You are here

Spatial statistical learning methods for estimating ambient air pollution

Principal Investigator: 

Utrecht University

Hoek and colleagues will prepare maps of modeled annual average air pollution across the Netherlands, validate the maps using new measurements from over 100 sites, and evaluate the performance of several exposure models. The investigators will conduct cross-comparisons to evaluate how different exposure assessment methods compare in their ability to predict long-term pollutant concentrations, with a particular focus on spatial variability of pollutants.

Funded under

Comparison of Long-Term Air Pollution Exposure Assessment Based on Mobile Monitoring, Low-Cost Sensors, Dispersion Modelling, and Routine Monitoring-Based Models

Gerard Hoek1, Kees de Hoogh2, Ulrike Gehring1, Roel Vermeulen1, Sjoerd van Ratingen3, Erik Tielemans3, Nicole Janssen3, Joost Wesseling3

1Institute for Risk Assessment Sciences, Utrecht University, the Netherlands; 2Swiss Tropical and Public Health Institute, Basel, Switzerland; 3National Institute for Public Health and the Environment, Bilthoven, the Netherlands

Background: Assessment of long-term exposure to (traffic-related) outdoor air pollution remains a major challenge for epidemiological studies. One challenge is the characterization of the spatial variation of the ambient concentrations of key traffic-related air pollutants including ultrafine particles (UFP), Black carbon (BC) and NO2 as these pollutants vary on a fine spatial scale. Recently epidemiological studies have used different approaches, including land use regression (LUR) models based upon mobile monitoring (UFP, BC), models based upon low cost sensors (PM2.5, NO2) or routine monitoring data and increasingly sophisticated hybrid models incorporating routine surface monitoring, satellite, chemical transport and land use data. Very little information is available about the relative performance of these different approaches to assess long-term exposure to traffic-related air pollution. Differences in performance may affect conclusions drawn from epidemiological studies applying different exposure assessment approaches.

Our study addresses these challenges by comparing commonly used air pollution exposure assessment methods both in terms of performance of assessing spatial variation and in air pollution effect estimates in selected epidemiological studies in the Netherlands. In most approaches, we will assess UFP, NO2, BC and PM2.5.

The specific aims of the proposal are: 

  1. Develop long-term ambient air pollution exposure estimates for selected epidemiological studies based upon low-cost sensors, mobile and stationary monitoring and deterministic dispersion modelling;
  2. Compare different exposure assessment methods in terms of their ability to predict spatial variation of long-term average concentrations using external validation data;
  3. Compare different exposure assessment methods in terms of air pollution effect estimates in selected epidemiological studies.

Methods: The project will generate and evaluate annual average air pollution maps using seven different exposure assessment methods, which differ in modeling approach (empirical LUR, deterministic dispersion models and hybrid models) and monitoring data used (low cost sensors, mobile monitoring). For all empirical models we will test three model development algorithms covering major families of modeling approaches: supervised linear regression, random forest and elastic net.

The predictions and performance of the models will be compared at 20,000 addresses across the Netherlands and tested on newly collected and existing external validation data. A unique feature of our project is access to an annual average UFP map in the city of Amsterdam for the year 2002-2004, allowing assessment of how well currently developed models predict past UFP exposure. Epidemiological analyses in three cohort studies in which we previously found associations with one of the evaluated models will be conducted to compare health risk estimates of the different exposure assessment methods. The studies include a national administrative cohort, a classical cohort study and a mature birth cohort in which we will assess mortality, cardiovascular disease incidence and lung function / asthma respectively.