Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
Contents lists available at ScienceDirect
Spatial and Spatio-temporal Epidemiology journal homepage: www.elsevier.com/locate/sste
Original Research
Assessing environmental inequalities in ambient air pollution across urban Australia Luke D. Knibbs a,⇑, Adrian G. Barnett b a b
School of Population Health, The University of Queensland, Brisbane, Australia School of Public Health and Social Work & Institute of Health and Biomedical Innovation, Queensland University of Technology, Brisbane, Australia
a r t i c l e
i n f o
Article history: Received 6 December 2013 Revised 23 December 2014 Accepted 19 March 2015 Available online 25 March 2015 Keywords: Air pollution Environmental justice Indigenous Socio-economic Australia
a b s t r a c t Identifying inequalities in air pollution levels across population groups can help address environmental justice concerns. We were interested in assessing these inequalities across major urban areas in Australia. We used a land-use regression model to predict ambient nitrogen dioxide (NO2) levels and sought the best socio-economic and population predictor variables. We used a generalised least squares model that accounted for spatial correlation in NO2 levels to examine the associations between the variables. We found that the best model included the index of economic resources (IER) score as a non-linear variable and the percentage of non-Indigenous persons as a linear variable. NO2 levels decreased with increasing IER scores (higher scores indicate less disadvantage) in almost all major urban areas, and NO2 also decreased slightly as the percentage of non-Indigenous persons increased. However, the magnitude of differences in NO2 levels was small and may not translate into substantive differences in health. Ó 2015 Elsevier Ltd. All rights reserved.
1. Introduction Environmental injustices occur when more disadvantaged populations bear a disproportionate burden of the adverse impacts of pollution or other environmental hazards (Brulle and Pellow, 2006). They have been the focus of numerous studies over the past 50 years, which have identified injustices associated with a range of contaminants among low-socioeconomic status (SES) communities and racial minorities (Clark et al., 2014; Jerrett, 2009). Environmental inequality is a closely-related but distinct concept that refers to differences in levels of contaminants among different population groups (Marshall, 2008). Air pollution is a particularly relevant environmental exposure due to its ubiquitousness in urban areas and because it is one the top 10 risk factors in the global disease burden ⇑ Corresponding author at: Public Health Building, The University of Queensland, Herston, QLD 4006, Australia. Tel.: +61 7 3365 5409. E-mail address: [email protected] (L.D. Knibbs). http://dx.doi.org/10.1016/j.sste.2015.03.001 1877-5845/Ó 2015 Elsevier Ltd. All rights reserved.
(Lim et al., 2012). Determining whether environmental inequalities in air pollution exposure exist can inform policy measures and interventions to reduce their impacts (Bell et al., 2005). Australia (population 23 million) has an advanced economy and its Human Development Index (HDI) was ranked second in the world by in 2013 (United Nations Development Programme, 2014). However, parts of Australia’s population are subject to pronounced disadvantages that are juxtaposed against its overall development. This is best exemplified by the greater than 10-year shortfall in life expectancy among Indigenous (i.e. Aboriginal and Torres Strait Islander) compared with nonIndigenous Australians (Australian Institute of Health and Welfare, 2012). Also, more socio-economically disadvantaged Australians exhibit higher prevalence of health risk factors (e.g. smoking) and experience poorer health than less disadvantaged persons (Australian Institute of Health and Welfare, 2012).
2
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
Despite the considerable socio-economic and racial gradients in health among Australians, there is a conspicuous absence of studies addressing environmental inequalities or injustices in Australia (Chakraborty and Green, 2014). This lack of basic information makes it impossible to determine if policy-based responses are required. To contribute towards filling this knowledge gap, we sought to assess if environmental inequalities in ambient air pollution exposure exist in Australia. 2. Methods 2.1. Population and socio-economic predictors We obtained population data based on the 2011 census from the Australian Bureau of Statistics (Australian Bureau of Statistics, 2011). The data were at ABS Statistical Area Level 1 (SA1), which is the smallest spatial unit for which specific census data (e.g. socio-economic variables) are released (Australian Bureau of Statistics, 2011). There are almost 55,000 SA1s across Australia and together they cover the entire country. Their median population is 385 persons (range = 0–6434), and their median size is 0.22 km2 (range = 0.002–329,722 km2). SA1s in urban areas are smaller than those in rural and remote areas (Australian Bureau of Statistics, 2011). We determined the total population density (per km2) in each SA1. The total number of people identifying as Aboriginal Torres Strait Islander or both (i.e. all Indigenous persons) in each SA1 was used to calculate the Indigenous population percentage and population density. We obtained the 2011 socio-economic indexes for areas (SEIFA) for each SA1 from the Australian Bureau of Statistics. SEIFA comprises 4 indexes: the index or relative socio-economic disadvantage (IRSD), index of relative socio-economic advantage and disadvantage (IRSAD), index of economic resources (IER) and index of education and occupation (IEO). The indexes are based on between 9 (IEO) and 25 (IRSAD) variables including income, education, employment, occupation, housing, mortgage and rent payments, English language skills, disability and single parent families (Australian Bureau of Statistics, 2013). The indexes are numerical scores based on weighted combinations of the input variables and are assigned to approximately 96% of all SA1s (Australian Bureau of Statistics, 2013). Depending on the index, lower scores can mean greater levels of relative disadvantage (with or without a corresponding lack of advantage), a lack of access to economic resources, or people who are unemployed and without qualifications. The technical basis and validation of the indexes is described elsewhere, and they are the standard metric used to evaluate socio-economic patterns in Australia (Australian Bureau of Statistics, 2013). 2.2. Air pollution data We used a recently developed and validated satellitebased land-use regression (LUR) model to estimate longterm ambient nitrogen dioxide (NO2) levels. The LUR
