ABSTRACT
INTRODUCTION
High-density high-rise cities have become a more prominent feature globally. Air quality is a significant public health risk in many of these cities. There is a need to better understand the extent to which vertical variation in air pollution and population mobility in such cities affect exposure and exposure–response relationships in epidemiological studies.
METHODS
We used a novel strategy to execute a staged model development that incorporated horizontal and vertical pollutant dispersion, building infiltration, and population mobility patterns in estimating traffic-related air pollution (TRAP*) exposures in the Hong Kong Special Administrative Region (HK SAR).
Two street-level spatial monitoring campaigns were undertaken to facilitate the creation of a two-dimensional land-use regression (LUR) model. A network of approximately 100 passive nitric oxide–nitrogen dioxide (NO–NO2) monitors was deployed for two-week periods during the cool and warm seasons. Sampling locations were selected based on population and road network density with a range of physical and geographical characteristics represented. Eight sets of portable monitors for black carbon (BC) and particulate matter ≤2.5 μm in aerodynamic diameter (PM2.5) were rotated so as to be deployed at 80 locations for a 24-hour period. Land-use, geographical, and emissions layers were combined with the spatial monitoring campaign results to create spatiotemporal exposure models.
Vertical air pollution monitoring was carried out at six strategic locations for two weeks in the warm season and two weeks in the cool season. Continuous measurements were carried out at four different heights of a residential building and on both sides of a street canyon. The heights ranged from as close to street level as practically possible up to a maximum of 50 meters (i.e., below the 20th floor). Paired indoor monitoring was included to allow the calculation of infiltration coefficients to feed into the dynamic component of the exposure model.
The final phase of model development addressed population mobility. A population-representative travel behavior survey (n = 89,358) was used to produce the dynamic component of the model, with time-weighted exposure estimates split between home and work or school. Transport microenvironment exposures were taken from published literature. Time–activity exposure estimates were split by age, sex, and employment status.
Development of the exposure model in distinct packages allowed the application of a staged approach to an existing cohort data set. Mortality risk estimates for an elderly cohort of 66,000 Hong Kong residents were calculated using increasing exposure model complexity.
RESULTS
The street-level (2-dimensional [2D]) LUR modeling captured important spatial parameters and represented spatial patterns of air quality in Hong Kong that were consistent with the literature. Higher concentrations of gaseous pollutants were centered in Kowloon and the northern region of Hong Kong Island. PM2.5 and BC predictions exhibited a north–south/west–east gradient, with higher concentrations in the northwest due to regional transport of particulate pollutants from Mainland China. While the degree of explained variance of the models was in line with other LUR modeling efforts in Asia, R2 values ranged from 0.46 (NO2) to 0.59 (PM2.5).
Exponential decay rates (k) were calculated at each monitoring location. While it was clear that k values were higher during the warm season than the cool season, no robust patterns were identified relating to the canyon physical parameters. Therefore, a single decay rate was used for each pollutant across the whole region for derivation of the 3-dimensional (3D) exposure layer (k = 0.004 and 0.012 for PM2.5 and BC, respectively). An alternative decay profile that capped decay at 20 meters above street level was proposed and evaluated. The electrochemical sensors deployed during the canyon campaigns did not exhibit the degree of interunit precision necessary to detect vertical variations in gaseous pollutants, and these results were excluded from the study.
We found that values of the median infiltration efficiencies (Finf) for both BC and PM2.5 were especially high during the cool season (91%). Finf values were somewhat lower during the warm season (81% and 88% for PM2.5 and BC, respectively), and we found a significant negative correlation between air conditioning use and Finf. The Finf for a mechanically ventilated office building was 45% and 40% during the cool and warm seasons, respectively.
Dynamic exposure estimates were compared against home outdoor estimates. As expected, the addition of an indoor component decreased time-weighted exposure estimates, which were balanced out to some extent by the inclusion of transport microenvironments. Overall, mean time-weighted exposures for the full dynamic model were around 20% lower than home outdoor estimates.
Higher levels of exposures were found with working adults and students than for those neither in work nor study. This was due to the increased mobility of people going to work or school. The exposures to PM2.5, BC, and NO2 were, respectively, 13%, 39%, and 14% higher for people who were under age 18, compared with people who were 65 or older. Exposure estimates for the female population were approximately 4% lower.
The availability of an existing cohort data set of elderly Hong Kong residents (n = 66,820) facilitated the calculation and comparison of mortality risk estimates for the different exposure models.
Overall, results indicated that the application of exposure estimates that incorporated infiltration, vertical, and to a lesser extent, dynamic components resulted in higher hazard ratios (HRs) than the standard street-level model and increased the number of significant associations with all-natural-cause, cardiovascular, and respiratory mortality outcomes.
CONCLUSIONS
The results from the study provided the first evidence that considering air pollution exposure in a dynamic 3D landscape would benefit epidemiological studies. Higher HRs and a greater number of significant associations were found between mortality and pollutant exposures that would not have been found had standard 2D exposure models been used. Dynamic models can also identify differential exposures between population subtypes (e.g., students and working adults; those neither in work nor study).
Improved urban building design appears to be stimulating the dispersion of local TRAP in street canyons. Conversely, Finf values found in naturally ventilated buildings were high, and residences provided little protection from ambient air pollution.
