Demographic Parity
A fairness criterion requiring that the proportion of positive predictions be equal across all demographic groups.
Demographic parity (also called statistical parity) is one of the most widely cited fairness definitions in algorithmic decision-making. It requires that the predicted positive outcome rate is equal across all protected demographic groups.
Formal definition: A model satisfies demographic parity if P(ŷ=1 | A=0) = P(ŷ=1 | A=1), where ŷ is the predicted outcome and A is the protected attribute.
When demographic parity is appropriate: It is most useful in contexts where the base rates of outcomes should be equal across groups — for example, where historical disparities are the result of past discrimination, not genuine differences in risk.
Limitations: Demographic parity can conflict with individual fairness and accuracy. A model that achieves demographic parity by making worse predictions for the majority group or better-than-warranted predictions for the minority group may not be considered fair by other metrics.
Contrast with equalised odds: Equalised odds requires equal true positive rates AND equal false positive rates across groups — a stricter criterion that accounts for base rate differences. It is generally preferred in high-stakes settings (credit, healthcare) over demographic parity alone.
In practice: Most enterprise AI governance frameworks require monitoring multiple fairness metrics simultaneously. Claris tracks demographic parity, equalised odds, disparate impact ratio, and individual fairness metrics in parallel, with configurable alert thresholds per metric.