You Can't Have All the Fairness Metrics: Choosing Between Parity and Equalized Odds

Demographic parity, equalized odds, and calibration can't all hold at once when outcomes genuinely differ by group — that's a mathematical certainty, not a modeling failure. Choosing among them means choosing which kind of error you'll tolerate, which makes it a values call for the product leader, not a metric a data scientist can pick unilaterally.

Quick answer: Demographic parity, equalized odds, and calibration mathematically conflict whenever a protected group's base rate differs from another's and the model isn't perfect. Pick the metric that matches the harm you most want to prevent, then document the tradeoffs you accepted — don't let the choice default to whatever the model shipped with.

Why You Can't Satisfy Every Fairness Metric at Once

The impossibility result, proven independently by Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan (2016) and by Alexandra Chouldechova (2017), shows that when a group's base rate for an outcome differs from another group's, no classifier short of a perfect one can simultaneously satisfy demographic parity, equalized odds, and calibration. You must choose.

Three technical properties are in tension, and you can typically hold any two of them at once:

• Demographic parity — equal selection rates across groups.
• Equalized odds — equal true-positive and false-positive rates across groups.
• Calibration — a given score means the same real-world probability regardless of group.

Hold two of these and the third typically breaks — except in the narrow case of equal base rates across groups, or a flawless model, neither of which describes most real deployments.

This isn't a limitation of your data science team, your labeled data, or your feature set. It's arithmetic. If two groups have different real-world rates of the outcome you're predicting — loan default, on-the-job success, recidivism — then forcing equal outcomes (parity), forcing equal error rates (equalized odds), and preserving equal score meaning (calibration) pull the decision threshold in three different directions at once.

This sits inside the broader discipline of responsible AI product management. For the full landscape this specific tradeoff fits into, see the complete guide to responsible AI for product teams.

Why base rates differ in the first place matters for how you think about this. A group's lower historical approval rate, lower average credit score, or lower callback rate is rarely a neutral fact. It's frequently the residue of unequal access to credit history, unequal access to referral networks, or biased historical labels the model was trained on.

Choosing a fairness metric doesn't fix that upstream inequality — it decides how your model responds to it.

The Three Fairness Definitions, Explained With a Lending Example

Demographic parity requires equal approval rates across groups, whatever the underlying qualification mix. Equalized odds requires equal true-positive and false-positive rates, so equally creditworthy applicants get equal treatment regardless of group. Calibration requires a given risk score to mean the same real default probability everywhere. Each protects a different population from a different harm.

Demographic Parity: Equal Outcomes, Whatever the Inputs

Imagine a lending model approving loans for Group A and Group B applicants. Demographic parity is satisfied when both groups are approved at the same rate — say, 40% — regardless of how their underlying default risk compares. It's the metric that most directly answers "does this group get shut out of credit?"

The cost: if the two groups genuinely have different average risk profiles, hitting equal approval rates means the threshold for approval is not the same score cutoff in each group. Some lower-risk Group A applicants get rejected to hit the target, or some higher-risk Group B applicants get approved to hit it — or both.

Equalized Odds: Equal Errors, Whatever the Outcomes

Equalized odds asks a different question: among applicants who would actually repay, are they approved at the same rate across groups? And among applicants who would actually default, are they rejected at the same rate across groups? It holds the error rates constant rather than the approval rate.

The cost: if Group B has a higher true default rate than Group A, equalized odds can still leave Group B with a noticeably lower overall approval rate — the model is treating equally-risky people equally, but the two groups aren't equally risky on average, so the aggregate numbers diverge. That divergence is exactly what triggers disparate-impact scrutiny under U.S. fair-lending law, even when the underlying error rates are identical.

Calibration: A Score That Means the Same Thing Everywhere

Calibration requires that among all applicants who get a "15% default risk" score, roughly 15% actually default — in every group. It protects the integrity of the number itself, which matters enormously once a loan officer or hiring manager starts reading scores literally.

