Generation Time

Generation time—also termed the generation interval—is the fundamental temporal metric in infectious disease epidemiology and public health modeling defined as the time elapsed between the chronological moment a primary case (the infector, A) is infected by a pathogen and the chronological moment that infector transmits the pathogen to infect a secondary case (the infectee, B) along a transmission chain. Formalized mathematically through the Lotka–Euler renewal equation by Alfred J. Lotka (1907) and William O. Kermack and Anderson G. McKendrick (1927), and integrated into modern outbreak inference by Jacco Wallinga and Marc Lipsitch (2007), Åke Svensson (2007), and Sang Woo Park, Jonathan Dushoff, and David J. D. Earn, the generation-time probability density distribution g(τ) provides the mathematical bridge connecting an epidemic's observable real-time exponential growth rate (r) during the initial growth phase (e^rt) to its Basic Reproduction Number (R₀) and Herd Immunity Threshold. Because the moment of microbial inoculation is rarely observable in clinical surveillance, epidemiologists frequently approximate the generation time using the serial interval (the time between symptom onset in the infector and symptom onset in the infectee); however, whenever a pathogen exhibits pre-symptomatic transmission or variable incubation periods, the serial interval and the generation time diverge in both variance and sign, requiring explicit deconvolution and dynamical truncation corrections.

1. Mathematical Relationship Between Generation Time (G), Incubation Periods (I), and Serial Interval (S)

Consider a directed transmission pair in which primary host A is infected at chronological time t_inf,A, develops clinical symptoms at time t_sym,A, transmits the infectious agent to secondary host B at time t_inf,B, and secondary host B subsequently develops symptoms at time t_sym,B:

1. Incubation Period of the Infector (I_A): I_A = t_sym,A − t_inf,A (always positive, I_A > 0).
2. Incubation Period of the Infectee (I_B): I_B = t_sym,B − t_inf,B (always positive, I_B > 0).
3. Generation Time (G or τ):

G = t_inf,B − t_inf,A Because a host cannot transmit a pathogen before acquiring it, the generation time G is biologically constrained to be positive (G > 0) in every transmission pair.

1. Serial Interval (S):

S = t_sym,B − t_sym,A = (t_inf,B + I_B) − (t_inf,A + I_A) = G + (I_B − I_A)

2. Variance Inflation and Negative Serial Intervals Under Pre-Symptomatic Transmission

From the identity S = G + (I_B − I_A), derived by Åke Svensson (2007) and Fine (2003), three statistical properties follow:

• Identical Expectations Under Independent, Identically Distributed Incubation Periods: If the incubation periods I_A and I_B are drawn from the same distribution with mean E[I], then the expected value of their difference is zero (E[I_B − I_A] = 0), so the mean serial interval equals the mean generation time: E[S] = E[G].
• Greater Variance of the Serial Interval: Assuming G and I_B are independent, the variance of the serial interval is Var(S) = Var(G − I_A) + Var(I_B). In the simple case where biological infectiousness is independent of symptom timing, Var(S) = Var(G) + 2 Var(I). Thus, the observable serial interval distribution always has a wider dispersion (larger variance) than the underlying true generation time distribution.
• Negative Serial Intervals (S < 0) During Pre-Symptomatic Shedding: Although the generation time is always positive (G > 0), whenever an infector transmits the virus prior to developing symptoms (t_inf,B < t_sym,A, as occurs in 40% to 50% of SARS-CoV-2 transmissions) and the infectee experiences a shorter incubation period than the infector (I_B < I_A − G), the infectee develops symptoms before the infector (t_sym,B < t_sym,A), yielding a negative serial interval (S < 0). Substituting raw serial intervals (or truncating negative serial intervals) directly into renewal equations distorts estimates of R₀.

During the early phase of an outbreak in a susceptible population, surveillance systems can directly measure the exponential growth rate (r) of daily incidence i(t) = i₀ e^rt (related to the epidemic doubling time T₂ by r = ln(2) / T₂). However, r alone has units of inverse time (days⁻¹) and cannot tell public health officials what fraction of transmission must be blocked by vaccination or non-pharmaceutical interventions—a decision governed by the dimensionless Basic Reproduction Number (R₀).

As demonstrated by Jacco Wallinga and Marc Lipsitch (2007), the mathematical link between r and R₀ is governed entirely by the intrinsic generation-time distribution g(τ) through the Lotka–Euler renewal equation. Let β(τ) denote the expected biological rate at which a primary case infected τ time units ago generates secondary infections in a wholly susceptible population. By definition:

• R₀ = ∫₀^∞ β(τ) dτ
• g(τ) = β(τ) / R₀ (the normalized probability density function of generation times, satisfying ∫₀^∞ g(τ) dτ = 1).

Because total population incidence at time t equals the integral over all past infectors infected at time t − τ multiplied by their transmission rate β(τ):

i(t) = ∫₀^∞ i(t − τ) β(τ) dτ = R₀ ∫₀^∞ i(t − τ) g(τ) dτ

Substituting exponential growth i(t) = i₀ e^rt and i(t − τ) = i₀ e^{r(t − τ)} and cancelling i₀ e^rt from both sides yields the Wallinga–Lipsitch / Lotka–Euler Moment-Generating Function Formula:

1 / R₀ = ∫₀^∞ e^−rτ g(τ) dτ = M_g(−r)

where M_g(−r) is the Laplace transform (or moment-generating function evaluated at −r) of the generation-time distribution g(τ).

