How to Read the Right Edge Without Fooling Yourself
The final question is how to interpret the newest part of an epidemic curve when the data are still arriving. The subject is nowcasting as a bridge between incomplete reports and time-sensitive public-health judgment; the focus is the distinction between a raw decline, a reconstructed current metric, and a downstream transmission signal. That distinction gives readers a practical rule: do not ask only whether the line is falling—ask what the line measures, how complete it is, how the missing part was estimated, and what the evidence can actually support.
A falling edge is first a claim about information
An epidemic curve is often drawn as if each point were a settled count of events on a date. At the newest edge, that assumption is usually the least safe. A visit, hospitalization, case, or death may have occurred, but the corresponding report may not yet have entered the surveillance system. Older dates have had more time to receive late reports. Newer dates have had less. The curve can therefore fall toward the present because the record is progressively less complete, even if the underlying event process is not falling.
This is the direct answer to the title’s paradox. Transmission can still be rising while a preliminary curve looks as though it is declining because the curve is mixing two processes: events happening in the population and information about those events arriving in the data. A delayed report about Monday can increase Monday’s count on Wednesday. It is new information about an old event date, not necessarily a new event on Wednesday. If the chart is read without that distinction, the movement of information is mistaken for the movement of disease.
The answer remains conditional. A falling preliminary curve may also reflect a genuine fall in events, a change in who is being observed, or a reporting process that has changed in an unusual way. Reporting delay is a serious alternative explanation, not a universal rebuttal to every decline. That is why nowcasting matters: it does not tell the reader to distrust all data; it supplies a way to estimate how much of the recent metric is likely still missing, using evidence about how comparable records filled in.
The difference between a raw point and a nowcast is therefore a difference in the status of the claim. A raw recent count says what has been reported so far for an event date. A nowcast combines that partial observation with a model of reporting delay to estimate a more complete current level. The estimate can be more useful for the present than the raw point, but it is not a report that has already arrived. It remains an inference about the portion of the record that has not yet been observed.
That inference should be accompanied by uncertainty. CDC describes prediction intervals as a common way to show the most plausible range of values given current data and prior information. The range matters because several eventual counts may be consistent with the reports seen so far. It also prevents a point estimate from appearing as settled as an older, more completely reported date. The interval does not identify the exact reports that will arrive tomorrow, but it keeps unresolved possibilities visible today.
A reader can now see why the newest edge needs a different reading from the middle of the curve. A historical point may have been revised many times and changed little in the latest snapshots. A current point may be only partly reported and still capable of moving. Comparing their heights without accounting for their different ages treats the reporting process as if it were uniform when it is not.
The reconstruction also depends on evidence outside the visible curve. Historical snapshots show how counts for earlier event dates accumulated as reports arrived. Event dates and report dates reveal the elapsed time associated with those revisions. Together, they allow a model to estimate the reporting pattern that normally applies at a given age of the record. Without such revision history, the current line contains less information about how much is missing from its newest dates.
This is why a completed historical curve is not simply “the truth” and a current nowcast is not simply “a guess.” The historical curve is more settled because it has received more reports.