Journal of Exposure Science and Environmental Epidemiology (2013) 23, 259–267 & 2013 Nature America, Inc. All rights reserved 1559-0631/13
www.nature.com/jes
ORIGINAL ARTICLE
Development of temporally refined land-use regression models predicting daily household-level air pollution in a panel study of lung function among asthmatic children Markey Johnson1, Morgan MacNeill1, Alice Grgicak-Mannion2, Elizabeth Nethery1, Xiaohong Xu3, Robert Dales1, Pat Rasmussen4 and Amanda Wheeler1 Regulatory monitoring data and land-use regression (LUR) models have been widely used for estimating individual exposure to ambient air pollution in epidemiologic studies. However, LUR models lack fine-scale temporal resolution for predicting acute exposure and regulatory monitoring provides daily concentrations, but fails to capture spatial variability within urban areas. This study coupled LUR models with continuous regulatory monitoring to predict daily ambient nitrogen dioxide (NO2) and particulate matter (PM2.5) at 50 homes in Windsor, Ontario. We compared predicted versus measured daily outdoor concentrations for 5 days in winter and 5 days in summer at each home. We also examined the implications of using modeled versus measured daily pollutant concentrations to predict daily lung function among asthmatic children living in those homes. Mixed effect analysis suggested that temporally refined LUR models explained a greater proportion of the spatial and temporal variance in daily household-level outdoor NO2 measurements compared with daily concentrations based on regulatory monitoring. Temporally refined LUR models captured 40% (summer) and 10% (winter) more of the spatial variance compared with regulatory monitoring data. Ambient PM2.5 showed little spatial variation; therefore, daily PM2.5 models were similar to regulatory monitoring data in the proportion of variance explained. Furthermore, effect estimates for forced expiratory volume in 1 s (FEV1) and peak expiratory flow (PEF) based on modeled pollutant concentrations were consistent with effects based on household-level measurements for NO2 and PM2.5. These results suggest that LUR modeling can be combined with continuous regulatory monitoring data to predict daily household-level exposure to ambient air pollution. Temporally refined LUR models provided a modest improvement in estimating daily householdlevel NO2 compared with regulatory monitoring data alone, suggesting that this approach could potentially improve exposure estimation for spatially heterogeneous pollutants. These findings have important implications for epidemiologic studies — in particular, for research focused on short-term exposure and health effects. Journal of Exposure Science and Environmental Epidemiology (2013) 23, 259–267; doi:10.1038/jes.2013.1; published online 27 March 2013 Keywords: child exposure/health; criteria pollutants; epidemiology; exposure modeling; environmental monitoring
INTRODUCTION Both long- and short-term exposures to air pollution have been linked to a variety of adverse health effects.1 Historically, improvements in exposure assessment have contributed to an increased understanding of air pollution health effects, particularly with respect to early/subclinical and low-dose effects. However, the current methodologies still present limitations. Air quality measurements from regulatory monitoring stations are widely used to represent the exposure of subjects residing within large geographic areas surrounding these monitoring stations.2 However, urban air pollution levels, particularly mobile source pollutants, exhibit high spatial and temporal variability3–5 and ambient monitoring data are usually limited in the number of monitoring sites, as well as the species of pollutants measured. Thus, these methods do not adequately capture the fine-scale spatial and temporal variations required to accurately estimate exposure.
Concentrations and exposures to air pollution can be assessed via indoor, outdoor, and personal measurements collected by trained technicians.6–9 They can also be predicted using physical or mechanistic modeling-based approaches including atmospheric, indoor/outdoor/personal exposure, and hybrid models.10–15 However, technician-based air monitoring can be resource intensive and may impose a significant burden on study participants, while modeling studies typically require detailed information on emissions, building characteristics, and exposure factors, as well as significant technical expertise. Due to the increasing ability of geographic information systems (GIS) to extract and analyze land-use data, land-use regression (LUR) models have emerged as a widely used approach for capturing the local-scale spatial variability in community health studies.16–23 LUR models were recently recommended as a more accurate method for exposure analysis in a critical review conducted by the Health Effects Institute.1 However, despite
1 Air Health Science Division, Water Air and Climate Change Bureau, Health Canada, 269 Laurier Avenue West, Ottawa, Ontario, Canada K1A 0K9; 2Great Lakes Institute for Environmental Research, University of Windsor, 401 Sunset Avenue, Windsor, Ontario, Canada N9B 3P4; 3Department of Civil and Environmental Engineering, University of Windsor, Windsor, Ontario, Canada N9B 3P4 and 4Exposure and Biomonitoring Division, Health Canada, 50 Colombine Driveway, Tunney’s Pasture, Ottawa, Ontario, Canada K1A 0K9. Correspondence: Dr. Markey Johnson, Air Health Science Division, Water Air and Climate Change Bureau, Health Canada, 269 Laurier Avenue West, Room 3-024, Ottawa, ON, Canada K1A 0K9. Tel.: +1 613 952 3707. Fax: +1 613 948 8482. Email:
[email protected] Received 27 February 2012; accepted 17 October 2012; published online 27 March 2013
Development of temporally refined land-use regression models Johnson et al
260 their advantages of being practical, cost-effective, and easy to apply, LUR models have several current limitations including the inability to predict short-term (e.g., daily) exposures. LUR models typically reflect either seasonal or annual average concentrations. This paper introduces a method for estimating daily outdoor concentration at the household level for ambient air pollution by coupling data from LUR models and continuous regulatory monitoring data in Windsor, Ontario. We focused on nitrogen dioxide (NO2), a spatially heterogeneous pollutant emitted by both mobile and industrial combustion sources. We selected particulate matter (PM2.5) to represent a locally homogeneous regional pollution. Daily household estimates of NO2 and PM2.5 based on this approach were compared with daily measurements collected at the homes of 50 asthmatic children enrolled in the Windsor Ontario Exposure Assessment (WOEAS) Study.8 We also examined the implications of using modeled versus measured pollutant concentrations to predict daily lung function among asthmatic children enrolled in WOEAS. METHODS The analyses for this study were conducted in two parts. First, we coupled LUR and regulatory site monitoring data to estimate daily ambient pollutant concentrations for NO2 and PM2.5 at each home. We then examined the implications of using predicted versus measured outdoor residential concentrations of NO2 and PM2.5 in models of lung function among asthmatic children enrolled in the study.
