Multivariate Analysis of Factors Affecting COVID-19 Case and Death Rate in U.S. Counties: The Significant Effects of Black Race and Temperature

COVID-19 Cases and Deaths Data

COVID-19 confirmed case number and death number through April 14, 2020 were obtained for each U.S. county from the Center for Systems Science and Engineering (CSSE) Coronavirus Resource Center at Johns Hopkins University. This data source contains COVID-19 case number and deaths data from all 3,143 U.S. counties in 50 states and the District of Columbia (D.C.). Cases per 100,000 people and deaths per 100,000 people for each U.S. county were calculated using CSSE and census data. Counties were excluded from the analyses if they had fewer than 50 cases or their first case occurred fewer than 3 weeks prior to the end of the study. With these constraints, 661 counties from 50 states and Washington D.C were included in case analysis. Counties were excluded from death analyses if they reported fewer than 10 deaths or the first death had occurred fewer than 2 weeks prior to the end of the study. The deaths analyses include 217 counties from 37 states and D.C. (Table 1).

Data Sources for Covariates

Race demographics for counties was obtained from the County Health Rankings and Roadmaps Program database. Daily temperature data for counties was obtained from the National Oceanic and Atmospheric Administration. County temperature was calculated using mean temperature for a period starting 10 days before the first confirmed county case and through the most current date (April 14, 2020). Unacast social distancing data was obtained through a research agreement with the company. Data for other potential risk factors used in analysis was obtained through publicly available data sources. (Table 2).

Statistical Analysis

T-tests were used to compare differences between covariates for the highest quartile and lowest quartile cases per 100,000 and the top 50% versus the bottom 50% deaths per 100,000. Sequential regression modelling was used to demonstrate the effect of race and temperature on cases and deaths per 100,000. Logistic regression was used to compare counties in the highest quartile for cases per 100,000 to those in the lowest. There were an insufficient number of counties meeting the deaths per 100,000 requirements to run logistic regression, thus linear regression was used for deaths per 100,000 analyses. Results of linear regression for cases per 100,000 including all 661 counties are provided for direct comparison between cases and deaths.

The sequential model had four parts. Each model included all variables from the previous models. Model 1 was a univariate analysis. Model 2 added the following county macroeconomic and COVID-19-specific variables: population density, GDP/capita, COVID-19 tests per 100,000 people (state level data), COVID-19 cases/100,000 people (deaths analysis only), a marker for which State the county is in, average percent reduction in cellphone movement from day of first confirmed case to end of study, the percent of population living in overcrowded housing, and the number of days since the first confirmed COVID-19 case. Model 3 added these county demographics and environmental variables to Model 2: percent of population over 65, proportion of Black residents, percent of the population that is female, percent of population living in rural areas, the Food Environment Index (a measure of accessibility and affordability of healthy food), the rate of violent crime per 100,000 people, the average temperature from 10 days before the first case to the end of the study, air quality measured as the average annual ambient concentrations of PM2.5, percent of the population considered to be in fair or poor health, and the poverty rate. Model 4 was the full model and included medical comorbidities and access to health care variables in addition to all variables in Model 3. Specifically these new variables were: diabetes, obesity, physical inactivity, excessive drinking, and smoking were all reported as percent of the population; liver disease, hypertension, coronary heart disease, and chronic respiratory disease were reported as mortality per 100,000 people; and the patient to primary care physician ratio, the percent of people sleeping fewer than seven hours per night, the percent of the population without health insurance and the percent of the population who received the flu vaccine were also included

All statistical analyses were performed with Prism 8.0 (GraphPad Software, San Diego, CA). For all analyses, α = 0.05.

Impact of Race

In recent weeks, numerous headlines and research studies have brought race to the forefront as a potentially significant risk factor for increased COVID-19 incidence in the United States.³¹ African Americans have accounted for all but three COVID-19 related deaths in St. Louis, Missouri and three quarters of those in Milwaukee, Wisconsin.³² Further, recent analyses of patients hospitalized for COVID-19 in 99 U.S. counties spanning 14 states have demonstrated significantly increased hospitalization rates amongst Black Americans.^6,31,33 Our study corroborates and expands on these findings to suggest that Black Americans may be at significantly higher risk of COVID-19 infection and mortality nationwide. However, at this time, current theories to explain this disproportionate burden of COVID-19 morbidity and mortality in the Black population have remained speculative and limited by confounding variables. Recent studies on COVID-19 incidence amongst Black Americans have pointed to the fact that, compared to the general population of the United States, Black Americans have been shown to suffer from increased rates of chronic medical comorbidities such as cardiovascular disease, diabetes mellitus, hypertension, obesity, and chronic respiratory disease, all of which have been shown to convey increased risk of SARS-CoV-2 infection and lead to worse outcomes.^18-20 Studies have also cited the many inequities in structural variables, which now manifest in the disproportionate number of Black Americans who suffer from poverty, reside in densely packed areas with more environmental hazards, have decreased access to healthy food sources, and lack healthcare coverage and access.³¹ In our comprehensive multivariate analysis, we were able to control for these potentially confounding variables; however, after doing so, our data continue to demonstrate a disproportionate number of COVID-19 cases and deaths in counties with large Black populations in the United States.

