Forecasting COVID-19 Infections across the US

Project Description

Until a vaccine is developed for COVID-19, monitoring and control of the pandemic requires accurate situational awareness at geographic scales relevant to decision making. In the United States, this means that tools must be available to reliably and accurately assess the current situation at state and county levels. Situational awareness at the county level is particularly difficult, however. Most counties in the US are small (median population 25,000), and most daily counts of new infections are small (single digits). From day-to-day, the number of cases can appear quite erratic, and it is difficult for most people to determine if day-to-day differences represent genuine changes in the presence of COVID-19, or are simply the product of noise that is inherent in all small counts.

In these situations, Bayesian, hierarchical techniques are known to be effective for estimating small groups by pooling data across many jurisdictions. Using the Stan software, we developed a Nowcasting model, capable of delivering estimates of the rate of infection for all 3,143 counties and county-equivalents with a daily update cycle.