Spatial prediction of COVID-19 pandemic dynamics in the United States

Background The impact of COVID-19 across the United States has been heterogeneous, with some areas demonstrating more rapid spread and greater mortality than others. We used geographically-linked data to test the hypothesis that the risk for COVID-19 is spatially defined and sought to define which features are most closely associated with elevated COVID-19 spread and mortality. Methods Leveraging geographically-restricted social, economic, political, and demographic information from U.S. counties, we developed a computational framework using structured Gaussian processing to predict county-level case and death counts during both the initial and the nationwide phases of the pandemic. After identifying the most predictive spatial features, we applied an unsupervised clustering algorithm, topic modelling, to identify groups of features that are most closely associated with COVID-19 spread. Findings We found that the inclusion of spatial features modeled case counts very well, with overall Pearson’s correlation coefficient (PCC) and R 2 of 0.96 and 0.84 during the initial phase and 0.95 and 0.87, respectively, during the nationwide phase. The most frequently selected features were associated with urbanicity and 2020 presidential vote margins. When trained using death counts, models revealed similar performance metrics, with the addition of aging metrics to those most frequently selected. Topic modeling showed that counties with similar socioeconomic and demographic features tended to group together, and some feature sets were associated with COVID-19 dynamics. Unsupervised clustering of counties based on these topics revealed groups of counties that experienced markedly different COVID-19 spread. Interpretation Spatial features explained most of the variability in COVID-19 dynamics between counties. Topic modeling can be used to group collinear features and identify counties with similar features in epidemiologic research.