Neighborhood Ecologies, Cases, and Deaths during the Beginning of the COVID-19 Pandemic: Lessons for Current and Future Epidemics?

Abstract

Objectives:

The coronavirus disease 2019 (COVID-19) pandemic has disproportionately affected Black and Latinx communities. Ecologic analyses have shown that counties with a higher percentage of Latinx and Black people have worse COVID-19 outcome rates. Few ecologic analyses have been published at the neighborhood (census tract) level. We sought to determine whether certain sociodemographic neighborhood ecologies were associated with COVID-19 case and death rates in metropolitan Atlanta, Georgia.

Methods:

We used census data and principal-component analysis to identify unique neighborhood ecologies. We then estimated correlation coefficients to determine whether the neighborhood profiles produced by a principal-component analysis were correlated with COVID-19 case and death rates. We conducted geographically weighted regression models to assess how correlation coefficients varied spatially for neighborhood ecologies and COVID-19 outcomes.

Results:

We identified two unique neighborhood profiles: (1) high percentage of residents, Hispanic ethnicity, without a high school diploma, without health insurance, living in crowded households, and lower percentage older than 65 years; and (2) high percentage of residents, Black race, living in poverty, unemployed, and households receiving Supplemental Nutrition Assistance Program benefits. Profile 1 was associated with COVID-19 case rate (Pearson r = 0.462, P < 0.001) and profile 2 was associated with COVID-19 death rate (Spearman r = 0.279, P < 0.001). Correlations between neighborhood profiles and COVID-19 outcomes varied spatially.

Conclusions:

Neighborhoods were differentially at risk of COVID-19 cases or deaths depending on their sociodemographic ecology at the beginning of the COVID-19 pandemic. Prevention methods and interventions may need to consider different social determinants of health when addressing potential cases and deaths during future emergent epidemics.

Coronavirus disease 2019 (COVID-19) case and death rates indicate that the pandemic has disproportionately affected Black and Hispanic communities.^1,2 Ecologic analysis has shown that counties with a higher proportion of Hispanic and Black populations have worse COVID-19 outcome rates.^3,4 Although many ecologic analyses of COVID-19 outcomes have been done at the county level, fewer ecologic analyses have examined the neighborhood (census tract) level. Census tract–level analyses allow for a more precise estimation of the neighborhood environment and better elucidation of proximate factors that may influence COVID-19 outcomes. The beginning of the COVID-19 pandemic allows us to study how different neighborhood socioecologic factors influence case and death rates for a newly emergent infection. Traditional epidemiologic analyses treat socioecologic factors as independent predictors of health outcomes. In the real world, however, socioecologic factors tend to group together in neighborhoods. We, therefore, propose the term “ecology” to describe a combination of sociodemographic factors that may influence both infectious and chronic disease health outcomes in a neighborhood. Specifically, the knowledge gained from an analysis of neighborhood ecologies in the early part of the COVID-19 pandemic should allow for better informed and tailored interventions for newly emergent infection epidemics in the future.

We chose to apply principal-component analysis (PCA), a methodology commonly used at the individual level in the social sciences, but rarely used in epidemiologic ecologic analyses, to capture how ecologic variables may group together as a neighborhood ecology. Most existing studies use logistic regression models to determine the independent contribution of area-level sociodemographic factors on COVID-19 outcomes. Regression models account for the independent and joint contributions of variables to the outcomes of interest,⁵ but spatial analyses show that factors such as racial and ethnic composition, poverty, education, and household crowding overlap within and across communities.⁶ PCA, however, mitigates collinearity and allows for the grouping and reduction of contributing factors,⁷ a modeling approach that is more reflective of real-world neighborhood environments. In the context of crafting and implementing public health and policy interventions, PCA may provide more contextual information and opportunities for addressing social determinants of health.

In this study, we used PCA to identify neighborhood ecologies, or groupings of sociodemographic neighborhood characteristics, that emerged as having a pattern of occurring together, and possibly together synergize the health-promoting or health-damaging effects of living in a particular neighborhood. We then sought to identify whether these sociodemographic neighborhood ecologies were associated with COVID-19 case and death rates in metropolitan Atlanta from March 2, 2020 to November 3, 2020. The beginning of the COVID-19 pandemic in the United States provided a rare opportunity to model the role of neighborhood ecologies on population-level vulnerability to a newly emergent pathogen.

