A multivariate spatiotemporal model for tracking COVID-19 incidence and death rates in socially vulnerable populations

A framework of zero-inflated Bayesian negative binomial regression models for spatiotemporal data

“…In panel (b), the counties with moderate overall SVI have more deaths than the high and low group. It is different from the finding that counties with high SVI had more COVID-19 deaths in other research focused on the U.S. or other locations (Karmakar et al, 2021, Neelon et al, 2022. To further understand how the COVID-19 deaths spread across the three SVI categories, we show the map of the daily average deaths in four different months in Figure 8 Table 6 shows the estimated coefficients as well as the 95% credible intervals in the binary and count component.…”

“…Starting from July 2020, the death counts has increased for all counties in Florida, and counties in the moderate category had the most significant increase than others. We also apply a recently developed spatiotemporal model (not in a zero-inflated setting) (Neelon et al, 2022) which used B-splines to model temporal effects for the COVID-19 analysis. We notice that it could be difficult to set the location and number of knots appropriately.…”

“…This may be related to emergency orders of closing certain indoor spaces and outdoor spaces where groups of people congregated without physical distancing in July 2020, which are not included in the analysis. For comparisons, we show the fitting results using the spatiotemporal model proposed in Neelon et al (2022) in Figure 13. The posterior mean daily deaths do not capture the trends very well and the 95% CIs fail to cover most the true values for Hernando county in panel (a), Polk county in panel (c), and Miami-Dade county in panel (d).…”

“…Distance-based exponential or Matérn covariance functions are commonly adopted to specify spatial correlations in geostatistical data, while conditionally autoregressive models are often used for temporal effects (Cressie and Wikle, 2015, Banerjee et al, 2003 through a hierarchical modeling framework (Cressie and Wikle, 2015). Neelon et al (2022) proposed a cubic B-spline model to fit both temporal (overall time trend) and spatial (county-specific trend) random effects in a Bayesian framework, which depends on the number and locations of the interior knots and specification of the degree of basis functions.…”

A Marginalized Zero‐Inflated Negative Binomial Model for Spatial Data: Modeling COVID‐19 Deaths in Georgia

“…For convenient updates of the spatial effects for both the binary and overall mean parts, we partition expression (9) into univariate conditional prior for each vector 𝝓 ℎ (ℎ = 1, 2) given the remaining random effects (Neelon 2019;Neelon, Wen, and Benjamin-Neelon 2022;Neelon, Zhu, and Neelon 2015). Using the properties of conditional multivariate normal theory, the conditional prior for the spatial intercepts for the binary component, 𝜙 1𝑖 (𝑖 = 1, … , 𝑛), is given by…”