Temporally Refined LUR Models LUR modeling. Pollutant estimates of NO2 and PM2.5 were extracted for each participant’s home from seasonal and yearly average geospatial LUR models that were developed as follows. Fifty samplers were deployed throughout the city of Windsor in 2004 and 2005 with measurements collected over 2-week periods in February, May, August, and October. Site selection was described in detail by Wheeler et al.21 Briefly, the monitoring sites were selected to represent the variability in both the pollutants of interest and land-use variables considered as potential predictors, and to be representative of areas with greater population density. Sites were also selected to reduce potential autocorrelation, which were assessed via Moran’s I, as well as by a sensitivity analyses. LUR sites in 2005 included households participating in an exposure study that were selected based on geographic distribution, described in detail in Wheeler et al.8 Air pollution data from this monitoring campaign was combined with land-use predictors, for example, traffic and industrial sources that were generated using GIS. Stepwise regression analysis was used to develop annual and seasonal predictive equations relating these land-use characteristics with pollutant concentrations. These models have been evaluated in previous publications.24,25 Winter LUR estimates of NO2 used in these analyses were based on LUR models estimates for February 2004, summer NO2 estimates were the average of May and August 2004, and LUR estimates for PM2.5 were based on annual 2005 models.25 All model r-squares were 40.70 with the exception of May 2004, which was o0.50. Note that seasonal PM2.5 models were examined, but did not improve upon the annual model for predicting daily concentrations in the preliminary analyses. Regulatory site monitoring. Regulatory monitoring data for NO2 and PM2.5 were obtained from National Air Pollution Surveillance (NAPS) sites (60204 and 60211) operated by Environment Canada for the city of Windsor.26 Daily concentrations from two regulatory monitoring sites were calculated to correspond with the daily sampling times for each home. The daily concentrations from the two monitoring stations were averaged to produce the daily regulatory monitoring site concentration used in equation 1, below. Estimation of daily outdoor pollution concentrations at study homes. We estimated daily ambient NO2 and PM2.5 at study homes using concentrations predicted by LUR models combined with continuous regulatory monitoring (NAPS) data in Windsor, Ontario. We introduced temporal variability to LUR estimates at each home by applying a calibration factor equal to the ratio of the daily concentration at the regulatory monitoring sites for each sampling day and the seasonal average at the regulatory
monitoring sites (equation 1). Daily estimates for each home were calculated as the product of the seasonal (NO2) or annual (PM2.5) LUR estimate for each home and the daily calibration factor (equation 2). Calibration factor = daily average at regulatory sites / seasonal average at regulatory sites Predicted daily concentration = calibration factor * seasonal or annual LUR estimate
ðeq:1Þ
ðeq:2Þ
Household Monitoring We used outdoor residential measurements from the WOEAS in Windsor, Ontario to evaluate modeled daily concentrations. WOEAS conducted daily air pollution monitoring at the homes of 50 asthmatic children in 2006. Households were selected to ensure even spatial distribution across Windsor; Figure 1 illustrates the locations of participating homes. Collection and analysis of air monitoring data in these studies has been described in detail by Wheeler et al.8 Briefly, air monitoring data were collected at each home for a total of 10 days — including 5 days in winter (January–March) and 5 days in summer (July–August). Air sampling was started on Monday evenings and ended on the following Saturday. Technicians visited each home daily to maintain sampling equipment and administer questionnaires. Daily NO2 samplers and PM2.5 filters were deployed at approximately B1600 hours, and retrieved at 1600 hours on the following evening. Data were collected for 8 weeks per season, with a total of six homes sampled concurrently per week. Sampling equipment was setup at one outdoor location in the backyard of each home. Samplers were not setup near pollutant sources such as barbeques, driveways, or dryer vents. A metal rain shelter was used to protect samplers from wind and precipitation. Samplers. Ogawa passive samplers and pretreated filters purchased from Ogawa (Ogawa & Company, Pompano Beach, FL, USA) were used to measure exposure NO2.27 The NOx badge used carbonate-coated quartzfiber filter to trap NOx. Sampling time was B24 h for each badge. Nitritecoated filters were extracted using ultra-pure (Milli-Q) water and carbonate-coated filters were extracted with 0.09% (v/v) hydrogen peroxide. Extracts were analyzed using ion chromatography (IC) performed by Dionex DX-300 or DX-600 IC systems (Dionex, Sunnyvale, CA, USA). Nitrate and other anions in carbonate-coated filter extracts were analyzed using a Dionex-AS4A column with carbonate/bicarbonate eluent. Calibration checks were performed daily before analysis of field samplers and after every 10–15 samples using standards prepared from NISTtraceable standards, resulting in overall uncertainties in the order of 10%. PM2.5 data were collected using the R&P Chempass multi-pollutant sampler (Chempass System R& P/Thermo, Waltham, MA, USA). The PM2.5 target flow rate was 4.0 l/min, with an acceptable range of ±20%.28 Flow rates were assessed pre and post sampling every 24 h using a soap bubble flow meter (AP Buck, Orlando, FL, USA). The particle masses were collected on a 37-mm TeflonTM filter and measured gravimetrically at Health Canada’s Archimedes M3TM Buoyancy-Corrected Gravimetric Analysis Facility.29 The filters were preconditioned for 24 h before weighing inside a chamber with automated controls, which maintained environmental conditions at a constant air temperature of 21 oC (±0.5 oC) and constant relative humidity (RH) of 40% (±1%). Static electricity was removed by passing each filter across two Polonium-210 strip deionizers placed side by side in the chamber. The Archimedes M3TM chamber was designed to simultaneously monitor air pressure, temperature, dew point temperature and RH and record these parameters automatically. Gravimetric measurements made with a Mettler UMX2 microbalance (readability of 0.1 mg) were subsequently corrected for changes in air buoyancy determined at 1 min intervals as described by Rasmussen et al.30 Calculation of the laboratory detection limit (LDL) was based on three times the SD of 13–14 measurements of each of two reference filters per day. Quality assurance. LDLs were estimated as three times the SD of the laboratory blanks and field detection limits were defined as three times the SD of the field blanks. Field blanks comprised B10% of the total number of samples. Sample data was blank-corrected by subtracting the median of the field blanks; however, if 50% or more of the field blanks were below
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
& 2013 Nature America, Inc.
