Emissions vs exposure_ Increasing injustice from road traffic-related air pollution in the United Kingdom

This paper presents unique spatial analyses identifying substantial discrepancies in traffic-related emissions generation and exposure by socioeconomic and demographic groups. It demonstrates a compelling environmental and social injustice narrative with strong policy implications for the


Introduction
Air pollution has been recognised by the World Health Organization (WHO, 2014) as the world's "largest single environmental health risk", with 4.2 million premature deaths in 2016 resulting from exposure to ambient (outdoor) air pollution (WHO, 2018). In highly-populated urban areas, road transport is often a major local source of ambient air pollution, in particular contributing significantly to nitrogen dioxide (NO 2 ) concentrations. In the UK, almost a third of total NO x was generated by road vehicles, with 23% from cars and light duty vehicles (LDVs) (Wakeling et al., 2016). Across European Member States, there are widespread exceedences of the 40 µg/m 3 NO 2 annual mean limit value set by Directive 2008/50/EC on ambient air quality and cleaner air for Europe (the https://doi.org/10.1016/j.trd.2019.05.012 • Secondly, socioeconomic and age-profile data from the UK 2011 Census, contemporaneous with the NO 2 and NO x data, are used to analyse patterns of exposure, whereas the previous study used non-contemporaneous data from the 1991 Census. • Thirdly, Lower-layer Super Output Areas (LSOAs) (∼700 households), rather than wards (∼6600 households), are used to reduce variability in the data and allow for significantly more focused spatial analyses.
• Fourthly, privileged access to DfT's annual vehicle safety inspection ('MOT' test) digital dataset, provided under the EPSRC-funded MOT Project, allows average household NO x emissions for each LSOA to be estimated based on actual mileages from every private vehicle, up to 3.5 tonnes, in the UK (n ≈ 27 million). The ability to calculate actual mileages for each vehicle enables a considerably more robust emissions calculation than in the previous study, which used a Weibull function fitted to MEET (Methodology for Estimating Emissions from Transport) observations (TRL, 1999) to estimate distance driven by age of vehicle.
By using the actual mileages driven by each vehicle, it is possible to explore whether differences in emissions between poor and less poor areas are due to differences in the 'absolute cleanliness' of vehicles (in terms of per km emission factors) or due to differences in actual emissions resulting from how far the vehicles are driven. Linking robust emission estimates to the LSOA of the registered keeper, together with car-ownership and socioeconomic data, enables identification of who (on an areal basis, if not in terms of actual individuals) generate the most traffic emissions, and whether this leads to new patterns of environmental injustices.
In addition to using ambient NO 2 concentrations as the measure of environmental harm, the analysis presented here also uses NO x emissions within an area as a more direct measure of transport impact for the following reasons. Firstly, due to the methods by which the Defra concentration data are modelled, this only provides an indication of 'background' concentrations and is not able to represent where concentrations may be higher due to localised road sources and poor dispersion related to small-scale geographic features (e.g. buildings). Secondly, and more importantly, air pollution is not the only health risk associated with heavily trafficked areas. Increased volumes of traffic are associated with greater levels of noise (Babisch et al, 2005), diminished social integration (Appleyard and Lintell, 1972;Hart and Parkhurst, 2011), lower quality public (especially green) spaces (Woolley, 2004), reduced physical activity (Pikora et al, 2003) and greater risk of road traffic accidents (Dickerson et al., 2000). These risks all increase as levels of traffic increase, irrespective of local pollution dispersion characteristics. The use of NO x emissions from road transport therefore may provide a much better indicator of health risks caused by motor traffic than NO 2 concentrations.

Study area and scale of analysis
This research utilises LSOAs for England and Wales from the 2011 Census (ONS, 2016a). With 34,753 LSOAs (∼700 households (range 400-1200), ∼1600 residents (range 1000-3000) in each), this analysis presents a much finer spatial resolution than the ward analysis (∼8500 in England and Wales) used by Mitchell and Dorling (2003). LSOAs also have a much more consistent population size and, importantly, a greater degree of uniformity in terms of socioeconomic characteristics, thereby limiting the ecological fallacy inherent in area census data analysis (Openshaw, 1984), i.e. the risk of misrepresenting the socioeconomic characteristics of individuals, based on analyses undertaken on the LSOAs in which they reside, is reduced.

