London Hybrid Exposure Model: Improving Human Exposure Estimates to NO2 and PM2.5 in an Urban Setting.

Here we describe the development of the London Hybrid Exposure Model (LHEM), which calculates exposure of the Greater London population to outdoor air pollution sources, in-buildings, in-vehicles, and outdoors, using survey data of when and where people spend their time. For comparison and to estimate exposure misclassification we compared Londoners LHEM exposure with exposure at the residential address, a commonly used exposure metric in epidemiological research. In 2011, the mean annual LHEM exposure to outdoor sources was estimated to be 37% lower for PM2.5 and 63% lower for NO2 than at the residential address. These decreased estimates reflect the effects of reduced exposure indoors, the amount of time spent indoors (∼95%), and the mode and duration of travel in London. We find that an individual's exposure to PM2.5 and NO2 outside their residential address is highly correlated (Pearson's R of 0.9). In contrast, LHEM exposure estimates for PM2.5 and NO2 suggest that the degree of correlation is influenced by their exposure in different transport modes. Further development of the LHEM has the potential to increase the understanding of exposure error and bias in time-series and cohort studies and thus better distinguish the independent effects of NO2 and PM2.5.


INTRODUCTION
Epidemiological research has demonstrated associations between ambient air pollution and a range of health effects especially on the respiratory and cardiovascular systems. 1−11 Epidemiological methods include estimating PM 2.5 and NO 2 exposure metrics at a coarse spatial scale, using degrees of longitude/latitude, 12,13 monitoring stations as proxies, 2,10,14−16 or by considering the proximity of a subjects address to nearby roads. 17 These studies often ignore the strong spatial concentration gradients and variability that exist within urban areas. More detailed approaches have evolved which estimate pollution concentrations at individual addresses or at the centroid of small areas such as postcodes using geostatistical interpolation, 18 land-use regression, 19 and dispersion models. 20,21 However, these models do not take account of exposure within the home, at work/school, and in different travel microenvironments, despite significant differences between personal and outdoor exposures demonstrated using comparative measurement studies. 22−24 Consequently, these methods contribute to random exposure error, which is likely to bias health associations to the null 25 and systematic error that will bias health associations in either direction. Furthermore, the correlations between PM 2. 5 and NO 2 which tend to be observed using such methods make it difficult to investigate the independent associations of these two pollutants with health effects. 26,27 Among the first attempts to obtain more detailed exposure estimates was the EXPOLIS study framework by Kousa et al. 28 who evaluated the temporal and spatial exposure of the population of Helsinki, Finland in different microenvironments. More recently, Dhondt et al. 29 developed a population exposure modeling approach taking into account population mobility in two urban areas in Belgium, concluding that large differences in health impacts occur when assessments neglect population mobility. Recent studies have highlighted the importance of exposure to air pollution during trips in urban areas. De Nazelle et al. 30 found that while a number of individuals in Barcelona spent 6% of their time in transit, it contributed 11% of their daily exposure to NO 2 . Likewise, Houston et al. 31 concluded that although travel comprised 5% of a participant's day, it represented 27% of their daily exposure to particle-bound polycyclic aromatic hydrocarbons. Other studies have reported that commuters in urban areas receive 12% of their daily PM 2.5 exposure and up to 30% of their black carbon inhaled dose while traveling. 32−34 Evidence of this kind has led the Health Effects Institute (HEI) to suggest activitybased or hybrid exposure models, which combine space-timeactivity data, personal measurements, and air quality models, come closest to a "best" estimate of human exposure. 35 Here we describe the London Hybrid Exposure Model (LHEM), calculating air pollution exposure at population level, while accounting for each individual's movements, during the day. We summarize the population's travel behavior and then compare LHEM exposure estimates with exposure at the resident's address, to assess exposure misclassification. We also assess whether the LHEM provides a better basis for investigating the independent effects of NO 2 and PM 2.5 on health. 