An official website of the United States government
Here's how you know
Official websites use .gov
A
.gov website belongs to an official
government organization in the United States.
Secure .gov websites use HTTPS
A lock (
) or https:// means you've safely
connected to the .gov website. Share sensitive
information only on official, secure websites.
As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with,
the contents by NLM or the National Institutes of Health.
Learn more:
PMC Disclaimer
|
PMC Copyright Notice
. 2025 Oct 1;2025:229.
Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects
Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects
This Statement, prepared by the Health Effects Institute, summarizes a research project funded by HEI and conducted by Dr. Kees de Hoogh at the Swiss Tropical and Public Health Institute and colleagues. Research Report 229 contains the detailed Investigators’ Report and a Commentary on the study prepared by the HEI Improved Exposure Assessment Studies Review Panel.
There remain important limitations and challenges when estimating long-term exposure to outdoor air pollution for epidemiological studies. To address these challenges, the Health Effects Institute issued Request for Applications 19-1 to develop and apply novel, scalable approaches to improve assessments of long-term exposure to outdoor air pollutants that vary highly across space and time.
de Hoogh was one of five investigators funded under this Request for Application. de Hoogh and colleagues assessed whether exposure estimates that accounted for mobility would improve exposure assessment and potentially reduce exposure measurement bias in health studies in Switzerland and the Netherlands. The study addressed an important outstanding issue because most health studies to date have assessed long-term air pollution exposure estimated as outdoor concentrations at participants’ residential locations only.
APPROACH
de Hoogh and colleagues simulated participants’ mobility patterns using agent-based modeling. Agent-based modeling is a computer modeling approach that simulates actions and interactions among people, things, places, and time, and allows for the integration of individual and population behaviors. Their simulations were informed by existing travel survey and census data from 2010 to 2019 in Switzerland and the Netherlands. The investigators developed 13 activity profiles to represent different time–activity patterns (e.g., commuting) using characteristics available (e.g., employment status) from the three epidemiological studies. Except for a subset of participants in one cohort, workplace locations were unknown and were estimated. They repeated the simulations between 50 and 1,000 times and evaluated the model using GPS location data, which were collected over 2 weeks in tracking campaigns that included nearly 700 participants in the two countries. The GPS location was determined using a mobile phone app and a dedicated tracking device.
The investigators developed nationwide hourly concentration maps of nitrogen dioxide and fine particulate matter, and they linked those to the time–activity data to estimate exposure. In Switzerland, the investigators used annual average concentration maps for 2016, derived from previously published spatiotemporal models, and applied a temporal adjustment approach to obtain hourly estimates. In the Netherlands, hourly land use regression models were developed specifically for this study using regulatory monitoring data from 2016 to 2019, stratified by season and weekday type (weekday or weekend day).
What This Study Adds.
The study evaluated whether long-term exposure estimates that account for people’s mobility would improve exposure assessment, using novel agent-based modeling.
The investigators compared residential-only and mobility-enhanced estimates of exposure to nitrogen dioxide and fine particulate matter in three adult cohort studies in Switzerland and the Netherlands.
Similar associations were reported between residential and mobility-enhanced exposure estimates and various health outcomes, and the estimates themselves were highly correlated.
The results suggest that the bias in epidemiological studies based on outdoor concentrations at residential locations might be small and that accounting for mobility might not be important.
The investigators then conducted epidemiological analyses using three adult population-based cohort studies: the Swiss National Cohort (SNC), the European Prospective Investigation into Cancer, Netherlands (EPIC-NL), and the Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults (SAPALDIA). They compared health effects estimates based on residential versus mobility-enhanced exposure estimates. For all cohort participants, the annual average residential exposure was assigned at the baseline home address (1993–2011, depending on the cohort). The mobility-enhanced exposure was assigned based on the most appropriate activity profile per participant, depending on their sex, employment status, and socioeconomic status. The investigators used single-pollutant Cox proportional hazards models for mortality and morbidity outcomes in the SNC and EPIC-NL and repeated cross-sectional linear regression analyses for lung function and blood pressure outcomes in SAPALDIA. All results were adjusted for age, sex, and individual and area-level socioeconomic status. Results from the analyses of EPIC-NL and SAPALDIA were further adjusted for various lifestyle factors. An additional cross-sectional analysis was conducted in a subset of SAPALDIA with known workplace addresses to investigate the importance of known versus estimated workplace location.