model is described in detail by Knibbs et al. (2014). Briefly, it uses satellite observations of tropospheric NO2 columns, land use, roads, and other predictors to estimate ground-level NO2 across Australia, and captures 81% of spatial variation in annual NO2 levels between 2006 and 2011 (absolute RMS error = 1.4 ppb). The LUR model is useful for assessing within-urban gradients in NO2, which made it well-suited to the aims of this study. We focused on NO2 because it is a strong indicator of traffic and other combustion-related pollution (e.g. industrial processes, coal-fired power generation), is a major component of ambient air pollution, and exhibits greater spatial heterogeneity than other air pollutants (Briggs et al., 1997; Jerrett et al., 2005). For these reasons, NO2 has been used as proxy in previous environmental inequality studies aimed at air pollution (e.g. Clark et al., 2014; Havard et al., 2009; Padilla et al., 2014; Yanosky et al., 2008). We used the LUR model to predict average NO2 concentrations during 2006–2011. Predictions were made at the centroid of each census mesh block (Knibbs et al., 2014), which is a standard method to estimate population exposures to NO2 using LUR (e.g. Novotny et al., 2011; Hystad et al., 2011). There are approximately 350,000 mesh blocks across Australia and they are the spatial unit that constitutes each SA1 (with no overlap), but unlike SA1s no census data are released for mesh blocks due to their small size (Australian Bureau of Statistics, 2011). The mean NO2 concentration at each SA1 was estimated using the concentrations predicted at the mesh block centroids within it. We used ArcGIS (version 10.0) to process our data. We restricted our analysis to include just the major urban areas in Australia, as defined by the Australian Bureau of Statistics (Australian Bureau of Statistics, 2011), and we only included SA1s with a non-zero total population and valid socioeconomic indexes. These criteria resulted in approximately 20,000 SA1s being dropped. Our final sample had 34,866 SA1s, covered approximately 10,100 km2, and incorporated 69.1% of the Australian population. The major urban areas we included were located near the capital cities of Australia’s 8 states and territories. We focused on major urban areas because they have higher and more heterogeneous levels of NO2 than rural and remote areas (Knibbs et al., 2014). 2.3. Analysis We aimed to find the best set of predictors of estimated NO2 at each SA1 from the 4 socio-economic indexes and the following area population variables: (1) nonIndigenous population density per km2; (2) Indigenous population density per km2; (3) total population density per km2 (4) percent Indigenous, and; (5) percent nonIndigenous. We note from our previous work (Knibbs et al., 2014) that there are many features of the environment that are potential predictors of NO2 (e.g. roads, impervious surfaces, industrial activity). However, our aim was not to produce a model highly predictive of NO2; instead, we were specifically interested in the role of the selected socio-economic variables. Other strong predictors of NO2 may be on the casual pathway between the
3
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
selected socio-economic variables and NO2, for example a lower socio-economic index may be associated with a higher density of nearby roads. However, our focus was on identifying potential environmental inequalities and we therefore aimed to find if the socio-economic variables were associated with NO2, rather than how the association occurs. Because of the very strong positive skew in total population density we used a log-transform to reduce the influence of a few very high densities. The dependent variable (estimated NO2 concentration) had a spatial correlation that we had to adjust for in our regression models (Supplement, Fig. S1; Tables S1 and S2). We did this using a generalised least squares model with correlated errors assuming an exponential function of latitude and longitude. This model took significant computing power, even on a supercomputer. The model could only be run using sub-samples of the data of size 4000. Larger samples took more than 72 h to run and/or exceeded the maximum memory size available. We therefore took 10 random samples of the complete data (n = 34,866) of size 4000 and ran the same model 10 times. These random samples tended to have a small number of locations with high Indigenous population density. This made it difficult for the models to estimate the effects of the indigenous population variables. To address this we included an extra sample of 400 more locations with a high Indigenous density (over 4%). We used R (version 3.1.0) to perform our analyses. The model fit statistic was the total mean square error. This was estimated by comparing the predicted and observed values for each model across all 34,866 locations. We combined the model parameters from the 10 samples using a multivariate Bayesian meta-analysis fitted in WinBUGS (version 1.4.2). We used the following staged procedure to select the best combination of predictors: (1) Find the best socio-economic index variable from the four available using the minimum mean square error. The association between the socio-economic index and pollutant is tested for both linear and non-linear associations using linear association and a natural spline with two and three degrees of freedom. (2) Find the best population variable from the five above using the minimum mean square error. As above the association is both linear and non-linear. (3) Starting from a model with the above two variables, fit every non-included variable and check the minimum mean square error.
Table 1 Descriptive summary statistics for all variables. Note: ppb = parts per billion; IRSAD = index of relative socio-economic advantage and disadvantage; IRSD = index of relative socio-economic disadvantage; IER = index of economic resources; IEO = index of education and occupation. Variable (units)
Mean
S.D.
Min.
Max.
Annual average NO2 2006–11 (ppb) Non-Indigenous population density (persons/km2) Indigenous population density (persons/km2) Total population density (persons/km2) Percent non-Indigenous (%) Percent Indigenous (%) IRSAD score IRSD score IER score IEO score
8.0
2.8
2.3
30.1
2983
3716
0.5
155456
36.1
70.8
0
2289
3212
4177
0.5
186148
93.8 1.4 1019 1014 1007 1025
4.8 2.6 96.7 95.5 98.4 100
2.3 0 432 306 404 632
100 96.6 1246 1193 1290 1375
Supplement (Fig. S2 and Table S3). Because of the extremely strong correlation between IRSAD and IRSD score we dropped IRSD score, as IRSAD includes information on both relative advantage and disadvantage, rather than disadvantage only. Fig. 1 shows the results of the variable selection process for socio-economic indexes expressed in terms of their error sum of squares. The best model used the IER score with two degrees of freedom as judged by the error. The results for two and three degrees of freedom were similar, but we used two degrees on the principle of parsimony. Fig. 2 shows the results of variable selection for population variables. Non-Indigenous population density was dropped because it had an extremely strong correlation with total population density. The best population predictor as judged by the error was the percentage of nonIndigenous persons as a linear variable. Similar fits were achieved using the percentage of Indigenous persons. We considered additional models using the remaining socio-economic and population variables, and found that none offered a substantial improvement on the base
To help with model convergence all predictors were standardised.
3. Results Table 1 summarises the descriptive statistics for each variable in the model. Histograms and correlations for the dependent and independent variables are in the
Fig. 1. Boxplots of error sum of squares for selecting the best socioeconomic index. Note: IEO = index of education and occupation; IER = index of economic resources; IRSAD = index of relative socioeconomic advantage and disadvantage.