We have demonstrated that the creation of effective advanced exposure models is possible in Asian cities without an undue burden on resources. We recommend that vertical exposure patterns be incorporated in future epidemiological studies in high-rise cities where the floor of residence is recorded in health record data.
INTRODUCTION
An increasing proportion of the world’s population lives in densely populated urban landscapes. Further, cities in developing countries are projected to absorb nearly all of the future global population growth (United Nations 2015). Many of the world’s megacities identified as having the most severe air pollution problems are characterized as high-rise cities with large numbers of people living in tall, densely clustered buildings. Despite this, current TRAP exposure estimates are strictly two-dimensional; neither LUR nor dispersion modeling methods are currently able to account for vertical profiles in air pollution and pollutant behavior. Given this limitation, epidemiological studies of TRAP suffer from exposure misclassification, and there is a lack of information on exposures to inform risk assessment for land-use and transportation planning.
Most epidemiological studies of the health effects of TRAP estimate exposure based on 2D residential location and do not consider population mobility or time–activity patterns during the day (Smith et al. 2016). In fact, the very actions that lead to TRAP require population mobility and suggest interactions between movement patterns and air pollution levels (Khreis and Nieuwenhuijsen 2017; Spalt et al. 2016). Recent advances in the use of travel smart cards and data mining approaches have facilitated detailed spatiotemporal analysis of individual travel behaviors, providing a much richer and more precise level of detail than is typically available from routine travel surveys (Jensen et al. 2010). For example, a recent analysis of London mobility patterns, based on the aggregation of individual-level Oyster travel smart card data, indicates dramatic population movement into the city core during morning rush hour; lesser, but highly complex movement within the core during the normal workday; and movement back to the periphery at the end of the workday (Gordon 2012). Similar patterns occur in most large cities and can lead to exposure misclassification if not considered (Nyhan et al. 2016; Smith et al. 2016). Simulations based on a travel survey and a regional air quality model (which would not incorporate TRAP exposure gradients) suggest that the incorporation of mobility can affect exposure estimates by as much as 30% (Marshall et al. 2006). Further simulations, including those applied to epidemiological effect estimates derived from LUR models that incorporate mobility, indicate a bias of effect estimates toward the null when mobility is not considered (Setton et al. 2008).
Thus, the development of exposure models that include 3D variability in TRAP and that incorporate population mobility could dramatically reduce exposure misclassification. Further, such models would allow for scenario analysis to assess the impact of changes in transportation patterns and land use on exposure.
A number of published studies have investigated pollutant behavior within street canyons in Asian cities. Chan and Kwok (2000) investigated vertical dispersion of particulate matter in an open street and a street canyon in Hong Kong. An exponential decay function was used to describe vertical dispersion in PM. Vardoulakis and colleagues (2002) also proposed an exponential decay to describe variation in gaseous pollutants within a street canyon in Paris. Li and colleagues (2007) carried out monitoring at four heights within a canyon in Shanghai, China, to study particle size distributions, which they found varied significantly with height. The investigators in these studies were primarily interested in pollutant behavior and did not attempt to apply their results to health studies or population exposure estimates. Wu and colleagues (2014) investigated the impact of residential height above street level on population exposure in Boston, Massachusetts, U.S.A., downwind of a highway using a mobile monitoring platform and hoist. They found very little variation in PM2.5 concentration with height.
As high-density, high-rise cities become increasingly common in Asia, it is important to understand how modeling spatial variation and exposure may differ from European and North American cities where most LUR modeling has been focused (Hoek et al. 2008). European and North American cities have lower pollution and building densities and are likely to have fewer small-scale dispersed pollution sources than the high-density, high-rise cities (Cohen 2004).
Hong Kong, a coastal city in southern China, is one of the most advanced examples of a high-density, high-rise city with significant air quality issues. Being one of the most densely populated regions in the world, Hong Kong has an average population density of 6,690 people/km2 with a total population of 7,240,000 as of mid-2014 (Government of Hong Kong Information Services Department 2015). Because of the clustering of developments and mountainous terrain, less than 25% of the total territory of 1,104 km2 is developed, leading to extremely high population densities in some areas (Government of Hong Kong 2015; Government of Hong Kong Information Services Department 2015). The clustering effect is further enhanced by the prevalence of high-rise buildings in Hong Kong.
Because of a well-developed network of vehicle flow and pollution monitoring sites, an established public transport system used by 98% of the population, and a current government administration keen on supporting research into their air quality issues, Hong Kong represents an ideal development site for TRAP modeling in high-density, high-rise Asian cities. Such methodology can then be used to inform future modeling in other Asian cities whose information networks are not as well developed.
The study benefitted from an existing cohort of 66,000 elderly residents, with residential floor number recorded as part of the address and a detailed health record database (Schooling et al. 2016). This cohort allowed the study methodology to be evaluated.
The study was a multidisciplinary collaboration between research groups in Hong Kong, the United Kingdom, and Canada with international reputations for air pollution exposure and health assessment research. The team had demonstrable knowledge and experience in each of the study’s main themes — LUR modeling, TRAP monitoring and characterization, population mobility, statistical evaluation for population studies, and high-density urban landscapes.
DISCUSSION AND CONCLUSIONS
The Hong Kong D3D study had the overarching aim of creating and evaluating an advanced TRAP exposure model methodology that incorporated population mobility and residential height above street level, using Hong Kong, a densely populated Asian city, as a case study. Our principal hypothesis was that the inclusion of dynamic and vertical components in TRAP exposure models applied to Asian cities would lead to increased confidence in associated health outcomes.