KEY RESULTS
The mobility-enhanced exposure estimates were highly correlated (correlation > 0.95) with residential-only estimates for nitrogen dioxide and fine particulate matter among the cohort participants. In all cohorts, slightly lower mean air pollution exposures and contrasts (interquartile range) in individuals’ exposures were estimated for the mobility-enhanced exposures compared to the residential exposures. For both exposures, the contrast was higher for nitrogen dioxide than for fine particulate matter, consistent with the larger spatial variation of nitrogen dioxide.
The residential and mobility-enhanced exposure estimates showed similar associations with various health outcomes within the cohorts. Hence, the investigators concluded that using mobility-enhanced exposure estimates compared to residential-only estimates did not reduce bias in epidemiological studies.
Positive (adverse) associations were observed most clearly in the SNC (Statement Figure). The investigators found that exposure to nitrogen dioxide and fine particulate matter was associated with an increased risk of natural mortality and that exposure to nitrogen dioxide, but not exposure to fine particulate matter, was associated with an increased risk of cardiovascular mortality. They reported a lack of association and, in some instances, even negative associations for the various health outcomes in the EPIC-NL and SAPALDIA cohorts. No consistent difference in the association was found with exposures using the known versus estimated workplace locations in the subset of SAPALDIA participants for which workplace locations were available.
Association between nitrogen dioxide (NO2) and fine particulate matter (PM2.5) and mortality in the Swiss National Cohort.
INTERPRETATION AND CONCLUSIONS
In its independent review, the HEI Review Panel considered the study to be well-motivated and determined that it effectively leveraged a wealth of air pollution and health data across two countries. The Panel found that the study had several strengths. First, the use of agent-based modeling was considered a novel and useful approach to account for mobility in pollutant exposure estimates for large populations. Second, linking time–activity data to hourly air pollution models enabled the investigators to capture diurnal variations in air pollution. Third, the Panel recognized the value of the application of the various exposure estimates in relation to health outcomes in three cohort studies. In particular, the health analysis for a large population (3.5 million participants) that included all Swiss adult citizens 30 years or older was considered informative.
Although the Panel broadly agreed with the investigators’ conclusions, some limitations should inform the interpretation of the results. The evaluation of agent-based modeling using tracking data was limited because of the small convenience sample, although the Panel appreciated the comparison of two different tracking devices. The Panel was concerned about the uncertainty associated with the estimated workplace locations. They also noted that the exposure estimates did not incorporate infiltration rates of pollutants into homes or other buildings. Furthermore, the study did not account for exposure induced by different modes of transportation, nor did it examine pollutants other than nitrogen dioxide and fine particulate matter.
Some uncertainties were noted in the health analyses of EPIC-NL, such as the temporal mismatch between the period captured by the exposure model (2016–2019) and the exposure window most relevant for epidemiological purposes (1993–2013). Moreover, the lack of associations in the EPIC-NL and SAPALDIA cohorts hampered the assessment of the influence of mobility and known (rather than estimated) workplace location on health estimates. Replication of findings in other settings is needed, given these limitations and the potentially limited generalizability of the study populations and time–activity patterns in relatively well-off Western European countries.
Overall, the study contributed new knowledge to exposure assessment for epidemiological research and generated findings that will be of broad interest and value to a wide audience. The study found slight differences in residential and mobility-enhanced exposure estimates for nitrogen dioxide and fine particulate matter, resulting in similar associations between exposure estimates and health outcomes. The results suggest that the exposure measurement bias in epidemiological studies based on outdoor concentrations at residential locations might be small and that accounting for mobility might not be an important consideration.
Although this is reassuring, further research is needed — especially research about other pollutants and in other locations and populations — to examine fully the added value of collecting mobility information on a large scale in air pollution cohort studies.
Res Rep Health Eff Inst. 2025 Oct 1;2025:229.
Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects
1
Swiss Tropical and Public Health Institute, Allschwil, Switzerland
2
University of Basel, Switzerland
3 Department of Physical Geography (Geo), Utrecht University, the Netherlands
4 Department of Geography, University of Bayreuth, Germany
5
Institute of Risk Assessment Sciences (IRAS), Utrecht University, the Netherlands
This Investigators’ Report is one part of Health Effects Institute Research Report 229, which also includes a Commentary by the Improved Exposure Assessment Studies Review Panel and an HEI Statement about the research project. Correspondence concerning the Investigators’ Report may be addressed to Dr. Kees de Hoogh, Swiss Tropical and Public Health Institute, Kreuzstrasse 2, CH-4123, Allschwil, Switzerland; email: c.dehoogh@swisstph.ch. No potential conflict of interest was reported by the authors.