Census data
Socioeconomic and demographic data (covering every household in England and Wales) for all LSOAs (ONS, 2016b) were analysed against air pollution data following the methodology set out by Mitchell and Dorling (2003). Quantile plots were constructed using R (v. 3.2.3) (R Core Team, 2015).

Socioeconomic data and analysis
There are many different ways of assessing poverty (e.g. income, deprivation), however their temporal and spatial inconsistencies can prove problematic in undertaking socioeconomic analyses. A number of social equity and air pollution studies and datasets were reviewed to determine the most suitable poverty metric for this study (Hick, 2013;Treanor, 2013;Mitchell and Norman, 2012;Berthoud and Bryan, 2010;Nolan and Whelan, 2010;Gordon, 2006;Walker et al., 2005;Samet and White, 2004;Hillyard et al., 2003;Mack and Lansley, 1985). Initial investigation tested the utility of deprivation and income proxies: the 2010 English Indices of Deprivation (IMD) (Department for Communities and Local Government, 2010) and estimated median household income (Experian, 2011). Although the pattern of the results of these datasets are comparable, the variability (large error margins on the Experian data) and spatial coverage (independent deprivation indices for England and Wales) were not appropriate for this study. Instead the 'percentage of households in poverty' metric was adopted, based on the University of Bristol's Poverty and Social Exclusion (PSE) Unit (http://www.poverty.ac.uk) Breadline Britain Index (BBI), and following Mitchell and Dorling (2003).
The BBI was originally based on analysis of the 1990 Breadline Britain survey data (Gordon and Pantazis, 1997). The percentage of households 'Below the Breadline' (i.e. in poverty) are determined based on census data e.g. housing tenure, employment, health, household composition. The Breadline Britain survey was repeated by PSE in 1999, to reflect changing patterns of poverty, and updated to include new 2001 Census variables. A revised BBI was created for each LSOA by totaling the variables (Dorling et al., 2007).
For this updated analysis, advice was sought from academics involved in the creation of the PSE index 1 on the suitability of applying the 1999 variable coefficients to the 2011 Census data. Although data on 'households without sole use of amenities' was absent from the 2011 Census, and other determinants and symptoms of poverty may have changed, it was judged that this was the best available methodological approach. It is important to note that, as with measures like the IMD, a lower level of poverty (or deprivation) does not necessarily correlate directly with greater affluence or wealth. For example, the fact that an area only has a small number of households that are below the breadline, does not indicate anything about the wealth of the remaining houses (DCLG, 2015).

Car availability analysis
Using R, LSOAs showing 2011 UK Census data on availability of cars and vans were sorted into deciles by car/van availability within the following categories: proportion of households with availability of 0, 1, 2, 3, and 3+ cars/vans.

Pollution data and analysis
Air pollution data were acquired from two sources. Firstly, modelled annual mean UK ambient background data for NO x as NO 2 and PM 10 for 2011 (Defra, 2018), split into attributions to 20 and 23 different source sectors respectively, allowing just the road transport components to be extracted. Secondly, UK National Atmospheric Emissions Inventory (NAEI) (2017) ASCII files of 2011 UK NO x emissions, split into 13 different source categories, from which the road transport component was extracted. Where the term 'total' concentrations/emissions is used this refers to the sum of concentrations/emissions from all sources; 'road NO x /NO 2 /PM' relates to only that portion attributable in the datasets to road transport sources.
The ASCII files of NO x emissions data were converted to point vector data (via raster) using ArcGIS (10.3). Both the NO 2 and NO x datasets, used by Defra to report against the AAQD (2008/50/EC), are available at a 1 km 2 grid resolution. LSOAs vary in spatial area dependent on population density, with more rural LSOAs comprising > 1 km 2 grid cell while more densely-populated LSOAs may cover < 1 km 2 grid cell. Annual mean NO 2 was proportionally averaged across each LSOA, weighted by area, against the LSOA fraction intersecting with each 1 km 2 grid cell, e.g. if quarter of a LSOA fell within a grid cell with 40 µg/m 3 and the other threequarters fell within another with 60 µg/m 3 the NO 2 concentration attributed to the LSOA would be (25% × 40 µg/ m 3 ) + (75% × 60 µg/m 3 ) = 55 µg/m 3 . Conversely, NO x emissions were area-weighted based on the fraction of grid cells covered by each LSOA, e.g. if a LSOA fell across half of a grid cell with 10 t/km 2 and the whole of another with 15 t/km 2 the NO x emission attributed to the LSOA would be (50% × 10 t/km 2 ) + (100% × 15 t/km 2 ) = 20 t/km 2 . Without finer resolution data, with which to reveal the degree of variability within these areas, this approach assumes uniformity of average NO 2 concentrations across each LSOA and NO x emissions over each individual grid cell.

