Checkpoint: A Nowcast Is a Reconstruction, Not a Guess
When a public-health dashboard shows a sharp drop at its newest edge, two easy explanations compete: transmission has fallen, or the data are wrong. The more useful judgment is to ask what kind of data are being seen and how far they have moved through the reporting process. Nowcasting makes a third possibility precise: the curve is a partial observation of recent events, and a model can reconstruct the likely current level while keeping the unresolved part visible.
Read the curve as four connected layers
The first layer is the event process: visits, hospitalizations, cases, or deaths occurring on particular dates. Those events are what public-health users ultimately want to understand. The second is the reporting process: information about those events arriving after a delay. A chart organized by event date can therefore show a lower value at its newest edge simply because the newest dates have had less time to receive reports.
The third layer is estimation. A nowcast combines the reports currently available with evidence about how reporting has unfolded for comparable event dates in the past. It estimates the portion of the current record that may not yet be visible. The fourth layer is decision: a public-health team may use the estimate to assess burden, allocate attention, or investigate a possible change, but the decision must carry both the estimate’s uncertainty and the condition of the reporting system.
These layers are connected, but they are not interchangeable. An event is not a report. A report is not a completed event total. A nowcast is not a direct observation. And a decision based on a nowcast is not proof that the underlying disease process has already moved in the direction suggested by the point estimate. The value of the model comes from keeping the links between layers explicit.
A useful test begins with a simple visual impression: the newest bar is shorter than the bars before it. The reader should ask four questions. Is the bar organized by the date the event happened or the date it was reported? How old is that event date compared with the dates to its left? What do archived snapshots show about how much reporting usually arrives at that age? And is the displayed value a raw count, a nowcast, or a downstream measure built from an estimate?
Those questions turn a curve into an evidence trail. If the bar is a raw event-date count, its small size may reflect incomplete reporting. If it is a nowcast, the reader should look for the uncertainty range and the data process behind it. If it is an R t estimate, the reader must ask what observed metric and what additional timing assumptions feed the calculation. The same downward shape can have different meanings depending on which layer the chart displays.