Development of temporally refined land-use regression models Johnson et al
261
Figure 1. Land-use regression (LUR) sampling sites and study homes in Windsor, Ontario. This map shows the homes of the 2006 health study participants for whom the temporally scaled LUR models were developed. The 2004 and 2005 sites were used for the initial LUR model development for nitrogen dioxide (NO2) and particulate matter (PM2.5). the detection limit, no blank correction was made. Samples below the LDL after blank correction were replaced with 1/2 LDL. The QA/QC program has been described in detail by Wheeler et al.8
Health Data Lung function. WOEAS recruited 50 participants from the Windsor Children’s Respiratory Health Study.31,32 Participation was limited to children between the ages of 10 and 13 who had physician diagnosed asthma. Health data collection was concurrent with air monitoring for all study participants, with daily lung function and questionnaire data collected for a total of 5 days in winter and 5 days in summer. Participating children provided peak expiratory flow (PEF) and forced expiratory volume in 1 s (FEV1) measurements twice daily — once after waking in the morning and again before going to bed (before taking any respiratory medication). The analyses in this paper were limited to morning measurements because their collection time was more consistent. Parents and children were instructed by trained respiratory therapists in the use of hand-held flow meters. FEV1 was measured using PiKo-1 electronic PEF/FEV1 meters, which are ATS/EU scale compliant (Ferraris Medical, Louiseville, CO, USA). For each measurement, participants were instructed to collect three flow readings. As described by Dales et al.,31,32 the maximum of the three readings was used for the analyses reported in this paper. Study design and protocols were approved by Health Canada’s Research Ethics Board and all personal information is protected according to the Access to Information and Privacy Acts. Covariates. Potential confounders and indoor sources were assessed through survey data provided by participants including baseline and daily questionnaires, as well as time activity diaries. Environmental tobacco smoke (ETS) was reported in number of smoking minutes per day. Height and weight were collected in November—December 2005 as part of the Windsor Children’s Respiratory Health Study.33 Variables coding for height and weight were log transformed. Temperature and RH were measured at each home using Smart Reader Plus 2 monitors (ACR Systems, Surrey, BC, Canada) as described in Wheeler et al.8 Health models were adjusted for mean daily temperature and RH.
Statistical Analyses Statistical analyses were performed using SAS 9.2 (SAS Institute, Cary, North Carolina, USA). We generated descriptive statistics for predicted and & 2013 Nature America, Inc.
measured residential pollutant concentrations, lung function, and potential confounders. Correlation analyses were performed using Spearman’s rank correlation based on the distribution of the data. Geospatial analyses and map development were performed using ESRI ArcInfo 9.3 (Redlands, CA, USA). Graphs were generated using Statistica 10.0 (Statsoft, Tulsa, OK, USA) Model evaluation. To evaluate the temporally refined LUR models, we compared daily ambient concentrations measured at the homes with daily concentrations produced by the temporally refined LUR models through generalized linear mixed effect models. Daily household measurements were specified as the dependent variable, while daily estimates based on temporally scaled LUR models and regulatory monitoring were specified as independent predictors. The random effects were participant (household) ID. These models were stratified by season. For NO2, the covariance matrix was specified as a first-order autoregressive (ar(1)) covariance structure for summer and a variance components (vc) covariance structure for winter. For PM2.5, the covariance structure was specified as compound symmetry for both seasons. Model covariance structure was selected based on the Akaike Information Criterion and likelihood ratio testing. For example, ar(1) provided a significant improvement over vc for NO2 summer models based on likelihood ratio tests (Po0.05); for winter models, there was no significant improvement, therefore, the vc covariance structure was used. Health analyses. For the epidemiologic analyses, we also specified mixed linear multivariate models with household ID as a random effect, and with a covariance structure of ar(1)—based on likelihood ratio testing. Because health models included winter and summer data, heterogeneity within the R matrix was specified by season. Fixed effects included the predicted or measured outdoor pollutant concentrations, as well as age, sex, height, weight, and ethnicity, as well as daily use of asthma medication and daily exposure to ETS. Health models examined associations between air pollution and morning lung function using air pollution estimates from the same day; therefore, each lung function measure was collected in the middle of each sampling day. Exclusion criteria. Pollutant analyses were limited to households with outdoor measurements and LUR estimates at the home (N ¼ 47). We excluded two homes that fell outside the LUR model area and one home that lacked NO2 measurements. Two participants were excluded from the lung function analyses because 4–5 daily observations for these
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
Development of temporally refined land-use regression models Johnson et al
262 participants were identified as influential points in the mixed model analyses (defined as an absolute value of the studentized residual (r) statistic42.0) and because the measured lung function values on those days fell within the 5th percentile for one of the participants and within the 95th percentile for the other. Daily observations that were influential but were not repeated for a single participant were included in the final models. Inclusion of influential observations tended to move the results toward the null hypothesis; however, models with and without these single influential observations were similar. Siblings were also excluded from the analysis. Descriptive statistics for participant characteristics were limited to the children included in the health models (N ¼ 45).
