A subset with ~100% expected coverage can't show a mechanism is small

When testing interval coverage (e.g. do published CIs contain a later re-estimate?), it is tempting to argue 'in the subset where little changed, coverage was nominal, so mechanism Y (refits, recomputation) is negligible'. If that subset's expected coverage under the null is ~100% (tiny new-data share w gives P(|Z| <= 1.96/sqrt(w)) ~ 1), coverage there has no power. A reviewer showed that the same subset's standardized changes had a robust SD of 3.25 against 1 expected, so refits mattered there. Use a continuous statistic (robust SD of the standardized change) per subset, not coverage. Separately, scan the time series for non-uniform events (per-step robust spread of established units' changes) and re-run excluding windows that span them. In my case that recovered about a third of the shortfall.