Skip to content

Transmission

This page explains how new cholera infections are generated each day in laser.cholera. The model carries two parallel transmission pathways — direct human-to-human contact and an environmental (water-reservoir) pathway — and combines them additively into a single per-patch force of infection that drives the susceptible-to-exposed transition. The page assumes a working familiarity with compartmental SIR-style models and with the parameter groups documented under Reference / Parameters; it does not walk through configuration steps (see the how-to pages for that). After reading you should understand which terms in the implementation correspond to which biological story, why both pathways are needed for cholera in particular, and which knobs in the parameter dictionary control each pathway's contribution.

The force of infection and the S to E draw

At every tick, each patch \(j\) accumulates two force-of-infection contributions — one from infectious humans (Lambda in the code) and one from the environmental reservoir (Psi). A susceptible individual in patch \(j\) is exposed during a tick with probability

\[ P(\text{S} \to \text{E})_{j,t} \;=\; 1 - \exp\!\bigl(-(\Lambda_{j,t} + \Psi_{j,t})\bigr) \]

The two pathways are sampled separately inside the model — first HumanToHuman draws new infections against 1 - exp(-Lambda), then EnvToHuman draws against 1 - exp(-Psi) using the already-decremented susceptible pool. Drawing them in two passes is numerically equivalent to drawing once against the combined hazard, and it gives the model clean separate counts for incidence_human and incidence_env (both also summed into a combined incidence recorder).

The pipeline order matters. Susceptible runs first and writes S[tick + 1] (carrying forward, then applying non-disease deaths and births). HumanToHuman and EnvToHuman both consume S[tick + 1] rather than S[tick], so each draws against the susceptible pool that remains after natural mortality (and, for EnvToHuman, after the human pathway has already taken its share). Environmental runs last and advances the reservoir W[tick + 1] based on the day's infectious counts; that updated reservoir is what next tick's EnvToHuman will see.

The human-to-human pathway

The direct pathway treats transmission as a contact process between infectious and susceptible people, structured by patch and modulated by seasonality and mobility. The per-tick, per-patch force of infection is

\[ \Lambda_{j,t} \;=\; \frac{\beta^{hum}_{jt} \cdot \bigl(I^{eff}_{j,t}\bigr)^{\alpha_1}}{N_{j,t}^{\alpha_2}} \]

where \(I^{eff}_{j,t}\) is the effective infectious count seen by patch \(j\), built from both local and visiting infectious individuals:

\[ I^{eff}_{j,t} \;=\; (1 - \tau_j)\,(I^{sym}_{j,t} + I^{asym}_{j,t}) \;+\; \sum_{i \neq j} \pi_{ij}\,\tau_i\,(I^{sym}_{i,t} + I^{asym}_{i,t}) \]

Several pieces of physical intuition are baked into this formula.

Symptomatic and asymptomatic cases shed equally for direct transmission. The model sums Isym + Iasym with no weighting; the asymmetry between the two cohorts shows up only in the environmental pathway (different shedding rates zeta_1 vs. zeta_2).

Mobility couples patches. The fraction tau_i of patch \(i\)'s infectious population is treated as "mobile" and is redistributed across destinations according to the gravity-model coupling matrix pi_ij (built once at construction time from latitudes, longitudes, mobility_omega, and mobility_gamma; see mobility). The remaining 1 - tau_i stays put. The implementation in humantohuman.py:144-148 uses the slightly opaque (vec * pi_ij.T).T.sum(axis=0) idiom to keep the matrix indexed [source, destination]; the result is that an outbreak in a hub patch spills into well-connected neighbours even before any of its own residents become infected.

Seasonality multiplies the baseline. beta_jt_human is a static (nticks, npatches) matrix built once at construction by get_daily_seasonality — it is the product of the per-patch baseline beta_j0_hum and a Fourier bracket \(1 + a_{1,j}\cos(2\pi t/p) + b_{1,j}\sin(2\pi t/p) + a_{2,j}\cos(4\pi t/p) + b_{2,j}\sin(4\pi t/p)\). Zeroing the four amplitudes leaves a flat envelope at the baseline; zeroing beta_j0_hum itself disables the entire pathway. See seasonality for the harmonic in detail.