4
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6 Table 2 Parameter estimates for the best model from the Bayesian meta-analysis. Note: ns (, k) = natural spline, kth parameter. Parameter Intercept ns(IER_score, 1) ns(IER_score, 2) Percent non-Indigenous
Fig. 2. Boxplots of error sum of squares for the best population variable. Note: ATSI = Aboriginal and Torres Strait Islander persons (i.e. Indigenous Australians); PC = percent.
model with IER score as a non-linear variable and the percentage of non-Indigenous persons as a linear variable (Supplement, Fig. S3). Therefore, we concluded that it was the best model (R2 = 6.4%). Fig. 3 gives the final predictions from the best model. It shows a decrease in predicted NO2 levels with increasing IER scores for all but the bottom 5% of SA1s. Predicted NO2 also decreased slightly as the percentage of non-Indigenous persons increased. The caveat to this result is that most areas have a high percentage of non-Indigenous persons with a small number of areas having a low percentage, hence the predictions interpolate an area where there is no data. Parameter estimates are in Table 2 and show that both parameters were statistically significant and the 95% credible intervals do not contain zero. 4. Discussion We assessed if environmental inequalities in ambient air pollution are present in Australia by identifying
Mean
S.D.
Lower
Upper
6.08 0.55 0.83 0.01
0.08 0.10 0.04 0.00
5.93 0.35 0.91 0.01
6.23 0.75 0.76 0.00
z 80.6 5.6 21.1 6.6
associations between socio-economic and population variables and pollutant levels. We found that the best socioeconomic and population predictors of ambient NO2 levels estimated by a satellite-based land-use regression model in urban Australia were the index of economic resources (IER) score and the percentage of non-Indigenous persons. While both were statistically significant they only explained 6.4% of the spatial variation in NO2 levels. However, IER score and the percentage of non-Indigenous persons were clearly the most important socio-economic and population variables associated with ambient NO2. We found that locations with higher IER scores had lower predicted NO2 across the vast majority of urban areas. The IER is based on 14 variables that together can be used to assess the financial aspects of socio-economic advantage and disadvantage, and higher scores occur in less disadvantaged areas (Australian Bureau of Statistics, 2013). A higher proportion of the population that identify as non-Indigenous was associated with slightly reduced NO2 levels. Indigenous Australians fare worse than their non-Indigenous counterparts across many indicators of social and socio-economic determinants of health, and this may also be the case for exposure to ambient NO2 (Australian Institute of Health and Welfare, 2012). Taken together, the results suggest some evidence of environmental inequalities in NO2 levels across urban Australia. However, the magnitude of the difference in NO2 levels was small (0.8 ppb between most and least-
Fig. 3. Predicted NO2 levels from the best model, which used the index of economic resources (IER) score as a non-linear variable and percent nonIndigenous as a linear variable (R2 = 6.4%). Density plots show the distribution of each variable. Note: ATSI = Aboriginal and Torres Strait Islander persons (i.e. Indigenous Australians).
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
exposed IER areas; 0.25 ppb between most and leastexposed percent non-Indigenous areas). While modest increases in air pollution can have relatively large public health impacts due to the large number of people exposed over long periods, the slight differences observed in this study may not translate into substantive differences in health. A recent Australian study identified that the toxicity and volume of industrial pollutant emissions are greater in areas that are more socio-economically disadvantaged and have a higher proportion of Indigenous persons (Chakraborty and Green, 2014). Our study used a different methodology and focused on a different exposure (ambient air pollution), but we also found some evidence to suggest environmental inequalities in pollutants in Australia. The extent to which our results are associated with adverse health effects is not known. Further studies to address these questions using more powerful study designs (e.g. cohort studies) that control for the numerous confounding influences could offer useful insights. Based on initial results that showed spatial correlation in NO2 concentrations (Supplement, Fig. S1; Tables S1 and S2), we used regression models with spatially correlated errors. Other studies have noted the need to account for spatial correlation in environmental inequality analyses to minimise the possibility of biased or misleading results (Yanosky et al., 2008; Havard et al., 2009; Padilla et al., 2014). These findings and ours demonstrate that assessing spatial correlation in environmental inequality studies is necessary, and that results based on the assumption of independent observations should be treated with caution. International studies have consistently identified environmental inequalities in proximity to air pollution sources or pollutant concentrations (e.g. Bell and Ebisu, 2012; Brainard et al., 2002; Chaix et al., 2006; Finkelstein et al., 2005; Grineski et al., 2007; Gunier et al., 2003; Havard et al., 2009; Jerrett, 2009; Marshall, 2008; Morello-Frosch et al., 2002; Padilla et al., 2014; Pearce et al., 2006; Pearce and Kingham, 2008; Perlin et al., 2001). These inequalities and pre-existing differences in risk due to health status have also been linked to greater adverse cardiovascular and birth outcomes among socioeconomically disadvantaged groups or racial minorities (e.g. Clark et al., 2014; Finkelstein et al., 2005; Ponce et al., 2005; Forastiere et al., 2007; Yanosky et al., 2008). Such findings are a concerning environmental manifestation of the inverse care law (Hart, 1971), where those who should ideally be best protected from air pollution are instead most exposed to it. Assessing environmental inequalities in pollutant exposures and health is important for identifying whether policy-level responses are required to protect against them and what the likely benefits of these actions will be. The main limitations of this study require consideration. First, we only examined NO2 and treated it as a proxy for ambient air pollution from vehicle and non-vehicle sources. Other pollutants may show different gradients. Second, there is potential for miscounting of Indigenous and non-Indigenous populations based on self-reporting in the census, although census data are of high quality and used broadly in research. Third, we used general
5
indicators of Indigenous persons (e.g. percent Indigenous and non-Indigenous) and did not have data to examine the associations among other racial groups in Australia. Fourth, we only focused on major urban areas and our study area did not include the 30.9% of the Australian population that live outside these areas. Finally, computational constraints imposed by our spatial model meant that we had to run 10 random selections of size 4000 of our study sample of 35,000. However, the results from each random model were relatively homogeneous. Strengths of this study include: (1) the use of a validated satellite-based LUR model with national coverage to predict NO2 concentrations across Australia (capable of explaining 81% of variation in NO2); (2) adjustment for spatial correlation in NO2 levels, and; (3) the use of a systematic approach and Bayesian meta-analysis to combine parameter estimates and select the best socio-economic and population predictors of NO2. In summary, we found small but statistically significant effects to suggest environmental inequalities in ambient NO2 levels in the major urban areas of Australia. Areas with higher IER scores, which indicate reduced disadvantage, were associated with lower NO2 levels, as were areas with a higher proportion of non-Indigenous persons. These findings provide a basis to pursue more comprehensive assessments of environmental inequality in air pollution across Australia. Acknowledgements Computational resources and services used in this study were provided by the High Performance Computer and Research Support Unit, Queensland University of Technology. L.D.K. is supported by an NHMRC Early Career (Australian Public Health) Fellowship (APP1036620). Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.sste.2015.03.001. References Australian Bureau of Statistics (ABS). Australian Statistical Geography Standard (ASGS): Volume 1 – Main structure and greater capital city statistical areas; 2011. http://www.abs.gov.au/ausstats/[email protected]/mf/ 1270.0.55.001 [accessed 23/12/14]. Australian Bureau of Statistics (ABS). Socio-economic indexes for areas (SEIFA); 2013. http://www.abs.gov.au/ausstats/[email protected]/mf/2033.0. 55.001 [accessed 23/12/14]. Australian Institute of Health and Welfare (AIHW). Australia’s health 2012; 2012. http://www.aihw.gov.au/WorkArea/DownloadAsset. aspx?id=10737422169 [accessed 23/12/14]. Bell ML, Ebisu K. Environmental inequality in exposure to airborne particulate matter components in the United States. Environ Health Perspect 2012;120:1699–704. Bell ML, O’Neill MS, Cifuentes LA, Braga ALF, Green C, Nweke A, et al. Challenges and recommendations for the study of socioeconomic factors and air pollution health effects. Environ Sci Policy 2005;8:525–33. Brainard JS, Jones AP, Bateman IJ, Lovett AA. Modelling environmental equity: access to air quality in Birmingham, England. Environ Plann A 2002;34:695–716.
