Beyond the Smoothed Surface: Combining Bayesian Spatial Modelling and Density-Based Clustering to Detect Hidden Pockets of Incomplete Childhood Vaccination in Mandera and Mombasa Counties, Kenya
Collins Kigen Sing’oei *
Department of Mathematics and Actuarial Science, The Catholic University of Eastern Africa (CUEA), Nairobi, Kenya.
Gladys Njoroge
Department of Mathematics and Actuarial Science, The Catholic University of Eastern Africa (CUEA), Nairobi, Kenya.
Hellen Waititu
Department of Mathematics and Actuarial Science, The Catholic University of Eastern Africa (CUEA), Nairobi, Kenya.
*Author to whom correspondence should be addressed.
Abstract
Aims/Objectives: To demonstrate the complementary value of Bayesian spatial modelling and density-based clustering for detecting hidden pockets of incomplete childhood vaccination in contrasting Kenyan county settings.
Study Design: Secondary cross-sectional spatial analysis of household survey data.
Place and Duration of Study: Mandera and Mombasa counties, Kenya, using data from the 2022 Kenya Demographic and Health Survey (KDHS).
Methodology: The descriptive sample included 223 children aged 12–23 months across 61 enumeration areas (Mandera n = 161; Mombasa n = 62). Incomplete vaccination was defined as failure to receive all eight nationally recommended vaccine doses. A Bayesian BYM2 spatial model was fitted using R-INLA and compared with a non-spatial hierarchical baseline using model-fit criteria. Residual spatial autocorrelation was assessed using Moran’s I. HDBSCAN density-based clustering was applied to 193 complete cases, with sensitivity analysis across minimum cluster sizes of 3, 5, 7, and 10.
Results: The BYM2 model improved fit relative to the non-spatial baseline (deviance information criterion 127.40 vs. 203.29) and reduced residual spatial autocorrelation (Moran’s I: 0.402 to −0.075). Smoothed risk was high in Mandera (mean 0.807) and low in Mombasa (mean 0.257). HDBSCAN identified four clusters, including five Mombasa areas with 100% incomplete vaccination, while BYM2 estimated a mean risk of 0.288 for this cluster.
Conclusion: The methods provided complementary information. Their combined use distinguished the broad geographic risk gradient from a local pocket that was not apparent from the smoothed surface alone, supporting joint use for county-level and local prioritisation.
Keywords: Childhood vaccination, incomplete vaccination, spatial epidemiology, Bayesian spatial modelling, BYM2, HDBSCAN, density-based clustering, geospatial analysis, immunisation coverage, Kenya