2.2. Trip Route Simulation. The LTDS data set includes start and end coordinates, times of trips, and mode of transport but does not describe the routes taken. These have been simulated in a number of different ways. Before undertaking trip route simulation, the data was cleaned to remove missing or clearly incorrect data, for example, unrealistically quick journeys, missing origin or destinations, and missing journey sections. This removed 9% of the data, leaving a final data set of 45,079 people taking 98,770 trips and a total of 340,754 stages (representing different transport modes within the same trip). To calculate the routes, start and end coordinates, times, and transport mode of each stage were processed in the R Statistical Computing Environment 38 using the RPostgreSQL 39 extension. Specifically, the start and end points of each stage were formed into URL strings and processed by a number of different routing application programming interfaces (APIs). The Open Route Service API 40 was used to simulate walking trips (shortest-path), the Project OSRM API 41 to simulate car trips (quickest-path), Google Directions 42 to simulate cycling (quickest-path), and the TfL Journey Planner 43 to simulate public transport trips (overground train, the London Underground, the Docklands Light Railway, and bus). These APIs returned routes between the start and end locations, which were stored in a PostGIS database, and split into points for each minute of the journey using linear interpolation. For periods between trips, people were assumed to stay indoors at the previous destination point. The resulting data set gave the location and environment (indoor, walking, cycling, driving, etc.) of each of the 45,079 LTDS respondents on a minute-byminute basis over 24 h.

MATERIALS AND METHODS
2.3. Exposure Estimates. 2.3.1. Outdoor Air Pollution Predictions. Exposure to outdoor air pollution was provided by CMAQ-urban, 44 which couples the Weather Research and Forecasting (WRF) meteorological model, 45 the Community Multiscale Air Quality (CMAQ) regional scale model, 46 and the Atmospheric Dispersion Modeling System (ADMS) roads model. 47 The CMAQ-urban model output was processed to give annual average hour of day NO x (NO + NO 2 ), NO 2, NO, and PM 2.5 concentrations in 2011, across a 20 m × 20 m grid over the UK ( Figure S1, SI). CMAQ-urban was run without bias correction and has previously been submitted to the UK Model Intercomparison exercise run by the UK Government department DEFRA, 48 performing well against other urban and regional models. In Appendix C and Figure S2 (SI) we summarize the CMAQ-urban model evaluation, which has r values of 0.9 (NO X /NO 2 ) and 0.77 (PM 2.5 ). The bias for NO 2 and PM 2.5 is approximately 3% and −10%, respectively.
2.3.2. In-Building Exposure. To calculate indoor exposure to outdoor sources of PM 2.5 and NO 2 using the LHEM, we apply indoor/outdoor (I/O) ratios for domestic properties to the outdoor CMAQ-urban. As all I/O ratios are less than 1, this means that exposure indoors is less than outdoors at the same location. The modeling methods are described in full in Taylor et al. 49 but briefly; models were run for 15 building types, derived from the English Housing Survey 50 and Geoinformation Group classifications, 51 which represent 76% of the known London housing stock. The building physics model (Energy Plus 8.0 52 ) estimated dwelling I/O ratios for PM 2.5 and NO 2 using background air infiltration and exfiltration; indoor and outdoor temperature-dependent window opening, representing realistic occupant behavior; and deposition rates and penetration factors (detailed in Appendix D, SI). Air infiltration and exfiltration were modeled by assigning permeabilities− typical of the age, type, and construction of the buildings, allowing the model to account for infiltration due to wind pressures and buoyancy effects. Buildings were modeled using a Test Reference Year (TRY) weather file for London, representing "typical" weather conditions, and derived from a 30 year baseline, with results output hourly for a year-long simulation. The ratios were summarized by hour of the day and weekday/Saturday/Sunday to match the temporal resolution of the CMAQ-Urban data set and then averaged for each postcode in Greater London. I/O ratios for offices were assumed to have the same value as houses in the same postcode. Figure 1 shows a map of daily average PM 2.5 ratios (although hourly ratios are used in the model). A similar map of 24 h average I/O ratios for NO 2 is given in Figure S3 (SI). For the postcodes in London where there was insufficient housing stock information to calculate ratios (white areas, Figure 1), the London average was used (0.31 for NO 2 and 0.56 for PM 2.5 ).