This increased disease burden did not seem to be explained by previously proposed mechanisms, indicating that other modulating factors should be considered and novel interventions designed appropriately. One potential domain includes additional sociocultural variables that have only now become relevant within the context of the pandemic. For example, African Americans make up a large percentage of the healthcare, transportation, government, and food supply industries, job sectors that have now been deemed “essential” services in light of the SARS-CoV-2 pandemic.³⁴ Despite the majority of Americans currently living in counties with some form of social distancing mandate, a recent announcement by the U.S. Surgeon General indicated that fewer than one in five Black Americans have a job that gives them this flexibility to work from home, compared with more than a third of white and Asian American workers.³⁵ The potentially fatal consequences of this fact can already be seen in the case of the New York City transit system. The city’s transit workforce, which has been among the hardest hit by the virus with more than 2,000 cases, 59 deaths, and over 4,000 employees in quarantine, is comprised mostly of ethnic and racial minorities, with African Americans accounting for nearly half of transit workers despite making up just 24% of the city’s overall population.^36,37 Furthermore, despite an 87% reduction in overall ridership since the pandemic began, the nearly one million remaining passengers–most of whom lack the luxury of social distancing due to essential employment–are predominantly low-income people of color.³⁸

However, this theory is complicated by the fact that Hispanic workers also makeup a disproportionate percentage of the essential workforce, yet our data do not suggest an increased burden of COVID-19 cases and mortality amongst counties with greater Hispanic populations.³⁵ A potential explanation may be that Black workers may be more likely than other racial and ethnic minority groups to work essential jobs that also require close proximity and frequent contact with others, such as bus drivers, postal workers, and grocery store clerks, which would convey additional risk of COVID-19 infection.^39,40 However, more work is needed to elucidate potential mechanisms to explain the differing disease burdens between these vulnerable populations.

One potential consideration is that non-Hispanic Black populations in the United States have been shown to suffer from vitamin D deficiency at rates much higher than any other ethnic or racial group, with recent data suggesting that non-Hispanic Black and Hispanic populations in the United States suffer from vitamin D deficiency at rates of 24 and 4 times more than those seen in the white population, respectively.⁴¹ In this comprehensive analysis of serum 25-hydroxyvitamin D (25(OH)D) levels collected from more than 26,000 adults in the United States between 2001-2010 as part of the National Health and Nutrition Examination Survey (NHANES), 71.9% of non-Hispanic Black individuals were found to suffer from vitamin D deficiency (defined as serum concentrations of less than 50 nmol/L) even after controlling for other potential predictors.⁴¹ Of the many sociodemographic, behavioral, and clinical variables studied, being a racial or ethnic minority was the strongest predictor of vitamin D deficiency.⁴¹

Vitamin D has been suggested as a potentially mitigative factor in the COVID-19 pandemic due to its important modulatory effect on immune response.^42,43 Furthermore, vitamin D deficiency has been implicated in numerous adverse health conditions such as acute respiratory syndromes, tuberculosis, cardiovascular disease, autoimmune disease, and some cancers.⁴⁴ This higher prevalence of vitamin D deficiency in darker-skinned individuals is thought to largely be due to increased melanin pigmentation, which absorbs significantly more radiation from sunlight and thereby reduces the available ultraviolet-B radiation that is needed to trigger natural vitamin D production in the skin.⁴⁵

Impact of Temperature

The potential role of vitamin D as a modulator of COVID-19 burden is further supported by our other significant finding regarding temperature, which suggests a protective effect of warmer climates. Several recent reports have also linked higher temperature to decreased virus spread, while others have found no significant effects.^9-14 Here, we demonstrate an independent effect of temperature that results in reduced COVID-19 cases, but not mortality, across U.S. counties. If warmer temperatures do, in fact, play a role in mitigating disease spread, it is reasonable to expect a potential seasonal trend in global cases and mortality. As temperatures begin to warm in the Northern hemisphere during the summer months, we may begin to see a decrease in disease burden; however, we should prepare to see a resurgence in COVID-19 incidence when temperatures again decrease in the fall and winter months, as has been exhibited in previous pandemics like the Spanish flu.⁴⁶ Likewise, countries in the Southern hemisphere should be prepared to experience seasonal trends in the opposite direction.

The mechanism through which temperature may execute its mitigative effects remains unclear, though it has already been posited that increased vitamin D levels due to greater sun exposure may result in better immune response against the SARS-CoV-2 virus, as it has been shown to do within the context of other viruses.^42,43 Notably, when temperature is substituted out of our final model for other vitamin D proxies, such as average daily sunlight and latitude above and below 37 degrees, the results are also highly significant in the same direction. Including these variables together in the model resulted in multicollinearity, which suggests that they are most likely measuring the same effect. Thus, in this study, average temperature prior to the first reported COVID-19 case was chosen to be the vitamin D proxy because it was the continuous variable that contained the most specific and validated data at the county level.

