Street-view greenspace distribution across racial/ethnic, neighborhood income, and individual education subgroups

Abstract

Background:

The maldistribution of greenspaces across Black, Hispanic, and low-income communities can contribute to health disparities. It is unclear whether the interaction of race/ethnicity and socioeconomic status may explain the maldistribution of greenspace, or whether the maldistribution varies by type of greenspace.

Methods:

Applying deep learning algorithms to street-view images, we calculated percentages of specific types of residential greenspace (i.e., %Trees, %Grass) for each Multi-Ethnic Study of Atherosclerosis participant (N = 5,858; 2000–2002). We used multilevel analysis of individual heterogeneity and discriminatory accuracy to quantify inequities in greenspace type by intersecting stratum of race/ethnicity (Black, Chinese American, Hispanic, and White), education (high school, some college, and bachelor’s degree), and neighborhood socioeconomic status (NSES; low, moderate, and high). Models adjusted for age, sex, individual income, and study site.

Results:

The mean %Trees was 19.0 (SD 8.8) and the mean %Grass was 5.1 (4.6). Distribution of %Trees varied across strata, for example, 13.1% (95% confidence interval [CI] = 9.1, 23.8) for Hispanic participants in the lowest education and NSES group versus 20.5% (14.0, 30.4) for Hispanic participants in the highest education and NSES group. Patterns were similar among corresponding strata of Black and Chinese American participants. However, the lowest %Trees among White participants was in the highest NSES and education stratum (20.6, 95% CI = 14.8, 31.5). About 16% of the variability of %Trees and 11% of the variability of %Grass was explained by intersecting stratum of race/ethnicity, education, and NSES.

Conclusion:

Maldistribution of greenspace types may be explained by combinations of race/ethnicity, education, and NSES subgroups, as opposed to each factor alone.

What this study adds

• Street-view imagery paired with deep learning algorithms allowed us to examine specific types of green space.

• We found intersecting combinations of race/ethnicity, educational, and neighborhood socioeconomic status (NSES) demonstrated substantive “discriminatory accuracy” in predicting individuals’ residential street-view exposure to greenspace beyond individual-level estimates.

• We found 16% of individuals’ variability in the street-level view of trees and 11% of variability in the street-level view of grass were attributable to these intersecting strata.

• Finally, we found White participants at all education and NSES levels had higher mean exposures to residential, street-view percentages of trees and grass relative to Black, Chinese American, or Hispanic participants.

Background

Greenspace, inclusive of vegetation in natural as well as urban landscape areas,¹ has been associated with a range of health benefits by providing opportunities for physical activity and social interaction, reducing exposure to noise, air pollution, extreme heat, or stress.^2,3 Environmental injustice encompasses spatial maldistribution of harmful environmental exposures for marginalized groups and inequitable distribution of beneficial environmental exposures like greenspace.^4–6 From the late-19th century until the 1960s, racial/ethnic minorities in the US were explicitly and deliberately marginalized through practices and policies like Jim Crow laws and redlining designed to enforce residential and social segregation of racial/ethnic groups from Whites.^7,8 Segregation as a mechanism of structural racism concentrated marginalized communities in areas of economic and environmental deprivation and disinvestment.^7,8 Generations after such practices and policies were formally outlawed, historically segregated areas remain entrenched by racial/ethnic income and wealth disparities, lower educational attainment, persistent poverty, higher risk of environmental hazards, and lower availability and quality of greenspace compared with predominantly White, higher socioeconomic communities.^7–17 And while correlated, income and education can influence environmental exposures through different pathways and differently by racial/ethnic groupings, suggesting it is important to account for education separately as a measure of socioeconomic status and lived experience.^18,19

Principles of intersectionality, originally developed by Kimberlé Crenshaw, characterize how systems of oppression act in combination across multiple dimensions of social identity and privilege/disadvantage to harm or benefit groups differently.²⁰ This intersectional framework is increasingly being used to investigate how structural racism and social class combine to drive disparities in health and in environmental exposures that may be partially responsible for health disparities.^18–22 However, there remains a gap in environmental injustice research to incorporate intersectional approaches in investigating the inequitable distribution of residential greenspace.^5,6,23

Previous research on greenspace commonly used satellite-based indices like the Naturalized Difference Vegetation Index (NDVI) that provides no information on the type of vegetation (e.g., tree vs. grass), nor on whether the vegetation is visible at the ground level. Deep learning algorithms applied to street-view imagery offer the opportunity to generate refined estimates of exposure to greenspace, both from (a) a ground-based view and (b) allowing for identification of specific types of greenspace like trees, grass, plants, and flowers.^24,25

To address the gaps outlined above, we used multilevel analysis of individual heterogeneity and discriminatory accuracy (MAIHDA) to investigate how markers of structural racism and socioeconomic identity additively and/or interactively predict residential greenspace distribution. Following the MAIHDA approach, individuals are nested within combinations of intersecting social identity categories, allowing the quantification of whether residential greenspace distribution differs according to complex interactions of social strata. We used ground-level street-view greenspace metrics to quantify unequal distribution of residential greenspace by vegetation type for intersecting combinations of race/ethnicity, education, and neighborhood socioeconomic status (NSES).

