Article Text

Download PDFPDF
Lifetime cancer risks from hazardous air pollutants in US public school districts
  1. Sara Elizabeth Grineski1,
  2. Timothy Collins2
  1. 1 Sociology, University of Utah, Salt Lake City, Utah, USA
  2. 2 Geography, University of Utah, Salt Lake City, Utah, USA
  1. Correspondence to Dr Sara Elizabeth Grineski, Sociology, University of Utah, Salt Lake City, Utah, USA; sara.grineski{at}


Background Children are sensitive to the health impacts of environmental contaminants, but research assessing outdoor environmental exposures for children and schools is underdeveloped. There are no national-level studies examining geographical and social disparities in air pollution exposure for children in school districts. Focusing on school districts is important because they are meaningful decision-making entities for schools.

Methods Using data from the National Air Toxics Assessment, we spatially reallocated lifetime cancer risk (LCR) from hazardous air pollutants (HAPs) within US school district boundaries, and paired those estimates with school district level sociodemographic measures obtained through the Integrated Public Use Microdata Series National Historic Geographic Information System. We employed local Moran’s I to identify district-level hotpots and generalised estimating equations (GEEs) to quantify risk disparities.

Results We identified hotspots of elevated LCR from all sources of HAPs (called ‘total’). A regional hotspot extends throughout the southeastern USA and smaller regional hotspots are present in southern Arizona, southern California and in California’s central valley. School districts with higher proportions of children, children with disabilities, foreign-born children, black children and multiracial/other race children, and lower proportions of Native American children, had greater total LCR (p<0.001). The effect of poverty on total LCR (p<0.001) was nonlinear; the lowest and highest poverty districts had lower total LCR.

Conclusions Geographical and social disparities in LCR across US school districts may be affecting children’s health and future potential. This new knowledge can inform policy changes, as school districts can advocate for the environmental health of children.

  • children
  • environmental injustice
  • school districts
  • hazardous air pollutants

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.


Children are sensitive to the health impacts of environmental exposures, due to age-related patterns of exposure and unique biological vulnerabilities.1 2 In the USA, over 13 000 school districts are home to approximately 75.1 million children under age 18. Insufficient attention has been directed towards environmental health threats at school and there are no federal agencies charged with protecting school environmental health.3 Research focused on assessing environmental injustices in exposure to pollution for children and schools is also underdeveloped.3 4 Environmental injustice is ‘the inequitable and disproportionately heavy exposure of poor, minority and disenfranchised populations to… environmental hazards (Landrigan et al, p. 178).1 Many studies have documented the presence of environmental injustices.5–7 Children are of growing interest in this literature due to their lack of control over where they live and attend school.

Previous studies of children and environmental injustice have followed one of two approaches. Some have examined the percent of children in census tracts and found this variable to be associated with elevated exposure in some studies, but not others.8 9 While representing a fine spatial scale, census tracts are limited as units of analysis in that they are somewhat artificial, do not correspond with decision-making authority, and do not map to a sense of community.10 Others have analysed associations between school-level sociodemographics and environmental exposures and found that schools serving poor and minority students are exposed to higher levels of pollution than those serving affluent and white students.5 7 11–14 Only one of these studies has been national, and multivariate findings showed that increasing proportions of black, Hispanic and Asian/Pacific Islander (API) students at schools were associated with greater health risks from air neurotoxicants.14 These findings suggest a disturbing pattern that warrants additional study. While important, the study was limited by its reliance on school demographics, as only a small set of covariates could be examined, and by its consideration of only 24 hazardous air pollutants (HAPs).

