Default prior distributions for MOSAIC model parameters

A list of informative prior distributions for all MOSAIC model parameters, organised into global (single-value) priors and location-specific (per-country) priors for 40 SSA / N-Africa countries. Built by data-raw/make_priors_default.R (see that script's block comments for the literature sources and derivation of each prior). Current version: priors_default$metadata$version.

Usage

priors_default

Format

A list with 4 main components.

metadata: version, build date, a free-text description listing the changes that produced this version of the priors, build_date_start (the window start the priors were built for and the initial conditions estimated at; config_default's date_start must equal it), quiet_start_seeded (the locations whose E and I priors are the quiet-start seeding prior of est_initial_E_I) and imputed_window_fallback (the locations whose E/I window held no observed or reconstructed count and was read from their country-level reconstructions).
parameters_global: Named list of 25 global priors (single value shared across all locations). See Global parameters below.
parameters_location: Named list of 20 location-specific prior families. Each entry holds a location sublist keyed by ISO3 country code, so e.g. parameters_location$beta_j0_tot$location$ETH gives the prior for Ethiopia's total baseline transmission rate.
mu_jt: The prior for the reported case fatality ratio, which is not sampled but integrated out of the deaths likelihood per simulated path. See Reported case fatality ratio below.

Each leaf entry has the shape list(description, distribution, parameters) where parameters holds the distribution-specific hyperparameters (e.g. shape1/shape2 for beta, meanlog/sdlog for lognormal, mean/sd/a/b for truncated normal).

Global parameters (25)

Transmission and FOI structure:

alpha_2 – Exponent on N_jt in the FOI denominator (Beta). 1 = frequency-dependent transmission; 0 = density-dependent. A single global scalar (alpha_1 is now per-location – see below).
kappa – Half-saturation V. cholerae concentration at which the environmental dose-response is 50% (Lognormal; not a carrying capacity).

Environmental persistence (V. cholerae decay in water):

decay_days_short – Short-end decay timescale (Truncnorm, days).
decay_days_spread – Spread used to derive decay_days_long = decay_days_short + decay_days_spread (Truncnorm). decay_days_long itself is derived at sampling time and has no first-class prior.
decay_shape_1, decay_shape_2 – Truncnorm shape coefficients in the Beta-shaped mapping from suitability $\psi$ to decay rate (NOT natural-infection immunity decay).

Disease progression (rates, per day):

iota – Incubation rate E -> I (Lognormal; rate, not period). Prior median ~0.71/day.
gamma_1 – Symptomatic shedding-duration rate I_1 -> R (Lognormal). Prior median 0.1/day (~10-day shedding); 95% CI ~3.75 - 26.6 days.
gamma_2 – Asymptomatic shedding-duration rate I_2 -> R (Lognormal). Prior median 0.5/day (~2-day shedding); 95% CI ~0.91 - 4.39 days.
epsilon – Natural-infection immunity waning rate R -> S (Lognormal). Prior mean 3.9e-4/day -> ~7-yr immunity (King et al. 2008 + project's two-cohort re-fit).

Vaccination (OCV):

phi_1, phi_2 – One-dose / two-dose vaccine effectiveness at the moment of dose delivery (Beta).
omega_1, omega_2 – One-dose / two-dose vaccine-induced immunity waning rates (Gamma, per day).

Symptom expression, care-seeking, surveillance:

sigma – Fraction of infections that are symptomatic (Beta).
rho – Care-seeking rate: probability a symptomatic individual presents to surveillance (Beta; not a reporting fraction).
rho_deaths – Surveillance capture rate of true cholera deaths (Beta(36.95, 51.02); random-effects meta-analysis of Routh 2017, Shikanga 2009, Bwire 2013). Pinned at its mean, 0.42, by default: the engine's per-onset fatality probability is mu_jt * rho / (rho_deaths * chi_epidemic) and reported deaths are thinned by rho_deaths, so it cancels from reported deaths and sets only true (unreported) deaths.
chi_endemic, chi_epidemic – Positive predictive value among suspected cases during endemic vs epidemic phases (Beta).
delta_reporting_cases – Symptom-onset-to-surveillance-report delay (Truncnorm, days). Not infection-to-report – incubation is handled separately by the E compartment and iota. Deaths are reported on the same lag (a death is decided at symptom onset).