6
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
Briggs DJ, Collins S, Elliott P, Fischer P, Kingham S, Lebret E, et al. Mapping urban air pollution using GIS: a regression-based approach. Int J Geogr Inf Sci 1997;11:699–718. Brulle RJ, Pellow DN. Environmental justice: human health and environmental inequalities. Annu Rev Public Health 2006;27:103–24. Chaix B, Gustafsson S, Jerrett M, Kristersson H, Lithman T, Boalt A, Merlo J. Children’s exposure to nitrogen dioxide in Sweden: investigation environmental justice in an egalitarian country. J Epidemiol Commun Health 2006;60:234–41. Chakraborty J, Green D. Australia’s first national level quantitative environmental justice assessment of industrial air pollution. Environ Res Lett 2014;9:044010. Clark LP, Millet DB, Marshall JD. National patterns in environmental injustice and inequality: outdoor NO2 air pollution in the United States. PLoS One 2014;9:e94431. Finkelstein MM, Jerrett M, Sears MR. Environmental inequality and circulatory disease mortality gradients. J Epidemiol Commun Health 2005;59:481–7. Forastiere F, Stafoggia M, Tasco C, Picciotto S, Agabiti N, Cesaroni G, et al. Socioeconomic status, particulate air pollution, and daily mortality: differential exposure or differential susceptibility. Am J Ind Med 2007;50:208–16. Grineski S, Bolin B, Boone C. Criteria air pollution and marginalized populations: environmental inequity in Metropolitan Phoenix, Arizona. Soc Sci Q 2007;88:535–54. Gunier RB, Hertz A, Von Behren J, Reynolds P. Traffic density in California: socioeconomic and ethnic differences among potentially exposed children. J Expo Anal Environ Epidemiol 2003;13:240–6. Hart JT. The inverse care law. Lancet 1971;297:405–12. Havard S, Deguen S, Zmirou-Navier D, Schillinger C, Bard D. Traffic-related air pollution and socioeconomic status: a spatial autocorrelation study to assess environmental equity on a small-area scale. Epidemiology 2009;20:223–30. Hystad P, Setton E, Carvantes A, Poplawski K, Deschenes S, Brauer M, et al. Creating national air pollution models for population exposure assessment in Canada. Environ Health Perspect 2011;119:1123–9. Jerrett M. Global geographies of injustice in traffic-related air pollution exposure. Epidemiology 2009;20:231–3. Jerrett M, Arain A, Kanaroglou P, Beckerman B, Potoglou D, Sahsuvaroglu T, et al. A review and evaluation of intraurban air pollution exposure models. J Expo Anal Environ Epidemiol 2005;15:185–204.
Knibbs LD, Hewson MG, Bechle MJ, Marshall JD, Barnett AG. A national satellite-based land-use regression model for air pollution exposure assessment in Australia. Environ Res 2014;135:204–11. Lim SS, Vos T, Flaxman AD, Danaei G, Shibuya K, Adair-Rohani H, et al. A comparative risk assessment of burden of disease and injury attributable to 67 risk factors and risk factor clusters in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study. Lancet 2012;380:2224–60. Marshall JD. Environmental inequality: air pollution exposures in California’s South Coast Air Basin. Atmos Environ 2008;42:5499–503. Morello-Frosch R, Pastor M, Porras C, Sadd J. Environmental justice and regional inequality in Southern California: implications for future research. Environ Health Perspect 2002;110(Suppl. 2):149–54. Novotny EV, Bechle MJ, Millet DB, Marshall JD. National satellite-based land-use regression: NO2 in the United States. Environ Sci Technol 2011;45:4407–14. Padilla CM, Kihal-Talantikite W, Vieira V, Rossello P, Le Nir G, ZmirouNavier D, et al. Air quality and social deprivation in four French metropolitan areas – a localized spatio-temporal environmental inequality analysis. Environ Res 2014;134:315–24. Pearce J, Kingham S, Zawar-Reza P. Every breath you take? Environmental justice and air pollution in Christchurch, New Zealand. Environ Plann A 2006;38:919–38. Pearce J, Kingham S. Environmental inequalities in New Zealand: a national study of air pollution and environmental justice. Geoforum 2008;39:980–93. Perlin SA, Wong D, Sexton K. Residential proximity to industrial sources of air pollution: interrelationships among race, poverty, and age. J Air Waste Manage Assoc 2001;51:406–21. Ponce NA, Hoggatt KJ, Wilhelm M, Ritz B. Preterm birth: the interaction of traffic-related air pollution with economic hardship in Los Angeles neighborhoods. Am J Epidemiol 2005;162:140–8. United Nations Development Programme. Human development index and its components; 2014. http://hdr.undp.org/en/content/table-1human-development-index-and-its-components [accessed 23/12/ 14]. Yanosky JD, Schwartz J, Suh HH. Associations between measures of socioeconomic position and chronic nitrogen dioxide exposure in Worcester, Massachusetts. J Toxicol Environ Health A 2008;71: 1593–602.