This study advances environmental injustice research by identifying geographical hotspots of lifetime cancer risk (LCR) from HAPs and testing for social disparities in LCR across school districts nationwide. The use of school district boundaries is new and represents an important contribution. School districts are stakeholders and decision-makers with respect to environmental justice. District-level school boards can advocate for the environmental health of children. They make decisions relevant to environmental health, including selecting sites and designs for major school building projects, purchasing school buses and implementing indoor air quality initiatives.3 In addition to being organisations that enable learning, school districts are ‘geographic boundaries that serve as a magnifying lens… [for] issues of race and wealth’.15 Districts are on average larger than census tracts, but smaller than counties; specifically 5.6 times larger and 0.24 times smaller, respectively. This paper has two aims: to characterise (1) geographical and (2) social disparities in LCRs from HAPs at the school district-level in the USA.


Dependent variables

We began with the US EPA's 2011 National Air Toxics Assessment (NATA) census tract level LCR estimates16 from HAPs emitted by major point sources (eg, factories), on-road mobile sources (eg, cars/trucks) and all sources (total), which includes the sources previously mentioned as well as background, non-road mobile and smaller stationary sources. LCR estimates are based on dose-response values of 71 specific HAPs identified in the 1990 Clean Air Act Amendments.16 The EPA uses a multi-step methodology to generate estimates of LCR in each US census tract, based primarily on the 2011 National Emissions Inventory.17 A LCR of 100 in 1 million implies that 100 out of 1 million equally exposed people would contract cancer after being continuously exposed to the HAPs over their lifetime, beyond cancer cases due to other causes. While LCRs do not consider length of residence or daily movements, they are suitable to compare health risks associated with exposure to HAPs between places.5 6

School district boundaries for 2011 were obtained through the Integrated Public Use Microdata Series (IPUMS) National Historic Geographic Information System portal for the 50 US states, District of Columbia and Puerto Rico (see figure 1, n=13 020). We allocated LCRs from census tracts to district boundaries following an accepted approach.18 We rasterised census tract data that were originally in vector format. We used a pixel size of 500 m2 to increase transformational precision. We allocated pixel values to district boundaries by taking the maximum value for each pollution source. The districts had an average of 2818 pixels and a median of 1164 pixels. The range was 1 to 926 953 due to differences in the school district sizes. Said differently, the maximum value is the highest census tract-level LCR estimate among all of the tracts that intersect or are completely contained within the school district boundary.

Figure 1

School districts in the USA and Puerto Rico (n=13 020, 2011), including those districts with the highest overall total lifetime cancer risk (LCR) from hazardous air pollutants (HAPs). Note: There is only one school district in Puerto Rico and one school district in Hawaii.

We used the maximum for several reasons. It captures rather than attenuates variability between school districts, which provides more statistical leverage. More importantly, it addresses concerns about the limitations of averaging risk19 and is responsive to recent scholarship on toxic outliers. Measurement of inequality can be dampened by using averages since most pollution comes from a minority of sources.20 21 While industrial polluters in the USA unequally burden poor and minority communities, the inequalities become more dramatic when considering the toxic outliers.22 School districts with the highest total LCRs from HAPs in the USA are depicted in figure 1.

Independent variables

We used the same IPUMS portal to download demographic data from the 2010 Census and 2012 5-year American Community Survey (ACS) estimates, aggregated to 2011 US school district boundaries. The independent variables are child-specific, referring to persons <18 years of age. We draw eight sociodemographic indicators from the 2010 Decennial Census: (1) proportion of the total population that are children; proportions of children that are non-Hispanic (2) black, (3) American Indian, (4) Asian/Pacific Islander (API) or (5) multiracial/other race; (6) proportion of children that are Hispanic of any race; and (7) proportion of occupied housing units that are rented by households with children. (8) Population density of children was included as a control variable. We created four additional variables (not available in the census) using 2008–2012 ACS estimates, which are centred on 2010 and correspond with census data. These include (9) proportion of children in poverty and (10) that variable squared since the relationship between poverty and risk may be curvilinear23; (11) proportion of children who were foreign born; and (12) proportion of children that have a disability. These variables capture a range of sociodemographic characteristics and are measures of privilege/disadvantage, which are typically6 7 24 or occasionally25 26 considered in similar studies.

