Elsevier

Environmental Research

Volume 156, July 2017, Pages 201-230
Environmental Research

Comparison of spatiotemporal prediction models of daily exposure of individuals to ambient nitrogen dioxide and ozone in Montreal, Canada

https://doi.org/10.1016/j.envres.2017.03.017Get rights and content

Highlights

  • We predicted daily residential exposure to O3 and NO2 using four space-time methods.

  • We found substantial differences in daily estimates from these different methods.

  • The spatial patterns of agreement differed across pair of methods and pollutants.

  • For O3 but not NO2, greater disagreement was mostly near downtown and highways.

Abstract

Background

In previous studies investigating the short-term health effects of ambient air pollution the exposure metric that is often used is the daily average across monitors, thus assuming that all individuals have the same daily exposure. Studies that incorporate space-time exposures of individuals are essential to further our understanding of the short-term health effects of ambient air pollution.

Objectives

As part of a longitudinal cohort study of the acute effects of air pollution that incorporated subject-specific information and medical histories of subjects throughout the follow-up, the purpose of this study was to develop and compare different prediction models using data from fixed-site monitors and other monitoring campaigns to estimate daily, spatially-resolved concentrations of ozone (O3) and nitrogen dioxide (NO2) of participants’ residences in Montreal, 1991–2002.

Methods

We used the following methods to predict spatially-resolved daily concentrations of O3 and NO2 for each geographic region in Montreal (defined by three-character postal code areas): (1) assigning concentrations from the nearest monitor; (2) spatial interpolation using inverse-distance weighting; (3) back-extrapolation from a land-use regression model from a dense monitoring survey, and; (4) a combination of a land-use and Bayesian maximum entropy model. We used a variety of indices of agreement to compare estimates of exposure assigned from the different methods, notably scatterplots of pairwise predictions, distribution of differences and computation of the absolute agreement intraclass correlation (ICC). For each pairwise prediction, we also produced maps of the ICCs by these regions indicating the spatial variability in the degree of agreement.

Results

We found some substantial differences in agreement across pairs of methods in daily mean predicted concentrations of O3 and NO2. On a given day and postal code area the difference in the concentration assigned could be as high as 131 ppb for O3 and 108 ppb for NO2. For both pollutants, better agreement was found between predictions from the nearest monitor and the inverse-distance weighting interpolation methods, with ICCs of 0.89 (95% confidence interval (CI): 0.89, 0.89) for O3 and 0.81 (95%CI: 0.80, 0.81) for NO2, respectively. For this pair of methods the maximum difference on a given day and postal code area was 36 ppb for O3 and 74 ppb for NO2. The back-extrapolation method showed a higher degree of disagreement with the nearest monitor approach, inverse-distance weighting interpolation, and the Bayesian maximum entropy model, which were strongly constrained by the sparse monitoring network. The maps showed that the patterns of agreement differed across the postal code areas and the variability depended on the pair of methods compared and the pollutants. For O3, but not NO2, postal areas showing greater disagreement were mostly located near the city centre and along highways, especially in maps involving the back-extrapolation method.

Conclusions

In view of the substantial differences in daily concentrations of O3 and NO2 predicted by the different methods, we suggest that analyses of the health effects from air pollution should make use of multiple exposure assessment methods. Although we cannot make any recommendations as to which is the most valid method, models that make use of higher spatially resolved data, such as from dense exposure surveys or from high spatial resolution satellite data, likely provide the most valid estimates.

Introduction

The association between short-term variations in air pollution and health outcomes are most often investigated using grouped analyses of parallel time series or grouped case-crossover designs (Goldberg et al., 2003), but can also be investigated in panel studies (Buteau and Goldberg, 2016) and in longitudinal cohort studies (Goldberg and Burnett, 2005). In time series studies of associations between daily mortality, or other events from routinely collected data, and daily concentrations of pollutants from a fixed-site monitoring network in a circumscribed geographical area, the exposure metric that is often used is the daily average across monitors (Ozkaynak et al., 2013a, Ozkaynak et al., 2013b). It is thus assumed that all subjects have the same exposure on a given day. For air pollutants that are relatively spatially homogenous (e.g., fine particulate matter) these metrics can be used to describe changes in exposures over populations and thus more refined exposure assessment methods may not provide any more information on the associations (Baxter et al., 2013, Dionisio et al., 2016). For air pollutants that have greater spatial variability, such as nitrogen dioxide (NO2), which is a marker of traffic-related pollution (Health Effects Institute, 2010), the daily mean may provide reasonable estimates of the time varying component if it does not vary dramatically in space. An issue with many time series studies is that exposure is estimated from routine monitoring systems that are not dense enough to capture small-scale variability. Indeed, it has been found that the temporal variability of traffic-related air pollutants may differ spatially in metropolitan area (Dionisio et al., 2013), and stronger associations for respiratory outcomes have been found, for example, when using spatially-resolved estimates of carbon monoxide and nitrogen oxides at the American postal code level (5-digit ZIP codes) as compared to that estimated using fixed-site monitors (Sarnat et al., 2013).

