The model was implemented in the BUGS modelling language and run via JAGS 4.3.1. The complete model string is reproduced below for full reproducibility.
model {
# ---- Priors ----
beta0 ~ dnorm(0, 0.1)
beta_malaria ~ dnorm(0, 0.1)
beta_nuisance ~ dnorm(0, 0.1)
beta_nuisance_dry ~ dnorm(0, 0.1)
beta_temp ~ dnorm(0, 0.1)
beta_habit ~ dnorm(0, 0.1) T(0, 1)
# ---- Month 1: no lag term ----
logit(U[1]) <- beta0 +
beta_malaria * log(Malaria[1] + 1) +
(beta_nuisance + beta_nuisance_dry * is_dry[1]) *
log(Nuisance[1] + 1) +
beta_temp * Temp_Scaled[1]
Observed_Usage[1] ~ dbin(U[1], Pop[1])
# ---- Months 2 to N: autoregressive on latent U ----
for (t in 2:N_months) {
logit(U[t]) <- beta0 +
beta_malaria * log(Malaria[t] + 1) +
(beta_nuisance + beta_nuisance_dry * is_dry[t]) *
log(Nuisance[t] + 1) +
beta_temp * Temp_Scaled[t] +
beta_habit * U[t-1]
Observed_Usage[t] ~ dbin(U[t], Pop[t])
}
}
Data passed to JAGS:
| Variable | Description |
|---|---|
N_months |
Number of months in the analysis (21) |
Observed_Usage |
Integer count: round(LOESS proportion × Pop) |
Pop |
Monthly count of observed household members |
Malaria |
Mean Anopheles bites per trap-night (CDC) |
Nuisance |
Mean Culicine bites per trap-night (CDC) |
Temp_Scaled |
Monthly mean temperature, standardised (mean = 0, SD = 1) |
is_dry |
Binary season indicator: 1 = July–November, 0 = December–June |
Note that the latent probability vector U is a
stochastic node in the JAGS; it is not provided as input data. JAGS
estimates the full posterior distribution over U jointly
with the regression coefficients, so the binomial likelihood propagates
uncertainty in the latent usage probability through to all downstream
quantities.