The study had four main components: (1) the creation of a street-level LUR model for the Hong Kong region through an extensive seasonal sampling campaign; (2) the derivation of a canyon typology and associated vertical decay rate, through multiheight paired in–out sampling campaigns in several locations representative of population exposure; (3) creation of a series of dynamic model components incorporating infiltration efficiencies and population mobility utilizing an extensive travel behavior survey; and (4) application of a staged modeling approach to an existing cohort to compare mortality risk estimates.
In achieving this aim, several practical and methodological obstacles had to be overcome. On the whole, the original study plan was adhered to, but a few adjustments had to be made, most notably, the exclusion of gaseous pollutants from the vertical sampling campaign results because of sensor performance issues, and the simplification of canyon typology because there was no robust identifiable pattern in canyon TRAP dispersion patterns. Conversely, we achieved more than we anticipated in the creation of dynamic modeling components and the application to the existing cohort.
IMPLICATIONS FOR FUTURE USE OF LOW-COST AIR QUALITY SENSORS
The development of a new generation of relatively low-cost air pollution sensors has generated a great deal of interest, both in the research community and in public interest groups (U.S. Environmental Protection Agency 2016). This study used relatively high-cost (~4,000 USD per unit) active samplers for PM2.5, BC, and gaseous pollutants. While these samplers have advantages over passive samplers, we identified several major shortcomings that should be considered by others designing spatial sampling campaigns in Asian cities and elsewhere across the world. First, harsh sampling conditions — variously, high temperatures, intense rainfall, wind storms, high humidity, and high particulate levels, which are typical in tropical and subtropical climates — take their toll on sensitive electronics. Every active sampling unit we deployed required maintenance at least once during our campaigns, and several back-up units were required while repairs were carried out. Second, active samplers are more visible, heavier, and more expensive to replace than passive samplers. Safety and security are therefore major concerns, both to personnel and equipment. Our spatial sampling campaigns relied on collaboration with the Hong Kong EPD. Hong Kong is widely considered a very secure city, and our only losses were some passive samplers removed by concerned locals. No active samplers could be hung outside of buildings, necessitating the use of sampling tubing (for pumped PM samplers) and a manifold (for electrochemical samplers). Third, a high degree of interunit precision is necessary when deploying samplers in networks to detect spatial and vertical variations in TRAP. The development of refined methods of data scaling and ratification was required to achieve the necessary precision in each of the PM units. Such precision could not be achieved with the electrochemical gas sensors, which did not respond well to being regularly moved, and these data had to be excluded from further analysis.
In our experience, very careful experimental design is required if the low-cost electrochemical sensors currently available are to be used effectively in spatial exposure measurement campaigns where accurate representation of within-neighborhood variations is required. This is primarily due to issues with unstable baseline measurements creating bias of a magnitude greater than the spatial variation being investigated. While all instruments suffer from some degree of bias, this has been well characterized in more established monitors, and robust, demonstrably consistent methods for correction can be developed. The strong influence of a range of factors (including temperature, humidity, cross-gas interference, and signal noise), which combine to produce a complex pattern of interference in real-world conditions, makes consistent correction methods challenging to develop. Until such correction methods can be documented and demonstrated, the use of such electrochemical sensors has to be questioned closely before being incorporated into future studies.
DEVELOPMENT OF A 2D LAND-USE REGRESSION MODEL FOR HONG KONG
The street-level (2D) LUR modeling captured important spatial parameters and represented spatial patterns of air quality in Hong Kong that were consistent with the literature (Chiu and Lok 2011; Kok et al. 1997; Shi et al. 2016; Yu et al. 2004). Higher concentrations of gaseous pollutants were centered in Kowloon and the northern region of Hong Kong Island, consistent with the importance of motor vehicle traffic as a dominant source of local NO and NO2 (Tian et al. 2011). PM2.5 and BC predictions exhibited a north–south/west–east gradient, with higher concentrations in the northwest. This appears to be due to regional transport from Mainland China. A similar gradient in PM10 concentrations, noted previously in an analysis of the rooftop AQMS, was attributed to transport from Mainland China (Chiu and Lok 2011). Further, Kok and colleagues (1997) reported elevated BC concentrations in the western regions of the territory and similarly attributed these higher levels to regional sources. A recent LUR model of PM2.5 restricted to downtown Hong Kong indicated spatial patterns similar to those of the HK 2D model (Shi et al. 2016). For BC, the port was also an area of elevated predicted values. Yu and colleagues (2004) had noted the port as an important emission source affecting spatial distribution of BC levels with increases in background BC concentrations around the port, depending on the seasonal direction of the prevailing wind. Shipping lane variables were not, however, present in any of the final exposure models in the current study.
While the degree of explained variance of the models was modest, they were within the range seen with other LUR modeling efforts (e.g., Hoek et al. 2008). Given the complex urban morphology of Hong Kong compared with most of the European and North American cities, a somewhat reduced explained variance was expected. Compared with LUR models developed in other Asian cities where urban morphology, vehicle use, and building design may be similar, LUR models that were developed with dedicated sampling campaigns reported similar R2 values (shown in Appendix Table A.3, available on the HEI website).