Environmental shedding intensity (V. cholerae cells per person per day):

zeta_1 – Symptomatic shedding rate (Lognormal).
zeta_2 – Asymptomatic shedding rate (Lognormal). Also has a first-class prior, but is typically derived at sampling time as zeta_1 / zeta_ratio rather than drawn independently.
zeta_ratio – zeta_1 / zeta_2 (Lognormal; ~5-OOM literature spread retained intentionally). data-raw/make_priors_default.R truncates it below at 1 (parameters$lower = 1) so the derived zeta_2 never exceeds zeta_1; objects built before that change (v16.1 and earlier) carry no lower, and about 16% of their draws have zeta_2 > zeta_1.

Mobility (gravity-model exponents):

mobility_gamma, mobility_omega – Distance-decay and population-scaling exponents (Gamma).

Location-specific parameters (20)

Each carries a per-iso prior under parameters_location$<param>$location$<ISO3>. Distribution family is the same across isos for a given parameter (noted below); hyperparameters differ (except alpha_1, which uses a shared marginal across isos).

Transmission and contact:

alpha_1 – FOI mixing exponent on the infectious term (Beta; 0 = no mixing, 1 = well-mixed). Relocated from global to per-location (v15.16) with a shared informative Beta(28.4, 71.6) for every iso; the engine applies it elementwise per patch (a scalar alpha_1 is still accepted and broadcast for national/legacy configs).
beta_j0_tot – Total baseline transmission rate per location (Lognormal). beta_j0_hum and beta_j0_env are derived as p_beta * beta_j0_tot and (1 - p_beta) * beta_j0_tot; neither is directly sampled.
p_beta – Fraction of total transmission attributable to direct human-to-human contact (Beta).
tau_i – Daily departure probability per location, gravity model (Lognormal; the mass above 1 is negligible).
theta_j – WASH-coverage fraction per location (Beta).

Seasonality (Fourier coefficients, all fit to cases):

a_1_j, a_2_j, b_1_j, b_2_j – Annual + semi-annual Fourier coefficients (Normal). Earlier doc versions erroneously labelled b_1/b_2 as "for deaths" – all four are fitted to cases.

Initial conditions (per-compartment proportions of population at t=0):

prop_S_initial, prop_V1_initial, prop_V2_initial, prop_E_initial, prop_I_initial, prop_R_initial – All Beta priors per iso. Generated by est_initial_S/V1_V2/E_I/R() from demography, surveillance history, OCV campaigns, etc. sample_parameters() draws only V1, V2, E, I and R and sets S to the residual 1 - (V1 + V2 + E + I + R) (or 0, renormalising the others, when they sum to 1 or more), so the prop_S_initial prior is informational only: it never enters a simulation, and changing it does not change a calibration.

Case-reporting phase switch:

epidemic_threshold – Per-iso I_sym / N threshold above which the case-reporting PPV switches from chi_endemic to chi_epidemic (Truncnorm).

Environmental-suitability calibration (used by the LSTM $\psi$ pipeline):

psi_star_a – Gain on the logit-scale suitability score (Truncnorm).
psi_star_b – Logit-scale offset (Normal).
psi_star_z – Smoothing weight for the causal EWMA (Beta).
psi_star_k – Time-offset (lag) parameter (Truncnorm).

Reported case fatality ratio (`mu_jt`)

A top-level block, not a sampled prior. The engine reads the reported CFR as config$mu_jt (location x day); run_MOSAIC() integrates it out of each simulated path's deaths likelihood with logit mu_jt = logit mu0_jt + a_j + delta_{j,year}. The block holds:

location[[ISO3]]: year, logit_mean (the centre mu0, the value config_default$mu_jt is built from) and logit_se (the SE of the country-trend mean), from est_CFR_hierarchical()'s WHO-annual GAM.
sd_year: the GAM's country-year SD, the prior SD of each delta_{j,year}.
sd_product: residual error of the GAM centre against the observed reported CFR in the calibration window (sd(log) 0.19-0.32 over the 15-17 countries with at least 50 deaths, 2023-26); the location offset's prior SD is sqrt(sd_product^2 + mean(logit_se^2)).
tau: the GAM's between-country SD (reference only).

Countries

Location-specific parameters cover 40 SSA / North-African countries (ISO3 codes): AGO, BDI, BEN, BFA, BWA, CAF, CIV, CMR, COD, COG, ERI, ETH, GAB, GHA, GIN, GMB, GNB, GNQ, KEN, LBR, MLI, MOZ, MRT, MWI, NAM, NER, NGA, RWA, SEN, SLE, SOM, SSD, SWZ, TCD, TGO, TZA, UGA, ZAF, ZMB, ZWE.