Private vehicle emissions data and analysis
UK vehicles > 3 years old are required to have annual assessments, historically 'MOT' (Ministry Of Transport) tests. In 2010, the DfT started publishing the electronic records from these tests for every vehicle, including make and model, size of engine, type of fuel, registration date and recorded mileage. Date of first registration was used to estimate specific emission factors for each and every vehicle, which were multiplied by the annual distance driven by that vehicle (recorded mileage) to estimate a specific quantity of emissions attributable to every private vehicle (< 3.5 t) in Great Britain. These data were used to generate a profile for an average car within each LSOA and then, using Census data on household car ownership, NO x and PM 10 emissions (total and average) were calculated for all privately-owned vehicles in each LSOA. For a more detailed approach see Chatterton et al. (2015), however, note that subsequent to that account, two significant improvements in the dataset have been made which improve the accuracy of the results. Firstly, vehicles are now located by the LSOA of the registered keeper as opposed to the Postcode Area containing the Vehicle Testing Station. Secondly, it has become possible to account for vehicles < 3 years old (7.8% of private vehicles) in terms of their location and by calculating their annual mileage for 2011, by interpolating from the distance driven between their date of first registration and their recorded mileage at their first inspection test.
It is important to note that, contrary to most air pollution research, the analyses presented in this paper regarding emissions from private vehicles do not consider where these emissions happen. The novelty of this is specifically in the attribution of vehicle emissions to the location of the driver (or more precisely the registered vehicle keeper). A clear distinction is made regarding a) road transport NO x emissions (derived from the NAEI data), which reflect emissions occurring within an LSOA (that pollution to which residents are exposed) and b) NO x emissions from private vehicles registered within an LSOA, but not necessarily occurring within that LSOA (that pollution for which households are responsible). Fig. 1 shows the LSOA analysis of average annual mean NO 2 concentrations by age profile, using 2011 census and pollution data. The x-axis represents age deciles from low to high, i.e. age decile 1 denotes 10% LSOAs with the lowest proportion of an age group, whilst age decile 10 corresponds to 10% LSOAs with the highest proportion of an age group. Fig. 1 is comparable with the earlier Mitchell and Dorling plot, however concentrations of NO 2 have apparently reduced (by ∼10 µg/m 3 ), using the Defra data. (N.B. inaccuracies in these data are argued in the Discussion.) The comparative exposure between age-groupings, however, does not appear to have changed over the last 10 years, with areas with more under-fives and adults aged 20-44 having elevated NO 2 concentrations. On the other hand, areas with more over-45 s tend to have better air quality. Maximum variances between upper and lower age deciles are between ages 25-34 and 55-64 (Fig. 2). Fig. 2 examines the differential between the highest and lowest deciles in each age band in relation to both NO 2 and road NO x in each LSOA. The plots show a strong discrepancy between younger children (< 5 years) and adults (in their 20 s and 30 s) (higher NO 2 and NO x ) compared with older children (10-20) and adults (> 50) (lower NO 2 and NO x ). This pattern may reflect the tendency for income and wealth to increase with age, and the consequent ability of older (more affluent) families to choose areas with better air quality (Mitchell and Dorling, 2003). This new analysis reveals that this discrepancy is increased: NO 2 concentrations in areas with more young adults are two times the average (cf. ∼1.6-1.8 times (Mitchell and Dorling, 2003)) and there is a five-fold difference in NO x emissions (not previously analysed). This is a significant finding, particularly given the argument made above, that NO x emissions may be a better proxy for the range of health risks caused by motor traffic.    Barnes, et al. Transportation Research Part D 73 (2019) 56-66 emissions from private vehicles (left) and, similarly, PM emissions (right) also against households in poverty. Fig. 3 shows exposure of households to air pollution within an LSOA, regardless of the emitter (i.e. whether it is emitted by local residents, or by 'through traffic'). Fig. 4 utilises the MOT dataset and represents total private vehicle emissions per household within an LSOA, regardless of where the emissions are generated (i.e. in that locality, or on longer journeys). The relationship between areas with poorer households and NO 2 concentrations signifies that the more households in poverty in a LSOA, the greater their chances of being exposed to poorer air quality (Fig. 3, left). Indeed, LSOAs with the highest percentage of households in poverty have NO 2 concentrations greater than 50% higher than LSOAs with the least households in poverty. The trend also holds true for PM 10 (Fig. 3, right), although the differential between highest and lowest concentrations is less pronounced than for NO 2 , probably due to the relatively higher levels of regional background PM 10 .