Methods

Study sample

We used data from the Multi-Ethnic Study of Atherosclerosis (MESA), a longitudinal study that began enrollment in 2000–2002 (Exam 1) with 6,814 men and women who were 45 to 84 years of age, free of clinically recognized cardiovascular disease, and living in any of six US communities (Baltimore, Maryland; Chicago, Illinois; Forsyth County, North Carolina; Los Angeles County, California; New York, New York; and St. Paul/Minneapolis, Minnesota).²⁶ The MESA study includes a baseline and six follow-up exams spanning 2000–2002 to 2022–2024. For this study, we used Exam 1 data to assess intersectional inequities in the distribution of greenspace at baseline. Our final sample included 5,858 MESA participants for whom we had street-view image greenspace measures linked to participants’ Exam 1 residential geolocation (Figure S1; https://links.lww.com/EE/A387).²⁷

Intersectional social categories

To measure the intersection of exposure to structural racism with experience of social class, we created 36 intersecting strata from combinations of three social identity variables: baseline self-reported race/ethnicity (Black, Chinese American, Hispanic/Latino, and White), education (≤High school graduate/GED, some college, bachelor’s degree or higher) and NSES (tertiles: low, medium, high) (Figure 1). NSES was assessed using factor analysis including six indicators from the United States Census 2000 data at the census tract level:^29,30 median home value, median household income, household wealth, education, percentage of households with income >$50,000, and percentage of employed persons in executive, managerial, or professional occupations.³¹ Each participant was then assigned to only one of the 36 intersecting social identity strata (hereafter referred to as intersecting strata) based on their respective combination (Table 1, Figure S2; https://links.lww.com/EE/A387).

Figure 1.

Data structure for the MAIHDA approach. Adapted from Keller et al.²⁸. Bachelor’s+, bachelor’s degree or higher education; Hispanic, Hispanic/Latino; HS/GED, high school/general education diploma or less; MAIHDA, multilevel analysis of individual heterogeneity and discriminatory accuracy; MESA, Multi-Ethnic Study of Atherosclerosis; NSES, neighborhood socioeconomic status.

Table 1.

Number of individuals in each of 36 intersecting strata of race/ethnicity, education, and neighborhood socioeconomic status for Multi-Ethnic Study of Atherosclerosis participants. 2000–2002 (N = 5,858)

Stratum	Race/ ethnicity	Education	NSES	N	%
1	Black	High school/GED or less	Low	234	4.0
2	Black	High school/GED or less	Moderate	154	2.6
3	Black	High school/GED or less	High	62	1.1
4	Black	Some college of AA degree	Low	214	3.7
5	Black	Some college of AA degree	Moderate	212	3.6
6	Black	Some college of AA degree	High	120	2.0
7	Black	Bachelor’s or graduate degree	Low	170	2.9
8	Black	Bachelor’s or graduate degree	Moderate	214	3.7
9	Black	Bachelor’s or graduate degree	High	169	2.9
10	Chinese	High school/GED or less	Low	98	1.7
11	Chinese	High school/GED or less	Moderate	110	1.9
12	Chinese	High school/GED or less	High	61	1.0
13	Chinese	Some college of AA degree	Low	25	0.4
14	Chinese	Some college of AA degree	Moderate	56	1.0
15	Chinese	Some college of AA degree	High	63	1.1
16	Chinese	Bachelor’s or graduate degree	Low	40	0.7
17	Chinese	Bachelor’s or graduate degree	Moderate	87	1.5
18	Chinese	Bachelor’s or graduate degree	High	163	2.8
19	Hispanic	High school/GED or less	Low	506	8.6
20	Hispanic	High school/GED or less	Moderate	225	3.8
21	Hispanic	High school/GED or less	High	77	1.3
22	Hispanic	Some college of AA degree	Low	166	2.8
23	Hispanic	Some college of AA degree	Moderate	116	2.0
24	Hispanic	Some college of AA degree	High	61	1.0
25	Hispanic	Bachelor’s or graduate degree	Low	49	0.8
26	Hispanic	Bachelor’s or graduate degree	Moderate	38	0.6
27	Hispanic	Bachelor’s or graduate degree	High	48	0.8
28	White	High school/GED or less	Low	173	3.0
29	White	High school/GED or less	Moderate	191	3.3
30	White	High school/GED or less	High	120	2.0
31	White	Some college of AA degree	Low	147	2.5
32	White	Some college of AA degree	Moderate	251	4.3
33	White	Some college of AA degree	High	241	4.1
34	White	Bachelor’s or graduate degree	Low	104	1.8
35	White	Bachelor’s or graduate degree	Moderate	286	4.9
36	White	Bachelor’s or graduate degree	High	807	13.8

NSES indicates neighborhood socioeconomic status.