RESULTS This section describes (1) the evaluation of temporally refined LUR models using daily outdoor measurements collected at the household level, and (2) the implications of using modeled versus measured pollutant concentrations in the health analyses. Air Pollution Figures 2–3 show the distribution of daily ambient NO2 and PM2.5 measured at the study homes. NO2 displayed high temporal variability (variability between mean concentrations measured on different days) and spatial variability (variability between concentrations measured at study homes that conducted monitoring on the same days). In contrast, PM2.5 exhibited high temporal and low spatial variability. Daily NO2 concentrations predicted by the temporally refined LUR models explained 36–50% of the spatial variability and 37–42% of the temporal variability in daily ambient NO2 measured at the study homes (Table 1). Modeled NO2 concentrations explained a greater percent of the spatial variability in daily household NO2 measurements compared with daily measurements at the regulatory monitoring sites. Modeled concentrations and regulatory monitoring measurements were similar with respect to the percent temporal variability in household measurements explained for NO2. Modeled PM2.5 concentrations explained 72–83% of the spatial and 78–82% of the temporal variability in daily household measurements (Table 1). PM2.5 models were similar to regulatory monitoring data with respect to both the percent temporal and spatial variability explained for PM2.5. Figures 4–5 show scatter plots of predicted (temporally scaled LUR estimates) versus observed (household measurements) for NO2 and PM2.5. Daily LUR predictions were highly correlated with
daily measurements at the study homes, with correlation coefficients of 0.78 for NO2 and 0.88 for PM2.5. The correlation coefficients for the average daily concentrations measured at the regulatory monitoring sites and household-level measurements were 0.70 and 0.82, for NO2 and PM2.5, respectively. Although these results provide an alternate illustration of the relationship between predicted and observed concentrations, they should be interpreted with caution due to the autocorrelation between daily estimates for each home.
Lung Function Descriptive statistics for personal and household characteristics, as well as lung function are provided for the study participants in Table 2. Girls had higher mean FEV1 and PEF compared with boys in all of the multivariate pollutant models. Height was also positively associated with FEV1 (Po0.05). Daily exposure to ETS was associated with decreased PEF, and age was associated with higher PEF; however, these associations were only marginally significant (Po0.10). To further evaluate the temporally scaled LUR models, we examined the implications of using household measurements, temporally refined LUR models, and regulatory monitoring data to estimate daily exposure in epidemiologic models. Associations between air pollution and lung function were similar in models using measured versus predicted concentrations to predict exposure. Daily household measurements and temporally refined LUR estimates of daily NO2 at the home were associated with a 7–8% decrement in FEV1 per IQR, adjusting for age, sex, height, weight, and exposure to smoking (Po0.05), while daily NO2 measured at regulatory monitoring sites was not significantly associated with lung function (P ¼ 0.07). PM2.5 was not significantly associated with lung function in these models. In models adjusting for temperature, RH, and season as fixed effects, the associations between air pollution and lung function were no longer statistically significant for any of the three exposure metrics. However, there was evidence of a significant interaction between temperature and air pollution for both NO2 and PM2.5. In winter, decrements in lung function associated with NO2 and PM2.5 appeared to be greater at lower temperatures for both PEF (Po0.05) and FEV1 (Po0.10). Effect estimates for the interaction terms in these models were highly consistent across exposure metrics. There was also some evidence of a positive interaction between air pollution and temperature in summer, suggesting greater air pollution effects during hot weather events.
Winter
Summer
50
50 Median 25%-75% 5%-95% Outliers
Median 25%-75% 5%-95% Outliers 40
NO2 Concentration (ppb)
NO2 Concentration (ppb)
40
30
20
10
0 2006-01-23
30
20
10
2006-02-01
2006-02-10
2006-02-21
2006-03-02
2006-03-20
0 2006-07-03
2006-07-12
2006-07-21
2006-08-01
2006-08-10
2006-08-21
Figure 2. Daily outdoor nitrogen dioxide (NO2) measurements at study homes (N ¼ 47 homes). This figure shows the distribution of daily pollution concentrations of NO2 measured at the study homes. The box plots show the distribution of NO2 for homes that conducted sampling on the same day. Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
& 2013 Nature America, Inc.
Development of temporally refined land-use regression models Johnson et al
263 Winter
Summer
50
50
Median 25%-75% 5%-95% Outliers
Median 25%-75% 5%-95% Outliers 40 3 PM2.5 Concentration (g/m )
PM2.5 Concentration (g/m3)
40
30
20
10
0 2006-01-23
30
20
10
2006-02-01
2006-02-10
2006-02-21
2006-03-02
0 2006-07-03
2006-03-20
2006-07-12
2006-07-21
2006-08-01
2006-08-10
2006-08-21
Figure 3. Daily outdoor particulate matter (PM2.5) measurements at study homes (N ¼ 47 homes). This figure shows the distribution of daily pollution concentrations for PM2.5 measured at the study homes. The box plots show the distribution of PM2.5 for homes that conducted sampling on the same day.
Table 1.
