Benign vs harmful shift

A difference between the score distributions of the reference group (source) and the group under review (target) can be detectable without being harmful. Monitoring often needs a narrower question: did a score tied to the outcome move in a direction marked harmful? One score can miss a change another reveals.

Why equality tests can raise unhelpful alarms

Tests of equal score distributions ask whether source and target match; they do not ask whether a difference matters. With enough observations, even a small, harmless difference can produce a small p-value. If every such alarm prompts retraining, teams can spend effort without improving outcomes. Polyzotis et al. (2019) put it directly: distribution tests at ML scale are "too sensitive and also uninformative." The harm test takes a narrower view: it tests movement toward a tail you declared harmful in advance.

A test for no adverse shift

test_harm tests the null of no adverse shift in the direction you declared: target has not moved farther into the harmful tail. It does not test whether source and target score distributions are identical (Kamulete, 2022). The question is related to clinical noninferiority, but the test is not a clinical noninferiority test. It assumes no particular score distribution and uses no prespecified harm margin.

You define worse with one score — risk, error, confidence, outlier score — and a worse direction declared before testing. Higher risk is worse="higher". Lower confidence is worse="lower". Pick that direction from the score's meaning, never from the p-values.

A small p-value is evidence against the no-adverse-shift null in the direction you declared. A large p-value does not establish equivalence or prove that target is fine.

What that focus costs

test_shift uses ROC AUC to test whether a randomly chosen target score tends to be higher than a randomly chosen source score more or less often than chance. It is not an omnibus test: distributions can differ in shape while their AUC remains at chance (0.5). test_harm gives more weight to thresholds the source rarely exceeds. The ROC form is in How the harm test works. That focus is the point—and the trade-off. A tail test can miss a shift outside the harmful tail, including one you still care about. Run the two tests separately; a detectable shift is not automatically harmful.

How the two tests fit

• Use test_shift as a two-sided screen for rank-based separation.
• Use test_harm with a pre-declared worse when only movement toward the harmful tail warrants intervention.
• Both hold scores fixed and permute group labels. A small p-value means the observed statistic would be unusual if source and target labels were interchangeable—that is, unrelated to the pooled scores. It is not a measure of business impact.

See How the harm test works for the ROC derivation, Get started for the two-test workflow, Is the new drug good enough? for the same question with no model, and Monitor a credit model for risk, confidence, and error side by side.

References

Kamulete, V. M. (2022). Test for non-negligible adverse shifts. Proceedings of UAI, PMLR 180:959-968. Polyzotis, N. et al. (2019). Data validation for machine learning. Proc. MLSys 1:334-347.