What is Predictive Parity?
Predictive Parity is a group fairness criterion requiring positive predictive value to be equal across protected groups, so the proportion of observed positives among cases receiving a positive prediction is the same.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Use positive predictions as the denominator
For each group, calculate PPV=TP/(TP+FP) on cases whose outcomes are observable under the declared horizon. Publish the selected count and outcome-complete count; if a group receives no positive predictions, PPV is undefined rather than zero. Delayed repayment, diagnosis, or investigation labels require a fixed maturity window and a censoring policy before groups can be compared.
Separate threshold parity from score calibration
Equal PPV at one operating threshold does not show that a score of 0.7 has the same meaning across groups or that negative predictions are equally reliable. Pair thresholded Predictive Parity with reliability diagrams, proper scores such as Brier Score, negative predictive value where relevant, and uncertainty. Calibration can hold across score bins while a chosen threshold still creates different selection and error rates.
Expose the base-rate and error-rate tradeoff
Chouldechova's analysis shows that when observed prevalences differ, equal positive predictive value and balanced error rates generally cannot both hold for an imperfect risk score. This is not a license to treat observed prevalence as natural or immutable: investigate measurement, selection, access, and policy causes, then document which harms the chosen criterion addresses.
Key Characteristics
- Compares positive predictive value across protected groups
- Uses positive predictions, not all eligible cases, as the denominator
- Differs from selection-rate parity and conditional error-rate parity
- Is weaker than full score calibration or the complete Sufficiency condition
- Becomes undefined when a group has no positive predictions
- Can conflict with Equalized Odds when observed base rates differ
Common Use Cases
- Checking whether a positive fraud alert is equally reliable across groups
- Comparing confirmation rates among patients receiving a diagnostic alert
- Auditing whether approved risk flags have comparable observed meaning
- Monitoring PPV after a threshold or prevalence shift
- Evaluating alert-review workload when false positives impose group-specific costs
Example
Loading code...Frequently Asked Questions
How is Predictive Parity calculated?
For each group, divide true positive predictions by all positive predictions after outcomes have matured: `TP/(TP+FP)`. Report the positive-prediction denominator, label completeness, observation horizon, confidence interval, and pairwise or maximum gap. A missing denominator is undefined, not zero.
Is Predictive Parity the same as calibration?
No. Thresholded Predictive Parity compares PPV at one decision boundary. Calibration asks whether outcomes occur at the stated probability across score values. A score can be calibrated within groups while thresholded PPV or selection rates differ, depending on score distributions and thresholds.
How does Predictive Parity differ from Equalized Odds?
Predictive Parity conditions on the prediction and asks how often positive predictions are correct. Equalized Odds conditions on the observed outcome and compares true-positive and false-positive rates. Their denominators and operational harms differ, and both usually cannot be equalized when prevalences differ.
What happens if a group has no positive predictions?
Its PPV has a zero denominator and is undefined. Returning zero falsely describes evidence that does not exist. Report the missing support, inspect why no one was selected, and use a prespecified policy for aggregation, uncertainty, or withholding the comparison.
Can equal PPV still hide unfair treatment?
Yes. Groups can have equal PPV but different selection rates, false-negative rates, access to review, delay, or burden. Historical outcome labels may also reflect unequal opportunity. Pair PPV with process, label, error, utility, and contestability evidence.