The SIR model is a mathematical framework that divides a population into three groups — Susceptible, Infectious, and Recovered — and tracks how people move between them as a disease spreads. First formalized nearly a century ago, it remains the backbone of modern epidemiological modeling, used to estimate outbreak peaks, evaluate vaccination thresholds, and guide public health decisions. The model is deceptively simple in structure, but the insights it generates, and the ways researchers have extended it, touch almost every aspect of how we understand infectious disease.
Where the SIR Model Came From
The foundation was laid in 1927 by William Kermack and Anderson McKendrick, who published a paper asking a question that still drives epidemiology: does an epidemic end only when it runs out of susceptible people, or can the interplay of infection, recovery, and death cause it to burn out while many susceptible individuals remain?1Proceedings of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character. A contribution to the mathematical theory of epidemics Their answer was that epidemics can indeed fizzle out with plenty of people still uninfected, and the mathematics behind that answer became the SIR model.
Kermack and McKendrick made several simplifying assumptions that kept the math tractable. They assumed everyone in the population was equally susceptible, that a single infection conferred complete immunity, and that the population stayed roughly constant over the short timescale of an epidemic. These assumptions are obviously wrong for many real diseases, but they created a clean starting point that researchers have been refining ever since.
How the Three Compartments Work
The model’s logic is straightforward. Everyone starts in the Susceptible group. When a susceptible person encounters an infectious person, there is some probability of transmission, and the susceptible person moves into the Infectious group. After some period of illness, that person either recovers (gaining immunity) or dies, and in both cases moves into the Recovered group. The “recovered” label is a bit misleading — it really means “removed from the chain of transmission,” whether by immunity, death, or isolation.
Two rates govern everything. The transmission rate captures how quickly the disease spreads from infectious people to susceptible ones, and the recovery rate captures how quickly infectious people stop being infectious. The ratio of these two rates produces the model’s most famous output: the basic reproduction number, usually written R₀.
What R₀ Actually Means
R₀ represents the average number of new infections one sick person generates in a completely susceptible population. If R₀ is below 1, each person infects fewer than one other person on average, and the outbreak dies out. If R₀ is above 1, the outbreak grows. This threshold behavior is one of the SIR model’s most powerful predictions: there is a sharp dividing line between “the disease fades away” and “the disease becomes an epidemic.”
For the 1918 influenza pandemic, researchers applying SIR-type models to historical case data have estimated R₀ in the range of 2.0 to 3.0 at the city level, with some estimates varying by method and population studied.2PubMed Central. Comparative estimation of the reproduction number for pandemic influenza from daily case notification data A separate analysis found even wider variation: R₀ between about 1.3 and 3.2 in Maryland communities, but around 5.0 on ships where people were packed together.3PLoS ONE. Transmissibility of the Influenza Virus in the 1918 Pandemic The ship populations also saw their effective reproduction number drop below epidemic levels within about 10 days, compared to 30 to 40 days in the communities.3PLoS ONE. Transmissibility of the Influenza Virus in the 1918 Pandemic This illustrates something the basic SIR model makes clear in theory but that real outbreaks confirm: population density and contact patterns dramatically shape how fast and how far an epidemic spreads.
R₀ is not a fixed biological property of the pathogen. It depends on the population, the setting, the behavior of individuals, and even the weather. The same virus can have a very different R₀ on a crowded troop ship than in a rural town.
Herd Immunity and Vaccination Thresholds
One of the SIR model’s most consequential predictions is the herd immunity threshold: the fraction of the population that needs to be immune (through infection or vaccination) for the disease to stop spreading. The formula is simple — roughly 1 minus 1/R₀ — and it explains why diseases with higher R₀ values need higher vaccination coverage to control.
During the COVID-19 pandemic, this calculation became publicly debated in a way it never had before. One study assessed the vaccination coverage needed to establish herd immunity against SARS-CoV-2 across a range of R₀ values (from 1.1 to 10) and found that with R₀ between 3 and 10 and vaccine effectiveness of 70 to 100%, herd immunity was achievable, but only if vaccination coverage reached at least 90% worldwide and vaccine effectiveness against the Omicron variant reached at least 88%.4PubMed Central. Percentages of Vaccination Coverage Required to Establish Herd Immunity against SARS-CoV-2 Those are steep requirements, and the gap between what the model says is needed and what can actually be achieved is one reason herd immunity against COVID-19 proved elusive.
Fitting the Model to Real Outbreaks
The SIR model is only useful if its parameters can be estimated from real-world data, and this turns out to be harder than it sounds. You rarely observe all three compartments directly. You might have daily case counts (a proxy for new infections), hospitalization data, or death records, but you almost never have a clean time series of how many people are currently susceptible, currently infectious, and currently recovered.
