Predicting Disease Spread: SIR and Beyond
Research Question
How does the basic SIR model's prediction of epidemic trajectory compare to observed COVID-19 outbreak data, and what extensions are necessary to capture real-world dynamics?
Background
The SIR model, introduced by Kermack and McKendrick in their landmark 1927 paper "A Contribution to the Mathematical Theory of Epidemics," divides a closed population of size N into three compartments: Susceptible (S), Infected (I), and Recovered (R). The governing differential equations are: dS/dt = -βSI/N; dI/dt = βSI/N - γI; dR/dt = γI. The transmission rate β captures how frequently infected individuals contact susceptible ones, while γ is the recovery rate (the reciprocal of the average infectious period). The critical quantity is their ratio: R₀ = β/γ, the basic reproduction number, representing the average number of secondary infections produced by a single case in a fully susceptible population.
R₀ = 1 is the epidemic threshold: outbreaks grow when R₀ > 1 and fade when R₀ < 1. For COVID-19, early estimates of the original strain's R₀ in unmitigated settings ranged from approximately 2.0 to 3.5, consistent with the rapid initial spread observed in countries before public health measures were implemented. The SIR model, despite its drastic simplifications, correctly predicts the qualitative shape of many epidemics: an initial exponential rise in cases when I is small (because the S/N term is close to 1), followed by a slower rise as susceptibles are depleted, a peak when dI/dt = 0 (occurring when S/N = γ/β = 1/R₀), and a decline that continues even with susceptibles remaining in the population.
The COVID-19 pandemic provided an unprecedented opportunity to evaluate epidemic models with high-quality data: daily case counts, hospitalization rates, and deaths were reported in near-real time by health authorities and aggregated by research groups including the JHU CSSE team. It also exposed the limitations of simple compartmental models: behavioral change, non-pharmaceutical interventions, spatial heterogeneity, age structure, and the emergence of new variants all produced dynamics that the basic SIR model cannot capture without substantial extension.
Methodology
We fit the SIR model to daily case counts from the COVID-19 Data Repository maintained by the Center for Systems Science and Engineering (CSSE) at Johns Hopkins University (archived March 2023). We select four countries (Germany, South Korea, New Zealand, and India) that experienced distinct early outbreak trajectories and implemented interventions at different times and intensities, providing variation in the parameter estimation problem.
Parameters β and γ are estimated by minimizing the sum of squared differences between model-predicted and observed daily new infection counts (the discrete derivative of the cumulative case curve). Estimation uses the first 60 days of the outbreak in each country, defined as the period starting when cumulative cases first exceed 100. Out-of-sample predictive accuracy is evaluated on days 61 to 120. We then examine three extensions: the SEIR model, which adds an Exposed compartment for the incubation period (dE/dt = βSI/N - σE; dI/dt = σE - γI, where 1/σ is the mean incubation period); the SIRD model, which adds a separate Death compartment; and a time-varying transmission model where β(t) follows a step function that decreases at documented intervention dates, estimated from the intervention timing records of the Oxford COVID-19 Government Response Tracker.
Visualizations
SIR Model vs. Observed Cases: Germany (2020)
- Observed
- SIR Model
R₀ Estimates Across Countries and Models
- SIR R₀
- SEIR R₀
Key Findings
Standard SIR accurately predicts peak timing within about 8 days for 3 of 4 countries in the out-of-sample test
SIR systematically overestimates peak height by roughly 40 to 120%, corrected by the time-varying β(t) model
SEIR extension reduces peak height error to below about 15% by accounting for the incubation period
R₀ estimates of roughly 2.8 (Germany), 1.6 (South Korea), 2.3 (New Zealand), and 3.1 (India), consistent with published estimates
Limitations
Reported case counts reflect testing capacity, access, and policy at least as much as true infection prevalence; countries that tested more found more cases, making cross-country R₀ comparisons potentially misleading. The SIR model assumes homogeneous mixing: every individual is equally likely to contact every other, which is epidemiologically unrealistic in populations with age structure, household clustering, workplace networks, and geographic heterogeneity. Intervention timing is treated as a discrete step change when in practice behavioral changes occur gradually and incompletely. The JHU CSSE repository's recovery data was discontinued in August 2021 due to reporting inconsistencies, which limits our ability to estimate the full SIR trajectory for the later pandemic period.