The public health crisis created by the COVID-19 has uncovered the historical inequalities (1–4) between ethnic groups in certain countries, in particular in the UK and US, which are countries with ethnically diverse populations. These observations and consistent fatal outcomes in the minority ethnic groups (5, 6) have led to speculations about why patients from these groups are susceptible to infections, followed by severe complications. These trends could be due to different rates of COVID-19 infections, underlying health conditions, living conditions including housing density, having jobs as essential workers, access to health care, quality of care, and a mixture of multiple factors among these groups. The United States national data (7) from states and municipalities reports disproportionate COVID-19 infections, hospitalizations, and deaths among minority ethnic groups. Dobin and Dobin (8) showed that the infection rate is 4-fold for the Black and Hispanic population in selected counties in New York state. Moore et al. (9) observe a disproportionate number of COVID-19 cases among underrepresented racial/ethnic groups in the United States. Adhikari et al. (10) show that the racial and ethnic disparities in COVID-19 infections and deaths existed beyond those explained by income inequality.

The impact of COVID-19 varies widely across countries and even within a country or a region. For example, Sun et al. (11) showed a negative correlation between the number provincial COVID-19 cases and latitude, as well as altitude. Breen and Ermisch (12) use spatial autoregressive regression to show that the relation of COVID-19 mortality to social composition of geographical areas in England is distinct than that of non-COVID mortality. A number of factors including societal awareness and culture, public health measures, healthcare infrastructure, and more recently vaccination coverage are known underlie the variation for COVID-19 infection rates and adverse health outcome (13).

Although multiple studies have confirmed that black and Hispanic populations in the US are more vulnerable to COVID-19, to our knowledge, no data is available if they are equally susceptible across geographical locations. Stephens-Davidowitz (14) uses the search data from Google to show that there exists a wide variation in racism in the US within the 50 states. Thus, we surmise that the impact of the COVID-19 on various ethnicities may not be uniform in all regions of the US. Hence, we are interested in understanding if a geographical location plays a part in the variation of COVID-19 impact on minorities.

Emerging evidence highlights that comorbid conditions such as obesity, cardiovascular disease (CVD), and type 2 diabetes are directly linked to the severity of the COVID-19 disease (15–17). A meta-analysis including 76,993 patients with COVID-19 showed diabetes, CVD, smoking, malignancy, chronic kidney disease, hypertension, chronic obstructive pulmonary disease (COPD) are associated with poor prognosis (18). This conclusion was further supported by Richardson et al. (19), and Sun et al. (20). Using logistic regression (21) show that obesity was a risk factor for the severity of the COVID-19 disease. Furthermore, in a retrospective cohort study, Busetto et al. (22) conclude that despite their young age, overweight patients were more likely to need assisted ventilation and access to intensive care units than patients with normal weight. The connection between obesity and pulmonary function is well-established, e.g., Sharp et al. (23) observe that obese patients have significantly decreased total respiratory compliance. Moreover, Li et al. (24) find that reduction in functional residual capacity and diffusion impairment are the most common abnormalities in obese patients. Yan et al. (18) show that diabetic patients experienced more mortality than non-diabetic patients. Finally, just as in the case of the SARS epidemic (25), COVID-19 has disproportionately affected the older population (26). In fact, in the US, 92% of the COVID-19 recorded deaths till June, 2020 are in the age group 55 years and above (27). In summary, the main comorbidities for COVID-19 include obesity, diabetes, advanced age, hypertension, and cardiovascular disease.

However, these risk factors affect different ethnicities differently. For example, Paeratakul et al. (28) find that among obese individuals, the prevalence of hypertension was higher in NHB subjects than other groups. Sturm and Hattori (29) observe that the prevalence of obesity is about double among NHB than among Hispanics or NHW. Kuzawa and Sweet (30) note that NHB suffer from a disproportionate burden of CVD relative to NHW. Thus, we are motivated to understand whether or not these comorbidities affect different ethnicities differently.

