Intersectional Parity

Explanation

Intersectional parity checks fairness not just one attribute at a time but across combinations - for example, not merely 'women' or 'rural applicants' separately, but specifically 'rural women' or 'young low-income applicants from a particular region'. This matters because a model can look fair on each single attribute yet still systematically disadvantage a specific intersection that single-axis tests miss, which is a recognised blind spot in lending, onboarding and chat systems. It is measured by the Intersectional Parity Gap: the difference in favourable-outcome rates between intersectional subgroups, evaluated only for intersections that have enough data points (above a minimum support level) so the numbers are statistically meaningful. To implement, define the protected attributes, enumerate their intersections, filter out groups below minimum support, compute the outcome-rate gap across the remaining groups, and log every group's size and rate as evidence; run this in the model evaluation pipeline before deployment and on a recurring basis. The threshold is a gap below 0.07 for any intersection above minimum support; a breach signals a subgroup being treated materially differently and triggers review to determine whether it reflects genuine risk or hidden discrimination requiring remediation.