Neighborhood Composition and Air Pollution in Chicago: Monitoring Inequities With a Dense, Low-Cost Sensing Network, 2021

Abstract

Objectives. To evaluate the efficacy of a novel, real-time sensor network for routine monitoring of racial and economic disparities in fine particulate matter (PM_2.5; particulate matter ≤ 2.5 µm in diameter) exposures at the neighborhood level.

Methods. We deployed a dense network of low-cost PM_2.5 sensors in Chicago, Illinois, to evaluate associations between neighborhood-level composition variables (percentage of Black residents, percentage of Hispanic/Latinx residents, and percentage of households below poverty) and interpolated PM_2.5. Relationships were assessed in spatial lag models after adjustment for all composition variables. Models were fit with data both from the overall period and during high-pollution episodes associated with social events (July 4, 2021) and wildfires (July 23, 2021).

Results. The spatial lag models showed that racial/ethnic composition variables were associated with higher PM_2.5 levels. Levels were notably higher in neighborhoods with larger compositions of Hispanic/Latinx residents across the entire study period and notably higher in neighborhoods with larger Black populations during the July 4 episode.

Conclusions. As a complement to sparse regulatory networks, dense, low-cost sensor networks can capture spatial variations during short-term air pollution episodes and enable monitoring of neighborhood-level inequities in air pollution exposures in real time. (Am J Public Health. 2022;112(12):1765–1773. https://doi.org/10.2105/AJPH.2022.307068)

Recognizing health equity as a priority, public health researchers and practitioners are increasingly seeking to monitor and mitigate disparities in exposures to preventable causes of disease.^1,2 Air pollution in the form of fine particulate matter (PM_2.5; particulate matter ≤ 2.5 µm in diameter) is a leading environmental contributor to disease burdens³ and disparities in disease burdens.⁴ Adding to the urgency is the role of climate change in increasing overall PM_2.5 exposures through longer and more extreme wildfire seasons,⁵ although little is known about the extent to which these climate change–exacerbated pollution events affect inequities in exposures.

In the United States, public health agencies obtain PM_2.5 data from the Environmental Protection Agency (EPA) ambient air monitoring network. This network, implemented as a result of the Clean Air Act, has been credited with contributing to a reduction in PM_2.5 of approximately 70% since 1981.⁶ However, remote sensing data offer evidence that areas with the highest air pollution exposures in 1981 remain the most polluted areas more than 30 years later.⁶ Moreover, modeled estimates from EPA emissions inventories show that Black and Hispanic/Latinx people experience higher exposures than White people.⁷ These findings provide a rationale for monitoring systems that track not only pollution exposures but also disparities in these exposures.

Existing EPA data provide accurate information on air pollution exposures but are subject to limitations related to data coverage over space and time. In Chicago, Illinois, a city of 600 square kilometers, the EPA maintains 4 stations monitoring PM_2.5. Real-time hourly estimates of PM_2.5 from monitors at 2 of these stations are available through the public AirNow tool,⁸ which provides estimates of regional-level exposures; however, the data are too spatially sparse to allow inferences about more granular levels of exposure, which is a problem given that evidence from mobile monitoring campaigns shows air quality can vary significantly between neighborhoods and city blocks.⁹ Alternatively, the EPA does provide a “downscaler model” of PM_2.5 estimates at the census tract level,¹⁰ but the model cannot be used to track real-time exposures because the most recent data are from 2018.

Dense city-wide networks of low-cost sensors could complement existing regulatory networks, enabling routine real-time monitoring of spatial variations in environmental exposures. To date, however, cities in the United States have largely not designed and implemented their own dense sensing networks (with a few exceptions such as New York City¹¹). Instead, cities monitoring air quality at “hyperlocal” levels rely on data from crowdsourced and mobile monitoring initiatives, approaches subject to limitations. Crowdsourced networks are poorly suited to monitoring disparities as a result of systematic biases in sensor locations; for example, commonly used PurpleAir sensors are more likely to be located in White areas and areas of high socioeconomic status than in areas with environmental justice concerns.^12,13

Mobile monitoring campaigns, for which research-grade sensors are placed on moving vehicles, offer both city-wide coverage and insights on spatial variation⁹; if the vehicle fleet is small, however, this approach cannot compare multiple places at the same time or provide real-time insights for the city as a whole. Running larger campaigns would be prohibitively time and labor intensive for an urban public health department. Other common approaches (e.g., estimating PM_2.5 levels from satellite imagery, emissions inventories, or sophisticated chemical transport models) are subject to model-related uncertainties and would benefit from training and validation with additional data collected across diverse neighborhoods.⁵ There is thus an opportunity for dense real-time monitoring to complement existing regulatory networks for the specific purpose of monitoring disparities in exposures.