Patel et al. (31) note that economically disadvantaged people are vulnerable to COVID-19 due to a combination of factors. A time-series analysis conducted by Elgar et al. (32) reveals that income inequality is associated with a higher number of deaths due to COVID-19 in 84 countries. In particular, in the US, the states with higher income inequality experienced a higher rate of infection as well as the number of COVID-19 related deaths (33). This pattern could be because the comorbidities associated with COVID-19 are linked to poverty.

A longitudinal study involving 600,662 adults from Taiwan's National Healthcare Insurance database indicates that diabetes incidence is associated with poverty (34). This finding is particularly notable since the subjects from this study had access to universal healthcare. However, the subjects were from a ethnically homogeneous population. Thus, we are motivated to investigate the differential role of poverty among various races.

For this study, we choose Non-Hispanic white (NHW), Non-Hispanic Black (NHB), Hispanic, and Non-Hispanic Asian (NHA) as four ethnic groups. The risk factors we choose to focus on in this work are advanced age, obesity, cardiovascular disease, diabetes, hypertension, and poverty. We aim to investigate the following in this study:

To this end, we collected COVID-19 related infections and mortality data from various publicly available sources, as described in section 2.1. To investigate the geographical variation in the impact of COVID-19, we seek to quantify the difference between the distributions of infection rates, mortality rates, and populations of various ethnic groups in four geographical regions of the US (35). We describe our approach for this analysis in section 2.4.1. Finally, in section 2.4.3 we construct robust linear models with infection and mortality rates of various ethnicities as response variables.

Given the discussion in the sections 1.3, 1.1, and 1.4, we focus on the following factors in this work: obesity, diabetes, poverty, advanced age, hypertension, and cardiovascular diseases. We included the following ethnic groups in our work: NHW, NHB, Hispanic, and NHA. We excluded the groups American Indian/Alaska Native, and Native Hawaiian/Other Pacific Islander from this work due to lack of reliable and consistent data sets in the US, cf. (36–38). We used the American Community Survey (39) census database to collect the following information for each state: NHW, NHB, Hispanic, and NHA population and their respective percentage contribution to the population of each state. To investigate the geographical effect on COVID-19 prognosis, we used the classification of the US into the following categories: Northeast, South, Midwest, and West (35). COVID-19 infection and death counts between January 21, 2020, to September 30, 2020 from KFF Covid-19 data (40) were obtained. The time window roughly corresponds to the first pandemic wave experienced in the US. For each ethnicity E, and for a specific state S the variable Relative Infection % is defined as follow:

We note that the choice of “relative” infection percentage as a response variable is deliberate. This choice allows us to directly compare this number with the population share of that ethnicity in the region. Similarly, for each ethnicity E, and for a specific state S we define the variable Relative Mortality % as follow:

In our work, Relative Infections % and Relative Mortality % were considered as the response variables. For brevity, we write infection rate instead of Relative Infections %, and so on. The use of relative percentages allows a direct comparison with the population percentages of that ethnicity. For example, in a state with a 5% NHB population, relative mortality of 15% in the NHB community indicates disproportionately large mortality compared to the NHB population. The use of “relative” percentage is independent of the population of the state itself. The use of this measure also allows us to compare the impact on a certain ethnic group in two states with similar proportion of the minority population. As a concrete example, when we consider the states of California and Texas, both have a similar percentage of the Hispanic population, 39.5 and 40%, respectively. However, the relative mortality percentages for the Hispanic group in California and Texas are 48.3 and 56.1%, respectively. We collected the data on the percentage of people with age 60 or more in each state and ethnicity is obtained from the CDC dataset (41). Race and state-wise data were obtained from adults who reported being told by a health professional that they have diabetes (excluding prediabetes and gestational diabetes) using the America's Health Rankings (42). We used the body mass index (BMI) as a measure of obesity following (43) and define obesity as a condition of having a BMI of 30.0 or higher. The dataset (44) were used to obtain the obesity data from each state and for the races NHW, NHB, Hispanic, NHA. We acquired the percentage of adults whom a health care professional informed that they had a coronary heart disease, or myocardial infarction, or a stroke from AHR CVD data (45). This was gathered for each state and ethnicity of interest. We obtained the race and state-wise data on adults who reported being informed by a health professional that they have high blood pressure from AHR HBP data (46). The US Census Bureau defines the “poverty threshold” for a family with two adults and one child as $20,578 in 2019. We extracted the data from KFF Poverty data (47) on poverty defined by the “poverty threshold.” We obtained this data for each state and ethnicity of interest.