On the other hand, for panel studies and longitudinal analyses of cohort studies that incorporate time-dependent exposure of individuals, capturing not only temporal changes but also spatial variability in exposure is critical (Baxter et al., 2013). In some panel studies, personal monitoring has been carried out (e.g., Maikawa et al., 2016) so that spatial-temporal patterns for many pollutants were estimated. In large longitudinal studies, personal monitoring or dense monitoring networks that capture small-area variations may not be practical, so that it is often necessary to use data from existing fixed-site monitoring stations to predict statistically concentrations at times and places in which measurements were not made. Methods used in epidemiological studies to predict spatiotemporal exposure to air pollution include time-varying indicators of residential proximity to important sources (e.g., distance to high traffic-density roads (Brauer et al., 2008) or industries (Labelle et al., 2015, Lewin et al., 2013)), assigning the measurements from the nearest fixed-site monitor to participants’ residences (e.g., Basu et al., 2004; Brauer et al., 2008), and spatial interpolation using inverse-distance weighting (e.g., Brauer et al., 2008; Lin et al., 2015) or kriging (Beelen et al., 2009). Multivariate land-use regression modelling is another method of spatial interpolation that is usually derived from spatially dense monitoring campaigns (Crouse et al., 2009, Hoek et al., 2008, Jerrett et al., 2005), but unless multiple measurements are made through time these methods do not have a temporal component.

Advances in developing space-time models at refined scales include use of remote sensing data from satellites (Kloog et al., 2011), dispersion and atmospheric chemical models (Hennig et al., 2016), kriging with external drift (Ramos et al., 2016), and sophisticated hierarchical or hybrid models that can accommodate the unbalanced nature of monitoring data from different campaigns (Keller et al., 2015), or combine data from different sources (e.g., measurements from fixed-monitors and predictions from a land use regression model) in a Bayesian framework (e.g., Adam-Poupart et al., 2014; de Nazelle et al., 2010; Reyes and Serre, 2014; Vicedo-Cabrera et al., 2013; Xu et al., 2016; Yu et al., 2009).

The present analysis is part of a longitudinal study of the acute effects of air pollution that incorporates subject-specific information, including location of residences and medical histories of subjects throughout the follow-up. The purpose of this study was to develop and compare different prediction models using data from fixed-site monitors and other monitoring campaigns to spatiotemporally estimate concentrations of daily concentrations of ozone (O3) and NO2 to subjects in our cohort. Specifically, we used a variety of indices of agreement to compare estimates of exposure assigned from the following four methods: (1) assigning concentrations from the nearest monitor; (2) spatial interpolation using inverse-distance weighting; (3) back-extrapolation from a land-use regression model; and (4) a combination of a land-use and Bayesian maximum entropy model (Adam-Poupart et al., 2014).

Section snippets

Study population, period and geographic unit for exposure estimation

The cohort of subjects for whom we had health data and information about their residence was limited to residents of Montreal for the period from January 1, 1991 to December 31, 2002. Because of confidentiality restrictions, we were only provided the first three characters of the six-character postal codes of participants’ residential addresses during the follow-up. The first three characters of the Canadian postal code represent what is called a forward sortation area, and allows mail to be

Predicted concentrations of ozone and nitrogen dioxide

Table 1 presents descriptive statistics for the metrics used to estimate daily exposure to O3 (May-September) and NO2 (whole year) at the three-character postal code districts for 1991–2002. For O3, the four methods provided for 1836 days 8-h mean estimates for all 98 postal code districts. The nearest station, inverse-distance weighting, and BME yielded similar mean estimates (ranging from 27.91 to 30.00 ppb), but the BME had a lower standard deviation and a narrower range. The

