Two Clocks Behind Every Recent Count
The key to epidemic nowcasting is to separate two clocks that are easy to blend together: the clock of the disease event and the clock of the report about that event. The practical question is not merely when a case, visit, or hospitalization occurred, but how much of the information about that date had reached the surveillance system when the data were viewed. Once those clocks are visible, the apparent decline at the right edge becomes a measurable problem rather than a mysterious flaw in a dashboard.
An event date is not a report date
An event date answers, “When did this happen?” A report date answers, “When did the surveillance system record information about it?” Those dates can be separated by hours, days, or longer. The event may be a visit to an emergency department, a hospitalization, a diagnosed case, or a death. The report is the administrative or electronic trace that allows the surveillance system to count it.
That distinction matters because a time series can be organized by the first date while being observed through the second. If a chart groups emergency-department visits by the day of the visit, a visit that occurred on Monday belongs in Monday’s total even if the system does not receive the report until Wednesday. The Monday value is therefore not fixed when Monday ends. It is a temporary count that can rise as information arrives.
Now consider a sequence of daily snapshots. In the first snapshot, the system may have received many reports for earlier dates but only part of the reports for yesterday. In the second snapshot, yesterday’s count grows as late reports arrive. The newest event date still has the least time for reporting to catch up. A chart drawn from the first snapshot can show a descending edge even though the underlying events did not decline in that order.
This is not the same as saying that a report arriving on Wednesday represents a new event on Wednesday. It may be new information about Monday. The reporting clock has advanced while the event clock remains fixed. If a reader mistakes the movement of information for the movement of disease, the chart’s most recent slope will be interpreted in the wrong frame.
The distinction is also why a chart of reports by report date answers a different question from a chart of events by event date. A report-date chart can describe when the surveillance system received information. It may be useful for monitoring the operation of the reporting system, but it does not by itself show when the underlying visits or cases occurred. Nowcasting is concerned with event-based measures whose newest dates are still being filled in by the reporting process.
The age of a record is the hidden variable in this distortion. An event that occurred ten days ago has had ten days in which to be reported; an event that occurred yesterday has had one. Even if the reporting process behaves in exactly the same way on both dates, the older date will usually be more complete. That creates a directional pattern: the newest dates are not merely noisy versions of older dates but observations made with less reporting time. A true rise in events can be masked by that incompleteness, while a true fall can be exaggerated by it. The raw slope alone cannot tell which situation is present.
The delay is also attached to a setting, not just to an abstract disease. A hospital may transmit records on a different schedule from another hospital. An emergency-department visit may pass through a different workflow from a laboratory-confirmed case. A holiday or a change in staffing can alter when information arrives without changing when the underlying events happened. This is why the two-clock distinction is a starting point for measurement: it identifies the source of the distortion, but it does not yet say that one delay pattern can be used everywhere.
The same event record appears in different states of knowledge. The analyst does not need to pretend that the event happened twice; they need to understand how much reporting time has elapsed for its event date.
Historical snapshots reveal how the reporting clock behaves
A single current dataset tells an analyst what has been reported so far. It does not directly reveal how many reports are still missing. For that, the useful evidence is a sequence of archived snapshots: versions of the surveillance data saved at different times. Comparing them shows how counts for an earlier event date accumulated as reports arrived.
From these comparisons, a model can estimate a reporting-delay distribution. In plain language, this is a description of how likely an event is to have been reported after a given amount of time has passed since it occurred. At one elapsed time, only part of the eventual reporting may usually be visible. At a later elapsed time, more of it may be visible. The pattern can be used to judge how incomplete a newly observed event date is.
The logic is easier to derive than the terminology suggests. For a particular event date, start with the reports currently present. Then ask how much reporting has typically occurred by this age of the record. If historical snapshots show that a date is normally only partly reported at the current age, the observed count should not be read as its likely final level. The model uses the historical relationship between elapsed time and reporting completeness to estimate the portion that has not yet appeared.
A useful way to picture the archive is as a grid. Each row represents an event date. Each column represents a snapshot taken at a later point in time. The cells near the older edge may change little from one column to the next because most reports have already arrived. Cells near the newest edge can increase repeatedly as the same event dates receive additional reports. Reading down one column gives the incomplete curve available today; reading across a row shows how that date matured. The delay distribution is learned from those row-wise changes, then applied to the current column where the final cells are not yet visible.
