Written and maintained by CASRAI Editorial Board
Last updated
Short answer
Nowcasting, in infectious disease surveillance, is the use of statistical models to correct recent case counts for reports that have not arrived yet, so analysts can estimate what is happening now rather than what had been reported by today. CDC describes it as using models “to adjust for incomplete reported data based on historical reporting delays” (MMWR, 27 August 2026). This page covers the epidemiological sense of the word only.
Definition and scope
Surveillance data arrive late. A person falls ill, is tested, and is later reported to a health authority, so the most recent days in any case series are incomplete. The incomplete tail looks like a decline even when transmission is flat or rising. Gunther and colleagues describe nowcasting as adjusting daily case counts for “occurred-but-not-yet-reported events”, and the baselinenowcast R package documentation explains that recent counts are right-truncated and that a nowcast imputes the cases that will eventually be observed up to the present day.
Nowcasting is distinct from forecasting. A nowcast estimates the current, still-unfolding value of a quantity (daily cases, the effective reproduction number Rt); a forecast projects future values. The CDC South Carolina measles analysis used nowcasting to estimate real-time daily cases and Rt, and both outputs describe the present.
How it works in practice
Reference date, report date and the reporting triangle
Every case has a reference date (for example onset or diagnosis) and a report date. Their difference is the reporting delay. The CDC measles analysis defined delay as “the interval between rash onset and public health report date”, with a median of 2 to 3 days. Gunther et al. define it as the time between disease onset and official case reporting by a health authority.
The baselinenowcast vignette lays the data out as a reporting triangle: rows are reference dates, columns are delays, and the bottom-right corner is empty because those reports cannot have arrived yet. The delay distribution is estimated empirically from the filled part of the triangle and applied to the partial recent counts to estimate the final totals. Uncertainty comes from how wrong retrospective nowcasts would have been. By default half of the training data estimates the delay and the rest estimates uncertainty, so choosing a very large maximum delay leaves less data for the uncertainty step.
Missing onset dates and hierarchical models
Where onset dates are missing, the Bavarian COVID-19 analysis first imputed them (Weibull GAMLSS), then fitted a hierarchical Bayesian nowcast with negative binomial counts and a discrete-time hazard model for the delay, including time trends and weekday effects. Rt was then computed from nowcast samples with the Wallinga-Teunis method. The authors warn that “Ignoring such changes can severely bias the predicted number of disease onsets”, referring to changes in the delay distribution over time.
Three sourced approaches compared
| Source | Delay definition | Method emphasis | Output |
|---|---|---|---|
| CDC MMWR 75(33), South Carolina measles | Rash onset to public health report date | Delay correction using historical reporting; negative delays (contacts reported before rash onset) excluded | Real-time daily cases and Rt |
| Gunther et al., Biometrical Journal 2021, Bavaria COVID-19 | Disease onset to official report | Onset-date imputation, then hierarchical Bayesian nowcast | Onsets per day, then Re(t) via Wallinga-Teunis |
baselinenowcast (CRAN) |
Reference date to report date | Empirical reporting triangle, retrospective errors for uncertainty | Completed counts with uncertainty |
Where it matters for labs and clinics
Laboratory and clinic reporting is the first link in the delay chain: specimen collection, testing turnaround and report submission all feed the delay a nowcast has to correct. Teams that report to a health authority, or that read its dashboards, should know that the latest days are provisional. The same logic applies to sentinel networks; see sentinel surveillance, and for how counts are plotted by onset date see epidemic curves.
Checklist: when a nowcast can mislead
Drawn from the CDC measles analysis and the Bavarian study:
- Unstable delays. “Adjusting for reporting delays is most accurate when delays are relatively stable over time.” Changing delay patterns bias the estimate.
- Rapid growth. In South Carolina, nowcasts underestimated final case counts during fast growth in late December and January, when backfilling and backlogging reduced completeness.
- Right truncation ignored. The measles delay estimation did not account for right truncation, which may have contributed to underestimation.
- Early-identified cases. 362 cases (36.7%) found through contact tracing before rash onset were kept in the case analysis but excluded from the delay estimates.
- No gold standard for Rt. The authors note no objective reference exists to validate Rt trends, so accuracy was hard to evaluate quantitatively.
Common misunderstandings
- “Nowcast” means “forecast.” It does not; it estimates the present from incomplete data.
- A recent dip is a real decline. Right truncation alone can produce it.
- A nowcast replaces final data. It is an estimate with uncertainty that final reports supersede.
- One method fits all outbreaks. The CDC authors note other methods might perform better than theirs.
FAQ
What is nowcasting in epidemiology?
It is the model-based estimation of current disease trends from recent data that are still incomplete because of reporting delays.
What is the difference between a nowcast and a forecast?
A nowcast estimates the current value of a quantity, such as today’s cases; a forecast projects future values.
What is a reporting delay?
The interval between a reference event such as symptom onset and the date the case is reported to a health authority.
Are there open-source tools?
The baselinenowcast package is published on CRAN, with documentation on the reporting-triangle approach.
Related reading: Bayesian hierarchical models, outbreak investigation steps in hospitals.
General educational information on surveillance methods, not medical or public health advice.
Sources
- Miller PB et al. Use of Nowcasting to Estimate Real-Time Transmission Trends During a Measles Outbreak, South Carolina, October 2025 to March 2026. MMWR 2026;75(33):516-521. cdc.gov (read 2026-11-08).
- Gunther F et al. Nowcasting the COVID-19 pandemic in Bavaria. Biometrical Journal 2021;63(3):490-502. PMC7753318 (read 2026-11-08).
- baselinenowcast R package vignette, CRAN. cran.r-project.org (read 2026-11-08).
