Equalized Odds Explained

Fairness conditioned on the true outcome.

Equalized odds asks a sharper question than demographic parity: among people with the same true outcome, does the model treat groups the same?

The math

For each group, two rates must match:

True positive rate:  P(predicted positive | actually positive, group)
False positive rate: P(predicted positive | actually negative, group)

Equalized odds holds when both TPR and FPR are (approximately) equal across groups. Equal opportunity is the relaxed version that only requires TPR parity — "qualified candidates get selected at the same rate, whichever group they're in".

An example

A loan model approves 80% of creditworthy applicants from group A but only 60% of equally creditworthy applicants from group B. Demographic parity might still look fine (if group B applies less often), but equal opportunity is violated — creditworthy people in group B are being wrongly denied at twice the rate.

What you need to compute it

A ground-truth column in your uploaded data. Without true outcomes, TPR/FPR can't be calculated — the report will fall back to parity metrics and tell you so.

Which to prioritise

If your model's errors are costly to individuals (denied loans, rejected candidates, missed diagnoses), equalized odds and equal opportunity are usually the metrics that matter most — and the ones worth setting explicit thresholds for in your monitoring plan.

Didn't find what you need?

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