Assessment of exposure to visible greenspace features

We assessed residential greenspace exposure at 100, 500, and 1,000-m buffers based on nationwide street-view imagery processed by the deep-learning pyramid scene parsing network algorithm, providing an overall accuracy higher than 93% on pixel-level prediction.^24,25,32,33 Street-view images capture the visual domain along all streets in the continental US and provide a ground-level perspective of greenspace as individuals experience it. Specific features of greenspace were segmented out down to the pixel level, including natural features such as trees, shrubs, grass, plants, field, and flowers. MESA participants’ residential addresses were spatially joined over follow-up to annually updated street-view derived greenspace metrics from 2007 (the first date street-view images are available) to 2020. For residential addresses before 2007, including those for 2000–2002 used in this analysis, we assigned 2007 street-view data.³² We imputed missing greenspace measures so that if no street-view data were available for a particular year, we carried forward or backward data that was available immediately before or next in succession as relevant.^32,34 Main findings were based on 500-m residential buffer measures of %Trees (including trees and palms) and %Grass. In sensitivity analyses, we reported results for 100-m and 1,000-m buffer measures of %Trees and %Grass, as well as results for %Other greenspace (plant, flowers, and field), and %Total greenspace (trees, palms, grass, plants, flowers, and field).

Other variables

We adjusted our analyses for the following individual-level variables to create comparisons across groups standardized to these characteristics: baseline age in years (continuous), self-reported gender (male or female), total gross family income (≤$24,999, $25,000–$49,999, $50,000–$74,999, ≥$75,000), and MESA exam site.

Statistical analyses

To evaluate how intersecting strata predict residential street-view greenspace at baseline, we used two-level linear MAIHDA models to assess whether the distribution for each greenspace type can be explained by intersectional strata of race/ethnicity, education, and NSES. For each type of greenspace as an outcome, we estimated (a) a null model and (b) a main-effects model. The null model includes random effects for intersectional strata alongside fixed covariate effects of age, sex, individual income, and MESA site. The main effects model further adds fixed effects for race/ethnicity, education, and NSES.³⁵ Following Keller et al.,²⁸ the null model allows us to decompose total variance in greenspace type into variance that can be attributed to (i) differences between intersectional strata, and (ii) differences within intersectional strata. The null model is presented as:

$${\displaystyle \begin{array}{l}{\displaystyle y_{ij}={\gamma }_{00}+{\mu }_{0j}+e_{0ij}}\end{array}}$$ (1)

where in the case of %Trees as outcome, y_ij is the %Trees for individual (i) nested in the intersectional stratum (j), 𝛾₀₀ is the grand mean, 𝜇_0j is the residual for intersectional stratum (j), and e_0ij is the residual for individual (i) in stratum (j). The residual 𝜇_0j follows a normal distribution with ${\sigma }_{u0}^2$ indicating the variability among strata, and e_0ij follows a normal distribution with ${\sigma }_{e0}^2$ indicating the variability within strata.

From the null model in Eq. (1), we obtain: (i) the variance partition coefficient (VPC_null) in Eq. (2) as a measure of the proportion of variability of %Trees explained by the intersectional strata; and (ii) predicted greenspace type for each intersectional strata with 95% credible intervals (95% CIs).

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{V}\mathrm{P}{\mathrm{C}}_{\mathrm{n}\mathrm{u}\mathrm{l}\mathrm{l}}=\frac{{\sigma }_{u0}^2}{{\sigma }_{u0}^2+{\sigma }_{e0}^2}}\end{array}}$$ (2)

Similarly, from the main effects model, we calculated three main sets of parameters: (1) the VPC of the main-effects model (VPC_main-effects); (2) proportional change in variance (PCV) comparing VPC_null to VPC_main-effects; and (3) the strata-level residuals, allowing us to compare how the observed greenspace measure differs from the model prediction of greenspace type. In addition, to examine factors that could modify the association between intersectional strata and distribution of greenspace type, we a priori stratified our models by higher versus lower population density based on prior findings showing the distribution of greenspace in rural versus urban/suburban areas differed by NSES.¹⁰ We stratified by population density using census tract-level estimates from the 2000 United States Census and dichotomizing at the sample median value of 8,000 people per square mile.^29,30 See Table S1; https://links.lww.com/EE/A387 for participant numbers and percentages by population density and study site.

Our models were estimated using Bayesian Markov chain Monte Carlo methods with the R package “brms” (version 2.20.4).^36,37 Full model specifications in Supplementary Materials; https://links.lww.com/EE/A387: MAIHDA model specifications. Code used for analyses available in GitHub library https://github.com/SpatialHealth/GREEENS_MESA. All statistical analyses were performed in R Studio.