The mixing exponents shape how prevalence translates to risk. With alpha_1 = alpha_2 = 1 the formula is frequency-dependent — the relevant quantity is the infectious fraction \(I/N\), independent of how big the patch is. With alpha_2 = 0 it is density-dependent — the absolute count of infecteds drives risk. Intermediate values interpolate. The published MOSAIC parameterisations use alpha_1 close to (but slightly below) 1 with alpha_2 = 1, i.e. a frequency-dependent contact process with a mild sub-linear response to prevalence.

Once Lambda is assembled it is clipped at zero (a defensive guard against numerical noise in the seasonal multiplier) and used to draw new infections as Binomial(local, 1 - exp(-Lambda)), where local is the rounded count of non-emigrating susceptibles (1 - tau_i) * S. The result is added to E[tick + 1], subtracted from S[tick + 1], and banked into incidence_human.

The environmental pathway

Cholera also spreads through contaminated water, and laser.cholera models that with a per-patch reservoir \(W_j\) that accumulates from infectious shedding and decays at a suitability-dependent rate. The environmental force of infection is

\[ \Psi_{j,t} \;=\; \beta^{env}_{jt} \cdot (1 - \theta_j) \cdot \frac{W_{j,t}}{\kappa + W_{j,t}} \]

Three terms carry the biology.

WASH coverage gates exposure. theta_j is the fraction of the population in patch \(j\) with effective water, sanitation, and hygiene infrastructure; (1 - theta_j) is the fraction that can still acquire infection from contaminated water. The same factor appears on the shedding side in environmental.py ((1 - theta_j) * shedding), so improving WASH simultaneously reduces both contamination and exposure.

The reservoir saturates. The term \(W / (\kappa + W)\) is a Michaelis-Menten-style saturation: when \(W \ll \kappa\), exposure is roughly linear in \(W\); when \(W \gg \kappa\), it asymptotes at 1 regardless of how much contamination accumulates. kappa is the half-saturation constant in the same units as W.

Suitability modulates the per-tick rate. The static matrix beta_jt_env is built once at construction time from the baseline beta_j0_env and a per-patch suitability time series psi_jt:

psi_bar = psi.mean(axis=0, keepdims=True)
beta_jt_env = beta_j0_env.T * (1.0 + (psi - psi_bar) / psi_bar)

The deviation (psi - psi_bar) / psi_bar is centred on zero, so on average beta_jt_env equals beta_j0_env; days more suitable than the patch's annual mean amplify environmental transmission, less suitable days dampen it. Important numerical caveat: because the formula divides by psi_bar, setting psi_jt = np.zeros(...) produces NaN even when beta_j0_env = 0. The canonical "inert" value for suitability is psi_jt = np.ones(...), which gives psi_bar = 1 and collapses the bracket to 1.0. To switch the environmental pathway off, zero beta_j0_env and set psi_jt to ones; see §4 of the parameter reference for the full off-form recipe.

The reservoir's life cycle

The reservoir itself is owned by Environmental, which runs at the end of each tick (after both transmission components have drawn). Three things happen per patch:

  1. Carry forward: W[tick + 1] := W[tick].
  2. Decay: W -= Poisson(delta_jt[tick] * W), clamped not to exceed W so the reservoir cannot go negative even on extreme draws. The decay rate delta_jt is a static matrix derived from psi_jt via a Beta-CDF interpolator (map_suitability_to_decay) between 1 / decay_days_short (high decay, used when suitability is low — bacteria die quickly in hostile water) and 1 / decay_days_long (low decay, used when suitability is high — bacteria persist). decay_shape_1 and decay_shape_2 shape the Beta CDF that smoothly interpolates between these extremes.
  3. Shedding: symptomatic cases contribute (1 - theta_j) * Poisson(zeta_1 * Isym) and asymptomatic cases contribute (1 - theta_j) * Poisson(zeta_2 * Iasym). WASH coverage attenuates the deposit, mirroring the gating on the exposure side. zeta_1 > zeta_2 in the published parameterisations, so a symptomatic case is the more potent contaminator — by a large multiple — but asymptomatics matter because they outnumber symptomatics roughly 3-to-1 under the default sigma.