VERTICAL DECAY OF TRAP AND DERIVATION OF A CANYON TYPOLOGY
Our canyon sampling campaign was designed to capture variations in TRAP within canyons by height, time of day, and wind conditions. We selected canyons that had a range of aspect ratios, building configurations, and alignment to the prevailing wind. In practice, measurements did not show sufficiently consistent patterns in vertical pollutant concentrations to isolate the impact of many of these variables. We found no evidence of strong TRAP stagnation in typical weather conditions, even in high AADT narrow canyons such as NPC1. TRAP either was fully mixed within the canyon (decay rate approximately zero) or decayed rapidly over the lower floors.
It is important to note that our objective was to characterize population-level vertical exposure, not to produce a detailed explanation of emission dispersion patterns, such as that created by fluid dynamics models. While complex modeled eddies of the type identifiable through computational fluid dynamics are important for urban design, our results suggest that they are of less importance in urban canyons of the type typical in Hong Kong when assessing population exposure levels. Indeed, it may be that we found no evidence of stagnation because of improvements in urban street canyon design over the past 50 years. In Hong Kong, continuous slab-type buildings have largely been replaced by individual residential towers, many laid out in gridded estates with little or no vehicle traffic allowed within. New slab constructions are built at an angle, and many are pierced with voids at higher levels; both designs are intended to increase air flow. While these construction methods have had the primary aim of reducing heat stress (Deng et al. 2016; Ng 2009), they have also had the beneficial effect of increasing the dispersion of TRAP and other air pollution sources, such as cooking.
Our original intention was to produce a coupled street canyon typology with specific TRAP vertical decay rates for each canyon type. By creating such a classification system, street-level exposure estimates within high-density urban landscapes could be scaled vertically according to basic canyon geometry. Our results led us to the conclusion that derivation of such a classified system was not possible, and we assumed a single mean decay rate across the region. Widespread vertical sampling campaigns are challenging to execute, requiring large resources to identify suitable locations, recruit building and flat owners, and deploy and manage samplers. In our view, the addition of sampling in additional canyons would not have produced a substantially different result. However, uncertainties remain, and improvements could be made to subsequent vertical sampling campaigns. The addition of street-level sampling in all canyons would remove the necessity for temporally corrected model estimates, but would introduce additional practical considerations. Time-resolved measurements were not as informative as expected, and paired windward and leeward sampling was difficult to interpret.
An additional limitation in the vertical decay calculation was the separation of local and regional pollution components. Assuming that regional pollutant sources are well dispersed across the urban area, vertical decay occurs only in the local pollutant component. Thus, an improved version of Equation 2 would be:
where Cr is the regional component of concentration C. This approach would bring the k (decay rate) for BC and PM2.5 much closer together; BC being a primary component of local PM2.5. In this study, this improvement could have been achieved through a rooftop sampling campaign mirroring that of the street-level campaign. Emissions-based modeling methods would be able to make this separation more easily than the empirical LUR method that we employed. An attempt was made to incorporate this alternative decay pattern by capping decay to a height of 20 meters, but further measurements would be required to confirm this height estimate.
We found seasonal PM2.5 decay rates (k factors) of 0.009 and 0.001 for the warm and cool seasons, respectively, highlighting the dominance of regional PM2.5 sources during the cool season (see Table 7). For BC, the seasonal difference was 0.016 and 0.009. The single mean decay rates across the region for PM2.5 and BC were 0.004 and 0.012, respectively. Direct comparisons with these decay rates are difficult because of varying methodology; however, Chan and Kwok (2000) reported a decay rate for PM10 in Hong Kong of 0.017 from a campaign in the cool season. In common with Wu and colleagues (2002), we found little diurnal variation in decay rates.
BUILDING INFILTRATION EFFICIENCIES
A major advantage of the vertical monitoring campaign design was that paired in–out monitoring could be added relatively simply, allowing the calculation of infiltration efficiencies for the Hong Kong housing stock. We found that median Finf values for both BC and PM2.5 were especially high during the cool season (91%), indicating that residents were breathing only slightly lower levels of these pollutants indoors than was measured in ambient air. Median infiltration efficiencies were somewhat lower during the warm season (81% and 88% for PM2.5 and BC, respectively), and we found a significant negative correlation between air conditioning use and infiltration efficiencies of PM2.5 and BC. The MESA-Air study reported a median infiltration efficiency for PM2.5 across seven urban communities in North America of 62%, although the median for New York was 82% and was therefore similar to what we found in Hong Kong (Allen et al. 2012).
During the cool season, when PM2.5 concentrations are typically far higher in Hong Kong, residents were more likely to open their windows, leading to a greater infiltration of outdoor air. Therefore, higher ambient concentrations and higher infiltration efficiencies acted together to increase population exposure.
Infiltration efficiencies for the mechanically ventilated office building were 45% and 40% during the cool and warm seasons, respectively. While we only measured infiltration efficiencies in one such building, this is similar to those reported in other studies for occupied HVAC buildings (Chatoutsidou et al. 2015; Fisk et al. 2000). Only a very small proportion of high-value residences have mechanical ventilation, so few benefit from this protection. This finding has important socioeconomic implications for developing subtropical cities: those who can afford higher-specification homes are also more likely to have office jobs in similarly protected buildings. Conversely, these buildings have higher power requirements than naturally ventilated buildings and in many cases will contribute further to regional sources of PM2.5 through fossil-fuel–based electricity generation.