In this study, we deployed a low-cost sensor network built to monitor racial/ethnic and economic inequities in air pollution exposures. We conducted our research in the city of Chicago, building on previous efforts to document heightened exposures in environmental justice neighborhoods^14,15 as well as evidence that heightened air pollution exposures and social vulnerabilities are clustered on the south and west sides of the city.¹⁶ We evaluated data from July 2021 given that July historically has higher PM_2.5 readings in comparison with other months.¹⁷ In addition, there were 2 air pollution episodes in July 2021: on July 4, an expected pollution episode contributed to large but short-term increases in PM_2.5, and on July 23, an unexpected pollution event that corresponded to wildfires similarly contributed to heightened PM_2.5 levels over a short period of time. We examined spatial clustering in PM_2.5 in relation to the spatial clustering of sociodemographic variables; we further evaluated relationships between neighborhood-level sociodemographic composition and PM_2.5.

METHODS

Chicago is a diverse city characterized by roughly equal thirds of White, Black, and Hispanic/Latinx populations. Chicago is also one of the most segregated cities in the United States and has seen concerns regarding structural racism as a fundamental cause of inequitable pollution burdens.¹⁶

Air Pollution

Network design

Our work relied on air pollution data from a novel network of 115 sensors located on bus shelters across Chicago. We deployed the network during the summer of 2021 in collaboration with the city of Chicago, the academic Array of Things initiative, and JCDecaux Chicago, the local affiliate of JCDecaux SA—the world’s largest outdoor advertising company—which installed sensing devices on the city’s bus shelters. We also collaborated with the Environmental Law and Policy Center to support neighborhood environmental justice organizations in reviewing the network design.

The network was designed with the aim of monitoring local inequities in air pollution exposures. For 80 devices, we selected sites using a stratified random sampling design based on the approach of the New York City Community Air Survey.¹¹ Of the remaining devices, 26 were allocated to sites selected by community partners, and 9 were sited across 3 EPA regulatory monitoring stations (3 per station). Additional details on the network design can be found in Appendix A (available as a supplement to the online version of this article at http://www.ajph.org). In this analysis, we used only sites allocated through stratified random sampling.

Eclipse device

We designed devices used in the network to provide real-time measurements of air pollution in an urban setting (for full details on the hardware, see Daepp et al.¹⁸). Each device included a Sensirion SPS30 sensor, which collected PM_2.5 readings every 5 minutes, as well as sensors for relative humidity, barometric pressure, and temperature. Further details on the Sensirion SPS30 are provided in Appendix A. In addition, details on our calibration function to improve sensor accuracy to levels consistent with EPA recommendations for low-cost sensors are provided in Appendix B (available as a supplement to the online version of this article at http://www.ajph.org). Of note, daily average sensor values were highly correlated with daily averages from regulatory monitors surrounding Chicago (for details, see Appendix B and Appendix C, Figure C1, available as a supplement to the online version of this article at http://www.ajph.org).

FIGURE 1—

Inverse Distance Weighted Predicted PM_2.5 Levels Summarized to Community Areas and Average Values From Project Eclipse Sensors in Chicago, IL: (a) July 2–31, 2021; (b) July 4, 2021; and (c) July 23, 2021

Notes. PM_2.5 = fine particulate matter ≤ 2.5 µg/m³. There were 77 community areas and 77 sensors. Moran’s I = 0.74, 0.81, and 0.66 for parts a, b, and c, respectively.