Results

Our sample of 5,858 participants comprising 1,549 (26%) non-Hispanic Black individuals, 703 (12%) Chinese American individuals, 1,286 (22%) Hispanic/Latino individuals, and 2,320 (40%) non-Hispanic White individuals (Table 2). In our sample, 2,011 (34%) had the lowest education, 1,672 (29%) had some college education, and 2,175 (37%) participants had the highest education (Table 2). The Spearman correlation of %Trees with %Grass was 0.73, of %Trees with %Other Green was 0.46, and of %Grass with %Other Green was 0.43. %Trees and %Grass were overall highest among participants who were White, had the highest education, or lived in the highest NSES areas (Table 2). %Trees and %Grass were overall lowest for participants who were Hispanic, had the lowest education, or lived in the lowest NSES areas (Table 2). %Trees and %Grass were substantially higher for those residing in lower versus higher population density areas (e.g., 23.86% vs. 14.78% for Trees; Table 2). When individuals were combined into intersectional strata, the largest stratum included White participants with a bachelor’s degree or higher education living in a high NSES area (13.8%) (Figure S2; https://links.lww.com/EE/A387). The second largest group was Hispanic participants with the lowest education living in a low NSES area (8.6%). The smallest stratum was Chinese American participants with some college education living in a low NSES area (N = 25, 0.4%). All 36 strata had > 20 individuals, 97% had >30 individuals, and 86% had >50 individuals (Figure S2; https://links.lww.com/EE/A387).

Table 2.

Descriptive statistics of greenspace type at a 500 m buffer among Multi-Ethnic Study of Atherosclerosis participants. 2000–2002 (N = 5,858)

	Overall sample N = 5,858	%Trees	%Grass
Variables	N (%)/mean (SD)	Mean (SD)	Mean (SD)
Greenspace measures
%Trees	19.00 (8.76)	–	–
%Grass	5.14 (4.55)	–	–
Race/ethnicity (N%)
Black	1,549 (26.4)	18.58 (8.25)	5.70 (4.76)
Chinese American	703 (12.0)	16.93 (8.06)	3.88 (3.20)
Hispanic/Latino	1,286 (22.0)	15.43 (7.38)	2.94 (3.17)
White	2,320 (39.6)	21.89 (9.03)	6.37 (4.87)
Educational attainment (N%)
≤High school graduate/GED	2,011 (34.3)	16.58 (8.12)	4.34 (4.30)
Some college	1,672 (28.5)	19.64 (8.46)	5.56 (4.51)
Bachelor’s degree or higher	2,175 (37.1)	20.75 (9.04)	5.56 (4.72)
Neighborhood Socioeconomic status, NSES (mean/SD)a	0.30 (1.37)	–	–
NSES (N%)
Low	1,927 (32.9)	16.36 (8.67)	4.53 (4.60)
Moderate	1,940 (33.1)	19.69 (7.37)	6.06 (4.38)
High	1,991 (34.0)	20.88 (9.45)	4.84 (4.53)
Population density (mean/SD)	28,049 (44,736)	–	–
Population density (N%)
≤8,000 people per square mile	2,724 (46.5)	23.86 (7.82)	8.44 (4.23)
>8,000 people per square mile	3,134 (53.5)	14.78 (7.20)	2.27 (2.36)
Sex (N%)
Female	3,071 (52.4)	18.67 (8.62)	5.00 (4.48)
Male	2,787 (47.6)	19.36 (8.89)	5.30 (4.63)
Income (N%)
<$25,000	1,749 (29.9)	16.13 (8.02)	3.93 (3.80)
$25,000–$49,999	1,713 (29.2)	18.79 (8.26)	5.12 (4.43)
$50,000–$74,999	1,016 (17.3)	20.92 (8.14)	6.28 (4.81)
$75,000+	1,380 (23.6)	21.49 (9.55)	5.87 (5.01)
MESA exam site (N%)
Winston Salem, NC	850 (14.5)	28.62 (6.47)	12.48 (3.61)
New York, NY	993 (17.0)	12.56 (5.44)	0.40 (0.71)
Baltimore, MD	906 (15.5)	18.05 (8.01)	6.54 (3.43)
Minneapolis/St. Paul, MN	926 (15.8)	25.11 (4.61)	6.88 (1.95)
Chicago, IL	1,048 (17.9)	18.79 (8.38)	3.73 (3.20)
Los Angeles, CA	1,135 (19.4)	13.40 (6.08)	2.56 (1.55)
Age in years (mean/SD)	61.72 (10.11)	–	–
Age category (N%)
45–54 years	1,741 (29.7)	19.26 (8.61)	5.21 (4.53)
55–64 years	1,656 (28.3)	19.57 (8.81)	5.36 (4.79)
65–74 years	1,710 (29.2)	18.68 (8.81)	5.13 (4.51)
75–84 years	751 (12.8)	17.86 (8.74)	4.53 (4.11)

^aLower z-score value indicates lower NSES.

