Abstract / Summary
Abstract Background COVID-19 showed pronounced geographic variation across the United States, yet most county-level analyses summarize it in one cumulative outcome and fit a single global model, assuming the county-outcome association was stable in time and constant in space. We relax both. Methods We assembled a dataset covering 3,103 US counties, linking daily reported COVID-19 cases and deaths, January 2020 to March 2023, to 43 socio-environmental and resilience predictors assembled and checked against primary sources. Deaths were modeled as counts with a population offset using negative binomial regression, cumulatively and within seven calendar windows. Prediction was assessed by spatially blocked cross-validation, non-stationarity by spatial-lag, geographically weighted (GWR) and multiscale geographically weighted (MGWR) regression models and a coordinate-permutation test, and age by indirect standardization. Results County composition rather than county size carried the predictive information. Out-of-sample deviance explained for cumulative mortality was 0.455 for the 43 factors against 0.106 for population size alone, and for cases 0.214 against 0.005. The size-mortality association nonetheless reversed. Size alone gave an IRR per SD of log population of 1.478 (95% CI 1.410 to 1.550) between January and May 2020, and 0.802 to 0.883 in the six later windows (each P < 0.001). Adjusting for size and all 43 factors left the gradient non-significant in six of seven windows, so the size association does not persist under adjustment. MGWR fitted better than GWR (corrected AIC 5670.2 against 5772.9) cumulatively and in all seven windows, though neither improved out-of-sample prediction. On 950 coordinate permutations only ultraviolet exposure (Holm-adjusted P = 0.009) and household composition and disability (P = 0.017) exceeded chance variation under family-wise correction, and physical inactivity under false-discovery-rate control alone; the intercept and population-size surfaces did not. Conclusions County-level COVID-19 mortality is predicted out of sample by socio-environmental composition rather than county size, and its geography was not fixed: the size gradient reversed after the first wave, and a cumulative estimate is dominated by the later regime. This helps explain why single-period summaries have struggled to find consistent county-level predictors. Holm-adjusted permutation testing reversed the bandwidth-only reading for four of six surfaces.