2.3.3. In-Vehicle Exposure. For in-vehicle exposure, the pollutant concentration is derived by solving the mass balance equation 53 where C out is the outdoor concentration surrounding the vehicle, λ win and λ HVAC are the hourly air exchange rates from the windows and mechanical ventilation system, respectively, η is the filter removal efficiency taking values between 0−1, V g is the deposition velocity in m h −1 , A* is the internal surface area available for deposition, V is the volume of the vehicle, and Q is the in-vehicle particle emission rate in μg h −1 (defined as the product of resuspension rate and number of passengers not seated in the vehicle). The total surface area for deposition A* is calculated by adding the internal surface area of the vehicle A to the surface area of each passenger. 53 The ventilation conditions in cars were assumed to be constant throughout the journey with the assumption of no indoor sources, whereas the ventilation conditions were modified by the regular stops of buses and trains, with the model parameters reset at each stop, remaining constant between stops. Resuspension from people's movements was assumed to be the only indoor source of particles in buses and trains with no indoor sources for gaseous pollutants assumed. Details of the parameters used in the in-vehicle mass balance model are provided in Table S2 and Appendix E (SI).
Finally, for the London underground, concentrations were taken to be 94 μg m −3 for PM 2.5 and 51 μg m −3 for NO 2 . These figures were based upon PM 2.5 measurements conducted on London Underground platforms and trains during 22 journeys, over 450 min, using a TSI Sidepak AM510 56 (Benjamin Barratt, personal communication) and from NO 2 measurements on the Paris Metro. 57 2.3.4. Walking and Cycling Exposure. Exposure while walking or cycling was taken from time and location specific CMAQ-urban model concentrations.
By combining these methods, the LHEM calculates exposure minute-by-minute for 45,079 people, representing 6.8m Londoners. Using these data we calculate the population average daily exposure and the contribution from each indoor, in-vehicle and outdoor microenvironment. LHEM is able to simulate exposure to PM 10/2.5 , PM components, NO x /NO 2 , and O 3 , as well as the exposure to separate sources, although here we have limited our analysis to NO 2 and PM 2.5 .

RESULTS
3.1. Travel and Exposure to NO 2 and PM 2.5 in London. The results in Table 1 have been split into age categories, totalling 6.8m people. For each of these age categories results have been given for the number of people, the inner/outer London split, the percentage of the day spent in different microenvironments, their trip details during the day, and exposure to NO 2 and PM 2.5 in each microenvironment as a percentage of their total daily exposure. Missing data and data "not recorded" represented <0.01% of the total data set.
Results from the LHEM show that on average people spend over 95% of their time indoors, predominantly at home or at work, and that this proportion of time varies by only 2−3% across the age ranges and includes approximately 20% of people who, when surveyed, did not leave their house. Active travel (cycling and walking) takes up about 1−1.7% of the time during a typical day with inactive travel (driving, on the bus, underground, train, motorcycle) taking 1.4 to 3.6%. This translates into the number of trips being approximately 3 for children to almost 7 for adults, with mean journey times and journey lengths ranging between 12.8 min and 3.1 km for children and 26 min and 8 km for adults.
The percentage of daily exposure indoors to outdoor NO 2 and PM 2.5 sources ranges between 92% for children and 81% for adults and between 96% for children and 88% for young adults, respectively. Inactive travel in London results in 2.4 times greater exposure to NO 2 than active travel. Inactive travel results in 3.2 times greater exposure to PM 2.5 than active travel. People's trips contribute approximately 15% of their daily NO 2 exposure and 9% of their daily PM 2.5 exposure.
3.2. NO 2 and PM 2.5 Exposure and Exposure Misclassification. Exposure misclassification is an important consideration in estimating exposure to air pollution in health studies. Figure 2 (NO 2 left panel and PM 2.5 right panel) shows LHEM exposure estimates compared to the equivalent exposure estimated at the residential address.