Pollution and poverty
There is a slight upturn in the poorest decile in the left-hand Fig. 3 plot indicating that areas with the fewest households in poverty have slightly increased NO 2 concentrations compared to the next decile. However this upturn only represents < 1 µg/m 3 , much less than earlier analysis (Mitchell and Dorling, 2003), and may reflect variation in pollutant concentrations, changes in the determinants of poverty, or differences in the areal resolution and data sources.
Conversely to Fig. 3, Fig. 4 shows a strong negative correlation between emissions generation and poverty, signifying that households in the poorest areas emit the least NO x and PM, whilst the least poor areas emitted the most. The correlation presented in Fig. 3 indicates significant social inequality concerning exposure, however, Fig. 4 indicates that traffic pollution within those poorer areas is likely to be caused by those living in relatively more affluent areas, further compounding this environmental injustice. Furthermore, the new evidence presented in Fig. 4 contradicts the findings of Mitchell and Dorling (2003) who concluded that the poor were substantially contributing to the emissions to which they were exposed through owning older, more polluting vehicles.

Pollution and car ownership
The plots in Fig. 5 explore the relationship between household access to a vehicle and concentrations of NO 2 and confirm that those areas with households least able to access a vehicle also have the highest pollution concentrations, and conversely those areas with the highest household vehicle access have the lowest concentrations. Furthermore, as illustrated in Fig. 6, households in poverty are much less likely to have any car (left) let alone multiple cars (right). Some caution should be applied here as the PSE poverty calculation uses a proportion (18.4%) of 'households with no car' in its composition, therefore there may be some inherent correlation  J.H. Barnes, et al. Transportation Research Part D 73 (2019) 56-66 of this poverty metric with vehicle accessibility, however the weight applied to car ownership is less than in other deprivation indicators.

Poverty and emissions
Using the 'MOT' dataset, further analysis of household poverty was undertaken against the mean NO x emission factor for vehicles within each LSOA (Fig. 7). Similar analysis was also undertaken for PM and CO 2 (Fig. 8). Apart from a lower mean emission factor in the decile representing the fewest households in poverty, there is a strong negative correlation between poverty status and NO x Fig. 5. NO 2 concentrations against households without access to a car (left) and with access to 3+ cars (right). Error bars = 95% CIs. Fig. 6. Households in poverty against households without access to a car (left) and with access to 3+ cars (right). Error bars = 95% CIs. Fig. 7. Households in poverty against mean NO x emission factor (left) and percentage of private diesel vehicles (right). Error bars = 95% CIs. J.H. Barnes, et al. Transportation Research Part D 73 (2019) 56-66 emission factor (Fig. 7, left), indicating that, not only are households in poorer areas likely to own fewer vehicles, but that those vehicles also tend to produce lower, per km, NO x emissions than the average vehicle owned in less poor areas. This is consistent with the pattern seen in an analysis of PM emission factors (Fig. 8, left) and, to a lesser degree, CO 2 emission factors (Fig. 8, right). For NO x and PM, this relationship appears to be closely related to ownership of diesel vehicles, which have higher NO x and PM emissions factors than petrol vehicles, and which are less likely to be owned in poorer areas (Fig. 7, right). This finding contradicts Mitchell and Dorling's assumption (and perhaps perceived wisdom) that households in poorer areas drive more polluting vehicles. Also, while Fig. 9 (left) shows that households in poorer areas are more likely to have older vehicles, the differential in average vehicle age between the poorest and least poor areas is only 1.2 years. Furthermore, Fig. 9 (right) reveals, perhaps more intuitively, that poorer households tend to drive less far, which, together with the lower mean NO x emission factors (Fig. 7, left) results in lower total vehicle emissions from poorer households (Fig. 4, left). A statistical assessment of the relative importance of emission factor and distance driven on total NO x emissions (using the R package relaimpo (Grömping, 2015)) suggests that variation in distance driven accounts for 84% of variation in overall emissions compared to only 16% for the 'cleanliness' of the vehicle (as indicated by the emission factor).