Data cleaning and processing

From July 2 to July 31, we obtained 568 156 5-minute readings from 78 sensors. Following Lu et al., we implemented a 4-step quality control procedure.¹⁹ First, we removed sensor data that were deemed as malfunctioning, determined according to a moving 5-hour standard deviation of 0 (0% of readings). Second, we removed implausible readings by excluding values of 0 and values above the measurement range of 1000 micrograms per cubic meter per manufacturer specifications (0.003%). Third, using a 75% completeness criterion, we removed readings from hours with less than 9 (of 12) 5-minute measures and from days with less than 18 (of 24) hours of data (1.86%). Finally, as a secondary check for malfunctioning devices, we assessed the extent to which sensor readings were consistent with readings from neighboring sensors.

In addition, we performed a linear regression of daily average readings for a given index sensor and its neighbors within a 5-kilometer radius and removed all readings from the index sensor if the R² value was less than 0.6 (0.01% of readings, affecting only a single sensor). With these criteria, 557 457 (98.1%) of the original 5-minute readings remained.

Daily average PM_2.5 values were calculated for each sensor by initially averaging 5-minute readings for each hour within a specified day (up to 24 hours), calibrating hourly data, and then aggregating those hourly averages for each specified day. With the exclusion criteria, 77 sensors remained. Preliminary analyses assessing spatial clustering in the set of device days excluded suggested no substantial clustering (Appendix C, Figure C2, available as a supplement to the online version of this article at http://www.ajph.org).

Spatial interpolation

We used inverse distance weighting (IDW) to estimate PM_2.5 across 77 community areas for the entire study period and during 2 air pollution episodes (July 4, consistent with excess air pollution caused by fireworks,²⁰ and July 23, consistent with nationwide increases in air pollution caused by wildfires on the west coast²¹). Chicago is divided into 77 community areas covering an average of 7.8 square kilometers. These community areas have historically been used for planning and statistical purposes and remain largely consistent with residents’ contemporary perceptions of neighborhood boundaries.²²

IDW interpolation allowed us to predict PM_2.5 values across unknown points (i.e., reference points) on the basis of nearby points where PM_2.5 is known (i.e., monitoring points). The approach assigns values to reference points through a weighted average of the values at monitoring points; monitoring points closest to a given reference point have larger weights than monitoring points further away. Weights are defined by the inverse of the distance between each reference point and monitoring point and then raised to an arbitrary power that we set equal to 2 (i.e., the square of the inverse distance), a value supported by both empirical cross validation and previous literature.²³

We used IDW to create smoothed maps (rasters) of averaged estimated PM_2.5 at a grid cell resolution of 100 × 100 meters. We then aggregated these grid cells to compute community area–level averages such that a grid cell belonged to a given community area if its centroid fell inside of it. We further evaluated the robustness of our results to the use of a different interpolation approach, ordinary kriging; the 2 approaches produced similar neighborhood-level estimates (Spearman’s ρ = 0.86), and thus we used the IDW approach because of its interpretability and consistency with theoretical models of air pollution spread²³ and its widespread usage both in academic research^23,24 and by the EPA.²⁵

Sociodemographic Composition Variables

We used sociodemographic data from the 2015 to 2019 American Community Survey²⁶ to measure racial/ethnic composition and the percentage of households below poverty. Regarding racial/ethnic composition, we focused on percentages of non-Hispanic Black and Hispanic/Latinx residents because these are the 2 largest racial/ethnic minority groups in Chicago, accounting for 29.1% and 28.7% of the city’s population, respectively. Census tract–level measures were aggregated to community areas such that a census tract belonged to a given community area if its centroid fell inside of it.

Statistical Analysis

Initially, we computed descriptive statistics for each outcome (interpolated PM_2.5 values for July 2–31, July 4, and July 23) and composition variable. We also assessed the degree of spatial autocorrelation for each variable by calculating global Moran’s I statistics. Moran’s I values range from −1 to 1; values near 0 indicate no autocorrelation (i.e., randomness), positive values indicate clustering, and negative values indicate dispersion. We then mapped the distribution of PM_2.5 and each composition variable at the community area level. In addition, we assessed bivariate relationships between tract-level composition variables and PM_2.5 using Spearman correlation coefficients.