Within city variability is important in determining associations with health risks, and for PM 2.5 , the range in LHEM exposure is between 6.0 μg m −3 and 32.2 μg m −3 , larger than the range at the residential address (11. To establish more clearly the reasons for the differences in the exposure distributions of the LHEM and residential address methods, scatter plots of PM 2.5 and NO 2 are given in Figure 3,  Figure S3) has a similar pattern between central and outer London but has lower I/O ratios (0.11 to 0.59). Superimposed on the map is the London Underground network to aid orientation.

Environmental Science & Technology
Article divided into active travel, inactive travel, those who remain at home, and all three combined.
The correlation between exposure estimated using the LHEM and at residential address, shown in purple in Figure  3, is relatively weak, having Pearson's R values of 0.41 for PM 2.5 and 0.55 for NO 2 . For those who undertake inactive travel, shown in turquoise, there is also a weak relationship (R values of 0.56 for NO 2 and 0.39 for PM 2.5 ) due to the increased time spent in transit while exposed to high PM 2.5 and NO 2 concentrations. In contrast, there is a strong relationship between LHEM and residential address exposures for those staying at home, shown in red (R values of 0.91 for NO 2 and

Environmental Science & Technology
Article 0.86 for PM 2.5 ), and for those undertaking active travel, shown in green (R values of 0.77 for NO 2 and 0.79 for PM 2.5 ). Finally, for individuals whose dominant transport mode is the London Underground/DLR (not shown), there is no correlation between LHEM exposure and residential address exposure to PM 2.5 (R values of −0.01). In summary, the strength of the relationship between the LHEM and residential address exposure is determined by travel, with inactive travel modes contributing to the highest daily exposure estimates of PM 2. 5 and NO 2 and the largest exposure misclassification (shown in blue, Figure 3). The LHEM results also suggest that spending long periods of time indoors reduces exposure to outdoor sources of air pollution. I/O ratios (Figure 1 and Figure S3 (SI)) reduce between outer and inner London, due to newer building stock and the number of flats in large buildings, resulting in a small surface area available for infiltration of outdoor air. As a consequence there is greater protection from outdoor air pollution, which is highest in this central part of the city. However, greater air tightness will also increase the impact of exposure from indoor sources, currently not considered in the LHEM.

Pollutant Correlations in Exposure Estimates
. It is often difficult to establish independent health associations for NO 2 and PM 2.5 due to their strong correlation outdoors. Scatter plots of NO 2 vs PM 2.5 outdoor residential exposure results are given in Figure 4 (left panel) and show high correlation (R = 0.90). In contrast, the relationship between NO 2 and PM 2.5 exposure using the LHEM (Figure 4, right panel) shows a much more complicated picture, especially when broken down by dominant travel mode (Figure 3), with the majority of the people who stay at home having the lowest exposures ( Figure  3, in red), and those that travel the highest (Figure 3, green and blue/turquiose). Figure 4 (right panel) shows two distinct groups in the population: those who have high NO 2 exposure and low PM 2.5 exposure, and those who have low NO 2 and high PM 2.5 exposure. This relationship remains during typical and nontypical days (Appendix F and Figure S4, SI). Further work is needed to better understand these relationships, but when

Environmental Science & Technology
Article applied to health studies this contrast has the potential to examine the independent effects of NO 2 and PM 2.5 .
3.4. LHEM Sensitivity. While the outdoor exposure estimates of CMAQ-urban have been comprehensively evaluated using UK fixed site air quality data (Appendix C1, SI), there is little suitable observed personal exposure data with which to evaluate the LHEM exposure predictions. However, since transport and indoor microenvironments are important for exposure, we have tested the robustness of our results with five sensitivity tests using different input variables in these models. The results are presented as the percentage change in daily PM 2.5 exposure. Further work on the sensitivity of the PM 2.5 and NO 2 exposure is planned (see Table 2).
Test 1: New PM 2.5 I/O ratios were modeled with a lower particle deposition rate (0.125 h −1 instead of 0.19 h −1 ) and higher building penetration factor (0.82/1 instead of 0.8/1), resulting in ratios which were on average 20% higher, and consistent with those used in other measurement based studies 58,59 which include indoor sources.