Methods

We used COVID-19 test data for individuals under investigation as of November 3, 2020 from the Georgia Public Health Information Portal to identify people under investigation with positive COVID-19 test results and deaths attributed to COVID-19. We geocoded each case or death and aggregated the number of COVID-19 cases and COVID-19 deaths for each census tract in metropolitan Atlanta. Census tract–level COVID-19 case rate and death rate were calculated using total population for each census tract as a denominator based on the 2019 American Community Survey estimates. Census tracts with fewer than 10 COVID-19 cases were excluded from analysis for stable rates. Use of this deidentified secondary dataset was approved by the Georgia Department of Public Health. Consequently, our study was deemed exempt as secondary research for which further institutional review board approval was not required (project no. 200905).

Quintile groups for COVID-19 case and death rates were generated for descriptive analysis. We obtained 10 census tract–level socioecological measures from 2019 American Community Survey estimates: percentage of non-Hispanic Black people, percentage of Hispanic people, percentage of individuals without a high school diploma, percentage of individuals living under the federal poverty level, percentage of individuals who were uninsured, percentage of people aged 65 years or older, percentage of unemployed people, homeownership rate (proportion of occupied housing units that were owner occupied), percentage of crowded households (defined as the proportion of households in the census tract with more than one person per room, excluding bathrooms and kitchens), and percentage of households that received Supplemental Nutrition Assistance Program (SNAP) benefits. We selected these neighborhood characteristics based on the previously published work of our team and others who showed that these factors had been associated with COVID-19 case or death rates.

We applied PCA using the 10 census tract–level socioecological variables. The PCs were retained if they had an eigenvalue >1 and accounted for >70% of the total variance in the socioecological variables. PC scores for each retained PC, which is a linear composite of the optimally weighted observed variables, were calculated for each census tract. The rotated factor pattern (which shows how strongly the retained PCs and 10 socioecological variables are related) and communality (the proportion of variance in a socioecological variable that is accounted for by the retained PC) were reported.

We calculated and reported sociodemographic characteristics of the sample by COVID-19 case rate quintile groups and death rate quintile groups. We tested the correlation of the COVID-19 case rate with PC scores using Pearson correlation coefficients. Due to the highly skewed distribution of the COVID-19 death rate, we used Spearman’s rank-order correlation to test the correlation of the COVID-19 death rate with PC scores; no census tracts were excluded for having a zero-death count. All of the analyses were completed using SAS version 9.4 (SAS Institute, Cary, NC). We also created choropleth maps for PC scores using ArcGIS Pro (Esri, Redland, CA), and the categorization of PC scores was based on quartile groups.

To examine how the relationship between neighborhood ecology and COVID-19 case rates/death rates changed across space, we fitted geographically weighted regression models using each PC as the predictor and the COVID-19 case rate or log-transformed death rate (log transformation was used to normalize distribution in death rates) as the outcome, respectively. The geographically weighted regression generated coefficients of the relationship for each census tract, and we mapped these coefficients using ArcGIS Pro.

Results

Table 1 displays the distribution of the socioecological characteristics across quintiles of COVID-19 case and death rates for census tracts in metropolitan Atlanta. Characteristics including the proportion of the Hispanic population, the proportion living below the federal poverty level, the proportion of uninsured people, the homeownership rate, the proportion living in crowded households, and the proportion of households receiving SNAP benefits, showed a graded association across COVID-19 case rate quintiles. The percentages for these characteristics except homeownership rate increased across quintiles. Graded associations for death rates existed for the proportion of the Black population, the proportion of people without a high school diploma, the proportion who are unemployed, and the proportion of households receiving SNAP benefits. Although graded associations may not have existed across all of the characteristics, the highest quintile for both case and death rates had higher levels of social disadvantage than the lowest quintile for all measures.

Table 1.

