The Hong Kong University of Science and Technology · FACULTY OF ARTS & HUMANITIES

MAIE5103 Chap.4 Impossibility Results and Choosing a Fairness Metric

- one subject, every graph, every model, every mark
7 Chapters3-page Bible
Our own words - no uploaded lecturer files
Updated for this semester
Chapter 4 of 14 · MAIE5103

Impossibility Results and Choosing a Fairness Metric

Three conditions that cannot all hold

Kleinberg, Mullainathan and Raghavan proved that calibration within groups, balance for the positive class and balance for the negative class cannot hold together except in two special cases: equal base rates, or a perfect predictor. Chouldechova's companion result shows the same tension between equal precision and equalized odds.

The lecture does not treat this as a reason to give up on fairness. It treats it as a reason to make the choice explicit, write it down and defend it, instead of letting a default threshold make it silently.

The conflict in one line of Bayes' rule

The Beyond Bias lecture derives the conflict directly.

Precision for a group is the true-positive rate times the base rate, divided by that same quantity plus the false-positive rate times one minus the base rate. Hold the two error rates equal across groups and the base rate alone decides precision, so a group with fewer qualified members gets a less informative positive decision.

The algebra holds for any base rate strictly between zero and one with positive error rates, which is why the lecture calls the conflict algebraic rather than accidental.

Choosing by the harm

Equal rates are not equal counts, and equal counts are not equal welfare, so the lecture's decision aid starts from the harm. If the worry is qualified people being shut out, the true-positive rate measures it.

If it is wasted or wrongful follow-up of those selected, precision does. If it is overall participation, the selection rate does. Each choice protects someone and leaves something unmeasured, and a calibrated score still falls short of a justified decision, which also needs a valid target, its consequences and a fair procedure.

In this chapter

What this chapter covers

  • 01

    The Kleinberg, Mullainathan and Raghavan impossibility result

  • 02

    Chouldechova's precision and equalized-odds conflict

  • 03

    Deriving precision from Bayes' rule

  • 04

    Why base rates decide precision when error rates match

  • 05

    Equal rates, unequal counts and unequal welfare

  • 06

    Calibrated scores against justified decisions

  • 07

    Choosing a criterion by naming the harm

Worked example · free

Compute precision when error rates are equal

Q [6 marks]. AskSia-authored practice. A tenancy screen has a true-positive rate of 0.8 and a false-positive rate of 0.2 in two applicant groups. In Group R, 60% of applicants would be reliable tenants; in Group S, 30% would. Use Bayes' rule to find precision in each group, and explain what the difference means for a landlord reading a positive result. The marks shown are an AskSia study allocation, not a University marking scheme.
Precision rises with the base rate even when the error rates are identical.Precision rises with the base rate even when the error rates are identical. Holding TPR at 0.9 and FPR at 0.1 for both groups, the group with a 50% base rate gets 90% precision and the group with a 10% base rate gets 50%.Group X 90%Group Z 50%00.51base rate pprecision (PPV)
Precision rises with the base rate even when the error rates are identical. Holding TPR at 0.9 and FPR at 0.1 for both groups, the group with a 50% base rate gets 90% precision and the group with a 10% base rate gets 50%.
  • 2Write precision in terms of the two error rates and the base rate.
  • 2Compute precision for both groups.
  • 2Interpret the gap for a person reading a positive result.
Precision equals TPR times p divided by TPR times p plus FPR times one minus p. For R that is 0.48 divided by 0.48 plus 0.08, which is 0.48/0.56, about 85.7%. For S it is 0.24 divided by 0.24 plus 0.14, which is 0.24/0.38, about 63.2%. Equalized odds holds by construction, yet a positive result is a much weaker signal for Group S, because a larger share of its applicants are unreliable and the same false-positive rate therefore adds relatively more false flags. Equalising precision would require changing one group's error rates, which would break equalized odds.
Sia tip — If two fairness findings seem to contradict each other, check the base rates first; equal error rates and unequal precision is the expected pattern, not an error.
Glossary

Key terms

Impossibility Result
A theorem showing that calibration and balance for both classes cannot hold together except with equal base rates or a perfect predictor.
Positive-Class Balance
The condition that people whose true outcome is positive receive the same average score in every group.
Negative-Class Balance
The condition that people whose true outcome is negative receive the same average score in every group.
Perfect Predictor
A classifier with no false positives and no false negatives, one of the two special cases in which the fairness criteria stop conflicting.
FAQ

Impossibility Results and Choosing a Fairness Metric FAQ

Why can't an algorithm satisfy every fairness definition at once?

Because when groups have different base rates, the criteria place incompatible demands on the same confusion matrices. Holding true-positive and false-positive rates equal forces precision to differ, and forcing precision equal pushes the error rates apart. The two exceptions are equal base rates and perfect prediction. In practice the designer has to choose a criterion and should record and defend that choice.

What did the COMPAS dispute show about fairness metrics?

That two parties can both be correct about the same tool. The 2016 analysis found unequal false-positive rates between Black and White defendants, an equalized-odds failure, while the vendor's reply relied on measures in the predictive-parity family. With different base rates the two cannot both hold, so the real disagreement was about which criterion should govern, not about the data.

Is a calibrated score enough to make decisions fair?

No. Calibration says that a given score means the same probability in every group, which matters for pricing, ranking and trust in the number. A decision also needs a valid target, an account of its consequences and a fair procedure, and the lecture reminds students that base rates reflect measurement, history and access rather than anything intrinsic to a group.

Study strategy

Assessment move

Learn the precision formula until you can write it without looking, then practise one move with it: choose two base rates, hold the error rates fixed, and compute both precisions. Doing this three times makes the impossibility result intuitive, and it gives you a two-line calculation you can include in any written analysis where two audits of the same system appear to disagree.

Working through Impossibility Results and Choosing a Fairness Metric in MAIE5103? Sia is AskSia’s AI Arts and Humanities tutor — ask any MAIE5103 Impossibility Results and Choosing a Fairness Metric question and get a clear, step-by-step explanation grounded in how MAIE5103 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.

A+Everything unlocked
Unlocks this Bible + all 13 of your subjects at The Hong Kong University of Science and Technology, plus 1,000+ more Bibles across the AskSia library.
Sia - your MAIE5103 tutor, unlimited, worked the way the exam marks it
The full 3-page Bible + practice bank with worked solutions
Chrome extension - sync your LMS so Sia knows your deadlines
Bilingual EN / Chinese on every Bible and every Sia answer
$0.99 Trial
30-day money-back · cancel in one tap · how it works