Test 2: The car, taxi, and van ventilation settings for the invehicle model were changed from semiopened windows and air-conditioning off (natural ventilation setting) to windows closed and air-conditioning system in use (mechanical ventilation setting), resulting in a lower penetration of outside air into the vehicle.
Test 3: The number of people in buses, trains, and on the Docklands Light Railway (DLR) was increased to the maximum capacity of the vehicles (reflecting rush hour transport conditions). The number of active passengers was also increased (see Table S2, SI).
Test 4: Particle resuspension rates for the in-vehicle model were doubled from the base case value of 0.02 μg/min −1 to 0.04 μg/min −1 (from Ferro el al. 61 ).
Test 5: Pollutant deposition rates for the in-vehicle part of the model were increased from the mean values found in INDAIR 60 to the maximum values found in the same study (Appendix E, SI).
Results of the sensitivity tests show that altering the physical parameters and ventilation settings of the in-vehicle calculations (tests 2−5) had little effect on the overall population exposure with a maximum change in PM 2.5 exposure being −1.9% due to vehicle ventilation assumptions for adults, although they are likely to be more influential on the results of specific trips/ people. The parameter which had the largest effect was changing building I/O ratios, which for all age groups increased their exposure by >11.1%, demonstrating the importance of this microenvironment in exposure calculations.

DISCUSSION
Correctly estimating human exposure to air pollution remains an important source of uncertainty in public health research and one which can push existing associations toward the null, i.e. to incorrectly conclude that no association exists or to affect the precision of the association. The LHEM, introduced here, is the latest attempt 33,34,60,62 to establish a more comprehensive understanding of human exposure. There are many uncertainties to our model that require further investigation; however, using it, we find that our study population of 6.8m people, representing the population of London (>5 years of age), spends between 94.7 and 97.9% of their time indoors and 2.1− 5.3% in transit, depending on age and other variables. These time-activity patterns are broadly typical of other midlatitude urban populations, for example the proportion of the day in transit was 6% in a study in Barcelona 30 and 8% for the working population in Helsinki, Finland. 28 We find that travel contributes between 8.0 and 18.8% of our population's daily exposure to outdoor NO 2 sources and between 4.1 and 12.2% of outdoor PM 2.5 sources, with inactive travel contributing more than active travel. That people's trips contribute approximately 15% of their daily NO 2 exposure is similar to findings in Barcelona 30 and that approximately 9% of the daily PM 2.5 exposure comes from people's trips agrees with findings by Fondelli et al. 63 in Florence. The concentrations that we model for transport microenvironments (Appendix G and Table S3, SI) tend to have large ranges due to the temporal and spatial variability of the CMAQ-urban model used as input to the in-vehicle model. Results generally agree with studies that include measurements of these environments, 64−70 although there is a large variation reported in the literature. Time spent in indoor environments contributes the rest of the population's exposure of between 81.2 and 92% for NO 2 and 87.8 to 95.9% for PM 2.5 from outdoor sources.
When comparing the LHEM method to address-based exposure estimates, we find the average LHEM exposure is 37% lower for PM 2.5 and 63% lower for NO 2 . Results from other similar studies are difficult to find as we combine individual exposures for a large population and do not include indoor emissions. For NO 2 a study by de Nazelle 30 of a small number (36) of active volunteers moving around the city found subjects had 27% higher NO 2 exposure (including indoor sources) when a time-activity model was compared to residential exposure. Dhondt 29 found ambient NO 2 exposure to be between 5% lower and 15% higher when using dynamic methods with a model which has lower temporal and spatial resolution than the LHEM and a relatively simple transport exposure model. In addition the LHEM incorporates a larger range of activity levels and ages, including a number of participants who do not leave their home, and does not include indoor sources. For PM 2. 5 Burke, 71 when calculating exposure to ambient sources using a stochastic simulation method, found a median exposure to PM 2.5 of 7 μg m −3 compared to our mean of 8.5 μg m −3 .
