Reconstructing the Missing Right Edge

Nowcasting becomes concrete when we follow one event date through its reporting lifecycle. The focus here is right truncation: the fact that the newest part of a surveillance series contains events that have already happened but reports that have not all arrived. Understanding that lifecycle lets a reader see why earlier dates are revised, what a nowcast is estimating, and why the estimate must be accompanied by evidence about past revisions and a range of uncertainty.

From an unreported event to a more complete record

Right truncation describes a time series whose newest dates are incomplete because the reporting process is still catching up. The word “right” refers to the right-hand end of a chart—the part representing the most recent event dates. It does not mean that the event itself is uncertain in the same way. An emergency-department visit, hospitalization, diagnosed case, or death may already have occurred. What is incomplete is the record of how many such events have been reported for that date.

The difference becomes clearer if one date is followed across several snapshots. At the moment an event occurs, it may not yet be present in the surveillance data. After the first report arrives, the date has a partial record. Additional reports make the record more complete. Eventually, reporting may catch up enough that the observed value changes little, although “complete” here means complete enough for the surveillance process rather than perfectly known in an absolute sense.

Each transition changes the system’s knowledge of the old event date, not necessarily the number of events that occurred on the day a new report arrived. If a report about a Monday visit enters the system on Wednesday, Wednesday’s information has revised Monday’s count. It has not automatically added a new Wednesday visit. This is the lifecycle that a raw event-date curve hides: the event date stays fixed while the reported total moves through successive states of completeness.

The NowcastView in the diagram is an analyst’s view of a record while it is still incomplete, not a new biological state. It exists because decisions may have to be made before the record reaches the more complete end of its lifecycle. Later reports can then revise the event-date count and provide a comparison with the earlier estimate.

This lifecycle explains why the right edge is systematically different from the older part of the curve. Older event dates have had more opportunities to receive reports. Newer dates are observed earlier in their reporting lifecycle. If the event process is steady, the raw counts can still decline toward the newest date simply because each successive date has had less time to fill in. If the event process is rising, the same delay can conceal part of the rise. If it is falling, the delay can make the fall look steeper. Right truncation changes the visibility of a trend before it changes the trend itself.

The term also guards against a misleading interpretation of revisions. Suppose a date’s reported count increases between two data releases. That increase is evidence that the surveillance record became more complete; it is not, by itself, evidence that the disease suddenly became more transmissible on that earlier date. Conversely, if the newest date has a small preliminary count, the small value is evidence only of what has been recorded so far. It is not yet a reliable account of everything that occurred.

The right edge is not usually a single cutoff after which every count is incomplete and before which every count is settled. Completeness changes gradually with the age of a record. A date just outside the newest reporting window may still receive late reports, while a date at the visible edge may already contain some of its eventual information. The important pattern is the changing opportunity to report, not an arbitrary boundary drawn on the chart. This is also why the same correction should not be applied indiscriminately to every recent point: each date can be at a different stage of its reporting lifecycle.

There is a further distinction between missing records and corrected records. A missing record is an event that has not entered the dataset at all. A corrected record is information that was present but later changed because the surveillance system received an update or a more complete version. For the purpose of estimating a current event-date total, both affect what the analyst can see, but they need not have the same operational cause. An archive that preserves successive versions allows those patterns to be examined instead of treating every change as one undifferentiated kind of noise.

A nowcast operates at the point where the record is partial. It uses the current reported count and information about how similar records have filled in to estimate the likely level after more reporting. In that sense, the method does not “repair” the past by altering an event. It estimates the missing portion of the record so the current date can be compared more fairly with dates that are older and more complete.

The estimate remains tied to a particular metric and date. A nowcast of emergency-department visits estimates the incomplete emergency-department-visit count. It does not silently convert the date into an infection date, and it does not make a hospitalization count interchangeable with a case count. The lifecycle is about the reporting completeness of the chosen measure. Questions about the path from infection to that measure require another layer of reasoning.

Why archives and intervals are part of the reconstruction

To estimate the missing portion, an analyst needs more than the current line of counts. CDC identifies a time series for the metric, including its event date, as a basic requirement. It also emphasizes information about underreporting and reporting delays, including how those delays change over time and across locations. The strongest source of that information is a record of the same surveillance data at different points in time, with event and report dates available so that revisions can be followed.

Think of the archive as a record of the lifecycle just described. One snapshot shows what was known at an earlier time. A later snapshot shows which event dates gained reports. A still later snapshot shows whether the same pattern continued. The analyst can compare the count for a given event date at different ages and estimate how much reporting typically occurs by each elapsed time. That estimated pattern is a truncation distribution: a way of describing the portion of reporting that has ordinarily happened, and the portion that may still be missing, at the current age of the record.