DYNAMIC EXPOSURE MODELING
A population-representative travel behavior survey (n = 89,358) was used to further extend our exposure model to create a dynamic model comprising population mobility and derivation of time-weighted exposure estimates in different microenvironments. A staged approach was used to investigate the impact of each component on exposure estimates for the survey respondents. The vertical component of the model was not included as the survey did not record floor of residence.
Comparisons were made against the static outdoor exposure estimates. As expected, the addition of an indoor component decreased time-weighted exposure estimates, which were balanced out to some extent by the inclusion of transport microenvironments. Overall, mean time-weighted exposures for the full dynamic model were around 20% lower than the static outdoor estimates. The inclusion of diurnal factors had a greater impact on BC and NO2 exposure estimates than on PM2.5, because BC and NO2 tend to vary more during the day than PM2.5.
Smith and colleagues (2016) combined a nested dispersion modeling technique with building infiltration factors and travel behavior to create a dynamic exposure model for London. They found that the dynamic model produced estimated exposures 37% lower for PM2.5 and 63% lower for NO2 than the static ambient model. This difference is likely to be driven by the much lower mean infiltration efficiencies used for London (31% and 56% for NO2 and PM2.5, respectively).
If these differences were equally distributed across the population, then their inclusion would have little impact on health outcome analyses. A stratified analysis of population subgroups was carried out to test the hypothesis that the dynamic model increased variation in exposure estimates.
The stratified analysis confirmed this hypothesis. Higher levels of exposures were found with working adults and students than for those neither in work nor study. This was due to increased mobility, despite relatively low concentrations in office locations, particularly in BC estimates. The results consistently found higher exposures with persons under age 18, compared with other age groups. The exposures to PM2.5, BC, and NO2 were respectively 13%, 39%, and 14% higher for the under-age-18 population compared with the population of people who were age 65 or older. One explanation for this is that most students’ schools were located within the same TPU, and many commuted to school by walking. This pattern of increased exposure with longer travel time has been described by others in exposure monitoring studies (Chau et al. 2002; de Nazelle et al. 2013), and has been suggested to partially offset the physical activity benefits of walking (Hankey et al. 2012). We also assumed natural ventilation in schools, with higher infiltration rates than in office buildings. Spatial contrasts were amplified when accounting for diurnal variation in pollutants, as most subjects traveled during morning and evening rush-hour periods, indicating that population mobility is an important consideration beyond that of transport microenvironment effects.
We found the addition of additional exposure model components increased the gap between male and female exposures, with the female population having lower exposures to air pollution by approximately 4%. A study in Vancouver (Setton et al. 2010), which examined only the working population, found no significant difference in exposure by sex. However, a higher than 50% proportion of women in our survey data were in the nonworking category, which is likely to account for the different finding.
Many of our model results reflect those of Chau and colleagues (2002), who used portable samplers to examine exposure to PM10, NO2, and CO for different age groups in 20 different microenvironments in Hong Kong. They found that the Hong Kong population spent around 86% of the time indoors and around 8% (two hours) commuting. However, higher pollutant exposures were experienced during commuting, so commuting contributed a disproportionately high amount to the 24-hour average, particularly for NO2. They found high concentrations at restaurants, bars, and transport microenvironments, but low concentrations in offices. For both PM10 and NO2, concentrations in offices were much lower than in residential buildings. Concentrations monitored at schools were around four times those in offices. Out of 400 subjects sampled, the under-18 age group was found to have the highest exposures.
A key uncertainty in our transport microenvironment component arose from a lack of contemporary factors from recent monitoring surveys for all pollutants and modes of transport. Such surveys could be considered as part of the sampling requirements for development of similar dynamic models if published factors are not available.
STAGED EPIDEMIOLOGICAL ANALYSIS OF MORTALITY RATES IN AN ELDERLY COHORT
The availability of an existing cohort data set of elderly Hong Kong residents (n = 66,820) facilitated the calculation and comparison of mortality risk estimates for the different exposure models. We further incorporated results from an earlier study on the same cohort but used satellite-derived exposure estimates.
Overall, the results indicated that the addition of a vertical component to the exposure model modified the associations between long-term exposure to air pollution and mortality. The application of exposure estimates that incorporated infiltration, vertical, and, to a lesser extent, dynamic components produced narrower confidence intervals and increased the number of significant associations with all-natural-cause, cardiovascular disease, and respiratory disease mortality outcomes.
When considering only 2D exposure, PM2.5 was significantly associated with elevated risks of mortality only from all-natural and cardiovascular causes. When 3D and D3D exposure models were used, associations increased, and narrower CIs led to additional significant associations with cardiovascular subgroups and respiratory causes. Similarly, no significant associations were found for NO2 using the 2D model, but mortality from all natural causes, cardiovascular disease, and IHD became significant in the 3D model. Very little difference in associations was detected between the 3D and D3D model. This is because the population mobility of the elderly cohort was relatively modest since the cohort included only persons age 65 and older, producing little variation in exposure estimates.
The one anomaly occurred between BC and external causes of death where an unexplained significant association was reported. This may be a result of left-truncated exposure distribution or it may be due to high variability of estimates around a low mean. As expected, particulate pollutants displayed more health effects than did gaseous pollutants.