Data analysis

For the first aim, we conducted a hotspot analysis using the local Moran's I index to identify spatial clusters (e.g., groups of districts with high LCR) and spatial outliers (e.g., individual districts with relatively high LCR surrounded by those with lower LCR).27 We employed an inverse distance weighting function, which assumes that the strength of the spatial relationship diminishes proportionally to the inverse distance between two school districts.28 We applied a false discovery rate correction to the p values, which reduces the critical p value threshold to account for multiplicity and spatial dependency and results in the discovery of meaningful clusters.29

For the second aim, we specified three generalised estimating equations (GEEs) predicting total, point and on-road mobile LCRs from HAPs in school districts in the USA with ≥100 children. Districts with <100 children were removed to ensure stable proportions.14 Independent variables were standardised before entering them into the model. GEEs build from the generalised linear model and appropriately adjust for clustering.30 To define clusters of school districts, we used the US state of location and four ranked categories of urbanicity (i.e., degree to which the geographical unit is urban). This is because LCRs vary between states and between urban and rural areas. To create the urban cluster variable, we applied K-means cluster analyses to a ‘proportion urban’ values for each school district (using census data accessed through the IPUMS NHGIS portal aggregated to 2011 school districts). GEE models use an intracluster dependency correlation matrix30 and exchangeable was the best fitting specification. Exchangeable assumes compound symmetry or constant intracluster dependency.

To further refine model fit, we tested normal, gamma and inverse Gaussian distributions with logarithmic and identity link functions30 given that our dependent variables were not normally distributed. The best fitting specifications were inverse Gaussian with log link for total LCR and gamma with log link for point and on-road LCR. Results from the GEEs are not affected by multicollinearity based on variance inflation factor, tolerance and condition index criteria. As a sensitivity analysis, we also conducted the GEE analysis using the district-level average pixel LCR, which is the areal-weighted average for tracts that intersect or are completely contained within the district boundary.


Identifying geographical disparities in LCR from HAPs

The Moran’s I analysis identified geographical hotspots of LCR at the school-district level (see figure 2). In terms of total LCR from HAPs, a large regional hotspot extends throughout the southeastern USA. Regional hotspots are also present in southern Arizona, southern California and in California’s central valley. Individual hotspots are identified throughout the western, Midwestern and northeastern USA, and near Anchorage, Alaska. The pattern for point source risks is quite different. Smaller regional hotspots are found in the Rust Belt, mid-Atlantic and South. Numerous individual hotspots occur in California, Colorado, southern New Mexico and less densely across several western and Midwestern states. The patterns of regional and individual hotspots for on-road mobile LCR from HAPs are similar, with a few distinctions. Regional hotspots for on-road LCR are scattered east of the Mississippi and in school districts near Anchorage, Los Angeles and San Francisco, and in western Washington. Individual hotspots for on-road LCR occur throughout the Midwest, Intermountain West and Southwest.

Figure 2

Identifying spatial clusters and outliers of lifetime cancer risk (LCR) from hazardous air pollutants (HAPs) in US school districts (n=13 020, 2011). Findings for Puerto Rico and Hawaii are not depicted, since each has only one school district and both are geographically distant from other school districts.

Predicting LCR from HAPs

Table 1 presents descriptive statistics for all analysis variables for districts with ≥100 children. Table 2 presents GEE results predicting total and disaggregated source-specific LCRs from HAPs. The following summary is organised based on findings for each independent variable. As the proportion of children in a school district increased, total LCR was higher (b=0.02; 95% CI=0.01 to 0.03), as was point LCR (0.09; 0.02 to 0.15). There was a curvilinear relationship between the proportion of children in poverty and total (0.04 (main effect); 0.02 to 0.07 and −0.04 (squared term); –0.07 to –0.02), point (0.36; 0.19 to 0.52 and −0.38; -0.61 to –0.15) and on-road mobile (0.11; 0.06 to 0.15 and −0.0.15; -0.19 to –0.15).

