Bayesian modelling of malaria cases through the elicitation of prior distribution

Abstract

Climate change plays a critical role in shaping both the geographic spread and intensity of malaria outbreaks. Recent research demonstrates that climaterelated factors surpass epidemiological, socio-economic, and environmental contributors in driving malaria resurgence. Progress towards malaria elimination has been further hindered by the COVID-19 pandemic, which disrupted control programmes globally. In South Africa, despite implementation of the National Malaria Elimination Strategic Plan (NMESP), malaria elimination targets remain unmet, partly due to limitations in surveillance and predictive systems. This thesis addresses the underutilisation of subjective Bayesian methods in epidemiology, which incorporate expert knowledge into disease models. While objective priors are commonly used due to their simplicity, subjective priors informed by expert judgment offer richer, context-specific insights. However, the process of prior elicitation is often hindered by complexity and resource constraints. To overcome this, the thesis introduces a novel, efficient approach that combines the Analytic Hierarchy Process (AHP) with statistical validation techniques to incorporate expert knowledge into Bayesian models. The approach provides a simple, time and resource efficient framework for eliciting and applying subjective priors in Bayesian epidemiological models. This thesis utelised both primary and secondary data. Primary data were collected through a purposive sampling approach using a questionnaire to support expert elicitation, with participants recruited nationally and internationally. Secondary data included malaria case records (2018–2022) from the South African ii National Department of Health, climatological data from the South African Weather Services, and environmental data from EarthExplorer platform, covering the provinces of KwaZulu-Natal, Limpopo, and Mpumalanga. The major contribution of this thesis is the development of a robust Bayesian malaria prediction model that integrates both climate data and expert-informed prior distributions. Comparative analysis indicates that subjective priors outperform objective priors in enhancing model performance. These findings underscore the value of expert knowledge in improving predictive accuracy and guiding public health responses. The analysis reveals that malaria transmission intensifies in regions with temperatures between 20–30°C, rainfall ranging from 0–200 mm, and Normalised Difference Vegetation Index (NDVI) values of 0.5–0.8. Those conditions are associated with 200 to 1000 predicted malaria cases. Ehlanzeni (Mpumalanga), uMkhanyakude (KwaZulu-Natal), Vhembe, and Mopani (Limpopo) districts are identified as high-risk areas with elevated malaria counts. Malaria incidence peaks in summer and autumn, especially in regions experiencing temperatures of 12–20°C during the night, moderate rainfall (100–200 mm), and NDVI levels exceeding 0.6. This implies that malaria transmission intensify after cumulative rainfall creates optimal mosquito breeding conditions. These findings support the development of early warning systems and targeted vector control interventions. Future research should explore the integration of Bayesian machine learning and further evaluate the comparative performance of different priors across various epidemiological contexts.