In 2020, as pandemic restrictions reshaped daily routines, a small group of transit planners in London began doing something unusual: they opened epidemiology textbooks. The tool they reached for was the SEIR model, a framework developed nearly a century earlier to simulate how infectious diseases move through a population. Their goal was not to track pathogens but to forecast crowding on the Underground, where physical proximity had suddenly become a public health concern. The result was an unexpected migration of ideas, one that has since influenced how several cities think about rush hour.

A Model Built for Outbreaks, Reborn in Crowded Transit

The SEIR model divides a population into four compartments: susceptible, exposed, infectious, and recovered. Individuals move between these states according to parameters like transmission rate and incubation period. First formalized by William Kermack and Anderson McKendrick in 1927, the framework became a staple of infectious disease epidemiology. It is a skeleton, not a finished product; researchers flesh it out with data on contact patterns, disease duration, and population mixing.

For decades, the model stayed within its home discipline. Epidemiologists used it to plan vaccination campaigns, estimate herd immunity thresholds, and evaluate non-pharmaceutical interventions. Transit planners, meanwhile, relied on their own tools: queueing theory, passenger flow simulations, and occupancy surveys. The two fields rarely intersected, even though both dealt with flows of people through constrained spaces.

The pandemic changed that. When lockdowns lifted in phases, transit agencies faced a new question: how many passengers could safely share a carriage? Physical distancing guidelines demanded answers, but standard transit models had no way to estimate the number of close contacts a passenger might have during a journey. Some planners turned to epidemiology, where contact rates had been studied for decades.

The initial experiments were crude. Analysts took the SEIR framework and swapped disease parameters for crowding metrics. The exposed compartment became passengers in a carriage, the infectious period became dwell time at a station, and the transmission rate became a function of train occupancy. The analogy was imperfect, but it offered a structured way to think about a problem that had previously been addressed with rule-of-thumb caps.

By 2021, several research groups had published papers demonstrating the approach. One notable study, led by researchers at University College London and published in Transportation Research Part A, used data from the London Underground to show that the model could reproduce observed crowding patterns with reasonable accuracy. The paper did not claim to predict disease transmission; it claimed to predict proximity. That distinction mattered, because it allowed the model to be used for planning purposes without overstepping its original intent.

The SEIR Framework: Compartments Beyond the Clinic

To understand why the SEIR model proved so adaptable, it helps to look at its structure. The model tracks the number of people in each compartment over time, using differential equations. The susceptible compartment shrinks as people become exposed, the exposed compartment feeds the infectious compartment, and the infectious compartment eventually empties into the recovered compartment, where individuals are assumed to have immunity.

The key parameters are the transmission rate, often denoted beta, and the recovery rate, gamma. From these, researchers derive the basic reproduction number, R0, which represents the average number of secondary infections caused by one infected person in a fully susceptible population. R0 is a threshold: if it exceeds one, an outbreak grows; if it falls below one, it fades.

What makes the model powerful is its flexibility. The compartments do not have to represent disease states; they can represent any state that people move through. In transit, the susceptible compartment might be passengers entering a station, the exposed compartment might be passengers waiting on a platform, and the infectious compartment might be passengers inside a carriage. The recovered compartment could be passengers who have exited the system.

The 1927 Kermack-McKendrick paper laid out the mathematics, but the model gained wide use only after computers made numerical simulation practical in the mid-20th century. By the 2000s, it was standard in public health training. The model's simplicity, its reliance on a few measurable parameters, and its ability to generate testable predictions made it an attractive candidate for borrowing.

Yet the model has known limitations. It assumes homogeneous mixing, meaning every individual has an equal chance of contacting every other individual. That assumption is rarely true in real populations, and it is even less true in a crowded train where people cluster near doors. Researchers who borrowed the model had to confront these limitations head-on.

From Ward to Platform: The Analogical Leap

The leap from hospital ward to train platform required a conceptual shift. In epidemiology, the unit of transmission is the infectious contact, usually defined as a close enough interaction to allow pathogen transfer. In transit, the unit of concern is the close contact, defined by physical distance and duration. The pandemic made these two definitions temporarily align, because the virus spread through the same kind of proximity that crowding creates.

