Predictive Parity
What you'll learn: You'll understand how predictive parity ensures that model predictions carry the same meaning across different demographic groups.
The Core Idea
Predictive parity, also called the outcome test, requires that when your model makes a positive prediction, that prediction should be equally reliable regardless of which group the individual belongs to. In technical terms: positive predictive value (PPV) must be equal across all protected groups.
Positive predictive value answers: "Of all the cases we predicted positive, what fraction actually were positive?" It's also known as precision.
Why This Matters
Imagine a loan approval model. If it approves 100 applicants from Group A and 80 actually repay (PPV = 0.80), but approves 100 from Group B where only 50 repay (PPV = 0.50), then the model's "yes" means different things for different groups. That's unfair: Group B faces lower-quality predictions, and lenders acting on these predictions would face different risk profiles.
Predictive parity says: PPV(Group A) = PPV(Group B) = PPV(Group C) = ...
The Formula
For each group g:
PPV(g) = True Positives(g) / (True Positives(g) + False Positives(g))
Predictive parity holds when PPV is constant across groups.
The Trust Principle
This fairness criterion is about trust calibration. When someone receives a positive prediction, they should be able to trust it equally, regardless of their demographic background. It's particularly important in settings where stakeholders rely on prediction quality (medical diagnosis, credit decisions).
Key Takeaway: Predictive parity ensures that positive predictions have equal positive predictive value across all groups, making your model's outputs equally meaningful and trustworthy for everyone.