When a Falling Curve Is Not a Falling Epidemic
The subject of this series is epidemic nowcasting: the use of incomplete, recently reported disease data to estimate what is happening now. The specific question is deceptively simple: why can the newest part of an epidemic curve slope downward even when transmission is still rising? The answer supports a practical judgment for anyone reading a public-health dashboard: before treating a recent decline as evidence that an outbreak is easing, ask whether the reporting process has had time to catch up with the events themselves.
That does not mean every decline is an illusion. It means the newest points in the series are not measured under the same conditions as older points. The older portion has had more time to receive late reports. The right edge has not. A curve can therefore contain two different kinds of information at once: a more settled record of earlier events and a provisional, incomplete record of the most recent ones.
The newest numbers are not the newest events
A disease event and a report about that event have different dates. A person may visit an emergency department on one day, while that visit enters a surveillance system later. A hospitalization may be recorded after a delay caused by clinical documentation, data processing, or the transfer of information between organizations. The event belongs to the date on which it occurred; the report belongs to the date on which the system learned about it. A time series organized by event date will change when those late reports arrive.
This creates a simple but powerful distortion. Imagine a surveillance system displaying events by the day they occurred. On the first day, it has received only some of Monday’s reports and perhaps very few of Tuesday’s. On the next day, more Monday reports arrive, along with the first Tuesday reports. The Monday total is revised upward, while Tuesday may still look small. On a later day, Tuesday and Wednesday are revised in turn. If someone reads each snapshot as a completed account of its event date, the newest dates will appear to have fewer events than dates that are merely older.
That is why an unadjusted curve often bends down at its right edge. The slope may reflect the age of the records rather than a change in the disease process. Public-health data systems can show a preliminary fall simply because reports for recent events are still in transit. When the records fill in, the final reported pattern may look very different from the early snapshot.
The distinction matters most when a decision depends on the present rather than on a history that will be complete weeks from now. Hospital capacity, emergency planning, public communication, and outbreak response cannot always wait for every late record. Yet acting on raw recent counts treats missing reports as if they were evidence of fewer events. The problem is not that the data are useless. It is that their incompleteness is structured, and that structure must be recognized before the newest slope is interpreted.
The horizontal axis can conceal this problem. A chart organized by event date appears to describe when illness or care occurred, but each point is also a snapshot of how much the reporting system knew at the time the chart was made. A report timestamp answers a different question: when did information enter the system? If those two clocks are silently treated as one, an information delay is mistaken for a change in incidence. The curve is not lying; it is answering two questions at once, and the reader has to separate them.
The diagram separates the event process from the reporting process. Later reports do not necessarily represent new events on the old date; they can be new information about events that already happened. A nowcast is an attempt to use the pattern of that information flow before the record becomes complete.
What nowcasting changes—and what it does not
Nowcasting is a model-based estimate of a current disease metric after accounting for incomplete recent reporting. The metric might be a case count, an emergency-department visit, a hospitalization, or a death. The method does not pretend that the unavailable reports have been observed. It uses currently available reports together with historical reporting patterns to estimate how much of the recent record is likely still missing.
The basic idea can be made concrete without turning an example into a rule. Suppose historical data suggest that only 30 percent of emergency-department visits are reported on the same day as the visit. If a current snapshot contains 30 same-day reports, a simple intuition is that the eventual total may be larger because an expected share has not yet arrived. Real reporting is more complicated: the delay can vary by elapsed time, location, facility, season, and the way the surveillance system operates. A nowcasting model tries to represent those variations rather than applying one fixed multiplier to every date.
The result is an estimate, not a recovered fact. CDC describes nowcasts as outputs with uncertainty bounds, commonly expressed through a prediction interval: a range of plausible values given the data available and prior information. The interval is essential because the model is reasoning about events that have not yet been fully observed. A point estimate without its uncertainty can make an incomplete count look as definite as a completed historical count.
Uncertainty changes how a reader should describe a turn in the curve. A small movement in a point estimate is not automatically a meaningful change in current conditions, especially when the recent edge is still being reconstructed. The interval does not tell a reader which future report will arrive or guarantee that the model has anticipated an unusual disruption. It does, however, make the unresolved part of the estimate visible. That is a better basis for a cautious decision than a single precise-looking number.
This also defines the boundary of what the method says. A nowcast of hospitalizations is an estimate of hospitalizations, not a direct count of infections. A nowcast of emergency-department visits is not automatically a measurement of transmission. The observed metric can be useful for learning about an epidemic, but additional models and assumptions are needed to connect it to infections, and then to measures such as the time-varying reproduction number, R t. Nowcasting repairs one part of the timing problem; it does not remove every delay between infection and observation or settle every question about cause.
There is a genuine choice behind the method. One option is to exclude the newest days until reporting is nearly complete. That can reduce the distortion caused by late reports, but it also means analyzing the past when decision-makers need information about the present. The other option is to estimate the incomplete edge and accept a model-dependent range of uncertainty. Neither choice is free. Waiting gives more observed data but less timely knowledge. Nowcasting gives earlier situational awareness but requires confidence in the reporting patterns and assumptions used to reconstruct what has not arrived.
A better question for the right edge
The most useful change is not to stop looking at epidemic curves. It is to ask a better question of them. Instead of asking only, “Is the line going down?”, ask, “How complete are the reports for these dates, and what does the surveillance system usually learn later?” The answer requires evidence about the reporting process: records tied to event dates, report dates showing when information arrived, and—ideally—archived snapshots showing how earlier versions changed.
That evidence turns an apparent visual contradiction into a measurable problem. If reporting is predictably delayed, the missing portion can be estimated. If delays differ across locations, the model may need to account for site-specific behavior. If delays change over time, historical patterns may no longer transfer cleanly to the current period. If the system has suffered a disruption or changed its data definitions, a neat uncertainty interval cannot by itself guarantee that the reconstruction is sound.
A completed historical curve is therefore not “better” simply because it is older, and a current estimate is not “worse” simply because it is modeled. They answer different operational needs. The older curve offers more settled observation. The nowcast offers a reasoned estimate during the period when decisions are most time-sensitive. The right comparison is not certainty versus uncertainty, but delayed observation versus timely inference whose assumptions remain visible.
The four-part mental model is therefore straightforward:
- Events: what happened in the population, organized by the date of occurrence.
- Reports: what the surveillance system has recorded so far, arriving on a different timetable.
- Estimation: a nowcast that uses the reporting process to infer the likely current level and states uncertainty.
- Decision: an action or interpretation that must account for both timeliness and the limits of the estimate.
Keeping these layers separate prevents two opposite mistakes. The first is to read incomplete reports as completed events and conclude too quickly that transmission is declining. The second is to treat a nowcast as a transparent window onto reality and forget that it is still an inference. A responsible reading lies between those errors: recognize the raw right edge as provisional, use a model when timely information matters, and carry the interval and the data’s history into the conclusion.
The central lesson is not that an epidemic curve cannot fall. It is that the newest downward slope may describe the movement of information into the surveillance system rather than the movement of transmission through a population. To decide which explanation is more plausible, the next essential distinction is between the clock of the event and the clock of its report.