The reservoir is thus continuously recharged by current infectious cases and drained by decay; under endemic equilibrium these balance and \(W\) hovers near a steady state. A pulse of incidence drives \(W\) up sharply; once cases subside, decay (faster in unsuitable conditions) returns the reservoir toward baseline over days-to-weeks.

How the two pathways compose

The two pathways are additive in the combined hazard \(\Lambda + \Psi\), but because they are drawn in two consecutive binomial passes against the same susceptible pool, their effective contributions interact slightly through the depletion of \(S\) within a single tick. In practice the per-tick hazards are small enough that this is a second-order effect, and the implementation choice keeps the two incidence streams cleanly separable for diagnostics.

Under typical MOSAIC endemic-setting parameterisations the environmental pathway is the smaller of the two when measured by contribution to incidence — the human pathway dominates day-to-day transmission, while the environmental pathway provides a slower, spatially-anchored reservoir of risk that keeps cholera from going locally extinct between outbreak peaks. In epidemic regimes (or in patches with very low WASH coverage and high suitability) the balance can flip; the incidence_human vs. incidence_env recorders make this decomposition easy to inspect post-hoc.

Turning each pathway off

Both pathways have explicit off-forms that the validator accepts and the simulation honours:

  • Human pathway off: beta_j0_hum = np.zeros(npatches). The seasonal envelope beta_jt_human is built as beta_j0_hum * (1 + harmonic), so a zero baseline yields a permanently zero Lambda. The four seasonal amplitudes (a_1_j, b_1_j, a_2_j, b_2_j) can stay non-zero; they multiply zero and contribute nothing.
  • Environmental pathway off: beta_j0_env = np.zeros(npatches) and psi_jt = np.ones((nticks, npatches)). The ones value avoids the psi_bar division-by-zero noted above; the zero baseline kills beta_jt_env and therefore Psi. The reservoir W will still tick over (shedding minus decay), but no susceptible will ever be exposed via it.
  • Both pathways off: combine the two. The model still runs — births, deaths, vaccinations, vital dynamics all continue — but no susceptible ever transitions to exposed, so the disease compartments drain monotonically to zero as initial infectious cases recover.

These off-forms are useful for unit tests, for isolating one pathway's contribution in calibration, and for sanity-checking that a scenario configuration has actually wired up the pathway the user intended.

How it connects to the rest of the model

The per-tick pipeline (Susceptible first, then the other compartments, then Census, then HumanToHuman, EnvToHuman, Environmental, then DerivedValues and Analyzer) ensures three invariants for the transmission step:

  • S[tick + 1] is fully initialised (with mortality and births applied) before either transmission component runs.
  • N[tick] — the total patch population used in the denominator of Lambda — is the value Census computed for this tick, so \(\alpha_2\)-scaling sees the current population, not yesterday's.
  • W[tick] — used by EnvToHuman — is yesterday's reservoir state, because Environmental writes W[tick + 1] only after EnvToHuman has finished. This one-tick lag between shedding and exposure is intentional and matches the published MOSAIC formulation.

The Reference parameters that control transmission are grouped under human-to-human transmission (beta_j0_hum, the four seasonal amplitudes, p, alpha_1, alpha_2) and environmental transmission (beta_j0_env, psi_jt, theta_j, kappa, zeta_1, zeta_2, decay_days_short, decay_days_long, decay_shape_1, decay_shape_2). The mobility-coupling matrix pi_ij is built from geography-and-mobility parameters (latitude, longitude, tau_i, mobility_omega, mobility_gamma).

See also