Overview

A prediction interval expresses uncertainty about an incomplete count under the model’s representation of reporting; it does not automatically detect an interrupted pipeline or a changed data definition. A surprising edge may reflect predictable calendar or site effects, gradual reporting drift, an abrupt disruption, or a change in what the data mean. Archives, operational context, and cross-location evidence help distinguish these possibilities, while waiting for later reports trades timeliness for more settled observation.

Learning outcomes

A wide interval can reveal limited information, but a narrow interval is not a guarantee when historical reporting patterns no longer apply.
Operational disruptions and definition changes require context beyond the count; model uncertainty cannot identify or repair them.
Archives, operational knowledge, and location-specific evidence help distinguish ordinary variation from drift or breakdown.

Keywords

NowcastingPrediction intervalDisruptionDefinition changeArchivesComparability

What the Interval Cannot Fix

A nowcast can be carefully built and still become difficult to interpret when the reporting process changes underneath it. The focus here is diagnosis: separating ordinary variation that a model can represent from an abrupt disruption or data-definition change that historical evidence may not describe. That distinction helps readers use an uncertainty interval as information about an incomplete count without treating it as a guarantee that every failure in the surveillance system has been captured.

Not every surprising edge means the same thing

A surprising decline in a recent curve is a symptom, not a diagnosis. The first possible cause is ordinary variation in how reports arrive. CDC identifies day-of-week and holiday effects, as well as differences among population groups and health facilities, as patterns that can often be modeled. A hospital may report more slowly on a weekend, or one facility may consistently take longer than another to transmit information. Such variation can alter the newest counts without representing a sudden change in disease activity.

A second cause is gradual reporting drift. Staffing, reporting policies, or the operation of a surveillance system can change over time. The delay pattern learned from older records may become less representative of the current period, even if the change is smooth rather than dramatic. A model with a time-varying reporting function or site-specific effects may be able to follow some of that movement. But “can adapt” does not mean “will adapt perfectly.” The current data must contain enough information for the model to distinguish a changing reporting process from a changing event process.

A third cause is an abrupt operational disruption. CDC names cyberattacks, software updates, natural disasters, and changes in hospital operations or staffing as examples that can affect data storage, sharing, or reporting. These events can create a sudden gap or backlog in the reporting stream. A raw curve may fall because reports stopped arriving normally. A nowcast trained on ordinary historical delays may then face a condition for which its historical pattern is a poor guide.

These causes can look similar on a chart. A holiday effect, a slow change in reporting behavior, and a software problem can all produce fewer reports for recent event dates. The difference lies in the reporting process around the curve. Are the changes recurring on a calendar pattern? Do they occur at one facility or across many locations? Do later snapshots show the usual gradual catch-up, or is there an unusual gap? Has the surveillance system changed how it receives or defines the data? The curve alone cannot answer these questions.

stateDiagram-v2 [*] --> StablePattern StablePattern --> PredictableVariation: calendar or site effect PredictableVariation --> StablePattern: modeled pattern remains relevant StablePattern --> GradualDrift: reporting behavior changes GradualDrift --> CautiousAdaptation: time-varying or site-specific model CautiousAdaptation --> StablePattern: comparability returns StablePattern --> AbruptDisruption: operational failure AbruptDisruption --> CautiousInterpretation: historical delay may not apply StablePattern --> DefinitionChange: data meaning changes DefinitionChange --> Reconciliation: align historical and current structure Reconciliation --> CautiousInterpretation: comparability still needs checking

The states in the diagram are not a formal decision rule.

Published by DeepMox

FAQ

Why can a narrow nowcast interval still be misleading?

A narrow interval can still rely on a historical delay pattern that no longer describes the current reporting process. It expresses modeled uncertainty, not an inspection of every upstream condition.

How can you distinguish reporting drift from an abrupt disruption?

Compare revision timing and recurrence with operational information, affected locations, and whether later reports catch up normally. The curve alone cannot reliably separate gradual drift from a sudden failure.

Why do data-definition changes matter for nowcasting?

They can make historical and current counts refer to different quantities, so an interval cannot establish that the values remain comparable. Historical fields may need reconciliation with the current structure.

Previous10 minChoosing a Reporting-Delay Model

Next8 minHow to Read an Epidemic Curve’s Right Edge

What Prediction Intervals Cannot Fix in Nowcasting

What the Interval Cannot Fix

Not every surprising edge means the same thing

FAQ