Contents lists available at ScienceDirect
Spatial and Spatio-temporal Epidemiology journal homepage: www.elsevier.com/locate/sste
Original Research
Assessing environmental inequalities in ambient air pollution across urban Australia Luke D. Knibbs a,⇑, Adrian G. Barnett b a b
School of Population Health, The University of Queensland, Brisbane, Australia School of Public Health and Social Work & Institute of Health and Biomedical Innovation, Queensland University of Technology, Brisbane, Australia
a r t i c l e
i n f o
Article history: Received 6 December 2013 Revised 23 December 2014 Accepted 19 March 2015 Available online 25 March 2015 Keywords: Air pollution Environmental justice Indigenous Socio-economic Australia
a b s t r a c t Identifying inequalities in air pollution levels across population groups can help address environmental justice concerns. We were interested in assessing these inequalities across major urban areas in Australia. We used a land-use regression model to predict ambient nitrogen dioxide (NO2) levels and sought the best socio-economic and population predictor variables. We used a generalised least squares model that accounted for spatial correlation in NO2 levels to examine the associations between the variables. We found that the best model included the index of economic resources (IER) score as a non-linear variable and the percentage of non-Indigenous persons as a linear variable. NO2 levels decreased with increasing IER scores (higher scores indicate less disadvantage) in almost all major urban areas, and NO2 also decreased slightly as the percentage of non-Indigenous persons increased. However, the magnitude of differences in NO2 levels was small and may not translate into substantive differences in health. Ó 2015 Elsevier Ltd. All rights reserved.
1. Introduction Environmental injustices occur when more disadvantaged populations bear a disproportionate burden of the adverse impacts of pollution or other environmental hazards (Brulle and Pellow, 2006). They have been the focus of numerous studies over the past 50 years, which have identified injustices associated with a range of contaminants among low-socioeconomic status (SES) communities and racial minorities (Clark et al., 2014; Jerrett, 2009). Environmental inequality is a closely-related but distinct concept that refers to differences in levels of contaminants among different population groups (Marshall, 2008). Air pollution is a particularly relevant environmental exposure due to its ubiquitousness in urban areas and because it is one the top 10 risk factors in the global disease burden ⇑ Corresponding author at: Public Health Building, The University of Queensland, Herston, QLD 4006, Australia. Tel.: +61 7 3365 5409. E-mail address: [email protected] (L.D. Knibbs). http://dx.doi.org/10.1016/j.sste.2015.03.001 1877-5845/Ó 2015 Elsevier Ltd. All rights reserved.
(Lim et al., 2012). Determining whether environmental inequalities in air pollution exposure exist can inform policy measures and interventions to reduce their impacts (Bell et al., 2005). Australia (population 23 million) has an advanced economy and its Human Development Index (HDI) was ranked second in the world by in 2013 (United Nations Development Programme, 2014). However, parts of Australia’s population are subject to pronounced disadvantages that are juxtaposed against its overall development. This is best exemplified by the greater than 10-year shortfall in life expectancy among Indigenous (i.e. Aboriginal and Torres Strait Islander) compared with nonIndigenous Australians (Australian Institute of Health and Welfare, 2012). Also, more socio-economically disadvantaged Australians exhibit higher prevalence of health risk factors (e.g. smoking) and experience poorer health than less disadvantaged persons (Australian Institute of Health and Welfare, 2012).
2
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
Despite the considerable socio-economic and racial gradients in health among Australians, there is a conspicuous absence of studies addressing environmental inequalities or injustices in Australia (Chakraborty and Green, 2014). This lack of basic information makes it impossible to determine if policy-based responses are required. To contribute towards filling this knowledge gap, we sought to assess if environmental inequalities in ambient air pollution exposure exist in Australia. 2. Methods 2.1. Population and socio-economic predictors We obtained population data based on the 2011 census from the Australian Bureau of Statistics (Australian Bureau of Statistics, 2011). The data were at ABS Statistical Area Level 1 (SA1), which is the smallest spatial unit for which specific census data (e.g. socio-economic variables) are released (Australian Bureau of Statistics, 2011). There are almost 55,000 SA1s across Australia and together they cover the entire country. Their median population is 385 persons (range = 0–6434), and their median size is 0.22 km2 (range = 0.002–329,722 km2). SA1s in urban areas are smaller than those in rural and remote areas (Australian Bureau of Statistics, 2011). We determined the total population density (per km2) in each SA1. The total number of people identifying as Aboriginal Torres Strait Islander or both (i.e. all Indigenous persons) in each SA1 was used to calculate the Indigenous population percentage and population density. We obtained the 2011 socio-economic indexes for areas (SEIFA) for each SA1 from the Australian Bureau of Statistics. SEIFA comprises 4 indexes: the index or relative socio-economic disadvantage (IRSD), index of relative socio-economic advantage and disadvantage (IRSAD), index of economic resources (IER) and index of education and occupation (IEO). The indexes are based on between 9 (IEO) and 25 (IRSAD) variables including income, education, employment, occupation, housing, mortgage and rent payments, English language skills, disability and single parent families (Australian Bureau of Statistics, 2013). The indexes are numerical scores based on weighted combinations of the input variables and are assigned to approximately 96% of all SA1s (Australian Bureau of Statistics, 2013). Depending on the index, lower scores can mean greater levels of relative disadvantage (with or without a corresponding lack of advantage), a lack of access to economic resources, or people who are unemployed and without qualifications. The technical basis and validation of the indexes is described elsewhere, and they are the standard metric used to evaluate socio-economic patterns in Australia (Australian Bureau of Statistics, 2013). 2.2. Air pollution data We used a recently developed and validated satellitebased land-use regression (LUR) model to estimate longterm ambient nitrogen dioxide (NO2) levels. The LUR
