Overview
Epidemic nowcasting explains why a falling raw curve can reflect delayed reporting rather than falling transmission. The series separates event and report dates, reconstructs the right-truncated edge from archived revisions, carries prediction intervals through metric-to-infection and R_t analyses, and compares fixed, smoothed, hierarchical, and time-varying models. Timely estimates are useful, but conditional on reporting comparability and stable data meaning.
Learning outcomes
- A downward raw right edge can reflect incomplete reporting, so it is not by itself evidence that transmission is declining.
- Archived snapshots reveal how event-date counts mature and provide evidence for estimating missing recent reports.
- A nowcast estimates a chosen metric with uncertainty; it does not turn hospitalizations or ED visits into direct infection counts.
- Model flexibility can follow changing delays, but intervals cannot repair broken pipelines or changed data definitions.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why can an epidemic curve fall while transmission is rising?
Recent event dates may be underreported because reports arrive later, making the raw right edge look lower even when events are increasing. That is a serious alternative explanation, not proof that transmission is rising.
What does an epidemic nowcast estimate?
An epidemic nowcast estimates a recent chosen metric after accounting for reports that have not yet arrived. It remains a model-based estimate with uncertainty, not a completed observation.
What evidence makes a nowcast credible?
A credible nowcast uses event- and report-dated data, archived snapshots of past revisions, relevant delay assumptions, and visible uncertainty. Its historical evidence must still resemble the current reporting process.
Can a nowcasted hospitalization count measure transmission directly?
No. A nowcasted hospitalization count estimates hospitalizations; inferring infections or R_t additionally requires infection-to-observation timing and a generation-interval model.