AI Glossary / Governance, ethics and safety / Intermediate

Predictive parity

Definition

Predictive parity is a group fairness criterion requiring equal precision, also called positive predictive value, across the groups being compared. Among cases predicted positive, the proportion that is actually positive should be the same for each group.

What it means

The criterion asks how reliable a positive classification is for different groups. If a system flags cases for a particular condition, predictive parity compares the fraction of correct flags within each group. It does not compare how many people are flagged or how many actual positive cases are missed.

Precision is calculated as true positives divided by all predicted positives: TP / (TP + FP). The meaning of a positive label must be explicit. In some applications, positive means a beneficial selection; in others, it means a warning or an undesirable condition. The word itself does not establish whether the outcome is good.

Predictive parity is one formal criterion, not a complete ethical judgment. Its relevance depends on the decision, quality of the reference labels and consequences of errors. Equal precision can coexist with substantial differences in recall and other outcomes.

How it works

Define the groups, positive class and evaluation period. For each group, count true positives and false positives among the cases classified as positive. Divide to obtain precision, then compare the resulting rates with suitable attention to sample size and uncertainty.

The denominator matters. A group with no predicted positives has undefined precision under the usual formula, not automatically perfect precision. A group with only a few positive predictions may have a highly uncertain estimate. Missing or biased outcome labels can also make the comparison misleading.

Examine additional error rates and relevant consequences. Precision depends on the prevalence of the target condition as well as the classifier's true- and false-positive rates. When prevalence differs, common fairness criteria can conflict except in special cases. Choosing a criterion therefore requires substantive reasoning; an optimization routine cannot decide which tradeoff is appropriate on its own.

Three practical examples

• A review system compares the share of correct flags among flagged cases across specified user groups.
• A quality-assurance team checks whether a positive prediction has comparable reliability across operating settings, while separately measuring missed cases.
• An evaluator reports precision alongside recall and false-positive rates so stakeholders can see what a parity result does and does not establish.

Think of it this way

Think of two baskets labeled “ripe fruit.” Predictive parity asks whether the same proportion of fruit in each basket is actually ripe. It says nothing about how much ripe fruit was left outside either basket.

The analogy is limited because decisions about people involve rights, context and consequences that a sorting example cannot capture. It only clarifies the denominator in the precision calculation.

Worked example

Equal precision with unequal recall

In an invented evaluation, Group A has 100 actual positive cases. The model correctly flags 80 and incorrectly flags 20 actual negative cases. Precision is 80 / (80 + 20) = 80%, while recall is 80 / 100 = 80%.

Group B also has 100 actual positive cases. The model correctly flags 40 and incorrectly flags 10 actual negative cases. Precision is 40 / (40 + 10) = 80%, while recall is 40 / 100 = 40%.

The two groups have predictive parity in these observed counts. However, the model misses 20 actual positives in A and 60 in B. If identifying those cases is important, the recall gap matters. Equal precision alone cannot settle whether the system meets the intended service goal.

Illustrative case study

Fictional case: Cedar Review introduces a classifier that directs requests to a specialist queue. Its initial fairness report highlights equal precision across two groups and concludes that the process treats them equivalently.

A reviewer asks for the complete confusion counts. These reveal that one group has many more missed cases relative to its actual positives. The team revises the report to separate precision parity from coverage of the cases needing specialist attention. It examines both the model and the surrounding access process.

The example demonstrates a reporting correction, not a recommendation to optimize recall at any cost. The appropriate decision depends on the consequences and constraints.

Common misconceptions and mistakes

• Predictive parity means equal accuracy. Accuracy includes true negatives and uses a different denominator.
• Equal precision means equal recall. The worked example shows that these can differ substantially.
• Satisfying one fairness criterion proves nondiscrimination. A numerical criterion has a limited scope and does not replace contextual or legal assessment.

Business use

Businesses can use predictive parity when the reliability of a positive classification is a relevant concern. Report the actual counts, uncertainty and label definition so that users understand the comparison. Avoid presenting percentages without denominators, especially for small groups.

Consider what happens after a flag. Review capacity, error correction and the burden of false positives can affect the practical significance of the metric. Also examine those who are not flagged. A fair-looking positive queue can coexist with unequal access to assistance among people the classifier misses.

People also ask

Is predictive parity another name for precision parity?

In this common binary-classification usage, yes.

Does it require equal positive-selection rates?

No. Groups can have equal precision while very different fractions are selected.

Can it always coexist with equalized odds?

No. Different target prevalence and imperfect prediction can create incompatibilities; the assumptions and special cases matter.

Check your understanding

Group A has 18 correct positives among 20 positive predictions. Group B has 27 among 30. Do these observed counts satisfy predictive parity?

Answer: Both precision values are 90%, so they match on this sample. That does not establish equal recall, overall fairness or the certainty of future rates.

Related terms

• Precision
• Recall
• Equalized odds
• Demographic parity
• Fairness metric
• positive predictive value

Learn next

Read precision and recall. Build a small confusion table for each relevant group and identify which people appear in the denominator of each metric before comparing rates.

• Precision
• Recall

More in Governance, ethics and safety

• Practical significance - Practical significance is the importance of an observed or estimated effect in its real-world context.
• Pre-processing algorithm - A pre-processing algorithm is a defined procedure that prepares or transforms data before a model uses it.
• Prediction bias - Prediction bias is the difference between the average of a model's predictions and the average of the corresponding observed labels in a dataset.
• Preference alignment - Preference alignment is the process of adjusting an AI system's behavior to better match specified preferences about its outputs or actions.
• Presentation Bias - Presentation bias is a distortion in behavior or resulting data caused by how information is displayed.
• Privacy - Privacy concerns people's interests in how information about them and their lives is observed, collected, inferred, used and shared.

Sources

Google Machine Learning Glossary, predictive parity: https://developers.google.com/machine-learning/glossary/responsible-ai#predictive-parity. Hardt, Price and Srebro, Equality of Opportunity in Supervised Learning: https://arxiv.org/abs/1610.02413. Relevant primary glossary and research abstract text were reviewed for metric scope and limits of group-statistical criteria. All numerical examples and the fictional case are original illustrations.

Previous: Predictive modeling Next: Predictive state representation