Bayesian Spatiotemporal Small-Area Estimation of HIV Testing Uptake in Ghana, 2008-2022: Integrating Machine-Learning-Derived Geospatial Covariates with District-Level BYM2-RW1 Modelling of the Ghana Demographic and Health Surveys

Background: HIV testing is the entry point into the diagnosis, treatment and viral suppression cascade, yet in many low and middle income countries the household surveys used to monitor testing coverage are not powered for estimation below the regional level. We produced calibrated district level estimates of HIV testing uptake among women in Ghana across three Demographic and Health Survey (DHS) rounds and examined the spatial and temporal structure of the heterogeneity that remained once measured covariates were accounted for. Methods: We pooled individual recode and geo referenced cluster data from the 2008, 2014 and 2022 Ghana DHS (n = 4,769, 9,391 and 15,014 women respectively; outcome: ever tested for HIV, variable v781), aggregated to 261 level two administrative districts by survey round, and fitted a Bayesian hierarchical binomial model that combined a BYM2 conditional autoregressive spatial term, a first order random walk (RW1) temporal term, and three standardised covariates: WorldPop population density derived from a machine learning dasymetric algorithm, Malaria Atlas Project travel time to the nearest city, and cluster urban proportion, each extracted within buffers around cluster coordinates that matched the DHS displacement protocol. Inference used integrated nested Laplace approximation (INLA) implemented through R-INLA (Lindgren and Rue, 2015). Residual spatial structure was assessed with global and local Moran's I. Results: National crude testing prevalence rose from 20.7% in 2008 to 46.7% in 2014 and 53.8% in 2022. District sample sizes were small and unevenly distributed (2008 median n = 20 women per district; 91.8% of districts had fewer than 50), which is why model based smoothing rather than direct estimation was required. Urban cluster proportion was independently associated with higher testing odds (odds ratio [OR] 1.10, 95% credible interval [CrI] 1.04 to 1.16 per one standard deviation increase) and travel time to the nearest city with lower odds (OR 0.90, 95% CrI 0.84 to 0.96); population density showed no independent association once these two variables were included (OR 0.95, 95% CrI 0.89 to 1.02). The spatial mixing parameter of the BYM2 term (phi = 0.716, 95% CrI 0.498 to 0.887) indicated that around seven tenths of spatially attributable variance was structured rather than idiosyncratic. Global Moran's I on the fitted spatial effect surface was 0.616 (p = 6.6 x 10 to the power minus 60), and local indicators of spatial association identified a contiguous low uptake cluster across the northern regions together with three compact high uptake clusters in the south central corridor. Conclusions: Combining machine learning derived geospatial covariates with an explicit spatiotemporal Bayesian hierarchy exposes a persistent north to south gradient in HIV testing uptake that measured accessibility and urbanicity do not fully explain and identifies specific district clusters as priorities for targeted testing scale up.