Using the LHEM we conclude that the indoor environment is protective of exposure to outdoor emissions sources and that this is reflected in much lower exposure estimates than would be found using exposure outdoors at the residential address. The degree to which buildings are protective varies geographically in London and is pollutant and housing specific. Average NO 2 I/O ratios in the EXPOLIS and SAPALDIA studies 72,73 are found to be 0.76 and 0.545 respectively, and are higher than our range of 0.11 to 0.59 due to the influence of indoor sources. For PM 2.5 , Taylor et al. compared the ratios we use (range of 0.35 to 0.86) to existing studies and found good agreement. Furthermore, sensitivity tests demonstrated that

Environmental Science & Technology
Article even with increased I/O ratios, the differences between LHEM and residential exposure are still present. Finally a further limitation of the LHEM is the assumption that I/O ratios for commercial buildings are the same as domestic buildings of similar ages, morphologies, and in the same postcode. In addition, we have shown that transport microenvironments are important in correctly establishing exposure, with those who stay at home having an exposure which is closely correlated with outdoor residential concentrations and those who travel, in particular by car, bus, and by the London Underground being poorly correlated with outdoor concentrations at the residential address. Members of the population that travel by car, bus, and the London Underground are some of the most highly exposed, with small differences in the exposure across the four age groups used in this study.
There is also a more complex relationship between PM 2.5 and NO 2 LHEM exposure estimates compared to the correlated exposures at residential addresses. Within health studies this may have the potential to examine the independent health effects of NO 2 and PM 2.5 through the assessment of the differential degree of misclassification of personal exposure to each air pollutant. In future cohort studies in London, participants could complete questionnaires similar to those in the LTDS, which could then be combined with the LHEM to assess personal exposure, and have this linked to their health records. The LHEM could be used as a template for other cities, as it provides population exposure, not possible using alternatives such as personal monitoring. For time-series studies, Zeger et al. 74 illustrate how estimates of daily personal exposure in the population can be used to adjust the relative risk obtained from time-series studies, using daily ambient concentrations, for measurement error and to predict the likely bias occurring in multipollutant models. For health impact assessments, policies designed to reduce air pollution exposure by changing travel behavior, or by reducing specific microenvironment concentrations, are currently difficult to assess as the relevant exposure metric does not match those currently used in epidemiological studies. This situation can be improved by learning more about the relationships between different exposure metrics using the LHEM or by deriving concentration−response functions from future studies using the LHEM or an analogous model. As the LHEM is based upon CMAQ-urban predictions it is possible to predict exposure to components of PM, such as elemental, organic carbon, nitrate, and sulfate, rather than just PM mass, and also ozone.
The LTDS samples an individual's activity on a single day. The LTDS survey population has coverage of weekday and weekend activities throughout the year, and by combining all individuals and using weighting factors provided by TfL, we obtained an average daily NO 2 and PM 2.5 exposure for the London population in 2011.
The next stage of the LHEM is better predicted exposure indoors and outdoors from all sources, rather than to outdoor sources as we have here. To do this requires the inclusion of an indoor air pollution model with indoor sources, which given the large amount of time that people spend indoors is likely to alter LHEM exposure predictions. Second, a more comprehensive evaluation of LHEM is needed using personal exposure data, which have advanced rapidly in recent years, including the development of postprocessing methods to ensure high data quality. However, robust evaluation of models like LHEM is still limited due to the cost and size of monitors, as well as the practicality of obtaining large enough cohorts for the data to be reliable.
To date, we have not found an equivalent LTDS data set outside London; however, there is undoubtedly equivalent data in other cities throughout the world, enabling models such as LHEM to be developed. Although we recognize the significant ethical hurdles that this approach would entail, it is hoped that the LHEM, as the latest in this new breed of exposure models, will lead to linking public health records with personal exposure estimates, to better understand human exposure to air pollution and the associated health effects. Finally, the LHEM can play a part in testing air quality policy scenarios aimed at reducing exposure rather than lowering ambient concentrations, for investigating the exposure of different subpopulations and for assessing exposure indoors and outside. (2) Dockery, D.; Pope, C.; Xu, X. An association between air pollution and mortality in six US cities. N. Engl. J. Med. 1993, 329, 1753−1759. (