Although this report was produced with partial funding by the United States Environmental Protection Agency under Assistance Award CR–83998101 to the Health Effects Institute, it has not been subjected to the Agency’s peer and administrative review and might not reflect the views of the Agency; no official endorsement by it should be inferred. The contents of this report also have not been reviewed by private party institutions, including those that support the Health Effects Institute; therefore, it might not reflect the views or policies of these parties, and no endorsement by them should be inferred.
Large-scale epidemiological studies investigating long-term health effects of air pollution can typically only consider the residential locations of the participants, thereby ignoring the space–time–activity patterns that likely influence total exposure. People are mobile and can be exposed to considerably different levels of air pollution or air pollution mixtures when inside versus outside, commuting, recreating, or working. Neglecting these mechanisms in exposure assessment may lead to incorrect distributions of exposure over the population, which may, subsequently, lead to incorrect exposure–health relations in epidemiological studies. In this study, we investigated whether a more sophisticated mobility-enhanced exposure assessment would lead to different exposure predictions and health effect estimates compared with using a residential-based exposure.
Methods
Agent-based modeling (ABM*) was used to model mobility patterns in Switzerland and the Netherlands based on travel survey information. Hourly air pollution surfaces of nitrogen dioxide (NO2) and fine particulate matter (PM2.5) developed separately for the Netherlands and Switzerland, for weekdays and weekends, were overlaid with the ABM data to extract exposures. These air pollution exposures were assigned to two adult cohorts in Switzerland — the Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults (SAPALDIA) and the Swiss National Cohort (SNC) — and the European Prospective Investigation into Cancer study adult cohort in the Netherlands (EPIC-NL). Exposures were assigned based on (1) residential address location only (residential-based) and (2) residential and work address locations plus mobility (mobility-enhanced). In the case of SAPALDIA, known work address locations were available and additionally used. Associations with health outcomes (natural and cardiovascular mortality, coronary and stroke events, blood pressure, and lung function) in the three cohorts were investigated. To evaluate the performance of the ABM, we collected GPS readings from 489 participants in Switzerland and 189 participants in the Netherlands in tracking campaigns. The participants recorded GPS readings, using both a wearable GPS recording device and a mobile phone app while also recording their time–activity in the app diary.
Results
We successfully developed mobility-enhanced exposures for over 3 million participants, including an assessment of uncertainty. We found a good agreement between exposures estimated with the app and the GPS tracker, supporting the scalability of the approach. We evaluated the ABMs with GPS and time–activity data collected independently in tracking campaigns that included almost 700 participants from selected areas in the two countries. For these participants, the exposures based on GPS measurements versus those derived from ABM showed a moderate to good agreement (R2= 0.52–0.81). Within the three cohorts, when compared with exposure based on only the residential location, the mobility-enhanced exposure showed very high correlations (R > 0.95). Finally, the epidemiological analyses revealed very small differences in the associations across health outcomes for the different exposure estimates (mortality in SNC; cardiovascular morbidity and mortality in EPIC-NL; and lung function and blood pressure in SAPALDIA) within the three cohorts. In SAPALDIA, where the work address was known for a subset of individuals, a further comparison using the estimated work address in the ABM indicated little difference in mobility-enhanced exposures.
Conclusions
Our results suggest that the assessment of air pollution exposure at the residential address in epidemiological studies generally does not lead to substantial bias in health effects estimates. If time–activity patterns in other study areas differ greatly from the patterns analyzed in our study, differences between residential and activity-enhanced exposures may be larger. Despite the good agreement between residential and work locations, exposure research should continue to strive toward improving exposure assessment in large-scale epidemiological studies to minimize exposure misclassification.
1. INTRODUCTION
Air pollution has been associated with various adverse health effects.1 The World Health Organization (WHO) recently reported that, globally, an estimated 4 million deaths annually are due to long-term ambient air pollution.2 In Global Burden of Disease assessments, air pollution ranks number one as the most influential environmental exposure;3 the evidence for these assessments was derived primarily from epidemiological studies, especially those on long-term air pollution exposure. A panel appointed by HEI recently reviewed epidemiological studies of traffic-related air pollution, addressing exposure assessment issues in detail.4,5 The panel primarily assessed how well specific methods assessed the outdoor concentration, as well as the spatial alignment of outdoor exposure as assigned to the residential address (e.g., by assessing the spatial resolution of the address and exposure surfaces). Virtually all studies assigned outdoor concentrations to the residential address only, while a few studies in children incorporated exposures at the school address. Even fewer studies in adults have incorporated work address in the exposure assessment.