Spatial and temporal variability in household measurements explained by regulatory monitoring versus temporally refined LUR models
(N ¼ 47). Model description
Variance component (95% confidence interval) Between subject (spatial)a
Within subject (temporal)
% Variance explained Temporal
Spatiala
NO2 — Summer Baseline (intercept only) Regulatory site monitoring Temporally refined LUR
52.07 (39.00, 73.06) 31.73 (24.14, 43.58) 31.56 (24.13, 43.08)
9.27 (3.02, 118.25) 8.33 (3.56, 37.11) 4.65 (1.46, 73.89)
— 38.99 39.32
— 10.14 49.84
NO2 — Winter Baseline (intercept only) Regulatory site monitoring Temporally refined LUR
51.00 (41.84, 63.55) 29.57 (24.25, 36.88) 30.22 (24.77, 37.70)
5.35 (2.02, 36.25) 3.94 (1.66, 18.31) 3.41 (1.32, 21.07)
— 42.02 40.74
— 26.43 36.31
PM2.5 — Summer Baseline (intercept only) Central site monitoring Temporally refined LUR
50.77 (41.37, 63.80) 8.45 (6.90, 10.60) 8.98 (7.33, 11.26)
3.59 (1.02, 105.10) 0.82 (0.30, 6.66) 1.00 (0.39, 6.03)
— 83.35 82.31
— 77.08 72.12
PM2.5 — Winter Baseline (intercept only) Central site monitoring Temporally refined LUR
45.19 (36.73, 56.95) 10.51 (8.50, 13.33) 10.14 (8.22, 12.83)
1.76 (0.33, 4098.95) 0.18 (0.02, 9.36^10) 0.30 (0.04, 603914.00)
— 76.74 77.56
— 89.89 82.92
Abbreviations: LUR, land-use regression; NO2, nitrogen dioxide; PM2.5, particulate matter. a This study used a staggered sampling design in which household monitoring was conducted at six homes per week for 8 weeks in each season (see Figures 2–3). Due to this staggered sampling design, the variation in pollutant concentrations between homes (e.g., between subject variance) reflects primarily spatial, but also some temporal variability.
However, these interactions were not statistically significant for most of the summer models. Finally, overlapping confidence intervals suggest that air pollution health effects were not significantly different across exposure metrics; however, model diagnostics generally suggested a better fit for epidemiologic models using household-level metrics versus regulatory monitoring to predict lung function. Sensitivity Analyses The lung function results were sensitive to changes model specification. Health models stratified by season showed similar associations between air pollution and lung function; however, the associations did not achieve statistical significance in stratified models without inclusion of a temperature and air pollution & 2013 Nature America, Inc.
interaction term. Also, influential observations had a strong impact on statistical significance, but not on the direction of the effects reported. Furthermore, the health model results were generally consistent across exposure metrics regardless of model specification.
DISCUSSION Accurate assessment of both short- and long-term exposure is needed to fully evaluate the health effects of air pollution. Household-level pollution measurements and exposure modeling can be resource intensive and potentially burdensome; however, temporally scaled LUR models predicting air pollution can provide a cost-effective alternative. This study coupled daily monitoring
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
Development of temporally refined land-use regression models Johnson et al
264 Winter
Summer 50 Outdoor NO2 Concentration (ppb)
Outdoor NO2 Concentration (ppb)
50
40
30
20
10
40
30
20
10
0
0 0
5
10
15
20
25
30
35
40
2
4
Temporally Refined LUR NO2 estimate (ppb)
6
8
10
12
14
16
18
20
22
Temporally Refined LUR NO2 estimate (ppb)
Figure 4. Predicted versus observed daily nitrogen dioxide (NO2) in winter and summer (N ¼ 47 homes). This figure shows a comparison of predicted versus observed daily NO2 concentrations stratified by season.
Winter
Summer 50 Outdoor PM2.5 Concentration (µg/m3)
Outdoor PM2.5 Concentration (µg/m3)
50
40
30
20
10
0
40
30
20
10
0 0
5
10
15
20
25
30
35
40
10
0
3
20
30
40
50
60
3
Temporally Refined LUR PM2.5 estimate (µg/m )
Temporally Refined LUR PM2.5 estimate (µg/m )
Figure 5. Predicted versus observed daily particulate matter (PM2.5) in winter and summer (N ¼ 47 homes). This figure shows a comparison of predicted versus observed daily PM2.5 concentrations stratified by season.
Table 2.
Health and socio-demographic characteristics of study participants (N ¼ 45). N
Mean
Median
CV
Personal characteristics % Girls % Caucasian % Using asthma medication Age (years) Height (cm) Weight (kg)
45 45 45 45 45 45
37.78 93.33 17.78 10.78 139.03 52.00
— — — 11.00 145.00 41.00
Household characteristics % Households with exposure to ETSa
45
24.44
Lung function Mean FEV1 (l) Mean PEF (l/min)
45 45
2.21 287.31
2.21 284.33
Minimum
Maximum
— — — 7.37 20.29 66.89
— — — 10.00 40.00 26.00
— — — 13.00 164.00 146.00
18.48 20.13
1.33 181.00
3.20 445.29
Abbreviations: ETS, environmental tobacco smoke; FEV1, forced expiratory volume in 1 s; PEF, peak expiratory flow. a This percentage reflects homes with at least one ETS day during either the winter or summer sampling period. The health models adjust for daily ETS.
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
& 2013 Nature America, Inc.