Researchers have developed methods to extract SIR parameters from whatever data is available. One approach works with just the infection curve: the initial exponential rise gives you one parameter, the peak gives you another, and the final decline gives you a third, and from those three numbers, the full SIR dynamics can be reconstructed. This technique has been tested on both simulated data and real COVID-19 data from France.5The European Physical Journal Plus. An algorithm for the direct estimation of the parameters of the SIR epidemic model from the I(t) dynamics Other approaches use asymptotic methods and iterative numerical routines, and have been validated against both influenza and COVID-19 datasets.6PubMed Central. Analytical solutions and parameter estimation of the SIR epidemic model
The parameter estimation problem matters because small errors in the estimated transmission or recovery rate can lead to large errors in predictions about peak timing and total outbreak size. Getting the inputs right is often the hardest part of using the model.
When Immunity Does Not Last
The classic SIR model assumes that once you recover, you are immune forever. For some diseases, like measles, that is close enough to true. For many others, it is not. Immunity wanes. People become susceptible again. The model variant that accounts for this is called SIRS — susceptible, infectious, recovered, and then back to susceptible.
The SIRS framework predicts something the SIR model does not: endemic equilibrium, where the disease never fully disappears but settles into a steady background level. As R₀ rises above a threshold, the model transitions from a disease-free state to this persistent endemic state.7PubMed Central. Endemic oscillations for SARS-CoV-2 Omicron-A SIRS model analysis Research on simpler versions of the SIRS model confirms that shorter average immune periods lead to higher infection prevalence at equilibrium and longer persistence of infection in the population.8PubMed. The effect of waning immunity on long-term behaviour of stochastic models for the spread of infection
But here is where the details get consequential for public health planning. The standard SIRS model assumes immunity switches off abruptly after a fixed period, like a light going dark. In reality, immunity fades gradually. When researchers modeled gradual waning instead, they found that the long-term disease prevalence was higher, and significantly more vaccine was needed to achieve herd immunity. For parameters fitted to COVID-19, the critical amount of vaccine supply was about 50% higher when immunity waned linearly and more than 150% higher when immunity waned exponentially, compared to the classic SIRS assumption of abrupt loss.9PubMed Central. Extending susceptible-infectious-recovered-susceptible epidemics to allow for gradual waning of immunity That is a massive difference in resource planning, and it comes entirely from a modeling choice about how immunity decays.
Adding an Incubation Period and Other Compartments
The SIR model has no incubation period: people become infectious the instant they are infected. The SEIR model fixes this by adding an “Exposed” compartment for people who are infected but not yet infectious. For diseases like COVID-19, where the incubation period can be several days, this matters for predicting when the epidemic will peak and how high the peak will be.
Even within the SEIR framework, the shape of the incubation time distribution matters more than you might expect. Research has shown that the height of the infection peak, the timing of when it occurs, and the final total size of the epidemic all increase as the incubation period is subdivided into more realistic sub-stages.10Physica A: Statistical Mechanics and its Applications. The importance of the incubation time distribution in compartmental epidemiological models In other words, the standard approach of treating the incubation period as a single exponentially distributed step can systematically underestimate how bad an outbreak will get.
Multiple Strains and the Problem of Partial Immunity
Real pathogens mutate. Influenza drifts and shifts annually; SARS-CoV-2 produced Alpha, Delta, Omicron, and their sublineages. The basic SIR model handles one strain with lifelong immunity — a far cry from this reality.
Multi-strain SIR models tackle the problem by tracking multiple circulating strains simultaneously. A key question these models explore is what happens when a new strain emerges to which existing immunity offers no protection. In one multi-strain framework with selective immunity by vaccination, a newer strain introduced after the original has reached equilibrium can sweep through a population that thought it was protected, because vaccination and recovery from the first strain confer no cross-immunity to the second.11PubMed Central. The local stability of a modified multi-strain SIR model for emerging viral strains
When partial cross-immunity exists — meaning previous infection with one strain gives you some but not complete protection against another — the dynamics get richer. Models of cocirculating influenza strains with partial cross-protection show that strong enough cross-immunity can produce sustained oscillations, as the population’s immune profile overshoots in one direction and then corrects.12PubMed. The dynamics of cocirculating influenza strains conferring partial cross-immunity Whether two strains coexist or one drives the other to extinction depends on each strain’s R₀, the duration of temporary immunity, and the degree of cross-protection.13PubMed. A two-strain model of infectious disease spread with asymmetric temporary immunity periods and partial cross-immunity These are exactly the kinds of questions that mattered for predicting what would happen as Omicron displaced Delta.