To model the relationship between sociodemographic composition and PM_2.5, we created linear regression models adjusting for each socioeconomic composition measure over the entire study period and separately for the July 4 and July 23 air pollution episodes. We adjusted for all 3 measures to avoid overestimating the effects of any single variable. Because air pollution is spatially patterned, violating the linear regression assumption of independence,²⁷ we also fit a series of spatial lag models to account for spatial dependence.

Spatial lag models are similar to linear regression models, but they include in addition a lagged dependent variable reflecting the weighted average of PM_2.5 values across neighboring community areas. The coefficient associated with this lagged variable (ρ) quantifies the strength of spatial dependence. If ρ is greater than 0, this indicates that PM_2.5 values are positively related to those of neighboring community areas; a negative value indicates the inverse. A value of 0 indicates no dependence and renders the equation equivalent to a linear model.²⁸ Spatial lag models, relative to other spatial regression models such as spatial error models, are often used when researchers believe that spatial autocorrelation is caused by an underlying substantive process.^28,29 In the case of air pollution, we theorized that industrial zoning and related planning policies that resulted in clusters of pollution sources were ultimately caused by distal, structural processes of discrimination and residential segregation.³⁰

To evaluate the presence of autocorrelation in our regression models, we calculated Moran’s I values for residuals using first-order queen contiguity-based weights. Models incorporating queen contiguity-based weights yielded lower Akaike information criterion (AIC) values than preliminary models employing (1) rook contiguity-based weights and (2) the minimum distance for all community areas to have at least 1 neighbor and were thus chosen for our analysis. Pseudo P values for Moran’s I statistics were generated via a Monte Carlo simulation of 999 random replications. We considered autocorrelation to be present if pseudo P values were less than .05.

We used R version 4.1.0 (R Foundation for Statistical Computing, Vienna, Austria) in conducting our analyses; we used the gstat package to perform IDW and the spdep package to generate spatial lag models. All models were fit with community area–level data; as a sensitivity analysis addressing concerns regarding the modifiable areal unit problem, we replicated our main analyses at the census tract level.

RESULTS

Table 1 summarizes descriptive statistics across the 77 community areas. The average level of interpolated PM_2.5 during the study period was 13.16 µg/m³ (range = 11.3–14.6 µg/m³). Average interpolated PM_2.5 levels were slightly elevated on July 4 (14.2 µg/m³; range = 10.6–17.3 µg/m³) and doubled on July 23 (26.6 µg/m³; range = 24.1–28.3 µg/m³). The network design resulted in an average of 1 sensor allocated to each community area (range = 0–4).

TABLE 1—

Descriptive Statistics of Variables Across 77 Community Areas: Chicago IL, July 2–31, 2021

	Mean (SD)	Range
Interpolated PM2.5 (µg/m3)
July 2–31	13.2 (0.7)	11.3–14.6
July 4	14.2 (1.5)	10.6–17.3
July 23	26.6 (1.0)	24.1–28.3
Percentage
Black	38.1 (39.1)	0.4–96.5
Hispanic/Latinx	26.0 (26.9)	0.1–89.2
Households below poverty	19.6 (10.9)	3.5–53.9
Eclipse devices, no.	1.0 (0.9)	0.0–4.0

Note. PM_2.5 = particulate matter ≤ 2.5 µm in diameter.

The average community area was composed of 38.1% Black residents (range = 0.4%–96.5%) and 26.0% Hispanic/Latinx residents, with an average of 19.6% of households below poverty (range = 3.5%–53.6%). Consistent with known racial and economic segregation patterns in Chicago, as shown in Appendix C, Figure C3 (available as a supplement to the online version of this article at http://www.ajph.org), Black residents were clustered in areas on the west and south sides (Moran’s I = 0.71; P = .001); Hispanic/Latinx residents were clustered in areas on the northwest, southwest, and south sides (Moran’s I = 0.63; P = .001); and the percentage of households below poverty was generally larger along the outer edge of the city (Moran’s I = 0.51; P = .001).

Figure 1 illustrates the spatial distribution of interpolated PM_2.5 values during the entire study period and for each air pollution episode. Interpolated values of PM_2.5 displayed substantial spatial clustering, indicated by large Moran’s I values (overall: 0.74; July 4: 0.81; July 23: 0.66; all pseudo Ps = .001). Notably, higher PM_2.5 levels appeared to cluster primarily along the west side during the study period overall and the July 23 pollution episode, whereas higher levels appeared to cluster along the south side during the July 4 pollution episode.