Development of temporally refined land-use regression models Johnson et al
265 data from regulatory sites with household-level LUR estimates to predict daily outdoor concentrations at the homes of asthmatic children in Windsor, Ontario. Collection of daily household-level NO2 and lung function measurements facilitated a two-phase evaluation of the temporally refined LUR estimates. We evaluated these temporally refined (daily) LUR estimates at the household level using outdoor measurements collected at the homes. We also applied the temporally refined LUR estimates in a panel study of lung function to examine the implications of using alternate metrics to estimate exposure in epidemiologic models. Temporally scaled LUR estimates explained as much of the temporal variability in daily ambient residential concentrations of NO2 as daily measurements at regulatory monitoring data in Windsor. Temporally refined models also provided a modest (10–40%) improvement in the spatial variability explained for household-level NO2 compared with regulatory site monitoring. These results suggest that temporally refined LUR models improved upon regulatory monitoring for predicting measured ambient NO2 at the household level. As expected, ambient PM2.5 showed little spatial variation; therefore, daily PM2.5 models were similar to regulatory monitoring in the proportion of variance explained in daily household concentrations. These results suggest that regulatory monitoring was adequate for estimating outdoor residential concentrations of PM2.5 due to the low spatial variability in PM2.5 mass. Previous studies have combined historic air monitoring and land-use data with current LUR models to provide historic long-term concentrations at the household level.34–37 While these models facilitated retrospective exposure assessment, the models were limited to providing long-term estimates of air pollution. Similarly, a number of studies have developed temporally refined LUR models using techniques similar to those described in the current study to examine air pollution impacts during fetal development.16,38–43 However, these models were used to estimate monthly and trimester averages rather than daily exposures. Both Novotny et al.44 and Kloog et al.45 modeled daily pollutant concentrations using a combination of LUR and remote sensing. However, while remote sensing offers a powerful tool for estimating long-term pollutant concentrations, both studies had to address limitations due to the high percentage of missing daily values provided by remotely sensed measurements of aerosol optical depth and pollutant concentrations estimated from those measurements. Finally, regression models have also been specified with both temporal and spatial predictors to estimate spatial and temporally resolved concentrations of NO246 and PAH.47 The PAH model developed by Noth et al.47 explained 80% of the spatial and 18% of the temporal variability in daily PAH concentrations, while the models developed by Su et al.46 provided significant improvement over traditional LUR models for predicting seasonal concentrations. However, both models require accurate daily or hourly meteorological inputs to predict short-term concentrations, and further evaluation of the models developed by Su et al.46 is needed to assess their efficacy in predicting daily pollutant concentrations. We further evaluated the temporally refined LUR models reported in this paper by examining the implications of using these models to estimate exposure in a panel study of lung function among asthmatic children. Effect estimates for lung function based on modeled pollutant concentrations were consistent with effects based on household-level measurements for NO2 and PM2.5. In addition, there was some evidence to suggest that both NO2 and PM2.5 had a stronger impact on both FEV1 and PEF at lower temperatures in winter. The health model results reported in this paper were included to compare the efficacy of alternate exposure metrics in a health study and these results should be interpreted with caution due to & 2013 Nature America, Inc.
the small sample size of the study population and the low levels of air pollution measured in Windsor. That said, air pollution impacts reported in this paper are consistent with previously reported associations between traffic and lung function in the larger study population.31,32,48 All children enrolled in the study met the criteria of having been diagnosed with asthma by a physician. However, between recruitment in 2004 and collection of household monitoring and lung function data in 2006, many participants discontinued use of asthma medication and some parents no longer considered their children asthmatic. Although ambient air pollution was associated with a modest decrement in lung function for these participants, there may have been stronger impacts among children with current and severe asthma. Despite these limitations, our analyses suggest that temporally refined LUR models provided a costeffective alternative to household-level monitoring for estimating exposure in a health study. We used daily and seasonal air pollution concentrations from regulatory monitoring data to temporally refine LUR models. This method provides an easily accessible, low-cost approach for estimating short-term concentrations at multiple locations within a study area. However, this approach for estimating daily household-level concentration assumes that the spatial distribution of pollutant concentrations was temporally static within seasons—for example, that temporal changes in pollution were consistent throughout the Windsor study area. The use of seasonal LUR models as the basis for this exercise allowed spatial patterns in NO2 concentrations to vary between winter and summer. However, within seasons, we assumed that spatial patterns were fixed, and that the regulatory monitoring was a reliable indicator of temporal changes throughout the study area. This simplifying assumption may have resulted in misclassification of daily household-level NO2. For example, Liu et al.49 reported significant differences in spatial variability of NO2 over time in Switzerland. In contrast, Crouse et al.50 suggested that while overall pollutant concentrations vary temporally, the spatial variability in NO2 concentrations is fairly consistent over time in Montreal, Quebec. Finally, spatial analysis and LUR modeling suggest that there were seasonal differences in the spatial distribution of NO2 in Windsor.21,25 Although there is some evidence that spatial patterns in pollutant concentration may shift over the long term within an urban area, the temporal scaling in this paper were limited to a single calendar year and seasonal LUR models were used to account for differences in spatial distribution of NO2 between winter and summer within that calendar year. To our knowledge, there are no studies that have examined whether the spatial distribution of NO2 changes significantly within seasons. Coupling LUR estimates with regulatory monitoring data to provide temporally refined household estimates provided a better estimate of daily household-level measurements compared with regulatory data alone for a spatially heterogeneous pollutant (NO2), and was consistent with household-level measurements in predicting health risks for both NO2 and PM2.5. Although further evaluation is needed in a larger study with a more geographically diverse study population, these results suggest that LUR modeling can be combined with regulatory monitoring data to predict daily household-level exposure to ambient air pollution. These findings have important implications for epidemiologic studies — in particular, for research focused on health effects associated with short-term exposure to spatially heterogeneous air pollutants.
ACKNOWLEDGEMENTS We would like to thank the scientists and field technicians who contributed to this study including Li Chen, Hongyu You, Ryan Kulka, Keith Van Ryswyk, Corrine Stocco, and Angelos Anastassopoulos (Health Canada). We also received helpful feedback on
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
Development of temporally refined land-use regression models Johnson et al
266 the draft manuscript from Nina Dobbins and Rick Burnett. Finally, we would like to thank the household participants without whom this study would not have been possible.