Environmental health researchers have been aware that the true personal exposure to air pollution is experienced in multiple so-called microenvironments.6,7 Personal exposure can be assessed directly by personal exposure monitoring or indirectly by evaluating concentrations in key microenvironments and obtaining time–activity data.6 There is a large exposure science literature on both approaches. However, in large epidemiological studies, direct and indirect personal exposure has seldom been assessed. The main reasons for this are that personal exposure monitoring is too costly to perform in a large number of participants, and assessing long-term exposure requires a fair number of repeated samples per subject. Indirect exposure assessment is not often applied because most epidemiological studies do not have information on where people spend time other than the home. The lack of time–activity data beyond the residential address is related to the fact that very few studies have been designed to primarily investigate the health effects of air pollution. In recent large studies based on administrative databases, data on residential address is also the only location data typically available.8,9 The few studies that were designed specifically to investigate the health effects of outdoor air pollution, including the Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults (SAPALDIA) study in Switzerland and the Prevention and Incidence of Asthma and Mite Allergy birth cohort study in the Netherlands, did obtain more detailed data on work or school address.10,11
Epidemiological studies on air pollution and other environmental exposures have been criticized for not taking time–activity patterns into account.12,13 While it is understandable that epidemiological studies have relied on residential address, questions remain regarding the extent of exposure measurement error and how much bias is potentially introduced in health effects estimates by focusing on the residential address only.
The Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects (MOBI-AIR) project attempts to answer this question by investigating whether more sophisticated estimates of individual exposure that consider population mobility could decrease bias in health studies. We used agent-based modeling (ABM) to simulate mobility and commuting tracks in the Netherlands and Switzerland for a number of profiles and combined these with detailed space–time air pollution data to enable calculation of “mobility-enhanced” nitrogen dioxide (NO2) and fine particulate matter (PM2.5) exposure estimates for every individual in the selected cohorts. In parallel, a tracking campaign in both Switzerland and the Netherlands using a purpose-designed mobile app captured mobility data for almost 700 individuals. This information was used to compare exposures produced via ABM with those based on real activity data. Exposure estimates were used to assess associations with selected health outcomes in three large cohort studies: SAPALDIA, the European Prospective Investigation into Cancer Study in the Netherlands (EPIC-NL), and the Swiss National Cohort (SNC). The influence of residential versus mobility-enhanced exposures was also evaluated.
2. SPECIFIC AIMS
We tested our hypothesis — that more accurate estimates of individual-based air pollution exposure, including population mobility, reduce bias in health studies — through the following specific aims:
To derive spatiotemporal mobility patterns that can be scaled to the general population. This was achieved through the conduct of a tracking campaign in Switzerland and the Netherlands, on a large number of individuals, to collect mobility patterns to inform the ABM, exploiting routine time–activity survey data.
To quantify long-term individual traffic-related air pollution exposures by incorporating spatiotemporal mobility patterns (from Aim 1) using enhanced existing modeled air pollution surfaces for NO2, ultrafine particles (UFP), PM2.5 elemental composition (copper [Cu], iron [Fe], and zinc [Zn]), black carbon (BC) and fine particulate matter (PM2.5), achieving a range of individual exposure estimates.
To evaluate the gain in performance of progressively sophisticated air pollution exposures by computing the exposure measurement error by comparing the range of air pollution exposures from ABM with personal measurement datasets from past studies, and simulating exposure error using realistic error scenarios.
To determine whether more sophisticated air pollution exposures lead to reduced bias in studies on long-term health effects through direct investigation of the associations using three large cohorts in Switzerland and the Netherlands, and the different exposure estimates.
3. STUDY DESIGN AND METHODS
3.1 OVERALL DESIGN
From the outset, we made changes to the initial specific aims, which are reflected in the final design of the study, shown in Figure 1.
Design of MOBI-AIR with links between different tasks (in blue) and datasets (in grey).
We were forced to make one change related to Aim 1: Because the start of the COVID-19 pandemic coincided with the start of our study, in April 2020, the tracking campaigns were delayed severely, such that we could not, as planned, use the mobility patterns collected from the tracking campaigns to inform ABM. Instead, we compared ABM-simulated mobility patterns based on travel survey information (Task 1) to data collected in tracking campaigns (Task 2) carried out in Switzerland and the Netherlands. In the Netherlands, hourly air pollution surfaces for weekdays and weekend days were developed using specific land use regression models. In Switzerland, existing air pollution surfaces were rescaled using diurnal patterns measured by routine monitoring stations (Task 3). The ABM data were then overlaid with the hourly exposure surfaces to extract the exposures.
