Title: Bayesian statistical machine learning methods for mapping health and development metrics using big data
Authors: Chigozie Utazi - University of Southampton (United Kingdom) [presenting]
Abstract: Recent technological developments such as the use of Global Positioning Systems (GPS) in surveys and data capture have resulted in unprecedented scales and frequencies at which data are collected. In the global health and development (H\&D) arena, this development has led to a rapid increase in the availability of and access to geo-referenced information for monitoring H\&D metrics. Many nationally representative household surveys are now geo-coded and increasing numbers of survey clusters typically geo-coded enumeration areas - are being included in more recent surveys that can enhance the estimation of H\&D indicators at finer scales than administrative-level one areas. Mapping spatial big data using Bayesian model-based geostatistical approaches poses computational challenges. We develop statistical machine learning methods for high resolution mapping of H\&D indicators using spatial big data. The methodology is an amalgamation of subsampling and ensemble approaches for fitting spatial generalized linear models. We consider random and stratified subsampling and model ensembles analogous to Bayesian model averaging. The Bayesian method is implemented using the INLA-SPDE approach and applied to mapping vaccination coverage using Demographic and Health Survey data. The output maps highlight significant heterogeneities in coverage levels which are indispensable for program planning and implementation.