Demographic Parity versus Equalized Odds in hiring

WHAT THIS TESTS: Whether you can precisely distinguish two formal fairness criteria, understand that they generally cannot both hold at once, and reason about which fits a hiring context given its goals and data biases.

A GOOD ANSWER COVERS: Demographic parity, also called statistical parity, requires that the positive prediction rate be equal across groups, so the model selects the same fraction of, say, men and women, regardless of their qualifications. It conditions only on the protected attribute, not on the true outcome. Equalized odds is stricter and conditions on the actual label: it requires equal true-positive rates and equal false-positive rates across groups, meaning genuinely qualified candidates from each group are equally likely to be selected and unqualified candidates equally likely to be rejected. For resume screening, equalized odds is often more appropriate because it ties fairness to merit, equal treatment for equally qualified people, rather than forcing equal pass rates that could mean advancing less-qualified candidates to hit a quota. The important caveat is that the historical labels themselves, who was hired before, may be biased, in which case equalized odds can perpetuate that bias and demographic parity becomes a corrective worth considering. The choice is a value judgment, not purely technical.

COMMON WRONG ANSWERS: Treating the metrics as interchangeable, or claiming a model can satisfy both simultaneously, which is generally impossible when base rates differ. Ignoring that biased ground-truth labels undermine equalized odds.

LIKELY FOLLOW-UPS: Why can't you satisfy both at once? What if your hiring labels are themselves discriminatory? How does this connect to equal opportunity, the true-positive-only variant?

ONE CONCRETE EXAMPLE: A resume screener passes thirty percent of men and twenty percent of women. Demographic parity says equalize those rates. Equalized odds instead asks whether equally qualified applicants get equal pass rates and equally unqualified ones equal rejection; if qualified women are passed less often, that is the violation to fix, assuming the qualification labels are trustworthy.