REFERENCES 1 Health Effects Institute 2010, Traffic-related air pollution: a critical review of the literature on emissions, exposure, and health effects. Special Report #17, 2010-0112. Available on-line at http://pubs.healtheffects.org/view.php?id=334 (accessed 02/11/2010). 2 Wilson JG, Kingham S, Pearce J, Sturman AP. A review of intraurban variations in particulate air pollution: implications for epidemiological research. Atmospheric Environ 2005; 39: 6444–6462. 3 Weijers EP, Khlystov AY, Kos GPA, Erisman JW. Variability of Particulate matter concentrations along roads and motorways determined by a moving measurement unit. Atmospheric Environ 2004; 38: 2993–3002. 4 Westerdahl D, Fruin S, Sax T, Fine P, Sioutas C. Mobile platform measurements of ultrafine particles and associated pollutant concentrations on freeways and residential streets in Los Angeles. Atmospheric Environ 2005; 39: 3597–3610. 5 Isakov V, Touma J, Khlystov A. A method of assessing air toxics concentrations using mobile platform measurements. J Air Waste Manag Assoc 2007; 57: 1286–1295. 6 Breysse PN, Buckley TJ, Williams D, Beck CM, Jo SJ, Merriman B et al. Indoor exposures to air pollutants and allergens in the homes of asthmatic children in inner-city Baltimore. Environ Res 2005; 98: 167–176. 7 Diette GB, Hansel NN, Buckley TJ, Curtin-Brosnan J, Eggleston PA, Matsui EC et al. Home indoor pollutant exposures among inner-city children with and without asthma. Environ Health Perspect 2007; 115: 1665–1669. 8 Wheeler AJ, Xu X, Kulka R, You H, Wallace L, Mallach G et al. Windsor, Ontario exposure assessment study: design and methods validation of personal, indoor, and outdoor air pollution monitoring. J Air Waste Manag Assoc 2011; 61: 324–338. 9 Williams R, Rea A, Vette A, Croghan C, Whitaker D, Wilson H et al. The design and field implementation of the Detroit Exposure and Aerosol Research Study (DEARS). J Exp Sci Environ Epidemiol 2009; 19: 643–659. 10 Jerrett M, Arain A, Kanaroglou P, Beckerman B, Potoglou D, Sahsuvaroglu T et al. A review and evaluation of intraurban air pollution exposure models. J Exp Anal Environ Epidemiol 2005; 15: 185–204. 11 Boothe V, Dimmick WF, Talbot TO. Relating air quality to environmental public health tracking data. In: Aral MM, Brebbia CA, Maslia ML, Sinks T (eds). Environmental Exposure and Health vol. 85. Wessex Institute Transactions on Ecology and the Environment: Southampton, UK, 2005, pp 43–52. 12 Isakov V, Graham S, Burke J, O¨zkaynak H. Linking air quality and exposure models. Air Waste Manag Assoc Environ Manag 2006; 26–29. 13 McConnell RB, Berhane K, Yao L, Jerrett M, Lurmann F, Gilliland F et al. Traffic, susceptibility, and childhood asthma. Environ Health Perspect 2006; 114: 766–772. 14 Isakov V, O¨zkaynak H. A modeling methodology to support evaluation of public health impacts of air pollution reduction programs. Proceedings of the 29th International Technical Meeting on Air Pollution Modeling. Aveiro: Portugal, 2007. 15 O¨zkaynak H, Palma T, Touma JS, Thurman J. Modeling population exposures to outdoor sources of hazardous air pollutants. J Expo Sci Environ Epidemiol 2008; 18: 45–48. 16 Brauer M, Hoek G, Van Vliet P, Meliefste K, Fischer PH, Wijga A et al. Air pollution from traffic and the development of respiratory infections and asthmatic and allergic symptoms in children. Am J Respir Crit Care Med 2003; 166: 1092–1098. 17 Ross Z, English PB, Scalf R, Gunier R, Smorodinsky S, Wall S et al. Nitrogen dioxide prediction in Southern California using land use regression modeling: potential for environmental health analyses. J Expo Sci Environ Epidemiol 2006; 16: 106–114. 18 Jerrett M, Arain MA, Kanaroglou P, Beckerman B, Crouse D, Gilbert NL et al. Modeling the intraurban variability of ambient traffic pollution in Toronto, Canada. J Toxicol Environ Health 2007; 70: 200–212. 19 Ryan PH, LeMasters GKA. Review of land-use regression models for characterizing intraurban air pollution exposure. Inhal Toxicol 2007; 19(Suppl 1): 127–133. 20 Hoek G, Beelen R, de Hoogh K, Vienneau D, Gulliver J, Fischer P et al. A review of land-use regression models to assess spatial variation of outdoor air pollution. Atmospheric Environ 2008; 42: 7561–7578. 21 Wheeler AJ, Smith-Doiron M, Xu X, Gilbert NL, Brook JR. Intra-urban variability of air pollution in Windsor, Ontario—measurement and modeling for human exposure assessment. Environ Res 2008; 106: 7–16. 22 Mukerjee S, Smith LA, Johnson MM, Neas LM, Stallings CA. Spatial analysis and land use regression of VOCs and NO2 from school-based urban air monitoring in Detroit/Dearborn, USA. Sci Total Environ, 2009; 407: 4642–4651. 23 Vienneau D, de Hoogh K, Beelen R, Fischer P, Hoek G, Briggs D. Comparison of land-use regression models between Great Britain and the Netherlands. Atmospheric Environ 2010; 44: 688–696.
