Five out of five is not the same evidence as five hundred
If a model detects all five positive cases in one subgroup, the observed sensitivity is 100%. The estimate still has substantial uncertainty because only five opportunities were observed.
Reporting the percentage without the denominator hides that limitation.
Use the denominator for the metric
Sensitivity depends on positive cases. Specificity depends on negative cases. A subgroup with a thousand total records can still contain very few positives, making sensitivity poorly estimated.
| Report | Why it helps |
|---|---|
| Relevant event counts | Shows the information behind the percentage |
| Uncertainty intervals | Communicates estimation precision |
| Missing-output frequency | Reveals who receives no usable result |
| Group definition | Makes the comparison interpretable |
Avoid two opposite mistakes
A small observed difference is not proof of equal performance. A large observed difference is not automatically a stable population disparity. Both need context, uncertainty, and an appropriate analysis plan.
Subgroup evaluation also needs meaningful categories and consideration of overlapping characteristics. When evidence is insufficient, say so and plan additional data collection or limits on claims. A blank evidence gap should not be filled with reassurance merely because the overall model score is strong.