We provide aspatial Spearman correlation coefficients in Appendix D, Table D1 (available as a supplement to the online version of this article at http://www.ajph.org), and focus the remainder of our results on regression models. Specifically, we focus on adjusted spatial lag models, as Moran’s I values were reduced to nearly 0 and pseudo P values were all above .05, indicating that spatial models sufficiently removed spatial autocorrelation; furthermore, these models achieved better fits than corresponding linear regression models. For the period from July 2 to July 31, Table 2 shows substantial evidence of a positive relationship between Hispanic/Latinx residential composition and PM_2.5 (B = 0.47; 95% confidence interval [CI] = 0.10, 0.84). Relationships between Black residential composition and percentage of households below poverty were also positive but marked with imprecision.

TABLE 2—

Adjusted Models of Sociodemographic Composition on PM_2.5 Across 77 Community Areas: Chicago, IL, July 2–31, 2021

	Linear Model	Spatial Lag Model
July 2–31
% Black, B (95% CI)	0.55 (−0.09, 1.18)	0.12 (−0.21, 0.45)
% Latinx, B (95% CI)	1.61 (0.93, 2.30)	0.47 (0.10, 0.84)
% of households below poverty, B (95% CI)	1.13 (−0.74, 2.99)	0.79 (−0.15, 1.74)
ρ, B (95% CI)	…	0.85 (0.75, 0.95)
Moran’s I (P)	0.63 (.001)	0.01 (.33a)
AIC	148.63	70.65
Log likelihood (df)	−69.32 (5)	−29.32 (6)
R2	0.29	0.75b
July 4
% Black, B (95% CI)	3.69 (2.73, 4.66)	1.13 (0.44, 1.82)
% Latinx, B (95% CI)	2.65 (1.61, 3.69)	0.67 (0.01, 1.33)
% of households below poverty, B (95% CI)	−0.14 (−2.98, 2.70)	0.24 (−1.46, 1.95)
ρ, B (95% CI)	…	0.76 (0.63, 0.88)
Moran’s I (P)	0.39 (.001)	0.01 (.4a)
AIC	213.47	155.76
Log likelihood (df)	−101.74 (5)	−71.88 (6)
R2	0.61	0.82b
July 23
% Black, B (95% CI)	0.17 (−0.82, 1.15)	0.13 (−0.42, 0.67)
% Latinx, B (95% CI)	1.20 (0.14, 2.27)	0.35 (−0.24, 0.94)
% of households below poverty, B (95% CI)	1.61 (−1.30, 4.51)	0.41 (−1.18, 1.99)
ρ, B (95% CI)	…	0.86 (0.75, 0.96)
Moran’s I (P)	0.61 (.001)	0.09 (.1a)
AIC	217.19	150.45
Log likelihood (df)	−103.60 (5)	−69.22 (6)
R2	0.10	0.63b

Note. AIC = Akaike information criterion; CI = confidence interval; PM_2.5 = particulate matter ≤ 2.5 μm in diameter. The sample size was 77.

^a Pseudo P value.

^b Nagelkerke pseudo R² value.

During the July 4 episode, the adjusted spatial model suggested positive, substantial relationships between PM_2.5 and both Black residential composition (B = 1.13; 95% CI = 0.44, 1.82) and Hispanic/Latinx residential composition (B = 0.67; 95% CI = 0.01, 1.33). The relationship between percentage of households below poverty and PM_2.5 was positive but marked with imprecision once again. Finally, during the July 23 episode, the adjusted spatial model showed that all sociodemographic composition variables had a positive relationship with PM_2.5 but were marked with imprecision.

Sensitivity analyses in which census tracts were used yielded results similar to those of our main analyses at the community area level, albeit with increased precision and substantial autocorrelation in adjusted spatial lag models (Appendix D, Table D2, available as a supplement to the online version of this article at http://www.ajph.org).