For each state, and each of the four ethnicity of interest (NHW, NHB, Hispanic, and NHA) we defined the variables: Age60+, BMI30+ (a measure of obesity), CVD, Diabetes, HBP, Poverty. For a state S, and an ethnicity E we defined the relative percentage of people with age 60 or over Age60+ as follows:

We use the variable name Age60+ instead of Relative Age60+ % for conciseness, and so on. We define the relative percentage variables Obesity, CVD, Diabetes, HBP, and Poverty in a similar manner.

We encountered a few irregularities during our data collection process in the format the data was made available by various states (36–38). For example, New York state does not provide ethnicity-wise COVID-19 infection and mortality data. Similarly, infection data for NHW communities only are available for North Dakota. Data availability of specific variables for different ethnicities also varied across states. Data for all variables for the NHW group could be obtained from 48 states whereas for the NHB and Hispanic communities such data were available from only 38 and 33 states, respectively. In our analysis, we included the states for which data for all the variables are available. Hatcher et al. (37) find that only 23 state in the US have complete data for American Indian and Alaska Native Persons.

The data for this study are state-level demographics based on four ethnic groups. We depict the relative infection % and relative mortality % for NHW, NHB, and Hispanic group in the map in Figure 1. As described in section 2.2 some states do not make the ethnicity-wise data public. The states with no color in the Figure 1 indicate that the ethnicity-wise infection and mortality data was not available in those states. We calculated state-wise descriptive statistics for the relative infection and mortality percentages and population comparing each ethnic group. We performed a descriptive analysis to explore the region-specific, state-wise characteristics of for the relative infection % and mortality % and population by calculating their medians, first and third quartiles, and presented in Figure 2.

Linear regression yields stationary and global regression coefficients. However, it is conceivable that these coefficients might have local variability. To find the geographical variability in the coefficients, we employed the geographically weighted regression (GWR) (55). Rather than producing global regression results, GWR yields “local” regression coefficients in terms of geographically varying functions. For our analysis, we used the infection rate and mortality rates as response variables and the variables obtained from section 2.4.3 as the explanatory variables.

For our coding, we used R language (version 4.0.0), along with the following libraries in our coding: readxl, dplyr, tidyr, FSA, ggplot2, car, qqplotr, nortest, pwr, spgwr, sp, sf, rgdal, rgeos, tmap, tmaptools.

The boxplots in Figure 2 summarize the relative impact of COVID-19 on various ethnicities across the four regions of the US and all regions as an aggregate. In Figure 2 we present various descriptive statistics of the population, infection, and mortality rates for the NHW, NHB, Hispanic, and the NHA groups across various regions. As noted in the section 2.2, not all the states are included in the analyses. Thus, the statistics shown in this plot do not correspond closely to those of the whole country. We describe the Kruskal-Wallis test results in Table 1. We see in Table 1 that the KW test for NHW is statistically significant in the Northeast and the West with p < 0.1. For the NHB group, the KW test is significant in the Northeast and the Midwest with p < 0.1. The KW test was statistically significant for the Hispanic group in “all four regions,” with p < 0.1 in the Northeast; with p < 0.05 in Midwest and the West; and p < 0.01 in the South. The NHA data yielded significant results with the KW test only in the South with p < 0.05.