model is described in detail by Knibbs et al. (2014). Briefly, it uses satellite observations of tropospheric NO2 columns, land use, roads, and other predictors to estimate ground-level NO2 across Australia, and captures 81% of spatial variation in annual NO2 levels between 2006 and 2011 (absolute RMS error = 1.4 ppb). The LUR model is useful for assessing within-urban gradients in NO2, which made it well-suited to the aims of this study. We focused on NO2 because it is a strong indicator of traffic and other combustion-related pollution (e.g. industrial processes, coal-fired power generation), is a major component of ambient air pollution, and exhibits greater spatial heterogeneity than other air pollutants (Briggs et al., 1997; Jerrett et al., 2005). For these reasons, NO2 has been used as proxy in previous environmental inequality studies aimed at air pollution (e.g. Clark et al., 2014; Havard et al., 2009; Padilla et al., 2014; Yanosky et al., 2008). We used the LUR model to predict average NO2 concentrations during 2006–2011. Predictions were made at the centroid of each census mesh block (Knibbs et al., 2014), which is a standard method to estimate population exposures to NO2 using LUR (e.g. Novotny et al., 2011; Hystad et al., 2011). There are approximately 350,000 mesh blocks across Australia and they are the spatial unit that constitutes each SA1 (with no overlap), but unlike SA1s no census data are released for mesh blocks due to their small size (Australian Bureau of Statistics, 2011). The mean NO2 concentration at each SA1 was estimated using the concentrations predicted at the mesh block centroids within it. We used ArcGIS (version 10.0) to process our data. We restricted our analysis to include just the major urban areas in Australia, as defined by the Australian Bureau of Statistics (Australian Bureau of Statistics, 2011), and we only included SA1s with a non-zero total population and valid socioeconomic indexes. These criteria resulted in approximately 20,000 SA1s being dropped. Our final sample had 34,866 SA1s, covered approximately 10,100 km2, and incorporated 69.1% of the Australian population. The major urban areas we included were located near the capital cities of Australia’s 8 states and territories. We focused on major urban areas because they have higher and more heterogeneous levels of NO2 than rural and remote areas (Knibbs et al., 2014). 2.3. Analysis We aimed to find the best set of predictors of estimated NO2 at each SA1 from the 4 socio-economic indexes and the following area population variables: (1) nonIndigenous population density per km2; (2) Indigenous population density per km2; (3) total population density per km2 (4) percent Indigenous, and; (5) percent nonIndigenous. We note from our previous work (Knibbs et al., 2014) that there are many features of the environment that are potential predictors of NO2 (e.g. roads, impervious surfaces, industrial activity). However, our aim was not to produce a model highly predictive of NO2; instead, we were specifically interested in the role of the selected socio-economic variables. Other strong predictors of NO2 may be on the casual pathway between the
3
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
selected socio-economic variables and NO2, for example a lower socio-economic index may be associated with a higher density of nearby roads. However, our focus was on identifying potential environmental inequalities and we therefore aimed to find if the socio-economic variables were associated with NO2, rather than how the association occurs. Because of the very strong positive skew in total population density we used a log-transform to reduce the influence of a few very high densities. The dependent variable (estimated NO2 concentration) had a spatial correlation that we had to adjust for in our regression models (Supplement, Fig. S1; Tables S1 and S2). We did this using a generalised least squares model with correlated errors assuming an exponential function of latitude and longitude. This model took significant computing power, even on a supercomputer. The model could only be run using sub-samples of the data of size 4000. Larger samples took more than 72 h to run and/or exceeded the maximum memory size available. We therefore took 10 random samples of the complete data (n = 34,866) of size 4000 and ran the same model 10 times. These random samples tended to have a small number of locations with high Indigenous population density. This made it difficult for the models to estimate the effects of the indigenous population variables. To address this we included an extra sample of 400 more locations with a high Indigenous density (over 4%). We used R (version 3.1.0) to perform our analyses. The model fit statistic was the total mean square error. This was estimated by comparing the predicted and observed values for each model across all 34,866 locations. We combined the model parameters from the 10 samples using a multivariate Bayesian meta-analysis fitted in WinBUGS (version 1.4.2). We used the following staged procedure to select the best combination of predictors: (1) Find the best socio-economic index variable from the four available using the minimum mean square error. The association between the socio-economic index and pollutant is tested for both linear and non-linear associations using linear association and a natural spline with two and three degrees of freedom. (2) Find the best population variable from the five above using the minimum mean square error. As above the association is both linear and non-linear. (3) Starting from a model with the above two variables, fit every non-included variable and check the minimum mean square error.
Table 1 Descriptive summary statistics for all variables. Note: ppb = parts per billion; IRSAD = index of relative socio-economic advantage and disadvantage; IRSD = index of relative socio-economic disadvantage; IER = index of economic resources; IEO = index of education and occupation. Variable (units)
Mean
S.D.
Min.
Max.
Annual average NO2 2006–11 (ppb) Non-Indigenous population density (persons/km2) Indigenous population density (persons/km2) Total population density (persons/km2) Percent non-Indigenous (%) Percent Indigenous (%) IRSAD score IRSD score IER score IEO score
8.0
2.8
2.3
30.1
2983
3716
0.5
155456
36.1
70.8
0
2289
3212
4177
0.5
186148
93.8 1.4 1019 1014 1007 1025
4.8 2.6 96.7 95.5 98.4 100
2.3 0 432 306 404 632
100 96.6 1246 1193 1290 1375
Supplement (Fig. S2 and Table S3). Because of the extremely strong correlation between IRSAD and IRSD score we dropped IRSD score, as IRSAD includes information on both relative advantage and disadvantage, rather than disadvantage only. Fig. 1 shows the results of the variable selection process for socio-economic indexes expressed in terms of their error sum of squares. The best model used the IER score with two degrees of freedom as judged by the error. The results for two and three degrees of freedom were similar, but we used two degrees on the principle of parsimony. Fig. 2 shows the results of variable selection for population variables. Non-Indigenous population density was dropped because it had an extremely strong correlation with total population density. The best population predictor as judged by the error was the percentage of nonIndigenous persons as a linear variable. Similar fits were achieved using the percentage of Indigenous persons. We considered additional models using the remaining socio-economic and population variables, and found that none offered a substantial improvement on the base
To help with model convergence all predictors were standardised.
3. Results Table 1 summarises the descriptive statistics for each variable in the model. Histograms and correlations for the dependent and independent variables are in the
Fig. 1. Boxplots of error sum of squares for selecting the best socioeconomic index. Note: IEO = index of education and occupation; IER = index of economic resources; IRSAD = index of relative socioeconomic advantage and disadvantage.
