation to judge whether the difference reflects genuine risk or bias, with remediation and sign-off required. Read it alongside error-based measures, since forcing parity when true base rates differ can reduce accuracy.