When we considered all four regions in the US together, the KW test was statistically significant for all ethnicities with p < 0.01 for NHW and Hispanic communities. The KW test was significant for the NHB and NHA when all regions were combined with p < 0.1.

We followed the significant KW tests with Dunn's multiple comparison test to identify factors differing in their distributions. We depict the results from the Dunn's test in Table 2. In particular, we obtained statistically significant results (with p < 0.05) in the South, Midwest, and West for the Hispanic population between the pairs ‘infection & mortality rates' and “infection rate and population share.”

The KW test provides evidence of geographical impact on various ethnicities. In this section we provide the results of correlation analysis between other risk factors. In Table 3 we see the Pearson correlations between the variables along with the 2-tailed significance values for the NHW group. The same statistics are provided in Tables 4, 5 for the NHB and Hispanic communities respectively. All the variables are strongly (p < 0.01) and positively correlated with every other variable for all ethnicities, with poverty being the sole exception. To be precise, for the NHW group, poverty is positively correlated with the infection rate (r = 0.35, p < 0.05), and diabetes (r = 0.29, p < 0.05). Poverty is either uncorrelated or weakly correlated with other variables in this study for all ethnicities.

In Table 6 we depict the linear models with infection rate and mortality rate as response variables for NHW. The first column shows preliminary models along with the VIFs for each explanatory variable. The second column depicts the final model obtained via the maximum VIF elimination algorithm described in section 2.4.3. The preliminary model with infection rates in the NHW community as the response variable accounts for 83% [R2=0.83,Radj2=0.80,F(6,41)=32.87,p<0.01] of the variability in the infection rates for NHW population. The final model for the NHW infection rates consists of obesity, diabetes, and poverty as the only explanatory variables. This model accounts for 82% [R2=0.82,Radj2=0.80,F(3,44)=65.06,p<0.01] of the variability in the NHW infection rates. The final NHW infection model and preliminary model both use 48 states.

The preliminary model with NHW mortality rates as the response variable accounts for 88% [R2=0.88,Radj2=0.87,F(6,411)=51.95,p<0.01] of the variability in the NHW mortality rates. The final model for the NHW mortality also consists of obesity, diabetes, and poverty as the only explanatory variables. This NHW mortality final model accounts for 7% [R2=0.78,Radj2=0.77,F(3,44)=53.48,p<0.01] of the variability in the NHW mortality.

In Table 7 we depict the linear models with infection rate and mortality rate as response variables for NHB. The preliminary model with infection rates in the NHB group as the response variable accounts for 81% [R2=0.81,Radj2=0.77,F(6,31)=17.51,p<0.01] of the variability in the infection rates for NHB population. The final model for the NHW infection rates consists of obesity and poverty as the only explanatory variables. This model accounts for 66% [R2=0.66,Radj2=0.64,F(2,37)=37.89,p<0.01] of the variability in the NHB infection rates. Note that the final NHB infection model uses 40 states instead of 38 states in the preliminary model. This discrepancy is because of the unavailability of data for the NHB community for all the explanatory variables, as discussed in section 2.4.3. The preliminary model with NHB mortality rates as the response variable accounts for 77% [R2=0.77,Radj2=0.73,F(6,31)=17.51,p<0.01] of the variability in the NHB mortality rates. The final model for the NHB mortality also consists of only obesity and poverty as the explanatory variables. This NHB mortality final model accounts for 67% [R2=0.67,Radj2=0.65,F(2,37)=199.9,p<0.01] of the variability in the NHB mortality rates.