Discussion

Accurate spatiotemporal estimation of concentrations of air pollutants is essential to further our understanding of the short-term health effects of ambient air pollution. It is challenging to assess participants’ exposure and different methods used to predict spatiotemporal concentrations of air pollutants have been developed and used in acute air pollution epidemiologic studies (Jerrett et al., 2005, Ozkaynak et al., 2013a, Ozkaynak et al., 2013b). In this paper, our main goal was to develop

Conclusion

We developed and compared different prediction models to spatiotemporally estimate daily ambient air pollution concentrations at the residences of participants of our longitudinal study. We showed that depending on the methods used to predict concentrations there could be substantial differences in the daily mean air pollutant exposure assigned to an individual using a residential postal code district. The back-extrapolation method extended to obtain daily exposure, which is based on a detailed

Funding source

This work was supported by the Canadian Institute for Health Research (Doctoral Award - Frederick Banting and Charles Best Canada Graduate Scholarship (201310GSD)) to Stephane Buteau.

Conflict of interest

none declared.

References (65)

  • G.Y. Lu et al.

    An adaptive inverse-distance weighting spatial interpolation technique

    Comput. Geosci.

    (2008)
  • J.D. Marshall

    Within-urban variability in ambient air pollution: comparison of estimation methods

    Atmos. Environ.

    (2008)
  • Y. Sellier

    Health effects of ambient air pollution: do different methods for estimating exposure lead to different results?

    Environ. Int.

    (2014)
  • J.-Y. Son

    Individual exposure to air pollution and lung function in Korea: spatial analysis using multiple exposure approaches

    Environ. Res.

    (2010)
  • P.F. Watson et al.

    Method agreement analysis: a review of correct methodology

    Theriogenology

    (2010)
  • J. Wu

    Comparing exposure assessment methods for traffic-related air pollution in an adverse pregnancy outcome study

    Environ. Res.

    (2011)
  • A. Adam-Poupart

    Spatiotemporal modeling of ozone levels in Quebec (Canada): a comparison of kriging, land-use regression (LUR), and combined Bayesian maximum entropy-LUR approaches

    Environ. Health Perspect.

    (2014)
  • D.G. Altman

    Practical Statistics for Medical Research

    (1991)
  • H.X. Barnhart

    An overview on assessing agreement with continuous measurements

    J. Biopharm. Stat.

    (2007)
  • R. Basu

    Comparing exposure metrics in the relationship between PM2.5 and birth weight in California

    J. Expo. Anal. Environ. Epidemiol.

    (2004)
  • L.K. Baxter

    Exposure prediction approaches used in air pollution epidemiology studies: key findings and future recommendations

    J. Expo. Sci. Environ. Epidemiol.

    (2013)
  • B.S. Beckerman

    A hybrid approach to estimating national scale spatiotemporal variability of PM2.5 in the contiguous United States

    Environ. Sci. Technol.

    (2013)
  • T. Bellander

    Using geographic information systems to assess individual historical exposure to air pollution from traffic and house heating in Stockholm

    Environ. Health Perspect.

    (2001)
  • J.M. Bland et al.

    Statistical methods for assessing agreement between two methods of clinical measurement

    Lancet

    (1986)
  • W. Bloss

    Chapter 3 Atmospheric Chemical Processes of Importance in Cities. Air Quality in Urban Environments

    (2009)
  • M. Brauer

    A cohort study of traffic-related air pollution impacts on birth outcomes

    Environ. Health Perspect.

    (2008)
  • D.J. Briggs

    Mapping urban air pollution using GIS: a regression-based approach

    Int. J. Geogr. Inf. Sci.

    (1997)
  • P.K. Choudhary et al.

    Measuring agreement in method comparison studies — a review

  • G. Christakos

    Modern Spatiotemporal Geostatistics

    (2000)
  • G. Christakos

    Temporal GIS: Advanced Functions for Field-based Applications

    (2001)
  • D.L. Crouse

    Postmenopausal breast cancer is associated with exposure to traffic-related air pollution in Montreal, Canada: a case-control study

    Environ. Health Perspect.

    (2010)
  • D. Deligiorgi et al.

    Spatial Interpolation Methodologies in Urban Air Pollution Modeling: Application for the Greater Area of Metropolitan Athens, Greece

  • Cited by (16)

    View all citing articles on Scopus
    View full text