Why the newest measles count can be the least complete

The right-hand edge of an outbreak chart often looks reassuring. It bends down towards the present. Yet the newest part can also be the part with the most cases still missing.

That was not a theoretical problem during South Carolina's measles outbreak from October 2025 to March 2026. A CDC study published on 27 August describes how delays between rash onset and a case reaching public-health records made provisional counts appear to fall while transmission was still increasing. A statistical nowcast helped the response team see through some of that delay, but it did not turn incomplete data into certainty.

The distinction starts with the calendar. A person develops a rash on one date. The case may reach a health department later, after care is sought, testing is completed or an investigation links the person to an outbreak. Recent onset dates have had less time to accumulate those reports. If a chart is drawn immediately, the latest bars can be shorter simply because the reporting pipeline has not caught up.

The South Carolina Department of Public Health sent updated, deidentified case lists to CDC about twice a week. The outbreak ultimately had 997 reported cases, the largest U.S. measles outbreak in roughly 30 years. For the cases used to estimate reporting delay, the median interval from rash onset to public-health report was two to three days. On average, 92.8% were reported within 14 days. Those averages were useful, but they were not constant.

CDC's model used the observed pattern of past delays to estimate how many recent cases were probably not yet visible. It also estimated the effective reproduction number, often written as Rt, with uncertainty intervals. Rt is a time-specific indicator of whether infections appear to be increasing, broadly stable or decreasing. It is not a personal risk score, and in this study it was not a guarantee about the next week.

This is the important difference between a nowcast and a forecast. A forecast looks ahead. A nowcast tries to describe the present more completely than the raw, still-arriving reports allow. The model did not replace confirmed case reports. It placed a transparent estimate, including a range of plausible values, over the incomplete edge.

In the first two runs, based on case lists from 19 and 23 December, the provisional data contained about three quarters of the final cases within the recent 14-day observation window and appeared to be declining. The nowcasts instead indicated a likely increasing trend. The report says those estimates contributed to South Carolina keeping staffing high over the winter holiday and beginning to recruit additional staff.

Then the reporting pattern changed. Cases rose rapidly around the holiday period, and records were backfilled through January. In the 6 January dataset, provisional reports contained only 25.7% of the final cases for the nowcast window. The model underestimated the eventual total and its 90% prediction intervals did not consistently contain the final daily counts. It still improved on the raw provisional total and identified the broad upward direction, but the size estimate was too low.

That failure matters because the method learns from the reporting system. When the delay pattern changes sharply, yesterday's pattern is a weaker guide to today's missing cases. The model also did not account for changes in contact around holiday gatherings. Nowcasting can correct for a familiar lag. It cannot automatically explain every new reason the data pipeline or the outbreak itself has changed.

The timing of the turn was difficult too. Final data showed daily cases peaking on 13 and 14 January. The model did not detect the decline in the 16 or 23 January data. By 27 January it estimated that the outbreak was likely decreasing. Later runs, when reporting completeness was higher, supported the judgement that the decline was real rather than another reporting artefact. That helped the response de-escalate during February and March.

There are further limits. More than a third of cases were identified through contact tracing before rash onset, so their report date came first. Those cases were kept in the outbreak analysis but excluded from the distribution used to estimate onset-to-report delay. Some records were missing rash-onset dates at first and were added later. The authors also note that there is no objective reference against which to validate an Rt estimate in real time, so its trend was compared with the final case curve.

This was one large outbreak with relatively frequent reporting and the essential dates needed by the model. The result does not establish that the same approach will perform equally well in every jurisdiction, disease or data system. CDC has published the code for partners, but the repository describes it as work in progress and still requires local data preparation and modelling judgement.

The wider U.S. measles context is active. CDC's public page listed 2,777 confirmed cases in 2026 as of 20 August and explicitly marked recent counts as preliminary and subject to change. It also explains why state and federal totals can differ: jurisdictions update on different schedules, and states generally hold the most current local information. Those national totals are separate from the completed South Carolina outbreak studied in the paper.

Measles is highly contagious and vaccine-preventable, which makes timely response important. But the central lesson of this study is about reading surveillance rather than judging an individual's care. A low final bar is not automatically evidence that an outbreak is ending. A nowcast can make the missing edge more visible, provided its uncertainty and its dependence on stable reporting are visible too.

The newest number deserves attention. It also deserves an asterisk.