One potential clinical implication of these findings is on the use of chloroquine and hydroxychloroquine therapy in treatment of the SARS-CoV-2 virus, which has shown some promise in early research.⁴⁷ Chloroquine has been shown to reduce the conversion of vitamin D from calcidiol to calcitriol, the former of which has been demonstrated to potentiate the immune system,⁴⁸ which may serve as a partial mechanism for the purported beneficial effects of chloroquine therapy on COVID-19 outcomes. Future research, both clinical and epidemiological, should focus on serum vitamin D levels directly to see if they demonstrate any significant effects on COVID-19 spread and morbidity. Such studies on the role of serum vitamin D levels on COVID-19 outcomes are already underway, including a retrospective review of patients with documented vitamin D levels at our institution as well as a prospective randomized, controlled clinical trial conducted by the University of Grenada that is investigating the role of vitamin D supplementation therapy versus placebo on outcomes in 200 COVID-19 patients.⁴⁹

Our results support the hypothesis that vitamin D supplementation, whether through behavioral or dietary interventions, may prove to be beneficial for SARS-CoV-2 outcomes in the United States. Though the potential benefits of vitamin D supplementation seem to be particularly relevant for racial and ethnic minorities, they should also be considered for the entire population, as studies have found that less than 30% of U.S. adults have vitamin D levels that are considered sufficient for optimal health outcomes.⁴¹ Though standardized data on worldwide vitamin D levels are currently lacking, existing meta-analyses suggest that vitamin D deficiency is also a rampant issue on a global scale and should be prioritized for future research in light of the current pandemic.⁵⁰ If the theorized effects of vitamin D on COVID-19 morbidity are supported by ongoing studies, vitamin D deficiency could emerge as a unifying theory for the findings herein and allow countries to mobilize resources in anticipation of seasonal trends and develop targeted interventions to mitigate risk.

Additional Factors

Other findings associated with increased COVID-19 cases and deaths included increased poverty, increased GDP per capita, increased pollution, and increased flu vaccination rate. Poverty and GDP are not typically thought to be positively correlated with each other, however, the correlations seen here may represent inequalities in health care versus and access to testing.For example, poverty may exacerbate case spread, while economic power increases testing rates, both of which would lead to increased cases.⁴⁷ Flu vaccination was included as a marker for access to healthcare and may be associated with greater testing rates. Additionally, flu vaccination is often mandated for healthcare workers, thus counties with higher flu vaccination rates may have more of its citizens having front line exposure to the virus.^52,53

Decreased pollution was associated with increased COVID-19 cases, which does not align with prior findings in studies that associated increased air pollution with increased mortality.^15,29 Our model also shows an inconclusive connection between air quality and COVID-19 deaths. However, air pollution data in this study is from 2014 and may not reflect current air pollution. Given increased social distancing, many counties with historically bad air pollution currently have significantly decreased pollution levels, which may have decreased pollution related COVID-19 cases.^54,55

Although the effect was occasionally significant in some steps of the sequential regression models, increased social distancing as measured by the average reduction in cellphone movement from the time of the first case to the end of the study was not a robust modulator of cases or deaths (data not shown). The most likely reason for this is due to the simultaneity bias of this metric. In particular, the metric used here does not differentiate between proactive counties and reactive counties. It is expected that less proactive social distancing will cause more cases and deaths (negative correlation) while higher numbers of cases and deaths induces fear, causing increased reactive social distancing (positive correlation). Since these effects work in opposite directions, this bias makes it more difficult to find an effect of social distancing. Studies that plan to model social distancing with numerous covariates should ensure their metric adequately delineates between proactive and reactive social distancing or employ an appropriate instrument to remove this simultaneity.⁵⁶

Limitations

The emerging nature of Covid-19 required inclusion of only counties with sufficient data. Many counties were excluded from analysis for lack of Covid-19 cases or lack of Covid-19 deaths. This limitation is unavoidable and waiting for sufficient data may deter public health response. This study is also limited by the fact the COVID-19 pandemic is still in progress in all of the counties included here and is at different stages. Thus, the number of cases and mortality statistics for the counties provides a snapshot of the current state, but may not, and likely will not, be reflective of the ultimate case and death tolls in these counties. Moreover, most counties are still in the early stages of accumulating deaths which axiomatically lags behind the number of cases. While our results of temperature and Black race on cases per 100,000 have been consistent for a number of weeks, the number of counties with at least 10 deaths has only recently become large enough to perform similar analyses. Temperature also demonstrates this lack of simultaneity, but changes more predictably and slowly than Covid-19 cases and deaths. Additionally, since we relied heavily on publicly available and easily accessible sources to create our database, we used sources that may be outdated or otherwise inconsistent with the actual value of various statistics during the past few months. Confounding effects are also possible, given that it is impossible to fully extricate socioeconomic, demographic, and even some environmental variables from each other. The stepwise approach to regression mitigates excessive confounding risk through sensitivity analysis and signifies that the findings are not error artifacts. Ecological fallacies may be present in the model for demographic and socioeconomic variables compared to individual cases and deaths; we report as granularly as practical to account for this.