4
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6 Table 2 Parameter estimates for the best model from the Bayesian meta-analysis. Note: ns (, k) = natural spline, kth parameter. Parameter Intercept ns(IER_score, 1) ns(IER_score, 2) Percent non-Indigenous
Fig. 2. Boxplots of error sum of squares for the best population variable. Note: ATSI = Aboriginal and Torres Strait Islander persons (i.e. Indigenous Australians); PC = percent.
model with IER score as a non-linear variable and the percentage of non-Indigenous persons as a linear variable (Supplement, Fig. S3). Therefore, we concluded that it was the best model (R2 = 6.4%). Fig. 3 gives the final predictions from the best model. It shows a decrease in predicted NO2 levels with increasing IER scores for all but the bottom 5% of SA1s. Predicted NO2 also decreased slightly as the percentage of non-Indigenous persons increased. The caveat to this result is that most areas have a high percentage of non-Indigenous persons with a small number of areas having a low percentage, hence the predictions interpolate an area where there is no data. Parameter estimates are in Table 2 and show that both parameters were statistically significant and the 95% credible intervals do not contain zero. 4. Discussion We assessed if environmental inequalities in ambient air pollution are present in Australia by identifying
Mean
S.D.
Lower
Upper
6.08 0.55 0.83 0.01
0.08 0.10 0.04 0.00
5.93 0.35 0.91 0.01
6.23 0.75 0.76 0.00
z 80.6 5.6 21.1 6.6
associations between socio-economic and population variables and pollutant levels. We found that the best socioeconomic and population predictors of ambient NO2 levels estimated by a satellite-based land-use regression model in urban Australia were the index of economic resources (IER) score and the percentage of non-Indigenous persons. While both were statistically significant they only explained 6.4% of the spatial variation in NO2 levels. However, IER score and the percentage of non-Indigenous persons were clearly the most important socio-economic and population variables associated with ambient NO2. We found that locations with higher IER scores had lower predicted NO2 across the vast majority of urban areas. The IER is based on 14 variables that together can be used to assess the financial aspects of socio-economic advantage and disadvantage, and higher scores occur in less disadvantaged areas (Australian Bureau of Statistics, 2013). A higher proportion of the population that identify as non-Indigenous was associated with slightly reduced NO2 levels. Indigenous Australians fare worse than their non-Indigenous counterparts across many indicators of social and socio-economic determinants of health, and this may also be the case for exposure to ambient NO2 (Australian Institute of Health and Welfare, 2012). Taken together, the results suggest some evidence of environmental inequalities in NO2 levels across urban Australia. However, the magnitude of the difference in NO2 levels was small (0.8 ppb between most and least-
Fig. 3. Predicted NO2 levels from the best model, which used the index of economic resources (IER) score as a non-linear variable and percent nonIndigenous as a linear variable (R2 = 6.4%). Density plots show the distribution of each variable. Note: ATSI = Aboriginal and Torres Strait Islander persons (i.e. Indigenous Australians).
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
exposed IER areas; 0.25 ppb between most and leastexposed percent non-Indigenous areas). While modest increases in air pollution can have relatively large public health impacts due to the large number of people exposed over long periods, the slight differences observed in this study may not translate into substantive differences in health. A recent Australian study identified that the toxicity and volume of industrial pollutant emissions are greater in areas that are more socio-economically disadvantaged and have a higher proportion of Indigenous persons (Chakraborty and Green, 2014). Our study used a different methodology and focused on a different exposure (ambient air pollution), but we also found some evidence to suggest environmental inequalities in pollutants in Australia. The extent to which our results are associated with adverse health effects is not known. Further studies to address these questions using more powerful study designs (e.g. cohort studies) that control for the numerous confounding influences could offer useful insights. Based on initial results that showed spatial correlation in NO2 concentrations (Supplement, Fig. S1; Tables S1 and S2), we used regression models with spatially correlated errors. Other studies have noted the need to account for spatial correlation in environmental inequality analyses to minimise the possibility of biased or misleading results (Yanosky et al., 2008; Havard et al., 2009; Padilla et al., 2014). These findings and ours demonstrate that assessing spatial correlation in environmental inequality studies is necessary, and that results based on the assumption of independent observations should be treated with caution. International studies have consistently identified environmental inequalities in proximity to air pollution sources or pollutant concentrations (e.g. Bell and Ebisu, 2012; Brainard et al., 2002; Chaix et al., 2006; Finkelstein et al., 2005; Grineski et al., 2007; Gunier et al., 2003; Havard et al., 2009; Jerrett, 2009; Marshall, 2008; Morello-Frosch et al., 2002; Padilla et al., 2014; Pearce et al., 2006; Pearce and Kingham, 2008; Perlin et al., 2001). These inequalities and pre-existing differences in risk due to health status have also been linked to greater adverse cardiovascular and birth outcomes among socioeconomically disadvantaged groups or racial minorities (e.g. Clark et al., 2014; Finkelstein et al., 2005; Ponce et al., 2005; Forastiere et al., 2007; Yanosky et al., 2008). Such findings are a concerning environmental manifestation of the inverse care law (Hart, 1971), where those who should ideally be best protected from air pollution are instead most exposed to it. Assessing environmental inequalities in pollutant exposures and health is important for identifying whether policy-level responses are required to protect against them and what the likely benefits of these actions will be. The main limitations of this study require consideration. First, we only examined NO2 and treated it as a proxy for ambient air pollution from vehicle and non-vehicle sources. Other pollutants may show different gradients. Second, there is potential for miscounting of Indigenous and non-Indigenous populations based on self-reporting in the census, although census data are of high quality and used broadly in research. Third, we used general
5
indicators of Indigenous persons (e.g. percent Indigenous and non-Indigenous) and did not have data to examine the associations among other racial groups in Australia. Fourth, we only focused on major urban areas and our study area did not include the 30.9% of the Australian population that live outside these areas. Finally, computational constraints imposed by our spatial model meant that we had to run 10 random selections of size 4000 of our study sample of 35,000. However, the results from each random model were relatively homogeneous. Strengths of this study include: (1) the use of a validated satellite-based LUR model with national coverage to predict NO2 concentrations across Australia (capable of explaining 81% of variation in NO2); (2) adjustment for spatial correlation in NO2 levels, and; (3) the use of a systematic approach and Bayesian meta-analysis to combine parameter estimates and select the best socio-economic and population predictors of NO2. In summary, we found small but statistically significant effects to suggest environmental inequalities in ambient NO2 levels in the major urban areas of Australia. Areas with higher IER scores, which indicate reduced disadvantage, were associated with lower NO2 levels, as were areas with a higher proportion of non-Indigenous persons. These findings provide a basis to pursue more comprehensive assessments of environmental inequality in air pollution across Australia. Acknowledgements Computational resources and services used in this study were provided by the High Performance Computer and Research Support Unit, Queensland University of Technology. L.D.K. is supported by an NHMRC Early Career (Australian Public Health) Fellowship (APP1036620). Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.sste.2015.03.001. References Australian Bureau of Statistics (ABS). Australian Statistical Geography Standard (ASGS): Volume 1 – Main structure and greater capital city statistical areas; 2011. http://www.abs.gov.au/ausstats/[email protected]/mf/ 1270.0.55.001 [accessed 23/12/14]. Australian Bureau of Statistics (ABS). Socio-economic indexes for areas (SEIFA); 2013. http://www.abs.gov.au/ausstats/[email protected]/mf/2033.0. 55.001 [accessed 23/12/14]. Australian Institute of Health and Welfare (AIHW). Australia’s health 2012; 2012. http://www.aihw.gov.au/WorkArea/DownloadAsset. aspx?id=10737422169 [accessed 23/12/14]. Bell ML, Ebisu K. Environmental inequality in exposure to airborne particulate matter components in the United States. Environ Health Perspect 2012;120:1699–704. Bell ML, O’Neill MS, Cifuentes LA, Braga ALF, Green C, Nweke A, et al. Challenges and recommendations for the study of socioeconomic factors and air pollution health effects. Environ Sci Policy 2005;8:525–33. Brainard JS, Jones AP, Bateman IJ, Lovett AA. Modelling environmental equity: access to air quality in Birmingham, England. Environ Plann A 2002;34:695–716.
