Quantifying performance disparity across subgroups

What's really being asked

This checks whether you know that aggregate accuracy can mask severe underperformance on a minority subgroup, especially when that group is small, and whether you can choose disaggregated, fairness-aware metrics.

The full answer

The core move is to stop looking at a single global number and slice every metric by the group of concern. For each group compute precision, recall, false positive rate, and false negative rate, because these reveal asymmetric harms, for example a model that misses far more true positives for the minority group has a high false negative rate there even if overall accuracy is high. Then compare across groups with fairness criteria: equal opportunity asks whether recall is equal across groups, equalized odds adds equal false positive rates, and demographic parity compares positive prediction rates. Quantify the disparity as the gap or ratio between groups. For visualization, use grouped bar charts showing each metric side by side per group so the gaps are obvious, present per-group confusion matrices, and overlay per-group ROC or precision-recall curves to compare discriminative power. A calibration plot per group can also show if predicted probabilities are reliable for each.

The mistakes people make

Reporting overall accuracy or a single aggregate metric, which is exactly what hides the problem. Computing only one metric like precision without recall, missing the type of error that harms the group. Not slicing by the subgroup at all. Choosing a fairness definition without stating which harm it targets.

What usually comes next

Which fairness definition fits this use case and why can they conflict? How would you fix the disparity? What if the group is too small for stable estimates?

A concrete example

A loan model is ninety-two percent accurate overall, but slicing shows recall of eighty-five percent for the majority group and fifty-eight percent for a minority group, meaning many qualified minority applicants are wrongly rejected. You plot a grouped bar chart of recall and false negative rate by group, making the twenty-seven-point recall gap unmissable, and you report the equal-opportunity difference as the headline fairness metric.