Explicit Closed-Form R₀ Solutions Across Classical Distributions

Depending on the shape and variance of g(τ) (with mean μ = E[G] and variance σ² = Var(G), or squared coefficient of variation κ = σ² / μ²), the exact same observed exponential growth rate r and the exact same mean generation time μ produce dramatically different values of R₀:

1. Delta Distribution (Zero Variance, σ² = 0; Every Infector Transmits at Exact Age τ = μ):

R₀ = e^rμ This represents the ** strict mathematical upper bound on R₀** for a given r and mean generation time μ (by Jensen's inequality, because f(τ) = e^−rτ is strictly convex).

1. Exponential Distribution (SIR Model with Constant Recovery Rate γ = 1/μ, No Latent Period; σ² = μ², κ = 1):

R₀ = 1 + rμ

1. SEIR Compartmental Model (Exponential Distribution for Latent Period of Mean 1/σ_L + Exponential Distribution for Infectious Period of Mean 1/γ, Total Mean μ = 1/σ_L + 1/γ):

R₀ = (1 + r / σ_L) (1 + r / γ)

1. Gamma Distribution (Realistic Empirical Fit with Mean μ and Squared Coefficient of Variation κ = σ² / μ² = 1/k, where k is the Gamma Shape Parameter):

R₀ = (1 + κ r μ)^1/κ

1. Second-Order Cumulant / Normal Approximation:

R₀ ≈ exp(rμ − (1/2) r² σ²)

This result reveals a critical insight of biostatistics: for a fixed growth rate r > 0 and mean generation time μ, a larger variance σ² in the generation time yields a smaller R₀. Why? Because when Var(G) is large, a subset of secondary infections occurs very early (small τ), and those early transmissions drive rapid epidemic doubling even when the average number of secondary cases per infector (R₀) is modest. Consequently, if an analyst mistakenly plugs the wider variance of the symptom-based serial interval Var(S) into the Lotka–Euler formula in place of the narrower true Var(G), they will systematically underestimate R₀.

As established by Champredon and Dushoff (2015) and Park, Sun, Champredon, Li, Bolker, Earn, Weitz, Grenfell, and Dushoff (2020–2021), the generation intervals measured empirically via contact tracing change systematically over the course of an epidemic even when the pathogen's biological infectiousness profile β(τ) remains completely constant:

1. Intrinsic vs. Forward (Cohort) vs. Backward Generation Intervals:
   • Intrinsic Generation Interval g(τ): The theoretical distribution of transmission times generated by an infector in a stationary, fully susceptible infinite population.
   • Backward (Infectee-Looking-Back) Generation Interval b_t(τ): When contact tracers interview a cohort of newly infected cases at calendar time t during the exponential growth phase (r > 0) and ask "When was your infector infected?", the cohort of potential infectors was growing exponentially as e^rt. Therefore, recently infected infectors (small τ) vastly outnumber older infectors (large τ) in the population! As a result, the observed backward generation interval during epidemic growth is exponentially tilted toward shorter intervals: b(τ) ∝ e^−rτ g(τ), artificially shrinking the observed mean generation interval below the true intrinsic mean μ. Conversely, during the epidemic decline phase (r < 0), older infectors outnumber new infectors, stretching the backward generation interval toward longer values.
   • Forward (Infector-Looking-Forward) Generation Interval f_t(τ): Following a cohort of infectors infected at time t forward in time, progressive depletion of susceptible contacts (S(t + τ) / N) inside households and local social networks — along with case isolation after symptom onset — truncates late transmissions (large τ), causing the realized forward generation interval to contract as an epidemic peaks or as contact tracing accelerates.
2. Real-Time Right-Truncation Bias: When epidemiologists analyze contact-tracing line lists in real time prior to the end of an outbreak at cutoff date T_cutoff, only transmission pairs whose secondary case has already been infected and developed symptoms before T_cutoff (t_inf,A + G + I_B ≤ T_cutoff) can be observed. Long generation intervals are systematically censored (right-truncated), requiring inverse-probability-of-censoring likelihood weighting or joint Bayesian renewal estimation (EpiNow2 / EpiEstim, developed by Anne Cori, Neil Ferguson, and Sebastian Funk) to recover unbiased estimates of g(τ) and the instantaneous reproduction number R_t.

1. Wallinga, J., & Lipsitch, M. (2007). "How generation intervals shape the relationship between growth rates and reproductive numbers." Proceedings of the Royal Society B: Biological Sciences, 274(1609), 599–604. https://doi.org/10.1098/rspb.2006.3754
2. Svensson, Å. (2007). "A note on generation times in epidemic models." Mathematical Biosciences, 208(1), 300–311. https://doi.org/10.1016/j.mbs.2006.10.010
3. Cori, A., Ferguson, N. M., Fraser, C., & Cauchemez, S. (2013). "A new framework and software to estimate time-varying reproduction numbers during epidemics." American Journal of Epidemiology, 178(9), 1505–1512. https://doi.org/10.1093/aje/kwt133
4. Park, S. W., Sun, K., Champredon, D., Li, M., Bolker, B. M., Earn, D. J. D., Weitz, J. S., Grenfell, B. T., & Dushoff, J. (2021). "Forward-looking serial intervals correctly link epidemic growth to reproduction numbers." Proceedings of the National Academy of Sciences, 118(2), e2011548118. https://doi.org/10.1073/pnas.2011548118