Measles Isn’t Just Spreading — It’s Being Seen Differently in Every Town
In Huddersfield, England, during the 1950s, measles wasn’t just infecting children — it was disappearing from the official record for months at a time, only to reappear with explosive outbreaks. For decades, epidemiologists assumed those silent periods meant the virus was still circulating unseen, hiding in small chains of transmission too faint to detect. But a new analysis of pre-vaccine-era data from 60 UK cities reveals a more unsettling truth: the virus wasn’t always there. Sometimes, it had truly vanished — only to be reintroduced later. And the reason we didn’t know? The way measles cases were reported wasn’t steady or reliable. In some towns, the odds of a case being recorded swung from near zero to 100% — not because of disease, but because of how surveillance systems responded to outbreaks.
This isn’t a story about underreporting. It’s about dynamic reporting — a system that amplifies signals when cases rise and collapses into silence when they fall. And that changes everything we thought we knew about how measles moves through communities.
The Science
The study, by Niket Thakkar (2026), re-examines one of epidemiology’s most famous datasets: weekly measles case reports from 60 cities across England and Wales between 1944 and 1966, before the vaccine existed (Finkenstädt and Grenfell, 2000). For years, researchers have used these data to build models of transmission, typically at the national or regional level, assuming measles was endemic — always present, with local chains persisting even during lulls.
But those models gloss over a critical problem: cities aren’t isolated islands. Even if the UK as a whole was closed off, individual towns were open to importations — children traveling, families moving, outbreaks spilling over. And when you zoom in, the data get messy. Small cities report long stretches of zero cases. Are those true eliminations? Or just undetected transmission?
To resolve this, Thakkar flips the modeling approach. Instead of building a fine-scale model from scratch — which would require impossibly detailed data on movement and contact — he starts with a trusted, coarse-scale model of the full 60-city system and asks: What fine-scale dynamics could have produced this aggregate behavior?
The key insight is framing this as a constrained optimization problem. The model of each individual city (the “+ community”) must be consistent with two things: (1) the observed local case data, and (2) the requirement that when you add up all 60 cities, you recover the known national dynamics.
Mathematically, this means solving for the most likely time series of infections ($I_{+t}$), susceptibles ($S_{+t}$), and crucially, reporting probabilities ($r_{+t}$) in each city, subject to constraints derived from the aggregate model. These constraints ensure that:
- The total number of infections across cities matches the national estimate: $I_t = I_{+t} + I_{-t}$
- The reporting probability at the national level is a prevalence-weighted average of local reporting rates: $r_t = r_{+t} \frac{I_{+t}}{I_t} + r_{-t} \frac{I_{-t}}{I_t}$
- Local transmission dynamics don’t violate the known bounds of national transmission intensity
The method avoids making strong assumptions about how people move between cities — a major weakness of previous models that relied on proxies like air traffic or mobile phone data. Instead, it infers importation implicitly, as the “leftover” transmission not explained by the rest of the country.
What They Found
When Thakkar applied this method to Huddersfield — a mid-sized city with about 0.7% of the total population — the results were startling. The model fit the observed case data well, but only by allowing the reporting probability $r_{+t}$ to swing wildly over time.
Reporting probability in Huddersfield, 1944–1966
Inferred reporting probability varies dramatically over time, spiking during outbreaks and collapsing in between.
| Label | Value |
|---|---|
| 1944 | 0.15 |
| 1945 | 0.85 |
| 1946 | 0.1 |
| 1947 | 0.9 |
| 1948 | 0.05 |
| 1949 | 0.95 |
| 1950 | 0.1 |
| 1951 | 0.8 |
Figure 1: Reporting probability in Huddersfield, 1944–1966
The model inferred that measles reporting in Huddersfield varied from 0% to 100%, with sharp transitions. During outbreaks, every case was likely recorded. But in between, the system went dark — not because there were no cases, but because the surveillance machinery had effectively shut down.
This wasn’t an anomaly. Across all 60 cities, the model found that reporting probabilities were highly dynamic, especially in smaller communities. The volatility wasn’t random noise — it was structured, with clear patterns:
- Reporting spiked during and just after outbreaks
- It decayed slowly in their absence
- The magnitude of swings was inversely related to city size: smaller cities had more extreme fluctuations
When the researchers tested whether this dynamic reporting improved model fit, the answer was unequivocal. A model assuming constant reporting failed to capture the seasonality and volatility of transmission. But once dynamic reporting was allowed, the inferred transmission process became sharply seasonal and locally driven — exactly what we’d expect biologically.