Discussion
Some 16 years after Mitchell and Dorling's initial work (2003) this study presents a decadal review of progress in equalizing air quality injustices in England and Wales. It has used more recent census and pollution data to investigate the contemporary relevance of the findings of the original study. These new analyses contribute to the current public debate on air quality health and environmental impacts and in so doing challenge previous assertions and assumptions.
The data presented here have identified relationships between car ownership, poverty, age and NO 2 exposure in Wales and England. By utilising emissions of NO x derived from both the NAEI and the MOT dataset, new data have been created that serve to illustrate the relationship between the area of domicile and pollutant emissions generated by residents' vehicle use. In addition, the quality of data presented here at LSOA level is at least four times finer than that available from ward level analysis.
The spatial, demographic and socioeconomic heterogeneity of population exposure to air pollution is well established (Mitchell and Dorling, 2003;Dietz and Atkinson, 2005;Gegisian et al., 2006;Pye et al., 2006;Briggs et al., 2008;Namdeo and Stringer, 2008). Fig. 8. Households in poverty against mean PM emission factor (left) and mean CO 2 emission factor (right). Error bars = 95% CIs. Fig. 9. Households in poverty against the average age of private vehicles per LSOA (left) and total distance driven per LSOA (right). Error bars = 95% CIs.
However, the environmental justice argument has been less clear given the difficulties with identifying responsibility for the creation of pollution. An important outcome of this work is the confirmation of the relationship of LSOAs with higher proportions of young families and households in poverty (those least able to move to cleaner areas) and both high NO x emissions and elevated concentrations of NO 2 . These areas also have a higher prevalence of non-car ownership and a lower prevalence of multiple-car households.
Although absolute concentrations of pollution may have reduced slightly over the decade between studies, the relationship appears to have strengthened with a much clearer correlation between poverty and NO 2 concentrations than previously seen. A more rounded representation of social and health impacts of road transport has been established through using the relationship between road NO x emissions and LSOAs. It is apparent that this relationship is stronger than that established using only NO 2 concentrations. This suggests that inequalities between the poorest and least poor areas may be even greater than first appreciated.
The new 'MOT' data analysis presented here adds a unique dimension to this understanding of environmental justice. By calculating emissions based on actual mileage and emission factors for all of the vehicles in England and Wales individually, this more robust analysis has also found that those LSOAs with more households defined as in poverty also have lower total private vehicle emissions per household (in addition to lower numbers of vehicles per household). This is due to households in poorer areas driving less far overall, as well as having vehicles with lower, per km, mean emission factors. Although households in poorer areas do tend to drive older vehicles, they are, on average, only just over a year older than those in the least poor areas. This may indicate that the least poor, multicar-owning households' second, third or fourth cars are likely to be even older. Furthermore, the greater proportion of diesel vehicles (which emit higher levels of NO x and PM per km) together with the greater mileages driven by vehicles, make households in the least poor areas the greatest contributors to road traffic emissions.
This work has confirmed that environmental injustice is occurring in England and Wales due to air pollution (emissions and concentrations) from road traffic. The research has shown that the most elevated levels of NO 2 and other pollutants occur in LSOAs containing the poorest households. These households contribute significantly less to air pollution than the least poor households. Furthermore, the economic circumstances of those living in the poorest households means that they are less capable of altering their situation, i.e. by moving to a less polluted area or purchasing newer, less polluting vehicles.
One of the limitations of the analytical approach presented here is that differentiation at scales below that of each broadly homogeneous LSOA is not possible and thus variability in the characteristics of a population, or their exposure to pollutants, cannot be represented. Another limitation of the method relates to the time household residents spend within the LSOA. Considerable amounts of time may be spent away from the LSOA attending, for example, work, school or leisure activities yet it is assumed that the concentration within the LSOA represents a good proxy for population exposure. Whilst UK policy (Defra, 2016) acknowledges residential location as a proxy for annual mean exposure it is clear that further work is needed to refine and improve the fraction of exposure associated with different activities. It is also important to note that those most vulnerable (i.e. young children, the elderly and the infirm) are more likely to spend longer at home than other sectors of the population.
The pollutant data utilised in this study are taken from UK national datasets used for compliance monitoring and other purposes. These are the best available datasets within the UK but there are, of course, questions relating to precision and accuracy amongst other features. Ongoing national action seeks to address these deficiencies. One final uncertainty relates to the emission factors used in the calculations where studies (e.g. Hooftman et al. 2018;O'Driscoll et al., 2018;DfT, 2016;Williams et al., 2016) have shown that real world emissions differ considerably from type approval data. These uncertainties and assumptions do not undermine the findings of this paper, although it is acknowledged that further methodological improvements will enhance the quality of outcomes. The spatial analysis of pollutant emissions and concentrations as NO x and NO 2 may well underestimate the actual emissions and concentration data for the most polluted LSOAs as monitoring data do not suggest the same degree of concentration reduction as the modelled data used here. The inaccuracies in emission factors at the scale of individual vehicles are also likely to mask a greater level of inequality, especially with regard to the underperformance of emission control technologies in newer vehicles that are more likely to be owned by households in the least poor areas.