Additionally, we decided early in the study to make a change to Aim 2 and focus only on NO2 and PM2.5 and not assess UFPs, PM2.5 elemental composition (Cu, Fe, and Zn), or BC. The reason for this was threefold: First, we were not able to develop models at the required hourly resolution because the monitoring data required to inform our air pollution models were sparse. Hourly monitoring data for UFP pollutants during the study period were available in up to five sites in the Netherlands, but not continuously at the same site, and some not hourly. BC was measured at 28 sites — still an insufficient number to inform the required hourly statistical air pollution models. The situation was similar in Switzerland: only nine stations measured elemental carbon in PM2.5 (daily values), and UFPs were not measured systematically. Second, NO2 and PM2.5 are key pollutants of interest, with the most evidence in terms of health effects and with set European limit values and WHO guidelines. Third, in general, the traffic-related exposures are highly correlated spatially (e.g., as found in the Effects of Low-Level Air Pollution: A Study in Europe [ELAPSE] for NO2, BC, Cu, Fe, and Zn14). UFPs are also highly correlated spatially with other traffic-related pollutants.15,16 The modeled exposures were assigned to two cohorts in Switzerland (SAPALDIA and SNC) and the EPIC-NL cohort in the Netherlands based on (1) residential address location only; (2) residential address and work address locations plus mobility (mobility enhanced); and (3) in the case of SAPALDIA, known work address location for the working population. For SNC and EPIC-NL, the work location was estimated, informed by national statistics. Associations with health outcomes in the three cohorts were investigated (Task 4), answering the question of whether exposure assessment, including mobility and work location exposures, resulted in different effect estimates compared with the commonly assessed residential exposure. Last, we evaluated the potential measurement error (Task 5) in our modeled exposures by comparing personal measurement data collected in the former EXPOsOMICS project with the ABM estimates. The EXPOsOMICS project aimed to develop new approaches to improve the exposure assessment of key pollutants, aiding the characterization of the exposome.17 Part of these personal exposure measurements was conducted in a number of study areas during 2013–2015, including in Switzerland and the Netherlands (see also Additional Materials Section S5.1). In addition, we compared the ABM-modeled exposures with the exposure taken from the tracking campaign with actual time–activity patterns. Figure 1 gives a schematic overview of how the tasks link up. The following sections provide more detail about the tasks tracked.
3.2 AGENT-BASED MODELING
ABMs are computer models used to capture the behavior of individuals and groups (agents) within an environment based on spatial and statistical demographic data.18 We used ABM to simulate travel behavior to better characterize individual exposure to traffic-related air pollution. Simulated tracks for commute, together with the location of the home and workplace, were overlaid with air pollution exposure surfaces of hourly temporal resolution to calculate time-weighted exposures.
Figure 2 shows the general ABM approach for a person or agent. From left to right, it describes the profile of a person or agent combined with their location-specific data (e.g., location of home), which is used to create an activity calendar and locations visited during each activity. With this information, the ABM calculates locations and routes visited by the person or agent over time. This is combined with air pollution data to calculate individual exposure. Simulations of diurnal patterns do not need to be run, for example, over an entire year; rather, they are run for particular time periods having approximately constant diurnal patterns in activity and air pollution (e.g., weekdays and weekend days). This can be extended to represent variation over seasons or months, depending on the level of detail in the available air pollution and time–activity data.
ABM framework for exposure assessment combining travel survey data, which inform activity schedule and spatial context models, to calculate the spatiotemporal context, overlaid with air pollution maps to provide mobility-enhanced exposures.
ABM involves the development and execution of three component models (Figure 2). A more detailed explanation will follow in Section 4.1. In short, the model input for each person includes their profile, which defines behavioral characteristics, as well as individual data constraining this behavior. When rich information on individual persons is available (e.g., small cohorts or surveys), agent behavior will be steered mainly by individual data; for cohorts with sparse information, typically larger or nationwide cohorts, the simulation will rely mainly on the profile of persons. The activity schedule component of the model generates, for each person, diurnal schedules of activities a1, a2,.., an. Uncertainties can be evaluated in Monte Carlo mode, generating multiple realizations of activity schedules (Figure 2). For each of a person’s activities, the spatial context model generates the geographical area where the activity takes place; in case of uncertainties, it generates the probability for each location of that person visiting the location. The spatial context model includes fixed locations (e.g., residential address), residences within an area defined by a certain buffer radius (e.g., to capture leisure activities around the residential address), and paths following the traffic infrastructure (e.g., commuting). By combining diurnal activity schedules and spatial contexts, the spatiotemporal context, which describes the location of the person over time, can then be generated for each person (Figure 2, right). Finally, the exposure assessment model intersects the spatiotemporal contexts with the spatiotemporal air pollution maps, resulting in the mobility-enhanced exposure. This is done for each person, pollutant, and Monte Carlo realization.
