Application of elastic net regression for modeling COVID-19 sociodemographic risk factors

Abstract

Objectives: COVID-19 has been at the forefront of global concern since its emergence in December of 2019. Determining the social factors that drive case incidence is paramount to mitigating disease spread. We gathered data from the Social Vulnerability Index (SVI) along with Democratic voting percentage to attempt to understand which county-level sociodemographic metrics had a significant correlation with case rate for COVID-19.

Methods: We used elastic net regression due to issues with variable collinearity and model overfitting. Our modelling framework included using the ten Health and Human Services regions as submodels for the two time periods 22 March 2020 to 15 June 2021 (prior to the Delta time period) and 15 June 2021 to 1 November 2021 (the Delta time period).

Results: Statistically, elastic net improved prediction when compared to multiple regression, as almost every HHS model consistently had a lower root mean square error (RMSE) and satisfactory R2 coefficients. These analyses show that the percentage of minorities, disabled individuals, individuals living in group quarters, and individuals who voted Democratic correlated significantly with COVID-19 attack rate as determined by Variable Importance Plots (VIPs).

Conclusions: The percentage of minorities per county correlated positively with cases in the earlier time period and negatively in the later time period, which complements previous research. In contrast, higher percentages of disabled individuals per county correlated negatively in the earlier time period. Counties with an above average percentage of group quarters experienced a high attack rate early which then diminished in significance after the primary vaccine rollout. Higher Democratic voting consistently correlated negatively with cases, coinciding with previous findings regarding a partisan divide in COVID-19 cases at the county level. Our findings can assist regional policymakers in distributing resources to more vulnerable counties in future pandemics based on SVI.

Fig 1. Variable importance plots for both pre-Delta and Delta time periods, organized from lowest to highest overall importance.

A: Pre-Delta. B: Delta.

Fig 2. Observed versus predicted COVID-19 case rate plot for both pandemic time periods, used to visualize model performance.

The “perfect” correlation line shows a perfect one-to-one relationship between the observed data and the fitted data, whereas each time period-specific line shows the observed correlation for each respective pandemic time period.