The important word is “typically.” The archive does not reveal the missing reports for today by looking them up. It supplies historical evidence about how reporting has unfolded. The current snapshot still contains uncertainty because today’s event date may not follow the historical pattern exactly. The method therefore combines what is observed now with what has been learned about the reporting process, rather than treating the archive as a hidden copy of the present.

A simple example shows why age must be part of the adjustment. Imagine two event dates with the same number of reports currently visible. One date is recent; the other is older. The older date has had more time to receive additional reports, so the same observed count can represent different degrees of completeness. If a model ignored age, it would assign the same meaning to both counts. A truncation adjustment instead asks how much reporting has normally accumulated by the age of each date.

The archive can be read in two directions. Looking forward across snapshots shows how a particular event date fills in. Looking down the event-date series at one snapshot shows the current curve, including its incomplete edge. The first direction supplies information about reporting delay; the second is the quantity public-health users want to understand. Confusing those directions turns the evidence about data arrival into a false description of disease activity.

This is also why the adjustment should be evaluated against later revisions rather than judged only by whether its current line looks smooth. A smooth curve may simply reflect a strong modeling assumption, while a revised curve may contain information the model did not anticipate. Comparing an earlier estimate with a later, more complete record asks the more useful question: did the reconstruction make reasonable use of what was known at the time? The source supports this iterative comparison, but it does not promise that historical agreement will survive a new reporting environment.

The model also has to respect the setting from which the evidence came. Reporting patterns may differ between hospitals, counties, states, or other locations. They may differ between hospitalization data and emergency-department data. They may change over time because reporting policies or operations change. Historical revisions are useful only when the historical process remains relevant enough to inform the current one. A delay pattern is not a permanent property of a pathogen detached from the people and systems that produce the data.

This is why the infrastructure behind the archive matters. A surveillance system may overwrite old records as they become more complete, perhaps to save storage space. If the earlier snapshots disappear, the analyst loses the trail showing how the counts changed. Even when revisions are preserved, they may not be shared publicly. In either case, a current dataset can look clean while providing too little evidence to learn how incomplete its newest dates are.

Archived data also make the reconstruction examinable. An analyst can take an earlier snapshot, estimate what the missing portion might be, and later compare that estimate with the more complete record. Such a comparison does not prove that the same adjustment will work under every future condition. It does provide a way to ask whether the model’s reconstruction was reasonably aligned with what later reporting revealed. Without a revision history, there is no equivalent observation-based check of the right-edge estimate.

The archive itself must be interpreted carefully. A later count can increase because reports arrived, but it can also be affected by a change in what the surveillance system includes or how it defines a field. The source emphasizes that reporting structures can change and that historical data may need to be modified to match current data structures. That issue is not a minor technical detail: if the quantity being counted changes, the difference between snapshots is no longer only a reporting-delay signal.

Uncertainty completes the picture. CDC describes nowcasts as estimates with bounds of uncertainty, commonly presented through a prediction interval. The interval represents a plausible range given the available data and prior information. It acknowledges that several values could be consistent with the reports seen so far and with the historical reporting process. A point estimate can be useful for situational awareness, but it should not be read as though the record had already reached its completed state.

The interval is especially important near the newest edge because the analyst is estimating a future revision to the current record. As the record moves from partial toward more complete, the observed count may change and the uncertainty about what remains may change with it. The interval lets a decision-maker distinguish a current estimate from the later observation that may eventually replace it. It also keeps the reporting process in view: uncertainty is not evidence that the event itself did not occur, but uncertainty about how much of the event count has been captured and how well the historical pattern applies now.

The interval also has a clear limit. It describes uncertainty within the model’s representation of incomplete reporting; it is not a universal shield against a broken data pipeline, a sudden operational disruption, or a definition change. If the current reporting process differs sharply from the archived process, the historical evidence may no longer describe the missing portion well. The interval may still be calculated, but its interpretation requires attention to the conditions that produced it.

This is the balance at the right edge. Waiting for a more complete record gives later reports time to arrive, but it delays knowledge about the present. Estimating too aggressively can create false confidence if the archive is thin or the process has changed. A responsible nowcast uses the historical lifecycle to make a timely estimate and keeps the unresolved part visible through uncertainty and context.

Right truncation is therefore not just a label for a chart pattern. It is a lifecycle problem: events occur, reports arrive unevenly, dates are revised, and an analyst must estimate the still-missing portion before the record settles. The key evidence is the history of those transitions. With that foundation, a nowcast can be understood as a reconstruction of an incomplete metric—not a guess detached from data, and not a completed observation disguised as one.