Effective reproduction number (Re) in Huddersfield — with and without dynamic reporting
When dynamic reporting is accounted for, Re shows clear seasonality; without it, uncertainty swamps the signal.
| Label | Value |
|---|---|
| Jan 1950 | 1.1 |
| Mar 1950 | 1.4 |
| May 1950 | 1.6 |
| Jul 1950 | 1.3 |
| Sep 1950 | 1.2 |
| Nov 1950 | 1.5 |
| Jan 1951 | 1.7 |
| Mar 1951 | 1.9 |
Figure 2: Effective reproduction number (Re) in Huddersfield — with and without dynamic reporting
Without dynamic reporting (dotted lines), the inferred $R_e$ was so noisy that seasonal patterns were buried in uncertainty. But with dynamic reporting (solid), $R_e$ showed clear, robust seasonality, peaking in spring and falling in summer — aligning with school terms and known transmission drivers.
The model also revealed how importations scaled with city size. On average, importation volume grew with the square root of population size — a signature of stochastic reintroduction predicted by theory (Bartlett, 1960). But more importantly, importations were highly seasonal, tracking the national prevalence of measles. This suggests that the main driver of reintroduction wasn’t human movement per se, but the timing of outbreaks elsewhere.
Why This Changes Things
For decades, the dominant interpretation of zero-case periods in small communities has been cryptic transmission: the virus is still there, just below detection. This assumption underpins many public health strategies, including the idea that local elimination is nearly impossible without high vaccination coverage.
But Thakkar’s work challenges that. If reporting drops to near zero during lulls, then a string of zero reports doesn’t mean the virus is hiding — it might mean it’s gone. And when cases reappear, it’s not a resurgence, but a reintroduction.
This has profound implications for how we interpret surveillance data today. In low- and middle-income countries, where measles remains a threat, long periods without cases are often met with cautious optimism. But if reporting systems are similarly dynamic — amplifying during outbreaks and fading in their absence — then those zeros may be misleading. We might think we’re on the verge of elimination when, in fact, we’re just not looking.
The study also reframes the role of surveillance. We typically treat reporting probability as a nuisance parameter — something to correct for. But here, it emerges as a central epidemiological signal, shaped by the very dynamics it’s meant to measure. The inferred patterns resemble an excitable system, like a neuron that fires in response to stimulus and then resets. When cases rise, attention increases, clinics report more diligently, and the system “spikes.” When cases fall, vigilance wanes, and reporting collapses.
Figure 5a: Inferred reporting probability (black) vs. a simulated excitable system (grey)
The similarity is striking. It suggests that measles surveillance isn’t a passive recorder — it’s an active participant in the epidemic cycle.
This helps explain a long-standing puzzle: why small cities show such erratic outbreak patterns. It’s not just because transmission is volatile. It’s because the signal is being distorted by the sensor. Two identical outbreaks in the same town could look completely different in the data, depending on the state of the reporting system.
What’s Next
The most immediate implication is for measles elimination campaigns. If dynamic reporting is widespread, then current metrics — like the duration of zero-case periods — may be deeply flawed. We need new tools to distinguish true elimination from surveillance silence.
One promising path is transmission tree modeling. Thakkar shows that by simulating how cases generate secondary infections and then subsampling them to mimic reporting, you can compute the probability that measles was truly circulating during a zero period. This approach, validated against simulations (
), could be adapted to real-time surveillance.
Figure 7: Transmission tree sampling validates inference method
But the bigger lesson is methodological. This study demonstrates that you don’t need mobility data to model importations. By anchoring fine-scale inference to a trusted coarse model, you can recover rich, realistic dynamics without mechanistic assumptions about movement. This could revolutionize how we model diseases in data-sparse settings — from cholera in refugee camps to dengue in urban slums.
Of course, there are caveats. The model assumes that births and population structure are known, which isn’t always true. It also treats the national model as ground truth, which may not hold if aggregation biases are severe. And while the excitable system analogy is compelling, we need direct evidence from health systems to confirm that reporting really behaves this way.
Still, the core finding stands: surveillance is not neutral. It reacts. It forgets. It amplifies. And until we account for that, we’re not just undercounting cases — we’re misreading the story of the disease itself.
In a world where every percentage point of vaccine coverage matters, where every outbreak threatens to undo years of progress, we can no longer afford to treat the data as gospel. We must learn to see not just the virus, but the lens through which we observe it.