Table 1

Descriptive statistics for child-specific analysis variables (n=12 516 school districts with ≥100 children)

Table 2

Results from generalised estimating equations† predicting maximum LCR‡ in US school districts (n=12 516 with ≥100 children§)

As the proportion of foreign-born children in a district increased, total LCR was higher (0.02; 0.01 to 0.02), as was on-road (0.04; 0.02 to 0.06) LCR. As the proportion of disabled children in a district increased, total LCR was higher (0.01; >0.00 to 0.01), as were point (0.09; 0.05 to 0.13) and on-road (0.01; >0.00 to 0.02) LCRs.

Racial/ethnic minority children faced greater LCR from HAPs with the exception of Native American children. Greater proportion of black children was associated with greater total (0.04; 0.03 to 0.06), point (0.22; 0.05 to 0.39) and on-road (0.08; 0.04 to 0.11) LCRs. School districts with higher proportions of Hispanic children had greater on-road LCR (0.09; 0.05 to 0.12), but lower point LCR (−0.12; −0.20 to –0.10). Greater proportion of API children predicted greater total (0.02; >0.00 to 0.04) and on-road (0.06; 0.04 to 0.08) LCRs. Lower proportion of Native American children was associated with greater total (−0.02; −0.03 to –0.01), point (−0.15; −0.20 to 0.10) and on-road (−0.07; −0.09 to –0.06) LCRs. Greater proportion of children of other/multiracial backgrounds was associated with greater total (0.03; 0.02 to 0.05) and on-road (0.11; 0.09 to 0.13) LCRs. Greater proportion of renter-occupants with children was associated with significantly lower LCR from on-road mobile sources (−0.03; −0.06 to <0.00).

The sensitivity analysis using average LCR from HAPs show that results were similar in terms of direction and significance (p<0.05), but inequalities were dampened when using average versus maximum, as expected.22 Both poverty terms lost significance (p<0.6), as did API (p<0.2), when predicting average total LCR. Children (p>0.1) and both poverty terms (p<0.1) lost significance when predicting average point LCR, while multiracial/other became significant (p<0.05). Children (p<0.001) became significant, but disability (p<0.1) and Hispanic (p<0.07) lost significance when predicting average on-road LCR.


US children are facing substantial exposure to cancer-causing HAPs in their school districts. These exposures will likely contribute to future cases of cancer and other serious health conditions. There are geographical disparities in LCR across school districts. The cluster analyses revealed elevated LCRs in the southeastern USA as well in many major US cities.

There are also social disparities in LCR from HAPs. Districts with higher proportions of children had significantly greater total and point LCRs. This indicates greater impacts in areas where populations have relatively high youth composition. We found positive and significant disparities in LCRs for districts with greater proportions of black children, which have also been found for school-based exposures.11 14 Greater proportions of Hispanic children was significantly associated with higher on-road LCR but lower point LCR. In Florida and California,7 31 inequalities in LCRs for black children were more substantial than they were for Hispanic children, although both experienced environmental injustice relative to white children.

A higher proportion of children from other/multiracial backgrounds was associated with greater LCR from HAPs in all three models (p<0.05 for total and on-road). The ‘other’ US Census category includes people’s open-ended responses like ‘biracial’ or a Hispanic/Latino subgroup like ‘Mexican’.32 Three-quarters of people who are ‘≥2 races’ are white and either black, other race, Asian or American Indian.32 While results for this variable are often ignored in similar studies, the multiracial/other population is important as growing numbers of Americans are multiracial; in 2010, ‘≥2 races’ was the third largest racial group in the USA after ‘white’ and ‘black’.32 This variable likely captures a wide range of people of colour who face discrimination, similar to people identifying with one minority group.