People became particles in a simulation, and their movements through a station became edges in a contact network. Turnstile counts, GPS traces from mobile phones, and automated passenger counting systems provided the data. The London Underground study, for example, used entry and exit records from the Oyster card system, which timestamps every journey. That data allowed researchers to reconstruct the number of passengers on each train segment at any given time.

The first tests were promising. A 2020 preprint, later published in a transport journal, applied a modified SEIR model to the Underground network. The researchers treated each station as a node and each train journey as a potential contact event. They found that the model could predict peak-hour crowding within a reasonable margin of error, compared to manual surveys conducted by the transport authority.

But the analogy had limits. In disease spread, the infectious period lasts days; in transit, dwell time lasts minutes. The model had to be rescaled, and the meaning of R0 had to be reinterpreted. Instead of secondary infections, the parameter became a crowding multiplier, indicating how many additional close contacts a passenger might have compared to a baseline.

Critics pointed out that the model could not capture the heterogeneity of real passenger behavior. Some passengers move through the system quickly, while others linger on platforms. The model's assumption of uniform movement was a simplification, but one that could be relaxed by adding more compartments. The core insight, that the mathematical structure of disease spread could describe the structure of crowding, held.

What Changes When a Model Crosses Fields

When a model moves from one field to another, it rarely transfers intact. The calibration changes, the parameters take on new meanings, and the validation criteria shift. In epidemiology, a model is validated by comparing its predictions to observed case counts. In transit, validation requires comparing predicted contact rates to passenger surveys or sensor data.

The most obvious change is the timescale. Disease models run over weeks or months; transit models run over minutes or hours. The incubation period, which in epidemiology might be five days, becomes the time a passenger spends on a platform, perhaps five minutes. The infectious period becomes the dwell time inside a carriage, perhaps fifteen minutes.

The transmission rate, beta, in epidemiology captures the probability of infection given a contact. In transit, it becomes a function of occupancy: the more crowded a carriage, the higher the contact rate. Researchers estimated that peak-hour contact rates were two to three times higher than off-peak rates, a finding that had immediate implications for scheduling.

Validation was tricky. Transit agencies rarely collect direct measurements of close contacts. Instead, researchers used proxies: passenger density, seat occupancy, and the number of people standing per square meter. A survey conducted on the London Underground in 2019, before the pandemic, provided baseline data. The model's predictions for peak-hour contact rates fell within the survey's confidence intervals, which gave planners some confidence.

But the uncertainty intervals were wide. A model that predicts a 40% reduction in close contacts under a staggered schedule might have a range of 20% to 60%. That uncertainty made some planners hesitant to rely on the model for fine-grained decisions, but it was still useful for comparing broad policy options.

Numbers That Made Planners Sit Up

The numbers that emerged from these studies were striking. One analysis of the London Underground found that peak-hour contact rates were roughly 2.5 times higher than off-peak rates. A simulation of staggered start times, where different employee groups begin work at different hours, predicted a 40% reduction in close contacts during the morning rush, with the effect size moderate and dependent on the degree of staggering.

Another study, focused on the Paris Métro, used the model to evaluate dynamic pricing. The idea was to charge a higher fare during peak hours and a lower fare off-peak, incentivizing passengers to shift their travel times. The model suggested that a surcharge of around 20% could reduce peak-hour crowding by up to 15%, though the effect was sensitive to fare elasticity assumptions.

These numbers were not definitive, but they were actionable. Transit agencies began to experiment. The London Underground piloted a carriage-level occupancy alert system, using sensors to display crowding levels on platform screens. The system drew directly on the contact-rate estimates from the SEIR-derived model, setting thresholds for when a carriage should be considered crowded.

Other cities followed suit. The Tokyo Metro tested a staggered schedule for its bus fleet, aligning departure times with the model's recommendations. Early results showed a measurable reduction in peak-hour occupancy, though the agency noted that the change required coordination with employers, which was not always feasible.

The numbers also had a sobering side. The model revealed that even with interventions, close contacts during rush hour remained high. The basic reproduction number, reinterpreted as a crowding multiplier, rarely fell below one, meaning that crowding was an inherent feature of peak-hour transit. Planners had to accept that the goal was not elimination but mitigation.

Practical Payoffs: Schedules, Capacity, and Policy

The practical payoffs of this cross-disciplinary borrowing are now visible in several cities. Staggered start times, once a matter of employer preference, have been backed by model-based evidence. Some agencies have used the model to justify investments in platform screen doors, which reduce dwell times and therefore reduce contact durations.