While air pollution is known to vary vertically and in street canyons (Berkowicz et al. 1996; Meroney et al. 1996; Vardoulakis et al. 2003), as well as between outdoor, indoor (Allen et al. 2012; Hystad et al. 2009), and transport microenvironments (Adams et al. 2001; Kaur and Nieuwenhuijsen 2009), few studies have taken these factors into consideration when investigating the long-term health effects of air pollution. This issue becomes more important in high-rise urban areas where activities take place in high buildings. Studies have included variables representing the 3D landscape in LUR models to improve street-level exposure estimates (Eeftens et al. 2013; Su et al. 2008; Tang et al. 2013). Wong and colleagues (2016) applied horizontal–vertical PM2.5 exposure estimates to assess cancer mortality by geocoding the vertical height of addresses, but they did not assess the epidemiological effects between the use of 2D and 3D exposures.
The range of associations was coherent with other cohort studies that looked at long-term exposure to air pollution and mortality. Associations were more pronounced with cardiovascular mortality, which is in common with findings in other study areas. The ESCAPE study (Beelen et al. 2014) found similar associations. HRs for 10-μg/m3 increases in PM2.5 were 1.15 (95% CI: 1.13–1.16) for all-natural-cause mortality and 1.31 (95% CI: 1.27–1.35) for IHD, consistent with our finding.
The principal limitation of this evaluation phase of the study was the use of a cohort not representative of the whole population. This meant that the dynamic components could not be fully tested. This presents a challenge as most health outcome data sets are age-biased in some way. Conversely, most published studies on dynamic exposure to air pollutants have been based on personal monitoring studies (e.g., Özkaynak et al. 2013; Steinle et al. 2013), which cannot easily be applied to cohorts to represent mobility and exposure patterns for the general population. Effort is required to bring together population-representative dynamic exposure methods with epidemiological data sets that have heterogeneous mobility patterns in order to fully test the impact of this component of exposure. We found that contrasts in exposure between dynamic and static models were greatest in pollutants with relatively high spatial variability; BC, NO, and NO2 are more influenced by traffic and other local emissions sources than is PM2.5. This conclusion was consistent with Setton and colleagues (2011), who compared mobility- and residential-based exposures of NO2 in Canada.
We also used modeled exposures that were back-extrapolated, for a comparatively long period, to match cohort data. However, previous studies have found high correlations among annual air pollution concentrations, even over a period of more than 10 years (Beelen et al. 2008), as major roads that influence air pollution exposure are likely to have been in place for the entire period.
IMPACT OF MODEL ASSUMPTIONS
As with all modeling exercises, we had to make a number of assumptions and simplifications in the development of the advanced exposure model. We were able to carry out a limited evaluation of some of these assumptions by utilizing street-level and rooftop monitoring results at five independent locations. These evaluations showed that while the model produced good results in some conditions, it performed less well in others. This is not a surprise; our aim was to develop a mean exposure model that could be applied in epidemiological studies to estimate long-term health effects.
While the model could be improved in a number of areas, the advantage of the resulting simple methodology proposed is that it could be applied within any urban area with a significant proportion of the population living above street level. With regard to the vertical component of the model, a decay profile could be applied to any street-level model, whether it is produced by LUR, dispersion modeling, or hybrid methods. Indeed, it would be straightforward to add a vertical component to appropriate existing epidemiological studies in cities other than Hong Kong where floor of residence is known in order to explore whether our findings are repeated.
APPLICATION OF TRAP EXPOSURE METHODOLOGY IN OTHER HIGH-RISE ASIAN CITIES
We aimed to create an incremental exposure assessment methodology without onerous demands on input data that could therefore be applied to other megacities across Asia and the developing world. Such input data could be specific to the study, gathered through monitoring campaigns, or readily available from previous comparable studies or accessible government data sets.
The demonstration of reasonable LUR models to describe spatial variability in pollutant concentrations in Hong Kong suggests this to be a viable modeling method for high-density, high-rise cities, which are especially common in Asia. Further, these results suggest the utility of model development using traditional sampling methods and relatively low-labor, low-data-intensive predictors. More complex urban development predictors, such as aspect ratio, that one might expect to be important in modeling a high-density, high-rise city, were not present in any of the final exposure models. This suggests that the added complexity in the spatial distribution of air pollutants in high-density, high-rise cities is reflected in the models’ performance rather than in the selection of predictors.
Our vertical monitoring campaign was intensive and resource heavy, but the results suggest that such a comprehensive campaign is not necessary to derive broadly applicable results in other cities. However, as described earlier in this section, there are significant practical challenges in carrying out vertical campaigns, particularly in cities less secure and collaborative than Hong Kong. Where building configurations are similar to those in Hong Kong, such as in many Mainland Chinese cities, stock decay factors may be used, although care must be taken when considering pollutants with large regional components, such as PM2.5. Our results suggest that severe stagnation, which would strongly affect decay rates, does not occur except in exceptional circumstances, either geographical or meteorological.
Our study benefited from a large travel behavior survey data set with population-representative sampling. Such large surveys are unlikely to be available in many Asian cities, but the importance of urban transport planning in densely populated cities is sufficiently high to make it likely that some form of survey exists in most. Several recent studies have demonstrated the use of mobile phone locational data in assessing population mobility for dynamic air pollution exposure assessments (de Nazelle et al. 2013; Dewulf et al. 2016; Nyhan et al. 2016). While these methods generate very large volumes of data, they have little context, requiring further assumptions about sex, age, and purpose of travel; therefore, they do not necessarily present a direct alternative to travel surveys. Where exposure estimates are applied to an epidemiological study, the cohort demographic may dictate the necessity of a dynamic component; we demonstrated that incorporation of the dynamic component did not improve associations in our elderly cohort.