ABM was executed for cohorts in Switzerland and the Netherlands. For all cohorts, we randomly assigned work locations constrained by travel survey data. Additionally, for SAPALDIA and the tracking cohorts, we used the known work locations to derive the commute routes.
3.3 TRACKING CAMPAIGN
We aimed to conduct a tracking campaign that would include 1,000 participants each in Switzerland and the Netherlands and would collect detailed information on the participants’ mobility patterns over a 2-week period using a purpose-designed mobile phone app and a GPS tracker. No health-related data were collected from the participants in the tracking campaigns. The Dutch part of the tracking campaign received an ethics exemption from the Medische Ethische Toetsingscommissie Utrecht. The Swiss part of the study was approved by the Ethikkommission Nordwest- und Zentralschweiz, as it recruited from the earlier-approved ongoing population-based SARS-CoV-2 Cohort Basel-Landschaft and Basel-Stadt (COVCO-Basel) study.
COVCO-Basel is a digital cohort in Northwestern Switzerland, specifically initiated to investigate the long-term impact of the COVID-19 pandemic and its containment measures on broad domains of health and well-being.19 Persons aged 18 years or older residing in Basel-Stadt or Basel-Landschaft for at least 5 years were eligible. At baseline, 12,724 participants entered the study between July 2020 and April 2021. Individuals who had previously expressed interest in participating in an environmental health study were invited to the MOBI-AIR tracking campaign. Finally, a total of 489 participants participated in the Swiss tracking campaign.
In the Netherlands, we started with a random population survey: 3,000 letter invitations were sent in September 2022. We used the Dutch address database Basisregistraties Adressen en Gebouwen in an attempt to select a random population; the database does not contain names or personal characteristics, so the letter was addressed to “the occupant.” Only 41 responded (1.4% response rate); this low response prompted new recruitment strategies through social media and a distribution of 500 leaflets in the city of Utrecht (capturing 73 additional participants). To improve response, we included a 25 Euro voucher as an incentive, based on advice from social scientists at Utrecht University working with surveys. Based on this experience, we delivered 5,000 more leaflets in the city and province of Utrecht in May 2023, in selected low and high-socioeconomic status neighborhoods within the city and selected towns around the city. This resulted in a total of 189 participants in the Dutch tracking campaign.
In both campaigns, but especially the Dutch, conforming to the data protection regulations was a major challenge, resulting in substantial delays and requiring extensive personnel efforts. A data protection impact assessment (DPIA) was needed and sent for approval to the single privacy officer at Utrecht University, after advice from the faculty-level privacy officer.
Initially, we planned to use the ExpoApp, developed by Donaire-Gonzalez and colleagues, which was specifically designed to track people and collect time–activity data.20 However, since the development of the ExpoApp, a change in its iOS and Android architecture meant that the GPS location functionality drained the phone battery too quickly, such that using this phone app was not an attractive option.
We therefore searched for other solutions, in collaboration with the EXposome Powered tools for healthy living in urbAN Settings project (EXPANSE, https://expanseproject.eu). This resulted in two methods used to track the study population. First, we used a tracker developed by SODAQ (https://sodaq.com). The device functionality includes a 20-second timestamp of individual location using GPS, cell tower, and Wi-Fi (Wi-Fi scraping), with the latter providing a more accurate indoor/outdoor indication. The device starts tracking only when movement is detected by a built-in accelerometer; otherwise, it takes a reading every 5 minutes when stationary. Second, we used a purpose-designed mobile phone app by Games for Health (https://gamesforhealth.net) to provide a daily time–activity questionnaire to the participants. The battery drain issue related to GPS data capture was solved before the full campaigns started; specifically, the iOS and Android architecture changed so that it became possible to turn on the GPS while using the app. This meant it was also possible to measure the location using a pin every 3–4 minutes. The tracking functionality of the app was thus a bonus (see S3.4A and S3.4B).
