Abstract

Estimating health and welfare at the neighbourhood level from surveillance data could have many social and economic applications, such as assisting local policymakers in organizing targeted activities to improve the health and well-being of their residents. However, health outcomes data often exhibit spatial correlation, rendering traditional statistical modelling methods impractical. While several spatial statistical models have been proposed to address this issue, they are prone to computational complexities. Machine learning techniques present attractive alternatives with potential benefits over traditional models, gaining recognition as solutions to small-area estimation challenges. Yet, they may not inherently recognize spatial context. Applying machine learning directly to health outcomes data without accounting for potential spatial autocorrelation could lead to biased outcomes. In this study, spatial context was integrated into machine learning techniques to efficiently capture the spatial variation of COVID-19 in contagion US and its associated risk factors. Beyond addressing spatial heterogeneity in COVID-19 data, the study utilized multiple indicators to create spatially weighted composite indicators to explain the disproportionate COVID-19 prevalence in small areas of the US in terms of demographic, socioeconomic, health access, vaccination, meteorological, and environmental risks.