Predictive spatial modeling of sociodemographic risk for COVID-19 mortality
TL;DR
Multivariate regression results for all regions across three time windows suggest that existing measures of social vulnerability for disaster preparedness (SVI) are predictive of a higher degree of mortality from COVID-19, and GWRF results indicate that a more nuanced modeling strategy is useful for determining spatial variation versus regional modeling approaches which may not capture feature clustering along border areas.
Abstract
In early 2020, the Coronavirus Disease 2019 (COVID-19) rapidly spread across the United States (US), exhibiting significant geographic variability. While several studies have examined the predictive relationships of differing factors on COVID-19, few have looked at spatiotemporal variation of COVID-19 deaths at refined geographic scales. The objective of this analysis is to examine the spatiotemporal variation in COVID-19 deaths with respect to socioeconomic, health, demographic, and political factors. We use multivariate regression applied to Health and Human Services (HHS) regions as well as nationwide county-level geographically weighted random forest (GWRF) models. Analyses were performed on data from three separate time frames which correspond to the spread of distinct viral variants in the US: pandemic onset until May 2021, May 2021 through November 2021, and December 2021 until April 2022. Spatial autocorrelation was additionally examined using a local and global Moran’s I test statistic. Multivariate regression results for all regions across three time windows suggest that existing measures of social vulnerability for disaster preparedness (SVI) are predictive of a higher degree of mortality from COVID-19. In comparison, GWRF models provide a spatially explicit evaluation of feature importance and prediction, identifying localized prediction and feature importance patterns, such as for obesity, that were not apparent in the regional regression models. Spatial autocorrelation indicates positive spatial clustering, with a progression from positively clustered low deaths for liberal counties (cold spots) to positively clustered high deaths for conservative counties (hot spots). GWRF results indicate that a more nuanced modeling strategy is useful for determining spatial variation versus regional modeling approaches which may not capture feature clustering along border areas. Spatially explicit modeling approaches, such as GWRF, provide a more robust feature importance assessment of sociodemographic factors associated with COVID-19 mortality.