There are distinct differences in terms of exposure to consider in Asian cities versus European and North American cities that can be capitalized upon to advance understanding of the health impacts of TRAP. These include relatively high infiltration efficiencies and population density, homogenous ethnicity, cohabiting extended families, greatly contrasting seasonal exposure levels, and the often unexplored potential for large-scale interventions. These opportunities make barriers such as data availability, quality, and access, unregulated emissions, and sometimes extreme occupational exposures worth challenging.
IMPLICATIONS OF FINDINGS
To date this is the first comprehensive study to investigate the health effects of traffic-related air pollution using detailed vertical and dynamic air pollution exposure assessment techniques. The results from the study provided evidence that considering air pollution exposure in a dynamic 3D landscape would benefit epidemiological studies. Significant associations were found between mortality and air pollution that would not have been found had standard 2D LUR or satellite exposure models been used.
We also identified differential exposures between population subtypes that would not be present in static exposure models, including higher exposures for those under the age of 18 and marginally higher exposures for male subjects. As more studies incorporate population mobility, such contrasts will become better defined, leading to increased variation in estimates across a population and between pollutants (Smith et al. 2016).
Improved urban building design appears to be stimulating the dispersion of local TRAP emissions in street canyons, including broken canyons, tower estates, and angled building layout. Importantly, we found no clear evidence of stagnation reaching the upper floors of buildings. The practice of setting aside lower floors of residential towers for commercial and leisure use means that most of the Hong Kong population is not exposed to undispersed TRAP emissions in their homes, where they spend the majority of their time.
Conversely, infiltration factors found in homes were close to 1, and residences provided little protection from ambient air pollution. This is particularly critical when considering regional pollutants, such as PM2.5, where height above street level makes little difference. There are also socioeconomic implications of this finding; those residents who can afford to live in mechanically ventilated buildings will have nearly half the exposure of those who cannot.
One of our stated aims was to create an incremental exposure assessment methodology that balanced exposure error with input data availability and that would be applicable to other megacities across Asia and the developing world. While there are several uncertainties associated with this study that could be improved in later iterations, we have demonstrated that the creation of effective advanced exposure models is possible in Asian cities without undue burden on resources.
A number of assumptions and simplifications had to be made in developing our dynamic three-dimensional exposure model. Yet it is intuitive that street-level TRAP exposure estimates will overestimate exposure for residents living in adjacent high-rise buildings. Our results provide evidence that this misclassification leads to a lower association between mortality and air pollution exposure. However, for vertical exposure patterns to be taken into consideration for epidemiological studies, the floor of residence must be recorded in health record data. While this requirement is likely to be difficult for the total population in most countries, it should be feasible within cohort studies. We recommend that the floor of residence be routinely recorded as part of basic participant personal details at recruitment and follow up.
ABOUT THE AUTHORS
Benjamin Barratt has a Ph.D. from King’s College London and an M.Sc. from the University of Portsmouth. He is a senior lecturer in the Division of Analytical and Environmental Sciences & Lau China Institute, King’s College London, U.K., and an honorary assistant professor, University of Hong Kong, HK SAR. Barratt has 15 years of experience conducting multidisciplinary research contracts within both academia and the public sector. He has specialized in characterizing pollutant sources and behavior using dedicated monitoring networks for the evaluation of air quality interventions, including the London Congestion Charging Scheme, Low Emission Zone. He has coauthored three HEI Research Reports, each examining population exposure in relation to TRAP. His current research focuses on the use of personal exposure monitoring to establish direct links between environmental stress and health outcomes. He leads the Exposure Science team within the Environmental Research Group and is an investigator within the MRC-PHE Centre for Environment and Health (www.environment-health.ac.uk). Barratt was principal investigator on the study and lead investigator for WP2.
Martha Lee has an M.Sc. from the University of British Columbia, Canada, where she is a research assistant. Her research interests are in outdoor air quality exposure and its associated adverse health effects. Lee was responsible for the fieldwork and development of LUR in WP1, which was her thesis project.
Paulina Wong has a Ph.D. from the University of Hong Kong and an M.Sc. from the University of Auckland. She is an assistant professor in the Science Unit at Lingnan University, Hong Kong, and was formerly senior research assistant in the Department of Geography, University of Hong Kong, HK SAR. Wong specializes in urban climate research with a focus on urban heat island effects and ecological modeling. She is proficient in geographic data processing, statistical analysis, and microclimate field monitoring and surveys. She has been a Fulbright scholar and is a certified geographical information system (GIS) professional. Her recent research addresses environmental impacts and their effects on health-related problems in Hong Kong. She has made use of mobile geospatial technologies and cloud computing strategies to assist the process of georeferencing. Wong provided GIS expertise, supervised the canyon fieldwork campaigns, and carried out analysis of vertical decay rates in WP2.
