Predictive Parity and Calibration
What you'll learn: How to verify that your model's confidence scores are equally trustworthy across different demographic groups.
The Core Idea
Imagine a weather app that says "70% chance of rain." If it rains on 7 out of 10 such days, the app is calibrated—its confidence matches reality. Now imagine this app is more accurate for city dwellers but less reliable for rural users. That's a calibration fairness problem.
Predictive parity asks: when your model predicts a certain probability (say, 80% chance of loan default), does that prediction mean the same thing across different groups? If the model says "80% risk" for Group A and Group B, we want roughly 80% of both groups to actually default—not 80% of Group A but 60% of Group B.
Calibration measures whether predicted probabilities align with observed outcomes. Calibration parity (closely related to predictive parity) demands this alignment holds equally well for all protected groups.
Why It Matters
A biased calibration means the model is overconfident or underconfident for certain groups. For instance:
- A hiring model might underestimate qualified candidates from Group A (low predicted scores, but high actual success rate)
- A medical diagnosis tool might overestimate risk for Group B (high predicted probabilities, but lower actual disease prevalence)
This creates unfair treatment even if other metrics like demographic parity are satisfied.
Measuring Calibration
You group predictions into bins (e.g., all predictions between 0.7–0.8) and compare the average predicted probability to the actual positive rate in each bin, separately for each demographic group. Large discrepancies signal calibration issues.
Key Takeaway: Predictive parity ensures your model's probability estimates are equally reliable across groups—a 70% prediction should mean 70% actual outcome rate for everyone, not just some groups.