The VPCs from the null models show that 16% of the variability in %Trees, and 11% of the variability in %Grass was attributable to between-intersectional strata differences (Table 3). While there is no hard rule or cutoff for determining what makes a meaningful VPC_null, the higher the VPC_null, the more explanatory the intersectional strata are for understanding individual differences in predicted greenspace type.³³ Figure 2 shows the predicted mean %Trees (panel A) and %Grass (panel B) for each intersectional stratum (specific estimates and 95% CIs provided in Tables S2 and S3; https://links.lww.com/EE/A387). While CIs overlapped, Black, Chinese American, and Hispanic participants in the lowest education strata overall trended with %Trees below the sample mean of 19.0% (SD 8.8), while %Trees for all strata of White participants were at or above the sample mean (Figure 2A). All strata of Hispanic participants had %Grass at or below the sample mean of 5.1% (SD 4.6), while most strata of White participants had %Grass above the sample mean (Figure 2B), although CIs overlapped.

Table 3.

Parameter estimates from multilevel analysis of individual heterogeneity and discriminatory accuracy of residential street-view greenspace measures among MESA participants at a 500 m buffer. 2000–2002 (N = 5,858)

	%Treesa		%Grass
	Null model Coefficient (95% CI)	Main-effects model Coefficient (95% CI)	Null model Coefficient (95% CI)	Main–effects model Coefficient (95% CI)
Fixed effects
Interceptb	27.02 (25.49, 28.50)	24.01 (22.16, 25.86)	11.91 (11.34, 12.49)	11.04 (10.15, 11.93)
Race/ethnicity
Black	–	Ref	–	Ref
Chinese American	–	1.22 (−0.30, 2.72)	–	0.81 (0.01, 1.59)
Hispanic/Latino	–	−0.80 (−2.27, 0.66)	–	0.19 (−0.59, 0.95)
White	–	−0.23 (−1.66, 1.23)	–	0.02 (−0.75, 0.77)
Educational attainment
≤High school graduate/GED	–	Ref	–	Ref
Some college	–	0.88 (−0.39, 2.13)	–	0.16 (−0.50, 0.83)
Bachelor’s degree or higher	–	1.17 (−0.06, 2.44)	–	0.17 (−0.51, 0.85)
Neighborhood SES (N%)
Low	–	Ref	–	Ref
Moderate	–	2.16 (0.92, 3.40)	–	0.87 (0.22, 1.54)
High	–	5.14 (3.87, 6.42)	–	0.73 (0.07, 1.40)
Random effects
Intersectional strata-level	2.72 (2.11, 3.51)	1.41 (1.00, 1.98)	0.86 (0.66, 1.12)	0.77 (0.57, 1.04)
Individual level	6.18 (6.07, 6.30)	6.18 (6.07, 6.30)	2.43 (2.39, 2.48)	2.43 (2.39, 2.48)
Measures of variance
VPCc	16.2%	5.0%	11.1%	9.1%
PCVd	–	73.5%	–	19.8%

The simple intersectional and main-effects models are additionally adjusted for age, gender, income, and exam site. The VPC indicates the proportion of individuals’ greenspace measure variance attributable to differences between intersectional strata.

^aCombined measures of trees, plants, grass, field, flowers, and palm tree.

^bIn the simple intersectional model: Population-Level Effects intercept estimate.

^cVPC: (variance at strata)/ (variance at strata + variance at individual level) × 100.

^dPCV: (simple intersectional model strata variance – main effects model strata variance)/simple intersectional model strata variance) × 100.

PCV indicates proportional change in variance; SES, neighborhood socioeconomic status; VPC, variance partition coefficient.

Figure 2.

Predicted street-view based %Trees (A) and %Grass (B) across strata in the main model. MESA 2000–2002, N = 5,858. MESA indicates Multi-Ethnic Study of Atherosclerosis; SES, socioeconomic status.

Results from the null models showed that expected trends of higher education and NSES in association with higher greenspace percentages varied by race/ethnicity group and by vegetation type. For example, Hispanic participants with the lowest education and NSES had %Trees of 13.1 (95% CI = 9.1, 23.8) versus 20.5 (95% CI = 14.0, 30.4) for Hispanic participants with the highest education and high NSES (Figure 2A, Table S2; https://links.lww.com/EE/A387). Contrary to this trend, %Trees for White participants overall varied little across NSES and educational strata, where notably the lowest %Trees for Whites as a group was observed for the highest NSES and education stratum (20.6, 95% CI = 14.8, 31.5) (Figure 2A, Table S2; https://links.lww.com/EE/A387). The hypothesized education and NSES trends were observed for %Grass, with intersecting strata of highest education and NSES associated with higher %Grass primarily for Chinese Americans (highest NSES and education: 6.4 [95% CI = 4.9, 7.3] compared to lowest NSES and education: 1.6 [95% CI = 0.8, 2.8]) (Figure 2B, Table S3; https://links.lww.com/EE/A387). Similar results were observed for 100 and 1,000-m buffer measures of %Trees and %Grass, as well as 500-m buffer %Total greenspace (Tables S2, S3, and S4; https://links.lww.com/EE/A387). Sensitivity analyses for 500-m buffer measures of %Other greenspace showed percentages ranging only from 0.6% to 1.4% across all strata, providing little variation for observation of trends (Table S4; https://links.lww.com/EE/A387).