The cost: calibration says nothing about whether the errors that do occur are evenly distributed across groups. A perfectly calibrated model can still produce more false denials for one group than another, because calibration constrains the meaning of the score, not the downstream consequences of acting on it.

A Hiring Example: How the Same Model Passes One Test and Fails Another

In a hypothetical resume-screening model, tuning for equal true-positive and false-positive rates across groups (equalized odds) can leave overall selection rates 10-15 points apart when the groups' underlying qualification rates differ — failing demographic parity even though every individual error rate is identical. The reverse threshold produces the opposite failure.

The numbers below are an illustrative, hypothetical scenario built to show the mechanism — not measured data from any real hiring system. Picture two applicant pools of 1,000 each: Group A has 400 truly qualified candidates (a 40% base rate), Group B has 200 (a 20% base rate), reflecting an upstream gap in access to the credentials the screen rewards.

Model tuned for equalized odds (same 75% true-positive rate and 10% false-positive rate in both groups):

• Group A: 300 qualified candidates correctly selected, 60 unqualified candidates incorrectly selected → 360 total selected, a 36% selection rate.
• Group B: 150 qualified candidates correctly selected, 80 unqualified candidates incorrectly selected → 230 total selected, a 23% selection rate.
• Result: identical error rates, but a 13-point selection-rate gap — demographic parity fails.

Model tuned for demographic parity (both groups selected at a flat 30% rate, 300 candidates each):

• Group A: 270 of 400 qualified candidates selected (a 68% true-positive rate), 30 unqualified candidates selected (a 5% false-positive rate).
• Group B: 180 of 200 qualified candidates selected (a 90% true-positive rate), 120 unqualified candidates selected (a 15% false-positive rate).
• Result: identical 30% selection rates, but a 22-point gap in true-positive rates and a 10-point gap in false-positive rates — equalized odds fails, and it fails in opposite directions for the two groups.

Notice the second scenario cuts both ways: Group A's genuinely qualified candidates face a harder bar (32% of them get rejected) while Group B's genuinely unqualified candidates face an easier one (15% still get selected). Neither group is uniformly better or worse off — the unfairness just moves to a different population and a different failure mode.

Mapping exactly where in an applicant's path this threshold gets applied — application, resume screen, interview invite, offer — is its own useful exercise. The general technique for it is covered in our guide to customer journey mapping.

A Decision Guide: Match the Metric to the Harm You Most Want to Prevent

Choose demographic parity when the harm you fear most is a whole group's structural exclusion from opportunity. Choose equalized odds, or its one-sided cousin equal opportunity, when the harm you fear most is unequal treatment of equally qualified individuals. Choose calibration when a human downstream will act directly on the score itself.

This is a jobs-to-be-done question as much as a statistics one. Each fairness metric is "hired" to prevent a specific harm for a specific stakeholder, and different people in the room are often optimizing for different jobs without realizing it — the same reframing that makes the complete guide to Jobs-to-Be-Done useful for untangling conflicting stakeholder requirements generally.

1. Check what the law already decided for you. If a regulator or your legal team applies a bright-line disparate-impact test — the U.S. EEOC's four-fifths rule for hiring selection rates, or a fair-lending exam under the Equal Credit Opportunity Act — something close to demographic parity may be the metric you're held to regardless of your statistical preference.
2. Ask who bears the cost of a false rejection. If the primary harm is an individually qualified person losing an opportunity, equalized odds or equal opportunity (which only constrains the true-positive rate) protects that person more directly than parity does.
3. Ask who reads the number. If the score itself reaches a human decision-maker — a loan officer, a hiring manager — who will interpret "72% risk" as a literal probability, calibration keeps that number honest across every group they'll see it applied to.
4. If you can't have your first choice cleanly, measure and disclose the gap on the metrics you didn't pick. An undisclosed tradeoff is what turns a defensible engineering decision into a discovery-phase liability later.