The tracking campaigns in the two countries were carried out in very similar fashions, via the steps shown in Table 1.
Table 1.
Steps Followed in the Tracking Campaigns from Recruitment to Feedback in Switzerland (CH) and the Netherlands (NL)
Steps
Switzerland
The Netherlands
1
Recruitment
Adult participants from the COVCO-Basel cohort, who expressed interest in participating in an environmental health study, were sent an invitation email via Research Electronic Data Capture (see S3.1A)
Participants were selected from the general adult population, using physical letters to a random sample of the Dutch population, social media calls, and home-delivered letters to selected towns in the province of Utrecht (see S3.1B)
2
Eligibility assessment
On agreeing, potential participants entered a website with a unique code and answered several questions to see if they were eligible. Exclusion criteria were as follows:
not capable of answering questions/handling devices
not strong enough to carry the sensor
not understanding the local language (German in CH and Dutch in NL)
not in possession of a smartphone (Android 5.0 or higher; iOS 11.0 or higher) and internet access
not capable of handling smartphone applications
3
Informed consent form
When eligible, the potential participant received an informed consent form by post that they must sign and send back (see S3.2A)
When eligible, the potential participant received an informed consent form by email that they must sign via Qualtrics, an online certified survey tool (see S3.2B)
4
Registration and availability
When the informed consent form was returned, the participant was registered in the system, and an email was sent with questions about available dates to participate
When the informed consent form was signed, the participant was registered in the system
5
Baseline questionnaire
When availability was confirmed, the link to the online baseline questionnaire implemented in Research Electronic Data Capture was sent by email (see S3.3A)
A baseline questionnaire was sent immediately after registration to be filled in via Qualtrics When the baseline questionnaire was completed, an email was sent with the availability for tracking questions in groups of registered participants (see S3.3B)
6
Sending device
Participants with completed baseline questionnaire received at the convenient measurement period start date, the device with instructions for device and app (time–activity questionnaire) activation by post, plus an email was sent (see S3.4A)
The device plus a letter with instructions for device and app (time–activity questionnaire) activation were sent by post, along with an email (see S3.4B)
7
Returning device
After 2 weeks, an email was sent reminding them to complete the time–activity questionnaire and stop the app, plus the instructions to send back the device
8
End of campaign
After receiving the returned device, a thank-you email was sent
After receiving the returned device, a thank-you email was sent with a voucher of €25
9
Feedback
After completing the 2-week tracking, each participant received a feedback letter with information about their total and activity-based NO2 and PM2.5 exposures, including general advice on how they can reduce their air pollution exposure (see S3.5a and S3.5b)
The association between long-term air pollution exposures and well-established health outcomes was analyzed in the adult cohorts listed in Table 2: blood pressure and lung function in SAPALDIA; incidence of stroke and coronary events in EPIC-NL; natural cause mortality in EPIC-NL and SNC; and cardiovascular mortality in SNC.
Table 2.
List of the Various Cohorts with Common Descriptive Demographical Information and Overview of Study Population Demographics at Baseline
CE = coronary event; CVD = cardiovascular disease; EPIC-NL = European Prospective Investigation into Cancer study, Netherlands cohort; SAPALDIA = Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults; SD = Standard Deviation; SNC = Swiss National Cohort.
a Age at assessment.
b Age at baseline.
c Low: primary school; middle: secondary school, middle school, or apprenticeship; high: technical college or university.
d Employed as yes/no.
e Not available for confounder adjustment, but used in agent-based modeling exposure assessment after combining several data sources to gap fill the information (see Fig S3.1).
The SAPALDIA cohort was initiated in 1991,21–24 with 9,651 randomly selected adults from eight areas in Switzerland. Since then, participants have been followed up with surveys in 2001/02, 2010/11 (SAPALDIA3), and 2017 (SAPALDIA4). In the current study, analyses were conducted using SAPALDIA3 and SAPALDIA4. Data obtained includes a detailed address history, with more recent follow-ups also obtaining information on work address. Participants received spirometric examinations on four occasions, blood pressure measurements on three occasions, and HbA1c measurements on two occasions; they were provided detailed information on cardiovascular and respiratory symptoms, as well as clinical cardiovascular, metabolic, and respiratory diagnoses, with associated medication intake. In addition, participants were characterized repeatedly and in detail on relevant sociodemographic and lifestyle characteristics, which allow for proper consideration of confounding and effect modification. Ethical approval was obtained for each survey and study area from the central ethics committee of the Swiss Academy of Medical Sciences and the Cantonal Ethics committees.