Results from the null model stratified by population density highlighted a key difference in greenspace trends for White participants in lower versus higher population density areas. We observed that lower population density areas White participants in the highest NSES and education stratum had the highest %Tree of all strata (29.6, 95% CI = 27.1, 31.8) (Figure 3A). However, among participants living in high population density areas, the highest %Trees was observed for White participants in the lowest NSES strata, regardless of education, while the hypothesized trend of higher %Trees with higher NSES was observed for Black, Chinese American, and Hispanic participants (Figure 3B). Stratified results by population density for %Grass in lower versus higher population density areas showed similar trends to our nonstratified results (Figure S3; https://links.lww.com/EE/A387).

Figure 3.

Predicted street-view based %Trees stratified by population density across strata in simple interaction model. MESA 2000–2002, N = 5,858. A, Lower population areas (<=8,000 persons/square mile) N = 2,724. B, Higher population density areas (>8,000 persons/square mile) N = 3,134. MESA indicates Multi-Ethnic Study of Atherosclerosis; SES, socioeconomic status.

Although not our primary focus, estimates from the main effects model indicated that %Trees and %Grass were higher for participants in neighborhoods with high NSES, compared to participants living in neighborhoods with low NSES (5.14, 95% CI = 3.87, 6.42 and 0.73, 95% CI = 0.07, 1.40 correspondingly for Trees and Grass; Table 3). The VPC_main-effects suggested that 5% of the variability in trees was explained by the interactive effect of intersecting strata, a reduction from the VPC_null of 16% that was explained by the additive effect of intersecting strata. Further, a PCV value of 74% suggested that most, but not all, of the between-intersectional strata differences in %Trees were explained by the main effects of race/ethnicity, education and NSES as additive rather than interactive effects (Table 3, Table S5; https://links.lww.com/EE/A387).

For %Grass, we observed a VPC_main-effects of 9% (a small reduction from the VPC_null of 11%). Further, a PCV of 20% indicated that most of the differences between intersectional strata in %Grass were explained by the interactive rather than additive effects of race/ethnicity, education, and NSES (Table 3, Table S6; https://links.lww.com/EE/A387).

Discussion

We used a MAIHDA approach to investigate how markers of structural racism and social identity predict maldistribution in residential greenspace. We hypothesized that greenspace types would differ both additively and interactively across intersections of race/ethnicity, education, and NSES in ways not apparent when assessing these social identity variables separately. We found that intersecting strata that included these variables demonstrated substantive “discriminatory accuracy” in predicting individuals’ residential exposure to greenspace beyond individual-level estimates, with 16% of individuals’ variability in trees and 11% of variability in grass attributable to these intersecting strata.

Overall, White participants at all education and NSES levels had higher mean exposure to %Trees and %Grass than Black, Chinese American, or Hispanic participants in MESA. Intersecting strata of higher education and NSES were associated with higher %Trees for Black, Chinese American, and Hispanic participants, while the lowest %Trees for Whites as a group was observed for the highest NSES and education. Further, when stratified by population density, we observed a contrast that in low population density areas, White participants in the highest education and NSES strata had the highest exposure to %Trees, but unexpectedly had the lowest %Trees in high population density areas. We found evidence for the association of higher education and NSES strata with a higher %Grass only for Chinese Americans. We also found evidence of interactive effects for intersectional strata of race/ethnicity, education, and NSES, suggesting that these social identities may combine in complex and beyond-additive ways to influence environmental injustice.

Our findings of intersecting strata of nonwhite race/ethnicity, lowest education, and low NSES strata associated with lower %Trees were largely consistent with results from a study by Klompmaker et. al. of greenspace across all United States census tracts.¹¹ Their study found that census tracts with larger proportions of Hispanic individuals and smaller proportions of non-Hispanic White individuals had less greenspace as measured by NDVI and NatureScore, a proprietary metric of nature.¹¹ Furthermore, Klompmaker et al. also found an association of lower socioeconomic status (SES) and lower percentages of non-Hispanic White individuals, separately, with lower NDVI and NatureScores in urban census tracts.¹¹ Our findings were also broadly consistent with a study by Casey et al.¹⁵ of NDVI in 59,483 urban census tracts in the US. Casey et al. found that for 2001 populations, each SD increase of percentage of Whites in the race/ethnicity composition of census tracts was associated with 0.021 (95% CI = 0.018, 0.023) units higher 2001 NDVI greenspace, while higher tract composition in any of the nonwhite race/ethnicity groups assessed was associated with lower NDVI measures.¹⁵

Our findings suggest that (a) most strata of Chinese Americans have lower %Trees and %Grass estimates relative to Whites (albeit with many overlapping confidence intervals), and (b) %Trees and %Grass consistently trend higher for Chinese Americans at increasingly higher NSES and education. These findings are separately consistent with findings of Klompmaker et al. that a higher percentage of nonwhites (including non-Hispanic Asian-identified people) in urban tract areas and lower SES census tracts were both associated with lower greenness measures (Klompmaker’s study did not investigate intersections of race/ethnicity and SES).¹¹ However, when investigating greenness of census tracts by percent of Asian residents, Casey et al.¹⁵ found no association with quintiles of Index of Concentration at the Extremes (ICE). While we are cautious to not suggest a causal interpretation due to the limitations of our study design, our findings may be suggestive of a more consistent “dose/response” pattern of higher education and NSES with higher tree and grass greenspace measures for Chinese Americans in our study than is observed for other race/ethnic groups in this sample.