Network Effects and Superspreading
The SIR model’s original assumption is that everyone mixes randomly with everyone else — what epidemiologists call homogeneous mixing. In reality, people have social networks. Some people have dozens of daily contacts; others have very few. And the structure of those networks fundamentally changes how epidemics behave.
On what are called scale-free networks — networks where a few individuals have vastly more connections than the typical person — the SIR model’s prediction of a clean epidemic threshold essentially vanishes. Research on infection dynamics in these networks found that the extreme variation in how connected people are means there is no clear R₀ value below which the disease reliably dies out, in stark contrast to the standard model’s clean threshold at R₀ = 1.14PubMed. Infection dynamics on scale-free networks
Superspreading is a related phenomenon. A systematic review and meta-analysis of SARS-CoV-2 transmission found that the dispersion parameter (which measures how unevenly transmission is distributed across individuals) was consistently small, indicating high variability. Low values of this parameter mean that a small number of infected people are responsible for a disproportionate share of all transmission, while most infected people infect nobody at all. Large outbreaks happen less frequently in this scenario, but when they do occur, they can be explosive.15PubMed Central. Superspreading, overdispersion and their implications in the SARS-CoV-2 (COVID-19) pandemic: a systematic review and meta-analysis of the literature The basic SIR model, which treats all individuals as equally likely to transmit, completely misses this clustering.
Age Structure and Contact Patterns
People of different ages mix with each other at different rates. Children have intense contact at school, working-age adults interact in offices and transit, and older adults may have fewer but more sustained contacts at home. An age-structured SIR model uses contact matrices — essentially tables of how frequently each age group interacts with every other age group — to account for these patterns. This approach was used to study the progress of COVID-19 in India, where the age distribution and social contact structure were plugged into an SIR framework to compute both R₀ and its time-dependent equivalent.16arXiv. Age-structured impact of social distancing on the COVID-19 epidemic in India
Estimating those contact matrices is its own challenge, especially during an outbreak when social behavior has already changed. Researchers have developed methods to infer age-specific contact patterns from epidemiological data during the early phase of an epidemic, and have applied these to reconstruct contact structure by age during the COVID-19 spread in Buenos Aires.17PubMed. On the Estimation of Contact Matrices for Age-Structured Models at the Onset of Epidemic Spread The practical payoff is that age-structured models can evaluate targeted interventions — closing schools vs. closing workplaces, for instance — in a way the basic SIR model cannot.
Modeling Interventions
The SIR framework has been widely adapted to evaluate both pharmaceutical and non-pharmaceutical interventions. During COVID-19, one of the most visible uses was modeling “flattening the curve” — reducing the transmission rate through social distancing and lockdowns to lower the peak of infections, even if the total number of infections might remain similar.
Researchers used SIR-based models to explore the feasibility of optimizing the duration, intensity, and trigger point of non-pharmaceutical interventions to minimize either the peak prevalence or the total attack rate of a simulated UK outbreak.18PubMed Central. Optimizing time-limited non-pharmaceutical interventions for COVID-19 outbreak control A separate study used an extended SEIR model with optimal control theory to compute the best intervention strategy for a scenario where no vaccine is ever found, finding that even purely non-pharmaceutical measures have an optimal timing and intensity that can be calculated.19PubMed Central. Beyond just “flattening the curve”: Optimal control of epidemics with purely non-pharmaceutical interventions
Vaccination modeling within the SIR framework has its own rich tradition. One approach that has drawn theoretical interest is pulse vaccination — vaccinating a large fraction of the population at regular intervals rather than continuously. Mathematical analysis shows that if the pulse size is large enough and the interval between pulses is short enough, the disease can be driven to eradication, and this can happen at lower overall vaccination rates than continuous vaccination would require.20PubMed. Pulse vaccination strategy in the SIR epidemic model Models incorporating pulse vaccination with time-delay effects confirm this: a sufficiently high vaccination rate per pulse eliminates the disease, while rates below a critical value allow the disease to persist.21PubMed Central. Analysis of an SIR epidemic model with pulse vaccination and distributed time delay
Deterministic vs. Stochastic Versions
The classic SIR model is deterministic: given the same starting conditions and parameters, it always produces the same epidemic curve. Real epidemics are not like that. Chance plays a role, especially when numbers are small. A single infected traveler arriving in a city might infect three people, or zero, depending on who they happen to sit next to.