Socioecological characteristics of census tracts in metropolitan Atlanta by COVID-19 case rate quintile group and death rate quintile groups as of November 3, 2020

	Q1 (n = 193)		Q2 (n = 194)		Q3 (n = 194)		Q4 (n = 194)		Q5 (n = 194)
	Mean	SD	Mean	SD	Mean	SD	Mean	SD	Mean	SD
COVID-19 case rate quintile groups
Black, %	24.0	27.8	29.7	28.7	36.1	30.7	40.3	31.2	32.3	26.3
Hispanic, %	6.1	5.6	7.4	7.6	8.7	8.7	11.2	12.0	21.7	19.0
No HS diploma, %	8.2	7.2	9.6	7.2	10.9	7.0	12.4	8.4	18.2	11.5
Poverty, %	12.5	13.7	12.9	10.6	13.4	9.5	14.9	10.3	17.9	10.0
No health insurance, %	9.2	5.8	11.7	7.1	12.6	6.7	15.1	7.7	19.9	10.6
Age ≥65 y, %	12.7	5.8	12.7	5.0	12.7	5.4	11.6	5.2	10.5	5.5
Unemployed, %	5.9	4.6	6.3	4.6	6.4	3.9	6.3	3.7	6.0	4.0
Homeownership rate, %	68.8	24.9	66.3	23.6	64.3	22.3	60.2	22.3	53.4	22.8
Crowded household, %	1.6	1.9	1.7	1.9	1.9	2.0	2.7	2.7	4.1	4.1
Households received SNAP, %	9.0	10.6	10.1	9.9	11.6	9.2	12.9	9.9	13.1	9.6
COVID-19 death rate quintile groups
Black, %	23.4	25.5	27.1	24.4	33.2	29.8	38.1	31.5	40.7	31.9
Hispanic, %	8.9	9.9	12.2	13.6	10.9	12.8	11.6	13.8	11.5	13.5
No HS diploma, %	8.4	7.9	10.8	9.5	11.3	8.3	13.9	9.1	14.9	9.2
Poverty, %	11.9	12.2	11.8	8.6	14.1	10.1	15.9	9.6	18.0	13.0
No health insurance, %	10.8	7.8	13.0	9.2	13.8	8.6	15.6	8.2	15.4	8.1
Age ≥65 y, %	10.8	5.7	11.6	5.7	11.5	4.8	12.2	4.9	13.9	5.6
Unemployed, %	5.0	3.7	5.4	2.9	6.4	4.3	6.8	3.9	7.3	5.2
Homeownership rate, %	64.4	25.6	64.5	23.5	64.2	23.8	61.0	22.9	58.7	22.6
Crowded household, %	1.9	2.7	2.3	3.0	2.2	2.4	2.7	3.1	2.7	2.9
Households received SNAP, %	7.4	8.5	8.8	7.7	11.4	9.5	13.8	9.8	15.2	11.7

COVID-19, coronavirus disease 2019; HS, high school; SD, standard deviation; SNAP, Supplemental Nutrition Assistance Program.

The PCA results are presented in Table 2. Two PCs were retained and accounted for 37% and 34% of variation in all of the socioecological variables, respectively. The proportion of Hispanic population, the proportion of people without a high school diploma, the proportion of uninsured people, the proportion of people aged 65 years or older (negative association), and the proportion of people living in crowded households demonstrated high (absolute value of factor loading >0.5) loadings only for PC1. The proportion of the non-Hispanic Black population, the proportion of people living below the federal poverty level, the proportion of unemployed people, and the proportion of households receiving SNAP benefits had high loadings only for PC2. The homeownership rate had high loading for both PC1 and PC2. PC1 and PC2 combined accounted for >70% of variation in the each of the socioecological variables (communality >0.7), except the percentage of individuals aged 65 years or older (communality = 0.28) and homeownership rate (communality = 0.59). PC1 score was significantly associated (r = 0.462, P < 0.001) with COVID-19 case rates but not COVID-19 death rates. PC2 score was significantly associated (r = 0.279, P < 0.001) with COVID-19 death rates but not with case rates (Table 3).

Table 2.

Rotated factor pattern and final communality estimates from PCA of census tract–level socioecological characteristics

Variables	Factor loading		Communality
	PC1	PC2
Black, %	−0.04	0.86	0.73
Hispanic, %	0.90	−0.23	0.86
No HS diploma, %	0.79	0.30	0.72
Poverty, %	0.46	0.78	0.83
No health insurance, %	0.85	0.31	0.82
Age ≥65 y, %	−0.53	−0.09	0.28
Unemployed, %	−0.02	0.86	0.74
Homeownership rate, %	−0.53	−0.55	0.59
Crowded household, %	0.84	0.15	0.73
Households received SNAP, %	0.28	0.86	0.82

Components 1 and 2 accounted for 37% and 34% of variation in all of the socioecological characteristics. HS, high school; PC1, principal component 1; PC2, principal component 2; PCA, principal-component analysis; SNAP, Supplemental Nutrition Assistance Program.