Given higher %Tree trends for higher education and NSES strata for nonwhite participants in our findings and for greenspace in Casey et al. and Klompmaker et al. studies noted above,^11,15 our finding of the lowest %Trees for Whites in the highest NSES/education stratum (and highest %Trees for lower NSES/education strata) is unexpected. However, as noted, in results stratified by population density, this trend held only in high population density areas. One hypothesis for these unexpectedly low %Trees measures for higher education Whites in higher NSES areas is the potential for smaller-area spatial heterogeneity in more dense urban areas not adequately captured by a higher-level Census tract socioeconomic metric like NSES.³⁸ Furthermore, our study includes participants from six sites across the US, thus, heterogeneity of geography, differing historic patterns and practices of segregation across different racial/ethnic groups, and more recent gentrification of historically segregated urban areas may further complicate and make more granular the spatiotemporal realities of greenspace distribution in higher population density areas.^15,38,39 Future studies could benefit by incorporating additional measures of economic segregation or inequality (ICE) to investigate these patterns more granularly.¹⁵

Our findings of intersecting strata of nonwhite race/ethnicity, lowest education, and low NSES strata associated with lower %Trees were largely consistent with two prior studies of racial/ethnic and socioeconomic inequities in greenspace distribution; however, these studies focused on accessibility of or distance to greenspace, not just on its availability or visibility. The first was a study in Atlanta, finding overall that African Americans and socioeconomically disadvantaged neighborhoods had worse access to greenspace relative to predominantly white neighborhoods.⁹ The second was a study of US nationwide census tract data in 2010 finding that higher residential percentages of Black and Hispanic people were associated with a lower percentage of greenspace coverage in both higher and lower population density areas.¹⁰

Our results stratified by population density were somewhat in concordance with findings of Wen et al.¹⁰ showing that low NSES was associated with a lower percentage of greenspace in urban/suburban areas and a higher percentage of greenspace in rural areas. However, in our findings, this lower greenspace with lower NSES was observed only for Whites. Our low versus high population density stratified results suggest a contrast with the Klompmaker et al.¹¹ finding that urban Census tracts in the highest median household income quintile (vs. lowest) had higher NDVI (44.8% of the SD [95% CI = 42.8, 46.8]). Our findings that %Trees and %Grass varied little across educational strata also contrasted with the finding in Klompmaker et al.¹¹ where census tracts with a larger proportion of more highly education population had higher NDVI and NatureScores (although this was for urban census tracts only).

Contrasts in our findings with other studies may be explained by our intersectional MAIHDA approach in analyzing race/ethnicity, education and NSES not as three separate domains, but rather as intersecting domains. In this intersectional approach, each strata combination represents what has been characterized as “social location,” or a unique combination of overlapping elements of privilege and disadvantage of each domain that shapes individuals’ lived experiences.^22,28 Our finding that 16% of the overall variability in individuals’ residential trees and 11% of the variability in grass attributable to these intersecting strata suggests that these intersecting strata are useful for explaining a substantive portion of how social location is associated with maldistribution of greenspace. When adding the strata variables of race/ethnicity, education, and NSES separately to the models as fixed effects, we found that 5% (for trees) to 9% (for grass) of the variation was still explained at the stratum-cluster level. In addition, by stratifying our intersectional analysis by participants in lower versus higher population density areas, we were able to illuminate that the reverse trend of higher NSES with lower greenspace measures for White participants only held for higher population density areas in our sample.

Differences between our results and those of other studies are likely also due in part to our analyses of greenspace by vegetation type, demonstrating heterogeneity in composition of residential greenspace, not just quantity of greenspace, across combinations of social identity strata. Differences in distribution of greenspace by vegetation type have important implications for health outcomes as demonstrated in a recent study by Rifas-Shiman et al.⁴⁰ Their study showed that higher percentage of trees was associated with better cardiovascular health (β = 2.4; 95% CI = 1.3, 3.5), behavioral (β = 2.8; 95% CI = 1.4, 4.3), and biomedical (β = 2.8; 95% CI = 1.0, 4.7) scores; higher percentage of other greenspace (plants, fields, and flowers) was associated with better biomedical cardiovascular health scores (β = 2.2; 95% CI = 0.4, 3.9); with no beneficial association found for higher percentage of grass.⁴⁰ And while our measure of percentage other greenspace used in sensitivity analyses did not show any trends across intersecting strata, our findings together with those of Rifas-Shiman et al.⁴⁰ highlight the importance of investigating disparities in greenspace exposures by specific types of vegetation and greenspace when possible.