Stochastic SIR models account for this randomness, and they produce a crucial insight that the deterministic model misses. In a deterministic SIR model, R₀ above 1 means the disease persists indefinitely; below 1, it goes extinct. In the stochastic version, the disease always eventually goes extinct, because random fluctuations eventually bring the number of infected individuals to zero. The difference is in how long this takes. When R₀ is above 1, the expected time to extinction can be enormously long — centuries or millennia in a large population — making the distinction academic in practice. But when R₀ is near 1 or the population is small, stochastic effects become the dominant story.22PubMed. Comparison of deterministic and stochastic SIS and SIR models in discrete time
Spatial Spread and Metapopulation Models
The SIR model treats a population as a single well-mixed unit, but diseases spread across cities, regions, and countries. Metapopulation models extend SIR by creating a network of population centers, each with its own set of SIR dynamics, connected by migration or travel. Individuals move between nodes based on a mobility matrix, and the disease can hop from one community to the next through these travelers.23PLOS Computational Biology. Connectivity, reproduction number, and mobility interact to determine communities’ epidemiological superspreader potential in a metapopulation network
A finding that should give modelers pause: the choice of how you model human movement between these population centers has a profound impact on the results. Research comparing different movement models within the same SIR framework found parameter regimes where switching from one plausible movement model to another completely changed epidemiological outcomes — whether the disease went endemic, how fast it spread, and which communities were hit hardest.24PubMed Central. Comparing metapopulation dynamics of infectious diseases under different models of human movement This means that even when the disease dynamics themselves are well understood, uncertainty about how people move can undermine predictions.
Machine Learning Meets Compartmental Models
A recent development that has generated substantial interest is the use of physics-informed neural networks (PINNs) to fit and forecast SIR-type models. Traditional approaches estimate parameters by running numerical solvers for the differential equations and adjusting parameters until the output matches observed data. PINNs do something different: they train a neural network to satisfy both the observed data and the mathematical rules of the SIR equations simultaneously.
One implementation used this hybrid approach for COVID-19, embedding an extended SIR model’s equations directly into the neural network’s training process, then using the trained network to generate future outbreak scenarios.25PubMed Central. A physics-informed neural network to model COVID-19 infection and hospitalization scenarios Another developed a split implementation of PINNs specifically for estimating how the transmission rate changes over time, reporting improvements in accuracy of up to an order of magnitude and a 20% speedup in computational time compared to a joint approach.26PLOS Computational Biology. A Physics-Informed Neural Network approach for compartmental epidemiological models
For multi-region forecasting, a model called PISID integrates a neural network module for encoding spatial and temporal patterns with an SIR module for epidemiological dynamics. Tested on real-world datasets, it achieved strong predictive performance with a compact architecture of roughly 27,000 parameters and fast training times.27PLOS ONE. Enhancing epidemic forecasting with a physics-informed spatial identity neural network The appeal of these approaches is that they let the data speak where the data is rich, while the SIR equations constrain the model in regions where data is sparse — keeping predictions physically plausible even when the neural network is extrapolating.
Behavioral Feedback and Vaccination Games
One of the SIR model’s blind spots is that it treats the transmission rate as something imposed from outside, like a law of physics. In reality, people change their behavior in response to epidemic conditions. They stay home when cases are surging. They stop wearing masks when cases drop. And they make individual decisions about whether to get vaccinated based on perceived risk, social norms, and trust in institutions.
Researchers have started embedding game theory into SIR-type models to capture this feedback. In one framework, individuals decide whether to vaccinate (cooperate) or free-ride on others’ vaccination (defect) based on a composite information index that includes disease incidence, vaccine factors, and the cooperative behavior of others.28PubMed Central. Behavioral vaccination policies and game-environment feedback in epidemic dynamics The dynamics this produces are messier than the classic SIR model’s smooth curves: vaccination rates oscillate as people respond to rising and falling case counts, creating feedback loops that the original model never anticipated.
SIR Beyond Epidemiology
The SIR framework’s appeal has extended well past infectious disease. Researchers have applied it to model how financial risk spreads through banking systems, drawing an analogy between financial institutions and organisms in an ecosystem where “infection” is the spread of credit risk or panic from one institution to its counterparties.29Quality & Quantity. On the dynamics of a SIR model for a financial risk contagion In one dual-layer network model, an SIR-SIR framework was used to capture how information spreading among entrepreneurs on a social network interacts with credit risk contagion on an enterprise association network, with the authenticity of the information and the willingness of people to spread it shaping how the “infection” of financial risk propagated.30Physica A: Statistical Mechanics and its Applications. Information authenticity, spreading willingness and credit risk contagion – A dual-layer network perspective
These cross-disciplinary applications work because the core SIR logic — a population of susceptible units, a contagious process, and some form of removal or recovery — maps onto any system where something spreads through contact. The spread of rumors, the adoption of technologies, the propagation of computer viruses, and the cascading failure of infrastructure networks have all been studied using compartmental models descended from Kermack and McKendrick’s original framework. The mathematics does not care whether the “pathogen” is a virus or a piece of misinformation; what matters is whether the contact structure and transmission dynamics fit the compartmental assumptions closely enough to produce useful predictions.