Details of Current Models

The datasets (i.e., components) comprising models 11.2 and 12.4 are described below. The main difference between the models is the inclusion of vaccination data (and subsequent re-weighting) in model 12.4. For both models, each component is assigned to an indicator (data slice) categorized into four domains: Infection Rate, Population Concentration, Intervention Measures, and Health & Environment. The expanded tables include a description of each component, the rationale for its inclusion, and a link to the public data source. Details of the methods used to build and assess the PVI are provided below.

The time-series data underlying the current model are available at github.com/COVID19PVI/data. The software used to generate PVI scores and profiles from these data is freely available at toxpi.org.

The PVI was developed to summarize and communicate complex, high-dimensional data related to communities’ vulnerability to COVID-19. The corresponding scorecards are human-centric and focus on impacts at the community level. Accordingly, data were apportioned into slices using a combination of quantitative modeling and prior knowledge of risk drivers.

To gauge the association of daily PVI values compared with cumulative death and case numbers, the rank-correlation between overall PVI values and the key vulnerability-related outcome metrics of cumulative deaths (Figure 1A/B) and cumulative cases (Figure 1C/D) was compared. The Spearman Rho values for the PVI (from January 13 through May 16, 2021) compared to outcomes 1, 7, 14, 21, and 28 days ahead of a given date are displayed. All daily rank-correlation estimates were highly significant (all P -values < 1e-14). The mean Rho values for each time horizon are displayed in blue text.

While the general PVI model communicates an integrated concept of vulnerability, purpose-built forecasting models were used to predict case and death numbers. Predictive modeling for both COVID-19 cases and deaths presents several unique challenges. Unlike other airborne viruses (e.g., influenza), testing methods and best medical practices for this contagion quickly evolved during the pandemic. Tests were initially limited to those presenting severe symptoms, testing practices differed by geographic region, and testing became more widely available as the pandemic progressed and resources became available. As a result, it is difficult to compare the observed number of cases per population and deaths per case across time or geographic areas. To accurately predict future cases and deaths, it was necessary to account for the fluid nature of the data.

Accordingly, the team developed a Bayesian spatiotemporal random-effects model that jointly describes the log-observed and log-death counts to build local forecasts. In this model, log-observed cases for a given day are predicted using known covariates (e.g., population density, social distancing metrics), a spatiotemporal random-effect smoothing component, and the time-weighted average number of cases for these counts. This smoothed time-weighted average is related to a Euler approximation of a differential equation; it provides modeling flexibility while approximating potential mechanistic models of disease spread. The smoothed case estimates are used in a similar spatiotemporal model to predict future log-death counts based on a geometric mean estimate of the number of observed cases for the previous seven days as well as other data streams. Figure 1 displays the resulting county-level predictions and corresponding confidence intervals.

Figure 1. Spearman correlation (Rho) estimates of daily PVI (by county) vs. Cumulative Deaths (A), Population Adjusted Cumulative Deaths (B), and % Deaths/Cases (C) for 1, 7, 14, 21, or 28 days into the future. Mean Rho values for each time horizon are displayed in blue.

Are You Ready to Get Started?

Visit the COVID-19 Pandemic Vulnerability Index (PVI) Dashboard