Limitations

This study has several limitations. We used cross-sectional data, so we did not evaluate how intersectional strata of race/ethnicity, education, and NSES may be associated with the maldistribution of greenspace over time. However, given the availability of longitudinal greenspace measures in the MESA study, future studies will have the opportunity to explore longitudinal associations. Also, while the use of street-view images (SVI) and deep learning algorithms has expanded and provided nuance to available metrics of greenspace exposure, the images are a snapshot of a location at a given time and may not provide an accurate representation of seasonal variability.³² And as noted, we imputed street-view data for participants’ residential locations from 2000 to 2002 using 2007 street-view image data. Future analyses should include quantitative bias analysis to assess the potential for measure error in pre-2007 imputed SVI measures due to temporal variation. In addition, the use of SVI-based residential measures of greenspace did not allow us to account for greenspace exposures that occur beyond the 500 m buffer of participants’ residences, introducing a potential for “neighborhood effect averaging problem”.⁴¹ In considering disparities in greenspace exposures, future studies could benefit from considering and incorporating mobility-based approaches to assessing greenspace exposures beyond just residential measures.^41–44 Our study was also limited in our use of a single variable that combines race and ethnicity into one of four mutually exclusive levels. We also note that our low (<8,000 persons/mi2) versus high (>8,000 persons/mi2) population density stratification may not be directly comparable to others’ use of rural (<1,000 persons/mi2) versus urban (>1,000 persons/mi2) population density distinctions. We also acknowledge that while our MESA sample is diverse and drawn from six diverse areas across the United States, our sample may not be representative of the whole United States, particularly given we have no truly rural areas represented. Future studies would benefit from considering a more expansive, nuanced approach to characterizing the complexities of racialized and ethnically marginalized experience.

Strengths

Our study also has numerous strengths. We believe ours is the first study to apply an intersectional MAIHDA approach to racial/ethnic and socioeconomic inequities in residential greenspace distribution. Our deployment of a MAIHDA approach allowed us to observe and compare trends in predicted measures of residential street-view greenspace at the intersections of racial/ethnic and socioeconomic characteristics rather than across each of these realms separately. Our study makes use of street-view image measures of residential greenspace, allowing us to observe inequities in greenspace distribution by type of vegetation. The null model VPCs demonstrate that a substantive portion of variability in street-view measures of trees and grass is due to combined additive and interactive contributions of intersecting strata that help to characterize combined influences of structural racism and socioeconomic factors and more holistically describe individuals’ lived experience. Furthermore, as demonstrated by Keller et al.,²⁸ a MAIHDA approach provides a more scalable, parsimonious method for accommodating combining multiple levels of intersecting social or experiential variables than traditional regression models, allowing better precision of estimates and greater ability to detect differences across strata. The contrast in some of our findings with others’ nonintersectional approaches to assessing the relationship of greenspace distribution with race/ethnicity and measures of SES, suggests added utility in a MAIHDA approach for investigating nuanced and multiplicative effects of systems of oppression like structural racism, classism, and socioeconomic segregation and deprivation in environmental justice research, including critical patterns of inequitable distribution of beneficial residential greenspace.

Conclusion

Residential greenspace is an important beneficial environmental exposure for people of all sociocultural and economic backgrounds and experiences, and is one important piece of the larger goal of environmental justice. Our results highlight evidence of the historic and ongoing systemic racism, segregation, and disparities driving the distribution of greenspace, and reinforce that efforts must continue towards meaningfully improving access to quality greenspaces in marginalized communities, particularly those of racial and ethnic minorities. Our findings of greater additive contributions of intersecting strata to residential tree measures and greater interactive contributions of intersecting strata to residential grass measures suggest that social identity strata combine in complex ways, but also by type of greenspace. In addition to the complexities of how multiple dimensions of social identities combine, these additive versus interactive differences by type of greenspace imply possible additional social, historical, and infrastructural factors that should be taken into account when pursuing interventions towards improving residential greenspace. Further environmental justice research and interventions on beneficial greenspace should consider not only disparities in overall quantity of greenspace, but also specific vegetation types, which we have also shown can differ across intersecting strata of social identity and socioeconomic status, especially given recent findings showing that health benefits can differ by vegetation type. For example, addressing greenspace inequities in marginalized communities by favoring tree planting over grass may be more a more beneficial combination of environmental justice and public health strategy over a “one type of greenspace fits all” approach.

Acknowledgments

We thank the investigators, staff, and participants of the MESA study for their important contributions. A full list of participating MESA investigators and institutions can be found at www.mesa-nhlbi.org.

Conflicts of interest statement

The authors declare that they have no conflicts of interest with regard to the content of this report.