The Table 8 depicts the linear models with infection rate and mortality rate as response variables for the Hispanic group. The preliminary model for infection rates in the Hispanic community accounts for 67% [R2=0.67,Radj2=0.60,F(6,26)=8.97,p<0.01] of the variability in the infection rates for the Hispanic population. The final model for the Hispanic infection rates consists of diabetes and poverty as the explanatory variables. This model accounts for 51% [R2=0.51,Radj2=0.48,F(6,25)=9.98,p<0.01] of the variability in the Hispanic infection rates. The preliminary model with mortality rates among the Hispanic community as the response variable accounts for 71% [R2=0.71,Radj2=0.63,F(6,25)=9.98,p<0.01] of the variability in the Hispanic mortality rates. The final model for Hispanic mortality consists of advanced age and poverty as the explanatory variables. Note that advanced age is the most significant explanatory variable in the final mortality model in the Hispanic group, whereas having diabetes was the most significant variable predicting infection in the Hispanic community. This final model accounts for 55% [R2=0.55,Radj2=0.53,F(2,39)=23.93,p<0.01] of the variability in the Hispanic mortality rates. We note that the final model for the Hispanic mortality includes 42 states, whereas the preliminary model has only 32 states due to lack of data availability.

Adjusted R² value for the regression model for NHW mortality was much higher (0.77) in comparison to NHB (0.65) and Hispanic (0.53). However, all six models showed statistical significance and satisfied normality tests for the residual values. Indeed, the Lilliefors normality test applied to the residuals obtained from each of these models revealed that the residuals were normally distributed with p>0.001. The histograms, and the QQ plots for the residuals are provided in Figure 3.

The geographically weighted regression yields coefficients for each risk factor for every state. We show the state-wise coefficients for the most significant explanatory variable for each ethnicity in Figure 4. Empty spaces for states in Figure 4 indicate that ethnicity-wise data was not available for those states for the corresponding risk factors. The p-values for the most significant explanatory variable and state-wise R² values obtained from the GWR model are included in Supplementary Figures S4, S5.

As discussed in section 2.2 and noted by other researchers (36, 38) there is a lack of consistency and availability of COVID-19 related data. Our study does not include data from the state of New York, since the state does not make the ethnicity wise data available. Moreover, the data we use is state wise statistics of the various risk factors. However, we note that the observations made using such data is consistent with those made by other researchers.

Although racial and ethnic disparities in COVID-19 infections and mortality are becoming increasingly clear from several studies based on available data, drivers of these disparate outcomes remain less understood at a national level. Our models, based on the nationwide data, indicate that “obesity” effectively encapsulates the effect of other co-dependent factors for NHW, and NHB populations (section 3.3). The link between COVID-19 infection severity and obesity is noted by Watanabe et al. (66) even in the early stages of the pandemic. Similarly, during the H1N1 pandemic of 2009 (67, 68) observed that obesity was associated with higher mortality.

Another implication from our work is that the Hispanic community is more susceptible to the COVID-19 infection. This observation is valid throughout the US. This situation could be remedied via public policy changes and awareness of the issue. The disproportionate impact of COVID-19 on the minority population is largely attributed to existing socioeconomic inequities. The low-income minority population are often compelled to work in an environment with higher risk of disease exposure, live in a crowded accommodation, and lack adequate access to healthcare. The government support to low-income families in the form of the CARES Act, Consolidated Appropriations Act, 2021, Department of Treasury US (69) and the American Rescue Plan Act of 2021 (70) are critical but might not be sufficient to fully mitigate observed disparity in infection and mortality rates. Our analysis indicates that certain subpopulations of the minority population are at higher risk of COVID-19 infection and mortality. Identifying these vulnerable subpopulations, such as Hispanics with diabetes or age over 60 years, and prioritizing additional attention to these populations could enable a more efficient allocation and utilization of resources. Increased effort toward educating and raising awareness on COVID-19 and associated risk factors could also be an effective method to develop community resilience. One potential avenue to improve awareness on COVID-19 will be through recruiting volunteers to educate the vulnerable population. For example, “Philly counts” (71), a program supported by the Philadelphia Department of Public Health, initially created for Census 2020, currently helps direct community engagement efforts for the COVID-19 vaccine. Extending similar initiatives to populations with major risk factors such as obesity could result in a major beneficial impact on overall COVID-19 burden.