This project implements a sparse conditional autoregressive Poisson model that predicts county COVID-19 death rates in the United States from county level features including:
- Percentage living in nursing homes
- Percentage Hispanic
- Percentage Black
- Percentage with at least high school education
- Median income
- Political lean from Democratic to Republican in the 2020 election
- Governor policy strictness
This work is intended as a supplement to our manuscript, The Evolving Roles of Partisanship and Vulnerability in the COVID-19 Pandemic.
We originally considered many more variables, but due to both the complexity of fitting a large spatial poisson model and the high correlation between many of the variables we have considered, we wanted to fit the spatial poisson model with only a small number of parameters. To select these in a parsimonious way, we tested parameters which appeared in the top 3 in terms of variable importance in any of the LASSO or spatial linear regressions presented in our main manuscript.
This is a correlation matrix of the variables originally considered:
The model results contained herein showed that the associations between the covariates mentioned above were robust even after accounting for spatial auto-correlation, or in other words the effect that counties which neighbor each other are expected to have similar outcomes.
We estimated separate models for periods 2 (June 1 2020 to September 30 2020) and 3 (October 1 2020 to February 12 2021).
Depicted below are the spatial effects estimated by the model in period 3.
Some counties were omitted from analysis due to missing data.
See the diagnostics.md file for a presentation of traceplots and the R-hat convergence diagnostics.
The full supplement detailing this work is available here, in supplement/supplement.pdf.
Spatial Models in Stan: Intrinsic Auto-Regressive Models for Areal Data