Epidemics: modeling growth, peak, and decline

Epidemiology has the most direct mathematical statement of "decay" of any field surveyed here: once a single number crosses a threshold, the system's decline becomes a guaranteed, computable fact.

The core idea

An outbreak is modeled by dividing a population into compartments — Susceptible, Infectious, Removed (recovered or deceased) — and writing differential equations for how people move between them. The rate at which new infections occur depends on how many susceptible people remain and how transmissible the disease is; the rate at which people leave the infectious pool depends on the recovery rate. Whether the epidemic is currently growing or decaying comes down to a single ratio: the effective reproduction number.

The formula

SIR model

dS/dt = −(β/N)·I·S dI/dt = (β/N)·I·S − γ·I dR/dt = γ·I

where S, I, R are the susceptible, infectious, and removed counts, N is the total population, β is the transmission rate, and γ is the recovery rate. The basic reproduction number is:

R₀ = β / γ

Rewriting the infectious-compartment equation makes the growth/decay condition explicit:

dI/dt = (R₀ · S/N − 1) · γ · I

This is the direct epidemiological analog of a decay-phase test: once the effective reproduction number (R₀ scaled by the remaining susceptible fraction) drops below 1, the infected count is in guaranteed monotonic decline, at a rate of γ(1 − R). There is no ambiguity about whether decline has begun — it's a direct consequence of the equation once R crosses the threshold.

SEIR model — adding an incubation compartment

dS/dt = −β·I(t)·S(t) dE/dt = β·I(t)·S(t) − ε·E(t) dI/dt = ε·E(t) − η·I(t) dR/dt = η·I(t)

where ε is the rate at which exposed individuals become infectious (the inverse of the incubation period) and η is the recovery rate (the inverse of the infectious period). The total population S+E+I+R stays constant throughout.

Effective reproduction number Rt — the renewal-equation estimator

R₀ describes a fully susceptible population at the start of an outbreak; Rt tracks how transmissibility evolves in real time as immunity, behavior, and interventions change. The field-standard estimator (Cori et al., implemented in the widely used EpiEstim package) is a renewal equation:

Rₜ = Iₜ / Σ_{s=1}^{t} I_{t−s}·wₛ

where Iₜ is the count of new infections on day t and wₛ is the generation-interval distribution — the probability that s days separate an index infection from the infection it caused. Rt is computed as a Bayesian posterior (using a gamma prior/posterior conjugate pair) and reported as a posterior mean with a 95% credible interval, not a bare point number.

How this field validates its predictions

Epidemiology's own methodology literature is unusually candid that idealized, synthetic-data analyses overstate the accuracy achievable on real surveillance data, even under best practice. Named error sources include uncertainty in the generation-interval distribution, imperfect case observation, reporting-delay and right-truncation effects, and the smoothing-window choice relative to sample size.

A calibration trap the field has caught itself in: a direct comparison study found that EpiEstim's 95% credible intervals were narrower than an alternative method's (Wallinga & Teunis) — but EpiEstim's intervals also failed to cover the true Rt more often, covering only 6 of 16 true simulated values in one comparison versus better coverage from the wider-looking alternative. A narrower interval looked more precise but was actually less well-calibrated. This is exactly the discrimination-vs-calibration trap described across every field on this site: a tighter-looking bound is not automatically a more honest one.

There is no single numeric "acceptable Rt error" threshold published field-wide. The convention instead is to report the posterior credible interval and separately audit its empirical coverage against known or simulated ground truth — the same calibration-check pattern used in survival analysis (see the medicine page).