Conclusion
Inequalities in exposure to air pollution across England and Wales have been identified in this work. These variations in exposure are stronger than those identified in 2003 and it can further be concluded that there are significant environmental justice issues with regard to generation of traffic pollution. This inequality in exposure has a further dimension of concern in that the young, who experience the most susceptibility to air pollution, have the least agency in their place of residence. Over the last decade, knowledge and understanding about the health risks from road traffic, particularly for nitrogen dioxide and noise, mean that the known consequences of these inequalities is even greater than when the initial study was undertaken.
This work helps to reinforce the arguments proposed by others (Hull, 2008;Docherty and Mackie, 2010), but not adequately accounted for in the delivery of UK air quality policy. That is, that the underlying air quality problems experienced by the population result from decisions made in land use, economic development and transport planning, that collectively result in very unequal risks of exposure, exacerbated by the poverty status of those in the poorest LSOAs. This inequality has two dimensions, relating firstly to the responsibility for emissions, and secondly to the exposure and vulnerability of citizens ; essentially in respect of traffic emissions the poor pollute the least and are polluted the most. We recommend that governments should take a more social approach to efforts to reduce air pollution as opposed to current approaches that rely predominantly on technological improvements and individual choices.
Future work to further develop the findings of this paper consists of two key streams of research. Firstly, it is intended to extend this analysis to incorporate household emissions from domestic energy consumption. The second, much more challenging stream, is to make links between the attribution of emissions to households presented in this paper, with the locations where these emissions are actually occurring on the road network. This is being done with a range of tools including transport models, Automatic Number Plate Recognition (ANPR) systems and on-board vehicle telematics. However, despite the ongoing research programme, the evidence presented in this paper with regard to the disparities between exposure to and generation of traffic-related pollution, serves to strengthen the environmental justice argument and more clearly lay the problem at the door of households in more affluent areas, who have access to more vehicles with higher average emissions and which are driven further, thereby generating an inequitable contribution to traffic-related pollution. This research has highlighted the environmental justice and social inequality issues of air pollution and it is clear that policies to remediate pollution would benefit by taking greater account of the differences between those who cause the problems and those who bear the costs. It is clear that Governments in the UK and elsewhere are struggling with devising and implementing effective air quality management policies and this work has highlighted an urgent area for policy attention and integration.