DISCUSSION

In this study, we examined differences in sociodemographic disparities in PM_2.5 across long-term monitoring periods versus during short-term air pollution episodes. We have provided evidence of spatial variation across neighborhoods in levels of PM_2.5 both across the study period and during specific high-pollution episodes characterized by social events (July 4) and wildfires (July 23). Exposures were substantially higher in areas with larger compositions of Hispanic/Latinx residents across the entire study period and substantially higher in areas with larger compositions of Hispanic/Latinx and Black residents during the July 4 episode. No sociodemographic composition variable was associated with PM_2.5 during the July 23 episode. Our results demonstrate the effectiveness of a city-wide, real-time sensing network for measuring ongoing and episodic neighborhood-level disparities in poor air exposures.

Evidence of heightened PM_2.5 in areas with relatively more Hispanic/Latinx and Black residents is consistent with literature documenting racial and ethnic disparities in PM_2.5³¹ as well as studies linking racial residential segregation to environmental disparities.^30,32 Our study adds to this body of literature in 2 key ways. First, we demonstrated differences in the groups affected in short-term episodes versus over longer time periods. Although neighborhoods with larger proportions of Hispanic/Latinx residents appeared to have the largest PM_2.5 burden overall, July 4 may be an especially harmful pollution event disproportionately affecting areas with larger proportions of Black residents. These results point to the need for more targeted interventions that consider both spatial and temporal contexts. Although similar findings might be obtained by downscaling models that combine regulatory monitoring data with chemical transport models or satellite data, a particular benefit of a real-time monitoring approach is its potential to make findings available to policymakers immediately after or even during a pollution episode, supporting mitigation efforts.

Second, our work offers evidence that city-wide high-pollution events such as the wildfire-related episode on July 23 may result in minimal variations across sociodemographic composition, making dense, real-time monitoring particularly beneficial if such events obscure disparities occurring during more typical PM_2.5 exposure days.

Limitations

Our study is subject to several key limitations. First, logistical delays in sensor deployment resulted in missing data over space and time. Of the 80 sensors intended for analysis in this study, 3 could not be deployed until the end of July; a slow rollout over the first week of monitoring as well as occasional intermittent sensor failures further limited the sample size to 59 sensors on July 4 and 71 on July 23. However, the data collected were robust against data quality issues, as our quality control procedure identified issues with less than 2% of the 5-minute readings collected and no spatial clustering was detected in the fraction of missing device days.

Second, low-cost sensors can exhibit low accuracy for regulatory purposes. We sought to address this limitation through calibration and focusing on the comparison of trends over time and of sensors with each other; however, exact estimates should be treated with caution because the results may have been affected by systematic sensor error. Our network thus cannot substitute for regulatory networks; rather, as a complement to sparse regulatory monitors, our approach can help prioritize monitoring and mitigation of disparities in air pollution. We were also able to reduce random error by aggregating data to daily average values; however, this methodological decision limited our ability to take advantage of the high temporal resolution of the sensor readings.

Third, the generalizability of our results may be limited by our siting of sensors at bus shelters. However, there were also benefits to the use of bus shelters; all devices were placed in similar contexts and at a consistent height, and, because technicians regularly visit these bus shelters, the siting approach facilitated maintenance that mitigated data loss.¹⁸ Moreover, bus shelters represent locations where people congregate and—importantly—breathe. Also, our analyses were limited to July 2021 and may not generalize to other months; however, our work does provide a framework for extending such a monitoring network over longer periods.

Finally, land use regression models could offer increased precision relative to IDW in estimating hyperlocal variations in PM_2.5. However, these models incorporate temporally invariant covariates (e.g., physical geography) that would smooth out key real-time fluctuations. A lack of ground-truth data beyond what the 4 EPA sites capture further limited our ability to validate our models.

Policy Implications

We have described the development of a dense, real-time sensor network that enables characterization of spatial variations in PM_2.5 at the neighborhood scale. Examining a period characterized by multiple short-term air pollution episodes, we showed persistent inequities throughout the study period as well as important variations in the groups most affected by different short-term events. It follows that interventions seeking to address inequities in air pollution exposures may need to contend with how inequities vary across time and specific pollution events. Our results show how low-cost sensors can be used in a large, urban setting for monitoring environmental inequities, offering an approach that can be reproduced by public health departments in other cities seeking to promote environmental justice.