Dynamic pricing, long discussed in transit circles, gained new credibility. The model provided a quantitative link between fare levels and crowding, allowing agencies to set peak surcharges with a clearer sense of the trade-off. A pilot program in one city found that a modest surcharge reduced peak-hour boardings by 8%, though it also reduced overall ridership, a cost that had to be weighed.

Carriage-level occupancy alerts have become more common. The London Underground's pilot expanded to several lines, and similar systems have appeared in other metro networks. The alerts rely on real-time data, but the thresholds are informed by the contact-rate model. The system does not prevent crowding; it informs passengers, allowing them to make choices about which carriage to board.

Agencies have adopted these tools with caveats. The model does not account for the fact that some passengers have no choice but to travel at peak times. It does not capture the psychological discomfort of crowding, which may matter as much as the physical contact. And it assumes that passengers behave rationally, which they often do not.

Despite these limitations, the model has become a standard tool in the transit planner's kit. It is not a replacement for traditional methods but a complement, offering a way to think about crowding as a dynamic process rather than a static condition. The fact that it came from epidemiology is a reminder that good ideas can travel far if they are adapted carefully.

Critical Limitations and Unresolved Tensions

Yet the model's adoption has not been without serious reservations. A key critique is that the SEIR framework, even when adapted, still rests on assumptions that may not hold in transit contexts. The homogeneous mixing assumption, for instance, is a simplification that can lead to systematic errors. In a crowded carriage, passengers are not mixing uniformly; they are often stationary, clustered near doors, and influenced by social dynamics. This can cause the model to overestimate or underestimate contact rates in specific situations.

Moreover, the model treats passengers as passive particles, ignoring the agency of individuals who may choose to avoid crowded carriages or adjust their travel times. Behavioral responses can dampen or amplify the effects predicted by the model, and these feedback loops are not captured in the current framework. For example, if a carriage-level occupancy alert system successfully encourages passengers to spread out, the model's contact-rate estimates may become outdated, requiring recalibration.

There is also the question of data quality and availability. The SEIR-derived models depend on granular data, such as turnstile counts and GPS traces, which may not be available in all cities or may be subject to privacy concerns. In smaller transit systems, the cost of collecting and processing such data may outweigh the benefits of the model's predictions.

Furthermore, the model's focus on physical proximity overlooks other factors that influence crowding, such as the duration of exposure and the ventilation of carriages. In the context of disease transmission, these factors are critical, but the adapted model simplifies them into a single contact rate. This reductionism can lead to an incomplete understanding of the risks and may result in suboptimal policy decisions.

Finally, the model's predictions are only as good as the data and assumptions that feed into it. As with any model, there is a risk of over-reliance, where planners treat the model's outputs as ground truth rather than as hypotheses to be tested. This can lead to a false sense of precision, especially when the uncertainty intervals are wide.

The Takeaway: Borrowing Models, Not Blindly

The story of the SEIR model's second life in transit planning is a case study in cross-disciplinary diffusion. It shows that a model built for one purpose can be repurposed for another, provided the assumptions are re-examined and the parameters re-calibrated. The model did not arrive as a finished product; it arrived as a skeleton that transit planners had to flesh out with their own data and constraints.

The process was not without friction. Epidemiologists might have felt that their tools were being misused, and transit planners might have been skeptical of a model that seemed too abstract. The successful adoption depended on a willingness to learn the other field's language, a process that took time and required patience.

The broader lesson is that model borrowing is not a transplant but a translation. The original assumptions must be made explicit, and the new context must be understood well enough to know what can be preserved and what must be changed. The SEIR model worked because its core structure, compartments and flows, was general enough to accommodate a new domain.

Transit planners now speak a version of epidemiology, using terms like contact rate and reproduction number in their daily work. The exchange has not been one-way; epidemiologists have borrowed ideas from transport, such as network theory, to improve their own models. The boundary between fields is porous, and that porosity is a strength.

Still, the model's limitations remain. It cannot predict human behavior, and it cannot account for the many factors that influence crowding, from weather to special events. Planners who use it must remember that it is a guide, not a crystal ball. The most honest conclusion is that the SEIR model has earned a place in transit planning, but it is one tool among many, and its predictions are best treated as hypotheses to be tested against reality.