6
L.D. Knibbs, A.G. Barnett / Spatial and Spatio-temporal Epidemiology 13 (2015) 1–6
Briggs DJ, Collins S, Elliott P, Fischer P, Kingham S, Lebret E, et al. Mapping urban air pollution using GIS: a regression-based approach. Int J Geogr Inf Sci 1997;11:699–718. Brulle RJ, Pellow DN. Environmental justice: human health and environmental inequalities. Annu Rev Public Health 2006;27:103–24. Chaix B, Gustafsson S, Jerrett M, Kristersson H, Lithman T, Boalt A, Merlo J. Children’s exposure to nitrogen dioxide in Sweden: investigation environmental justice in an egalitarian country. J Epidemiol Commun Health 2006;60:234–41. Chakraborty J, Green D. Australia’s first national level quantitative environmental justice assessment of industrial air pollution. Environ Res Lett 2014;9:044010. Clark LP, Millet DB, Marshall JD. National patterns in environmental injustice and inequality: outdoor NO2 air pollution in the United States. PLoS One 2014;9:e94431. Finkelstein MM, Jerrett M, Sears MR. Environmental inequality and circulatory disease mortality gradients. J Epidemiol Commun Health 2005;59:481–7. Forastiere F, Stafoggia M, Tasco C, Picciotto S, Agabiti N, Cesaroni G, et al. Socioeconomic status, particulate air pollution, and daily mortality: differential exposure or differential susceptibility. Am J Ind Med 2007;50:208–16. Grineski S, Bolin B, Boone C. Criteria air pollution and marginalized populations: environmental inequity in Metropolitan Phoenix, Arizona. Soc Sci Q 2007;88:535–54. Gunier RB, Hertz A, Von Behren J, Reynolds P. Traffic density in California: socioeconomic and ethnic differences among potentially exposed children. J Expo Anal Environ Epidemiol 2003;13:240–6. Hart JT. The inverse care law. Lancet 1971;297:405–12. Havard S, Deguen S, Zmirou-Navier D, Schillinger C, Bard D. Traffic-related air pollution and socioeconomic status: a spatial autocorrelation study to assess environmental equity on a small-area scale. Epidemiology 2009;20:223–30. Hystad P, Setton E, Carvantes A, Poplawski K, Deschenes S, Brauer M, et al. Creating national air pollution models for population exposure assessment in Canada. Environ Health Perspect 2011;119:1123–9. Jerrett M. Global geographies of injustice in traffic-related air pollution exposure. Epidemiology 2009;20:231–3. Jerrett M, Arain A, Kanaroglou P, Beckerman B, Potoglou D, Sahsuvaroglu T, et al. A review and evaluation of intraurban air pollution exposure models. J Expo Anal Environ Epidemiol 2005;15:185–204.
Knibbs LD, Hewson MG, Bechle MJ, Marshall JD, Barnett AG. A national satellite-based land-use regression model for air pollution exposure assessment in Australia. Environ Res 2014;135:204–11. Lim SS, Vos T, Flaxman AD, Danaei G, Shibuya K, Adair-Rohani H, et al. A comparative risk assessment of burden of disease and injury attributable to 67 risk factors and risk factor clusters in 21 regions, 1990–2010: a systematic analysis for the Global Burden of Disease Study. Lancet 2012;380:2224–60. Marshall JD. Environmental inequality: air pollution exposures in California’s South Coast Air Basin. Atmos Environ 2008;42:5499–503. Morello-Frosch R, Pastor M, Porras C, Sadd J. Environmental justice and regional inequality in Southern California: implications for future research. Environ Health Perspect 2002;110(Suppl. 2):149–54. Novotny EV, Bechle MJ, Millet DB, Marshall JD. National satellite-based land-use regression: NO2 in the United States. Environ Sci Technol 2011;45:4407–14. Padilla CM, Kihal-Talantikite W, Vieira V, Rossello P, Le Nir G, ZmirouNavier D, et al. Air quality and social deprivation in four French metropolitan areas – a localized spatio-temporal environmental inequality analysis. Environ Res 2014;134:315–24. Pearce J, Kingham S, Zawar-Reza P. Every breath you take? Environmental justice and air pollution in Christchurch, New Zealand. Environ Plann A 2006;38:919–38. Pearce J, Kingham S. Environmental inequalities in New Zealand: a national study of air pollution and environmental justice. Geoforum 2008;39:980–93. Perlin SA, Wong D, Sexton K. Residential proximity to industrial sources of air pollution: interrelationships among race, poverty, and age. J Air Waste Manage Assoc 2001;51:406–21. Ponce NA, Hoggatt KJ, Wilhelm M, Ritz B. Preterm birth: the interaction of traffic-related air pollution with economic hardship in Los Angeles neighborhoods. Am J Epidemiol 2005;162:140–8. United Nations Development Programme. Human development index and its components; 2014. http://hdr.undp.org/en/content/table-1human-development-index-and-its-components [accessed 23/12/ 14]. Yanosky JD, Schwartz J, Suh HH. Associations between measures of socioeconomic position and chronic nitrogen dioxide exposure in Worcester, Massachusetts. J Toxicol Environ Health A 2008;71: 1593–602.
