Nowcasting a Measles Outbreak in Real Time: South Carolina, 2025–2026

From October 2025 to March 2026, South Carolina had the largest measles outbreak in the United States in about 30 years, with 997 cases reported. Measles is highly contagious, and large outbreaks, often in communities with low vaccination coverage, have become more common.

Tracking an outbreak in real time is harder than it sounds. The South Carolina Department of Public Health (SCDPH) sent CDC updated case lists about twice a week, but recent cases always lag: people take time to seek care, specimens take time to test, and investigations take time to finish. Incomplete recent data can make an outbreak look as if it is fading when it is actually growing. CDC and SCDPH tried nowcasting to get around that, the first time CDC had used it during an active measles outbreak.

What nowcasting does

A nowcast uses past reporting delays to adjust incomplete recent data and estimate what the final numbers will turn out to be. CDC had used it on routine respiratory disease surveillance, but no one knew whether reporting during a measles outbreak would be consistent enough for it to work.

The model estimated two things:

• Daily case counts, with 90% prediction intervals: ranges expected to contain the final count.
• The effective reproduction number (Rt), the average number of people each infected person infects at a given time. Rt above 1 means an outbreak is growing, around 1 stable, and below 1 shrinking. The model reported the probability that Rt exceeded 1: 95% or more meant increasing, 60% to under 95% likely increasing, 40% to under 60% stable, 5% to under 40% likely decreasing, and under 5% decreasing.

The data

Each case record carried a rash onset date and, from December 19, 2025, the date SCDPH learned of the case. The gap between them, the reporting delay, had a median of 2–3 days, and on average 92.8% of cases were reported within 14 days of rash onset. Cases missing a rash onset date were left out of real-time analyses; they made up between 0% and 10% of any provisional list. And 362 cases (36.7%) were found through contact tracing before their rash appeared, so they were left out when estimating reporting delays.

The model was built with the EpiNow2 package in R, and its code is public for health departments and researchers to use.

How it performed

Late December: seeing growth the raw data hid. The first nowcasts, on December 19 and 23, 2025, used lists that contained 76.5%–77.2% of the eventual counts and seemed to show cases declining. The nowcasts instead captured the final counts within their prediction intervals and pointed to a likely increase. Partly because of them, SCDPH kept staffing high over the winter holidays and began hiring more staff. It was right to: from December 23 to January 6, the final cumulative count rose from 185 to 424.

January: underestimating the surge. On January 6 the provisional list held only 25.7% of the eventual counts, a lower share than expected as cases rose fast and many were filled in later. That day's nowcasts underestimated the final numbers and fell outside their prediction intervals, but they were still closer than the raw counts and correctly showed the outbreak growing. Cases peaked on January 13–14. The model missed the decline in the January 16 and 23 data, yet still improved estimates of total outbreak size by 10.5%–11.3% over the raw counts. On January 27 it judged the outbreak likely decreasing.

Late January on: confirming the decline. After January 23, with lists containing 72.1%–88.5% of the eventual counts, nowcasts captured cumulative and daily counts within their intervals and showed cases falling. That reassured SCDPH that the late-January decline was real rather than a reporting artifact, and let it scale down the response in February and March.

What made it work, and its limits

The model produced estimates within minutes of each new case list, feeding automated reports to SCDPH. Because it needed only two dates per case, it added little work for frontline staff, and SCDPH's contact tracing and investment in surveillance likely kept reporting fast and consistent.

The authors note several limitations:

• Transmission rose sharply at the end of December, probably from holiday gatherings, which the model did not account for, and in January cases were missing from provisional lists for a time. The nowcasts were least accurate then, though they still showed the right trend when the raw data did not.
• Other methods might handle rapid changes better. EpiNow2 was chosen because it estimates case counts and Rt together and could be set up quickly. The model also did not account for long delays not yet observed, which may have led it to underestimate cases while the outbreak grew.
• There is no objective yardstick for Rt, so its trends were judged by eye against the final data.

The takeaway

Every outbreak's data arrive late. When the essential dates are collected consistently, nowcasting can correct for those delays, separate real declines from reporting gaps and support decisions such as when to hire more staff, and it will become more useful as methods and data improve.

Sources

• Miller PB, Routledge I, Pollock ED, et al. "Use of Nowcasting to Estimate Real-Time Transmission Trends During a Measles Outbreak — South Carolina, October 2025–March 2026." MMWR 75(33). https://www.cdc.gov/mmwr/volumes/75/wr/mm7533a1.htm
• Model code: https://github.com/CDCgov/measles-nowcasting-2026; supplementary material: https://stacks.cdc.gov/view/cdc/258868
• The report's figures are not reproduced here because it was prepared jointly with the South Carolina Department of Public Health.