Increasing proportion of API children was significantly associated with total and on-road LCRs from HAPs. This finding contributes to a growing literature documenting distributional environmental injustices for API Americans.6 14 33 We found Native American children to be generally protected from cancer-causing HAPs, which mirrors a national neighbourhood-level study.6

We found that as proportion of foreign-born children in a district increased, LCRs from HAPs were higher (p<0.05 for total and on-road). Foreign-birth is generally underemphasised as a risk factor for pollution exposure. While disparities for foreign-born residents have been noted in a US-national county-level study,34 France,35 and El Paso (Texas),36 this indicator had not been examined previously in the context of children or schools. It is unclear how increased exposure may translate into health effects given the “immigrant epidemiological paradox”,37 which suggests that this area is ripe for future analyses.

The effect of poverty on LCR from HAPs was nonlinear with the lowest and highest poverty school districts have having significantly lower LCR in the three models. This is possibly because high poverty districts have fewer economic/pollution-generating activities while low poverty districts house residents with economic and political power to resist pollution.23

Our potentially most novel findings are associations between disabled children and significantly greater total, point and on-road LCRs from HAPs. The disability measure gauges hearing, vision, cognitive, ambulatory, self-care and independent living difficulties. Disability is rarely considered in distributional environmental injustice studies and while results do not show causality, they relate to a growing body of literature connecting pollution to disability.26 38

Limitations: The NATA excludes exposure from ingestion and skin contact and synergies between HAPs. The public health implications of LCR from HAPs quantified here should not be overstated since children do not remain in their 2011 school district throughout their entire lives, and the NATA assumes lifetime exposure. The EPA also warns that census tract-level NATA data are not designed to pinpoint specific risk values.16 The disability measure undercounts children with disabilities in that it tabulates ‘cognitive’, ‘ambulatory’ and ‘self-care’ disabilities only for children >5 years of age, and ‘challenges with independent living’ only for children >15. ACS data are subject to sampling errors that may bias results.39 Conducting the analysis at the district level influences findings since results can differ depending on choice of geographical unit.40 We selected school districts to increase policy relevance, as opposed to using more finely resolved census tracts. District-level results draw attention to the importance of considering air pollution in school siting and closure decisions within the districts. While examining district-level effects is valuable, future studies could employ child-level data which would enable the examination of intersectional risks between race/ethnicity and poverty.

The geographical and social inequalities in LCR revealed at the school-district level indicate the existence of highly unequally polluted environments which likely have disparate impacts on US children. The exposures may result in missed school days due to environmentally associated illnesses now, but could also cause cellular damage that will impact US children for decades to come. These disparate exposures in childhood may result in increased probabilities of suboptimal health and lower achievement in adulthood, and the perpetuation of intergenerational poverty.1 At the district level, elevated LCR from HAPs combines with additional challenges affecting already burdened school districts, including a long history of segregation and unequal funding. Nearly one-third of US children are enrolled in districts that are ≥75% non-white and for every student enrolled, non-white school districts receive $2226 less than white districts.15 More evidence-based research on long-term health consequences of hazard exposure with a focus on inequalities as well as decision-making guided by an ethical imperative to protect our children1 are needed to redress the injustices demonstrated in this paper.

What is already known on this subject

  • Only one environmental justice study has been conducted at the national level focused on US schools. Multivariate findings showed that increasing proportions of black, Hispanic and Asian/Pacific Islander students at schools were associated with greater risks from air neurotoxicants. These findings suggest a disturbing pattern nationwide that warrants additional study.

What this study adds

  • The previous national study was limited by its reliance on a limited suite of school demographics and hazardous air pollutant indicators. This study reveals geographical and social inequalities in lifetime cancer risks from hazardous air pollutants at the school-district level for the first time and uncovers risks for foreign-born and disabled children. The existence of these highly unequally polluted environments likely has disparate impacts on US children’s life trajectories.



  • Contributors SEG conceived of the study, conducted the statistical analysis and lead wrote the manuscript. TC provided input throughout all phases of the study and assisted in the writing of the manuscript.

  • Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.

  • Competing interests None declared.

  • Patient consent for publication Not required.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Data sharing statement Data are available in a public, open access repository.