Fig. 3.

Marginal effects of covariates on TB case notification rates, Blantyre, 2015–2017. Red: analysis 1 (all TB); blue: analysis 2 (microbiologically confirmed TB). Estimated by fitting a Bayesian spatial regression model with Poisson response, a k nearest-neighbours conditional spatial autocorrelation prior (with k = 6 for analysis 1 and k = 4 for analysis 2), with linear terms fitted for community health worker catchment area log₁₀ population density, adult M:F ratio, mean number of people per household, log₁₀ Cartesian distance from geographical centroid to the nearest TB clinic, percentage of population aged 15 years or older, mean percentage living on less than US $2 per day, offset term for log₁₀ population size, and with weakly informative prior on the population-level effects intercept (Gaussian: mean = 0, sd = 10), and predictor intercept (Gaussian, mean = 0, sd = 10). Lines are medians, estimated from 100 draws from posterior samples. Shaded areas are 95% credible intervals. Each marginal distribution estimated by holding all other variables constant at their median