Sufficiency and Separation
You'll learn two formal mathematical definitions that make fairness constraints precise: sufficiency (predictive parity) and separation (equalized odds).
Understanding These Independence Conditions
In the previous lessons, you've encountered fairness metrics like predictive parity and equalized odds. Now we formalize them as independence conditions using statistical notation.
Sufficiency (Predictive Parity)
Sufficiency means: given a prediction score, the actual outcome is independent of the protected attribute.
Formally: Y ⊥ A | Ŷ (Y independent of A given Ŷ)
In plain terms: If your model predicts someone will repay a loan with 80% confidence, that prediction should mean the same thing regardless of whether the person is in group A or group B. The calibration should be identical across groups.
Analogy: Think of a thermometer. If it reads 98.6°F, that reading should mean "healthy temperature" for everyone, regardless of age, gender, or ethnicity. The instrument's readings are sufficient for diagnosis without needing to know demographic information.
Separation (Equalized Odds)
Separation means: given the true outcome, the prediction is independent of the protected attribute.
Formally: Ŷ ⊥ A | Y (Ŷ independent of A given Y)
In plain terms: Among people who actually repay loans (Y=1), the model should predict success equally well across groups. Same for those who default (Y=0). This is exactly equalized odds—equal true positive rates AND equal false positive rates.
Analogy: A fair hiring process should select qualified candidates at equal rates across demographics, and reject unqualified candidates at equal rates too.
Why These Matter
These formal definitions help us reason mathematically about fairness constraints and understand why different fairness criteria can conflict (recall the impossibility theorem).
Key Takeaway: Sufficiency requires predictions to be calibrated identically across groups (independence of outcome from group given prediction), while separation requires error rates to be equal across groups (independence of prediction from group given outcome).