This also explains why the relevant probability is conditional on elapsed time. The question is not, “What fraction of all reports arrive late?” in the abstract. It is, “Given that this event date is currently this old, how much reporting has ordinarily occurred by now?” A date that is only one day old and a date that is two weeks old are at different points on the reporting clock. Treating them as if they had the same completeness would discard the central information the archive provides.
The correction is consequently not just an instruction to add a constant number to every recent date. It is a relationship between a current count and the age-dependent pattern of reporting. The relationship can be estimated from prior revisions, but its usefulness depends on whether those prior revisions resemble the current process. That is why the source emphasizes both the time series of the metric and information about how underreporting and delays vary over time and across locations.
The CDC gives a deliberately simple intuition: if only 30 percent of emergency-department visits are reported on the same day as the visit, the same-day count will leave an expected share still to arrive. That 30 percent is an illustration of the reasoning, not a universal reporting rate. Real delay patterns can change across locations, facilities, metrics, and periods. A credible correction must learn from the relevant reporting process rather than apply one fixed multiplier to every date.
This is where the reporting-delay distribution must be kept separate from the disease curve. The distribution describes how information enters the surveillance system. It does not say whether infections are growing, declining, or holding steady. A disease can be spreading while its reports arrive slowly; it can also be declining while reports continue to arrive late. The delay pattern is a measurement mechanism that affects what the analyst can see, not a biological explanation for the trend.
That separation is important even when the delay pattern was learned from periods with different disease levels. A high number of reports in a later snapshot can reflect either more events or more complete reporting of earlier events—or both. The event series supplies the subject being measured; the archive supplies clues about the observation process. Neither one can safely stand in for the other. A model that learns only the total number of incoming reports could confuse a change in disease burden with a change in reporting completeness.
It is also possible for the observation process and the event process to interact in practice. A surge in demand may strain the facilities or reporting staff responsible for entering records. A quiet period may permit faster processing. The available evidence therefore supports learning delay behavior from the surveillance system, but not assuming that the behavior is permanently independent of the epidemic. The more the current setting differs from the historical setting, the more carefully the result has to be interpreted.
The distinction protects against a second mistake as well. If the analyst sees a large increase in the count for an earlier date after a new snapshot, that increase does not necessarily mean transmission suddenly surged on the earlier date. It may mean that the reporting system caught up. Conversely, a small count for the newest date does not establish that few events occurred there. It may simply have had less time to receive reports.
Archived snapshots are therefore more than a convenient historical record. They are evidence about the surveillance system’s behavior. They show whether records usually fill in quickly or slowly, whether the pattern differs by place, and how much revision is typical at different ages of the data. Without them, the current line of counts cannot by itself tell the analyst which part of the right edge is missing.
Archives also make the correction testable over time. An analyst can use an earlier snapshot as if it were current, estimate the missing reports from the information then available, and later compare the estimate with the more complete record. The important point is not that such a comparison guarantees success. It is that the reporting process leaves a trail against which the reconstruction can be examined. A system that overwrites earlier versions to save storage may erase that trail, and a system that keeps revisions but does not share them makes independent scrutiny harder.
The archive must preserve enough meaning for the comparison to be fair. If a record changes its event definition between snapshots, a later increase may reflect a new inclusion rule rather than late reporting. If a field is renamed or a source begins supplying a new category, the analyst can no longer assume that an earlier row and a current row describe exactly the same quantity. The two-clock idea still applies, but the task is no longer only to measure elapsed reporting time. It also requires knowing whether the thing being counted has remained the same.
This is why a current line of counts cannot be judged in isolation. A reader needs to know what date the line represents, how old each point is, how the data were revised in the past, and whether the process that produced the reports has remained comparable. A delay model can be mathematically tidy while being attached to the wrong metric or the wrong period. The evidence is strongest when the data history and the definition of the event are both visible.
Several qualifications follow from this mechanism. The historical pattern must be relevant to the current metric and setting. A delay distribution learned from hospitalizations should not automatically be treated as the distribution for emergency-department visits. A pattern from one location may not describe another location. And a pattern that was stable in the past may not remain stable if reporting policies, staffing, or the data system changes. Historical snapshots provide the basis for estimation; they do not turn an estimate into a direct observation.
Once the two clocks are separated, “reporting delay” becomes a precise question: how does the count for an event date change as more time passes and more reports arrive? The next step is to follow that change from an unreported or partial record toward a more complete one, and to see how a nowcast estimates the missing right edge while it is still in motion.