Tackling public health data gaps through Bayesian high-resolution population estimation: A case study of Kasaï-Oriental, Democratic Republic of the Congo

Abstract

Most low- and middle-income countries face significant public health challenges, exacerbated by the lack of reliable demographic data supporting effective planning and intervention. In such data-scarce settings, statistical models combining geolocated survey data with geospatial datasets enable the estimation of population counts at high spatial resolution in the absence of dependable demographic data sources. This study introduces a Bayesian model jointly estimating building and population counts, combining geolocated survey data and gridded geospatial datasets. The model provides population estimates for the Kasaï-Oriental province, Democratic Republic of the Congo (DRC), at a spatial resolution of approximately one hectare. Posterior estimates are aggregated across health zones and health areas to offer probabilistic insights into their respective populations. The model exhibits a -0.28 bias, 0.47 inaccuracy, and 0.95 imprecision using scaled residuals, with robust 95% credible intervals. The estimated population of Kasaï-Oriental for 2024 is approximately 4.1 million, with a credible range of 3.4 to 4.8 million. Aggregations by health zones and health areas reveal significant variations in population estimates and uncertainty levels, particularly between the provincial capital, Mbuji-Mayi and the rural hinterland. High-resolution Bayesian population estimates allow flexible aggregation across spatial units while providing probabilistic insights into model uncertainty. These estimates offer a unique resource for the public health community working in Kasaï-Oriental, for instance, in support of a better-informed allocation of vaccines to different operational boundaries based on the upper bound of the 95% credible intervals.

Fig 1. Building and population counts observed in the 213 micro-census clusters located within the modelling environment categorised by settlement class.

Administrative boundaries sourced from OpenStreetMap (https://planet.osm.org) under the Open Database License (ODbL) (https://www.openstreetmap.org/copyright).

Fig 2. Posterior density distributions and boxplots for the parameters estimated in the building count (top) and population density (bottom) model components on a logarithmic scale.

The intercept $\mathit{\alpha }$ and variance $\mathit{\sigma }$ are estimated independently by settlement class $𝐬$. The multiplicative effects $\mathit{\beta }$ are applied to the covariate building count from settlement data (A) by settlement class. Covariates B to E represent the fixed, global multiplicative effects for the covariates distance to main roads (B), average dry matter productivity (C), distance to violence against civilians (D), and terrain slope (E).

Fig 3. Observed versus predicted building (left) and population (right) counts with 95% credible intervals (CIs) classified according to the respective settlement classes.

The diagonal black lines show a one-to-one relationship between observed and predicted values.

Fig 4. Mean gridded population estimates across Kasaï-Oriental (left) with a zoom-in of the provincial capital of Mbuji-Mayi (right).

Fully overlapping health zones (labelled) and health areas (unlabelled) are overlayed. Administrative boundaries sourced from OpenStreetMap (https://planet.osm.org) under the Open Database License (ODbL) (https://www.openstreetmap.org/copyright). Health zone and health area boundaries sourced from GRID3 (https://data.grid3.org/) under the CC BY 4.0 license (https://creativecommons.org/licenses/by/4.0/).

Fig 5. Mean population count estimates and uncertainty levels for Kasaï-Oriental, its health zones (labelled) and health areas (unlabelled).

The size of the circles represents the mean population estimate while the colour scale represents the uncertainty level computed as the difference between the upper and lower 95% CIs divided by the mean estimate.