24 Luginaah IN, Xu I, Fung K, Grgicak-Mannion A, Wintermute J, Wheeler AJ et al. Establishing the spatial variability of ambient nitrogen dioxide in Windsor, Ontario. Int J Environ Stud 2006; 63: 487–500. 25 Grgicak-Mannion A, Wintermute J, Xu X, Wheeler AJ, Luginaah I, Brook J. Geospatial land use regression modelling: seasonal and yearly variation of NO2, PM2.5 and BTEX in Windsor, Ontario. Proceeding of 2007 ISEA International Conference 2007. 26 Environment Canada: www.ec.gc.ca/rnspa-naps/default.asp?lang=En&n=5C0D33 CF-1(accessed 18 June 2012). 27 Sather ME, Slonecker ET, Kronmiller KG, Williams DD, Daughtrey H, Mathew J. Evaluation of short-term Ogawa passive, photolytic, and federal reference method sampling devices for nitrogen oxides in El Paso and Houston, Texas. J Environ Monit 2006; 8: 558–563. 28 Demokritou P, Kavouras IG, Harrison D, Ferguson ST, Koutrakis P. Development and performance evaluation of a sampler for the chemical composition analysis of PM2.5. J Air Waste Manag Assoc 2001; 51: 514–523. 29 Rasmussen PE, MacIntyre DJ, Guenette J. Buoyancy-Corrected Gravimetric Analysis System. United States Patent and Trademark Office. 2008, Patent. No. US 7,357,045 B2. Date: 15 April, 2008. 8 Claims, 3 Drawing Sheets http://patft.uspto.gov(accessed 18 June, 2012). 30 Rasmussen PE, Gardner HD, Niu J. Buoyancy-corrected gravimetric analysis of lightly loaded filters. J Air Waste Manag Assoc 2010; 60: 1065–1077. 31 Dales R, Wheeler A, Mahmud M, Frescura AM, Smith-Doiron M, Nethery E et al. The influence of living near roadways on spirometry and exhaled nitric oxide in elementary schoolchildren. Environ Health Perspect 2008; 116: 1423–1427. 32 Dales R, Wheeler AJ, Mahmud M, Frescura AM, Liu L. The influence of neighborhood roadways on respiratory symptoms among elementary schoolchildren. J Occup Environ Med 2009; 51: 654–660. 33 Liu L, Poon R, Chen L, Frescura AM, Montuschi P, Ciabattoni G et al. Acute effects of air pollution on pulmonary function, airway inflammation, and oxidative stress in asthmatic children. Environ Health Perspect 2009; 117: 668–674. 34 Gulliver J, Morris C, Lee K, Vienneau D, Briggs D, Hansell A. Land use regression modeling to estimate historic (1962-1991) concentrations of black smoke and sulfur dioxide for Great Britain. Environ Sci Technol 2011; 45: 3526–3532. 35 Chen H, Goldberg MS, Crouse DL, Burnett RT, Jerrett M, Villeneuve PJ et al. Backextrapolation of estimates of exposure from current land-use regression models. Atmospheric Environ 2010; 44: 4346–4354. 36 Clark NA, Demers PA, Karr CJ, Koehoorn M, Lencar C, Tamburic L et al. Effect of early life exposure to air pollution on development of childhood asthma. Environ Health Perspect 2010; 118: 284–290. 37 Mo¨lter A, Lindley S, de Vocht F, Simpson A, Agius R. Modelling air pollution for epidemiologic research—Part II: predicting temporal variation through land use regression. Sci Total Environ 2010; 409: 211–217. 38 Nethery E, Leckie SE, Teschke K, Brauer M 2008From measures to models: an evaluation of air pollution exposure assessment for epidemiological; studies of pregnant women. J Occup Environ Med 2008; 65: 579–586. 39 Aguilera I, Guxens M, Garcia-Esteban R, Corbella T, Nieuwenhuijsen MJ, Foradada CM et al. Association between GIS-based exposure to urban air pollution during pregnancy and birth Weight in the INMA sabadell cohort. Environ Health Perspect 2009; 117: 1322–1327. 40 Llop S, Ballester SF, Estarlich M, Esplugues A, Rebagliato M. Carmen In˜iguez. Preterm birth and exposure to air pollutants during pregnancy. Environ Res 2010; 110: 778–785. 41 Gehring U, Wijga AH, Fischer P, de Jongste JC, Kerkhof M, Koppelman GH et al. Traffic-related air pollution, preterm birth and term birthweight in the PIAMA birth cohort study. Environ Res 2011a; 111: 125–135. 42 Gehring U, van Eijsden M, Dijkema MBA, van der Wal MF, Fischer P, Brunekreef B. Traffic-related air pollution and pregnancy outcomes in the Dutch ABCD birth cohort study. Occup Environ Med 2011b; 68: 36–43. 43 Rudra CB, Williams MA, Lianne Sheppard L, Koenig JQ, Schiff MA. Ambient carbon monoxide and fine particulate matter in relation to preeclampsia and preterm delivery in western Washington state. Environ Health Perspect 2011; 119: 886–892. 44 Novotny EV, Bechle MJ, Millet DB, Marshall JD. National satellite-based land-use regression: NO2 in the United States. Environ Sci Technol 2011; 45: 4407–4414. 45 Kloog I, Koutrakis P, Coull BA, Lee HJ, Schwartz J. Assessing temporally and spatially resolved PM2.5 exposures for epidemiological studies using satellite aerosol optical depth measurements. Atmospheric Environ 2011; 45: 6267–6275. 46 Su JG, Brauer M, Bc Ainslie, Steyn D, Larson T, Buzzelli M. An innovative land use regression model incorporating meteorology for exposure analysis. Sci Total Environ 2008; 390: 520–529. 47 Noth EM, Hammond SK, Biging GS, Tager IB. A spatial-temporal regression model to predict daily outdoor residential PAH concentrations in an epidemiologic study in Fresno, CA. Atmospheric Environ 2011; 45: 2394e2403.
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
& 2013 Nature America, Inc.
Development of temporally refined land-use regression models Johnson et al
267 48 Cakmak S, Mahmud M, Grgicak-Mannion A, Dales RE. The influence of neighborhood traffic density on the respiratory health of elementary schoolchildren. Environment Int 2012; 39: 128–132. 49 Liu LJ, Tsai MY, Keidel D, Gemperli A, Ineichen A, Hazenkamp-von Arx M et al. Long-term exposure models for traffic related NO 2 across
& 2013 Nature America, Inc.
geographically diverse areas over separate years. Atmospheric Environ 2011; 46: 460–471. 50 Crouse DL, Goldberg MS, Ross NA. A prediction-based approach to modelling temporal and spatial variability of traffic-related air pollution in Montreal, Canada. Atmospheric Environ 2009; 43: 5075–5084.
Journal of Exposure Science and Environmental Epidemiology (2013), 259 – 267
