Demographic Parity Explained

The simplest fairness metric — and its limits.

Demographic parity (also called statistical parity) asks one question: do all groups receive the positive outcome at the same rate?

The math

For groups A and B:

P(prediction = positive | group A) ≈ P(prediction = positive | group B)

If 30% of male applicants are shortlisted and 15% of female applicants are, the demographic parity difference is 15 percentage points and the disparate impact ratio is 0.5 — far below the 0.8 screening threshold.

When it's the right lens

  • When the outcome allocates an opportunity or resource — interviews, loans, ad exposure for jobs and housing.
  • When historical data is itself suspect: if past hiring was biased, "accurate" predictions reproduce that bias, and parity is the corrective lens.

Its blind spots

  • It ignores qualifications entirely. A model can satisfy parity by selecting randomly within one group — equal rates, terrible decisions.
  • If genuine base rates differ between groups, forcing parity can *create* unfairness at the individual level.

That's why the report pairs it with equalized odds, which conditions on the true outcome.

Practical guidance

Treat a parity gap as a flag to investigate, not an automatic verdict: check the input features, the training data's history, and whether a legitimate factor explains the gap — then document what you found.

Didn't find what you need?

Guardia AI provides compliance tooling, not legal advice. For official regulatory text, see EU Regulation 2024/1689.