Globally Local: Hyper-local Modeling for Accurate Forecast of COVID-19

Abstract

AbstractMultiple efforts to model the epidemiology of SARS-CoV-2 have recently been launched in support of public health response at the national, state, and county levels. While the pandemic is global, the dynamics of this infectious disease varies with geography, local policies, and local variations in demographics. An underlying assumption of most infectious disease compartment modeling is that of a well mixed population at the resolution of the areas being modeled. The implicit need to model at fine spatial resolution is impeded by the quality of ground truth data for fine scale administrative subdivisions. To understand the trade-offs and benefits of such modeling as a function of scale, we compare the predictive performance of a SARS-CoV-2 modeling at the county, county cluster, and state level for the entire United States. Our results demonstrate that accurate prediction at the county level requires hyper-local modeling with county resolution. State level modeling does not accurately predict community spread in smaller sub-regions because state populations are not well mixed, resulting in large prediction errors. As an important use case, leveraging high resolution modeling with public health data and admissions data from Hillsborough County Florida, we performed weekly forecasts of both hospital admission and ICU bed demand for the county. The repeated forecasts between March and August 2020 were used to develop accurate resource allocation plans for Tampa General Hospital.2010 MSC92-D30, 91-C20