Table 3.

Correlation of COVID-19 case rate and COVID-19 death rate with PC1 and PC2

PC	COVID-19 case ratea		COVID-19 death rateb
	Coefficient	P	Coefficient	P
PC1	0.462	<0.001	0.090	0.005
PC2	0.011	0.748	0.279	<0.001

COVID-19, coronavirus disease 2019; PC1, principal component 1; PC2, principal component 2.

^aPearson correlation coefficients were used for COVID-19 case rate.

^bSpearman correlation coefficients (rank order) were used for COVID-19 death rate due to the highly skewed distribution.

Figure 1A is a map of the distribution for PC1 census tracts. PC1 tracts are clustered around major highways in the region. Figure 1B shows PC2 tracts that are clustered in South Fulton, South Dekalb, and North Clayton Counties. Figure 1C shows the COVID-19 cumulative case rates, with a map that shows high case rate tracts, and high PC1 score tracts have similar spatial patterns. COVID-19 death rates are shown in Figure 1D. Although high rates are scattered throughout metropolitan Atlanta, the high death rate tracts have a spatial pattern that is similar to those that have high PC2 scores.

Fig. 1.

Metropolitan Atlanta census tracts. (A) Score distribution of PC1. (B) Score distribution of PC2. (C) COVID-19 case rate through November 2, 2020. (D) COVID-19 death rate through November 2, 2020. COVID-19, coronavirus disease 2019; PC, principal component.

Figures 2A and 2B present the results of the geographic weighted regression model. Figure 2A illustrates how the relationships between PC1 and COVID-19 case rates varied across the census tracts in metropolitan Atlanta. The strongest impact of PC1 on COVID-19 case rates was observed at four outermost corners of the city; the impact of PC1 was weaker in the central corridor. Figure 2B shows the local relationships between PC2 and log-transformed COVID-19 death rates. PC2 was a stronger predictor of COVID-19 death rates in the northern, southwestern, eastern, and central (city of Atlanta and neighboring tracts) parts of metropolitan Atlanta.

Fig. 2.

Local coefficients from geographically weighted regression models. (A) Dependent variable: COVID-19 case rate, independent variable: PC1 score. (B) Dependent variable: log (COVID-19 death rate), independent variable: PC2 score. COVID-19, coronavirus disease 2019; PC, principal component.

Discussion

In this study, we applied PCA to explore neighborhood-level characteristics in Atlanta census tracts and their association with COVID-19 case and death rates. Our analysis identified two distinct neighborhood ecologies, each with different impacts on COVID-19 outcomes. Tracts with higher proportions of Hispanic populations, individuals without high school diplomas, uninsured individuals, and those living in crowded households tended to have higher case rates, whereas tracts with higher proportions of non-Hispanic Black populations, individuals living in poverty, unemployed individuals, and those receiving SNAP benefits tended to have higher death rates.

By using PCA, we aimed to overcome the limitations of traditional multivariate regression models, allowing us to understand how various neighborhood characteristics coexist to influence health outcomes such as COVID-19. We decided to take this approach to describe the ecology or phenotype of a neighborhood after we observed that although some factors such as uninsured rates and poverty levels were not associated with COVID-19 in Georgia counties when using multivariable regression, many of the areas with high COVID-19 case or death rates did, in fact, have high rates of poverty and low rates of health insurance coverage.⁶ Upon examining spatial patterns of demographics and social factors, it became clear that several potential predictors of COVID-19 outcomes were clustered tin the same counties. We hypothesized that this also may be the case at the census tract level—in other words, that there may be “neighborhood ecologies” that contribute to neighborhood risk or resilience and that make a neighborhood more or less susceptible to adverse COVID-19 outcomes. We chose PCA because of its ability to distinguish how values for variables tend to correlate with one another; the PCA score provides a way to express how much a particular census tract conforms to a particular factor (also known as ecology).