Why This Is a Values Decision, Not a Math Problem

No optimization run, dataset fix, or better algorithm resolves the conflict between these metrics, because the conflict is mathematical, not a bug. Someone has to decide, in advance and on the record, which population's errors the organization is more willing to bear — a legal, ethical, and business judgment that belongs with a cross-functional decision-maker, not a threshold buried in a model card.

The National Institute of Standards and Technology makes the same point institutionally. NIST's Towards a Standard for Identifying and Managing Bias in Artificial Intelligence (SP 1270) catalogs more than 20 competing mathematical fairness definitions in circulation and deliberately declines to endorse a single one — because the right choice depends on context, stakes, and who's harmed by which kind of error, not on which formula is more elegant.

Regulation Raises the Stakes on Getting This Documented

Both employment and creditworthiness systems fall under the European Union's high-risk AI classification in the EU AI Act, which means the fairness metric you pick — and your documented reasoning for picking it — can become a compliance artifact a regulator requests, not just an internal engineering note. Our EU AI Act map for product managers walks through the specific obligations that attach to high-risk classification, including documentation duties that make this exact decision auditable.

Practically, that means the choice needs a paper trail with four things in it:

1. Which metric you optimized for, named explicitly, not implied by the threshold you shipped.
2. Which harm you were protecting against, and which harms you accepted as a tradeoff.
3. Who was in the room when the decision was made — this shouldn't be a unilateral data science call, and shouldn't be a unilateral product call either.
4. When you'll revisit it, because a population, product surface, or regulatory environment that shifts can quietly invalidate a decision that was sound when you made it.

A recurring audit cadence — not a one-time decision — is the mechanism that catches that drift; our guide to AI fairness audits for product teams covers how to structure one. If the score or decision reaches an end user directly, how you communicate its confidence matters as much as how you computed it — a calibrated 15% risk score presented as false precision misleads people the same way an uncalibrated one does, which is the subject of our piece on honest confidence and AI uncertainty in the UX.

Documenting the Choice So It Doesn't Get Re-litigated Every Quarter

Whatever metric you choose, it will get questioned — by a new data scientist, a new stakeholder, an auditor, or your own future self eighteen months later asking why the model behaves this way for this group. The failure mode usually isn't picking the wrong metric once. It's re-litigating the same tradeoff from scratch every time someone asks, because nobody wrote down the reasoning the first time.

That's a documentation problem as much as a fairness problem, and it's the specific gap Prodinja's Journals are built to close. A Journal entry can capture a Reflection recording the fairness definition you chose, the harm you were optimizing against, who signed off, and what you knowingly traded away — so the next person who inherits the model finds a dated rationale instead of an unexplained threshold. It's a place to record a contested decision, not a system that makes the decision for you.

Key Takeaways

• The conflict is mathematical, not a bug. Kleinberg, Mullainathan, and Raghavan (2016) and Chouldechova (2017) independently proved you can't have demographic parity, equalized odds, and calibration together once base rates differ and the model isn't perfect.
• Demographic parity protects group-level representation but can require rejecting qualified individuals or approving unqualified ones to hit an equal rate.
• Equalized odds protects individual-level equal treatment among truly qualified applicants, but overall selection rates can still diverge sharply enough to trigger disparate-impact review.
• Calibration protects the meaning of the score for whoever reads it, but is silent on whether the resulting errors land evenly across groups.
• Match the metric to the harm, not the harm to the metric — structural exclusion favors parity, unequal individual treatment favors equalized odds, and a literal reading of the score favors calibration.
• Regulation sometimes decides for you. The EEOC's four-fifths rule and fair-lending disparate-impact tests can function as a legal floor regardless of your statistical preference.
• Document the choice, don't just make it. A dated rationale — including who decided and what was traded away — is what turns a defensible decision into a demonstrably defensible one later.