Why the Model Has to Move

A reporting-delay model is only as useful as its picture of how reports arrive. The focus here is the difference between a fixed historical pattern, a smoothed pattern that varies by site, and a reporting curve whose parameters are allowed to change over time. Understanding that difference helps a reader judge what each model is trying to represent, rather than treating a more elaborate model name as proof that its estimate is better.

Match the model to the reporting problem

There is no single thing called “the nowcasting model.” Models can operate at different layers and represent different parts of the problem. CDC distinguishes mechanistic models, which explicitly represent biological components of a disease system, from statistical models, which describe relationships among variables. That distinction is useful, but it is not a simple contest between biology and statistics. A model can use statistical evidence to describe reporting delays, biological timing to describe infection, or both in a semi-mechanistic design.

The first question is therefore: what relationship must the model preserve? If the immediate task is to estimate how many recent hospitalizations have not yet been reported, a model needs a credible representation of the reporting process. If the task is to infer infections and a transmission measure, it also needs the relationships that connect infection to the observed metric. Comparing those models as if they were interchangeable can produce a false choice. A delay-adjustment model and a full transmission pipeline may both use the word “nowcast,” while answering different questions.

CDC’s description of the Respiratory Virus Hospitalization Surveillance Network illustrates one practical arrangement. Its baseline model uses the distribution of delays observed in the past to project current delays. This is a sensible starting point when the reporting process is sufficiently similar to its history: past revisions provide evidence about what is likely to arrive next. The model does not need to guess each missing report independently. It uses the observed pattern of reporting delay to adjust the current, incomplete edge.

A fixed historical pattern has a clear strength: it is interpretable. A reader can understand that the model is asking how comparable records filled in previously and applying that evidence to the newest dates. It also has a clear dependency. If reporting behavior differs across locations or changes over time, a pooled or rigid historical pattern may describe the average past system better than the current one. The limitation is not that history is useless. It is that history must remain relevant to the process being estimated.

Generalized additive models, or GAMs, offer a statistical way to let the delay relationship bend rather than forcing it into one rigid form. CDC describes them as generalized linear models in which smoothing functions fit the delay distribution. A smoothing function can represent a changing relationship between elapsed time and reporting probability without requiring the analyst to specify one abrupt step for every point in time. In the nowcasting context, the result is a delay curve shaped by the data rather than a single correction factor.

The value of smoothing is not aesthetic.