Our study differs from previous neighborhood-level COVID-19 research in three key aspects: the use of PCA to identify neighborhood ecologies, inclusion of both racial and ethnic demographics, and consideration of both case and death rates simultaneously. Although other studies have examined associations between individual socioeconomic factors and COVID-19 outcomes, our approach provides a more comprehensive understanding of the collective impact of neighborhood characteristics.

One study of cases in Harris County (Houston), Texas, used PCA as a variable reduction method.⁸ The study found that a component with high poverty and low education was positively associated with COVID-19 case rates. Their study did not examine death rates, nor did it use demographic composition in the analyses. In our analysis, percentages of poverty and education were found to be related to different components, with only one component including educational attainment related to case rate. The differing results suggest that different groups of socioecological characteristics may be associated with case rates in different metropolitan areas. A recent study from southeastern Pennsylvania⁹ that looked at the independent associations of predictor variables (not PCA-derived ecologies) did not find an association between race and neighborhood COVID-19 incidence or mortality rates. The only race variable they used was percentage of White people, and they did not include ethnicity in the model. Potentially high rates among White Hispanics and low rates among non-Hispanic Whites may have canceled each other out in this study.

The neighborhood ecology approach facilitates a more nuanced approach to public health planning and intervention by identifying the complex interplay of cultural, linguistic, and social drivers of health within neighborhoods. In our study, different neighborhood ecologies were associated with census tract COVID-19 case and death rates. Areas with a high percentage of Hispanic populations and people without a high school diploma may be applied in the implementation of public health programming around prevention of cases using Spanish-language media and materials suited for lower reading levels. The high COVID-19 death rates associated with high poverty areas with large Black populations suggests the need to address healthcare access barriers associated with poverty such as transportation issues, but also the importance of building trust between local healthcare providers and the Black community. These complex findings present opportunities for community engagement and bidirectional communication related to mitigation efforts. It also should be noted that this approach highlights the heterogeneity of racial and ethnic populations and neighborhoods. The metropolitan Atlanta area is home to many middle- and upper-middle income predominantly Black neighborhoods, which were not associated with higher case or death rates. This heterogeneity would not be easily detected using traditional regression approaches, which may have identified an independent effect of percentage of Black residents but not differentiated by socioeconomic variables. The same may be true of Hispanic neighborhoods with more educated residents. The nuanced findings from this type of analysis provide public health authorities and community organizations with the data needed to provide culturally and linguistically appropriate services and information,^10–12 as opposed to one-size-fits-all approaches to COVID-19 mitigation.

We believe that this neighborhood ecology approach holds great promise in not just understanding how neighborhood environment influences COVID-19 rates but also in understanding the geographic distribution of other infectious and chronic disease outcomes. Using PCA to develop neighborhood ecologies is an important methodological contribution to the literature. Previous approaches to neighborhood deprivation indices^10,13 have been developed on a national level, but the method presented here allows for a local approach to identify the constellation of social determinants of health and demographic compositional ecologies present in a specific place. This method addresses one of the major weaknesses of national-level social deprivation indices, which assume that social environmental factors have a uniform impact on health and other outcomes across differing communities, spaces, and places.

It is important to note that our analysis has limitations. Ecological associations observed at the neighborhood level may not necessarily reflect individual-level associations.¹⁴ In addition, our study focused on cumulative cases across a period, and factors such as mask-wearing and social distancing, which can vary by geography within the same population, were not directly addressed. Different metropolitan areas may have different sets of ecologies than what we identified in Atlanta, and the same ecologies we identified in Atlanta may be differentially related to COVID-19 outcomes in other areas. This methodology could be applied in any setting to detect local neighborhood ecologies, however.

Our findings have significant implications for public health practice. Although the finding that disadvantaged neighborhoods are associated with poor health outcomes is not new, the ability to go beyond treating racial and ethnic groups and their neighborhoods as monolithic entities in interventions is a welcomed development, as well as the ability to investigate the particular neighborhood ecologies that may exist in different geographic areas. Tailored interventions are needed to address the specific needs of neighborhoods identified with higher case or death rates.

Conclusions

Our study highlights the importance of considering neighborhood-level factors in understanding COVID-19 outcomes. Further research should explore the concept of neighborhood ecologies in other geographic areas and their implications for various health outcomes beyond COVID-19.

Key Points.

• We identified two unique neighborhood profiles.

• We found that neighborhoods may be differentially at risk of coronavirus disease 2019 cases or deaths depending on their sociodemographic ecology.