Some metrics arenāt compatible with other metrics of fairness.
- Predictive accuracy of the best college students from SAT scores (yields to priviledged class)
Theorem
Predictive Parity and Classification Parity are not equal accuracy Alternatively, Calibration and Statistical Parity by Status are in tension.
Proof
https://www.marcellodibello.com/algorithmicfairness/handout/impossibility.html