Demographic parity gap — fairness as a number

The short answer. A demographic parity gap is the measured difference in a model's behaviour or performance between patient subgroups — by race, sex, age, language, payer. Its value is converting 'is the model fair?' from a debate into a number that can be measured locally, trended, and thresholded. Its limit: which gap matters, and how much is tolerable, remains a clinical and ethical judgment.

Fairness debates stall on abstraction; gaps are measurable. Pick the metric that matters clinically — sensitivity, say, for a screening model — compute it per subgroup on your population, and the difference is the gap: sensitivity 0.84 overall can coexist with 0.88 for one group and 0.71 for another, and the overall number will never tell you.

Three governance notes. Local, always — vendor fairness numbers describe the vendor’s population, which is the whole reason bias assessment is a local control. Trend it — gaps drift like everything else, and a gap opening over time is a drift signal the headline metric can miss. The number does not decide — statistics locate the gap; deciding what gap is tolerable, for which outcome, is the governance committee’s judgment to make and minute.

Where do you actually stand? The free RUAIH readiness score maps your organisation against the five focus areas in about eight minutes, and the published crosswalk shows how each control lands across RUAIH, CHAI and the NIST AI RMF.

Demographic parity gap — fairness as a number

Keep reading