Robert Tang has a Ph.D. and an M.Sc. from Imperial College London. He is a postdoctoral fellow at the School of Public Health, University of Hong Kong, HK SAR. While completing his Ph.D. at the MRC-PHE Centre for Environment and Health, Tang developed spatiotemporal air pollution models for exposure assessment using LUR, dispersion, hybrid techniques in GIS and Bayesian statistical framework. He was a research assistant for the EU FP7 EXPOsOMICS project, investigating the health effects of fine particulate matter. His current research is focused on air and noise pollution modeling; exposure assessments for epidemiological studies; and spatial analysis and long-term effects of air pollutants. Tang was lead investigator for WP3 and WP4.
Tsz Him Tsui has an M.P.H. from the Chinese University of Hong Kong and was formerly a research assistant at the School of Public Health, University of Hong Kong, HK SAR. Tsui specialized in population and global health during his M.P.H. studies. He led the canyon campaign fieldwork and contributed to the development of the dynamic model in WP3.
Wei Cheng has an M.Sc. from East China Normal University and is a Ph.D. candidate in the Department of Geography, University of Hong Kong, HK SAR. She is a geographer by training and specializes in cartography and geographic information science. She has done research on spatial analysis of crime in Shanghai and using remote sensing techniques to map vegetation and investigate evolution of urban river systems. She is pursuing a Ph.D. degree to investigate patterns of air pollutant dispersion in urban street canyons and visualization methods that can effectively display spatiotemporal movement of air pollutants within a 3D cityscape. Cheng developed the 3D visualizations in WP4.
Yang Yang holds a Master’s degree in biostatistics and epidemiology from Peking Union Medical College, which focused on the associations between air pollution in Beijing and related health effects. She also carried out data analysis for the project “Traffic related air pollution and the health impacts in Three Typical Chinese Cites.” Yang is an M.Phil. student at the School of Public Health, University of Hong Kong, HK SAR. She ran analyses relating to WP4.
Poh-Chin Lai has a Ph.D. and M.A. from the University of Waterloo. She is a professor in the Department of Geography and Honorary Deputy Director of the Geographical/Land Information System Research Centre, University of Hong Kong, HK SAR. Lai is an expert in the use of GIS information to identify associations between physical urban geography and environmental health risks. She has collaborated with international and local medical and public health professionals to develop and apply geospatial methodologies to understand spatial variation of health outcomes and health inequality. She has recently completed a study investigating relationships between incidences of tuberculosis and high-density urban residential characteristics, such as type of housing development, floor level of residence, and sky view factor. Previous studies have included the novel use of GIS systems to associate environmental factors with falls of older people and analyzing microclimate variation along a marathon course. These studies aim to examine effects of urban morphology on public health and the wellbeing of the people of Hong Kong. Lai provided local oversight and guidance for WP2.
Linwei Tian has a Ph.D. from the University of California–Berkeley, an M.Sc. from China CDC, and a M.B.B.S. from Shanxi Medical University, China. He is an associate professor at the School of Public Health, University of Hong Kong, HK SAR. Tian is an environmental epidemiologist with a focus on air pollution and health. He has been conducting field epidemiology and laboratory work on indoor air pollution and lung cancer in Xuan Wei County, which has the highest lung cancer rates among women in China. Using spatial analysis of coal use patterns and lung cancer rates in hundreds of villages, he has raised the hypothesis that crystalline silica (quartz) in coal smoke is an important risk factor in the lung cancer epidemic in rural Xuan Wei. Currently, he is working to quantify quartz and other carcinogens in coal smoke and to determine whether an exposure–response relationship can be found. Tian provided local oversight and guidance for WP3 and WP4.
Thuan-Quoc Thach has a Ph.D. from the University of Alberta and an M.Sc. from the University of Western Ontario. He is a scientific officer for the Department of Community Medicine, University of Hong Kong, HK SAR. Thach is a biostatistician with wide experience in development and application of epidemiological and statistical methods for assessing environment and health issues. He was a coinvestigator in three previous HEI-funded projects (HEI Research Report 154, Public Health and Air Pollution in Asia [PAPA]; Research Report 169, Effects of Short-Term Exposure to Air Pollution on Hospital Admissions of Young Children for Acute Lower Respiratory Infections in Ho Chi Minh City, Vietnam; and Research Report 170, Impact of the 1990 Hong Kong Legislation for Restriction on Sulfur Content in Fuel). Thach provided local oversight and guidance for the fieldwork campaigns in WP1 and WP2.
Ryan Allen has a Ph.D. and an M.Sc. from the University of Washington. He is an associate professor on the Faculty of Health Sciences at Simon Fraser University, Canada. Allen’s research interests are in the area of air pollution exposure assessment and epidemiology, with particular interests in the cardiovascular effects of air pollution and air quality in Asia. Allen led the infiltration efficiency analysis contained within WP2.
Michael Brauer has a Sc.D. from Harvard University. He is a professor in the School of Population and Public Health, University of British Columbia, Canada. Brauer has been a leader in the development of LUR models for TRAP and their application to epidemiological analyses. He developed the first LUR models for TRAP PM and developed the application of mobile monitoring, 3D urban morphometry, and meteorology to LUR models. He has a long history of research in the development of improved estimates of exposure to air pollution and in the assessment of exposure misclassification on epidemiological effect estimates. He has conducted an evaluation of LUR models and the incorporation of mobility on personal exposures, and an assessment of the impact of mobility on epidemiological effect estimates. He is also recognized as an expert on air quality in Asia and recently participated in the development of an LUR model for Delhi. Brauer was lead investigator on WP1 and provided scientific oversight of all work packages.