The EPIC-NL study consists of two cohorts: the Monitoring Project on Chronic Disease Risk Factors cohort and the Prospect cohort.25 The former consists of 22,654 men and women aged 20–65 years, randomly selected from the Dutch population in three towns in the Netherlands (Amsterdam, Doetinchem, and Maastricht). The Prospect cohort consists of 17,357 women aged 50–69 years, from the Dutch town of Utrecht and its vicinity, who participated in a breast cancer screening program. For both cohorts, participants were recruited between 1993 and 1997. Primary health outcomes were natural mortality and incidence of (specific) cardiovascular disease, following the ELAPSE study of which EPIC-NL was part.26,27 We specifically assessed the incidence of stroke and coronary (fatal and nonfatal) events, following ELAPSE definitions. Similar to SAPALDIA, EPIC-NL also included a rich amount of lifestyle variables, which were collected at baseline. We used the health and covariate database of the ELAPSE study to link in exposures. No further update of the follow-up was feasible for the purpose of this study.
The Swiss National Cohort (SNC) is a national longitudinal research platform, linking birth, mortality, and emigration data with census data.28,29 Because participation is mandatory, nearly all persons residing in Switzerland at the time of the 1990 and 2000 censuses are represented. Starting in 2011, annual censuses using community and cantonal population registries have been incorporated, including annual residential history. For each person, the SNC contains an individual, household, and building record, facilitating exposure assessment at the home location and basic confounder adjustment, including socioeconomic status. The SNC dataset used in MOBI-AIR included data from 2011 to 2018. Similar to EPIC-NL, we followed the ELAPSE study, in which SNC was included, to investigate natural and cardiovascular mortality.9
Table S3.1 presents the definitions of the health outcome for the SNC and EPIC-NL study, following the ELAPSE study.
3.5 EXPOSURE ERROR EVALUATION
We used two approaches to evaluate measurement error in the ABM approach. First, we compared the ABM-modeled exposures with the exposures generated with observed time–activity data in the tracking campaign. This evaluated the time–activity component of the model. Because the same ambient air pollution surface was used, it does not address the full uncertainty. Second, measured personal monitoring data were used to evaluate the mobility-enhanced exposures. The data were derived largely from the European EXPOsOMICS project.17 EXPOsOMICS included a detailed personal monitoring campaign from 2013 to 2015 among the participants of several European adult cohorts. Specifically, for the evaluation, we selected 35 participants from SAPALDIA in Switzerland, 41 participants from EPIC-NL, and participants in Amsterdam and Utrecht from outside the cohort.17 Those participants had complete data of three 24-hour measurements of PM2.5 over three different seasons for 1 year. A more detailed description can be found in the supplement (Section S5.1).
4. DATA ANALYSIS
4.1 DESCRIPTION OF ABM
We developed a set of diurnal activity profiles as behavioral input for the ABM (Table 3). For the activity profiles, we distinguished between workdays and weekend days. The profiles can be grouped in three main categories: (1) A residential profile that corresponds to the traditional exposure analysis and places a person at their residential address for the entire day. (2) A homemaker profile (i.e., a homemaker or someone without paid employment) that includes an activity outside and around the residential address, modeling, for example, leisure or shopping activities. As nationwide information on start time, duration, and maximum distance is unavailable, we assumed a 2-hour outside activity taking place between 08:00 and 23:00 for homemakers. We assumed that these activities took place in buffer sizes of 1, 5, or 10 km around the residential address as plausible alternatives. The average concentration values within the buffers were used to calculate the exposure during this period. (3) A commuter profile that includes two commute trips and an 8-hour period at the work location. Commuting to work was assumed to take place only on workdays and to start between 06:30 and 07:30. The maximum commute time was set to 2 hours.
Table 3.
Overview of the Exposure Assessment Methodologies and Description of Diurnal Activity Profiles Used As Behavioral Input for Agent-Based Modeling
Methodology
Weekdays
Weekends
Buffer
Sex
SEP
Residential-only (RES)
Residential
24 hours at home (average air pollution at home location)
-
F/M
-
Mobility-enhanced (ENH)
Homemaker 1
2-hour activity between 08:00 and 23:00 (average air pollution within a buffer around the home)
1,000 m
F/M
-
Homemaker 2
+
5,000 m
Homemaker 3
Remaining time at home (average air pollution at home location)
10,000 m
Commuter 1
Commuting between 06:30 and 07:30 (average air pollution along the route, mode of transit depends on distance from home to work)
2-hour activity between 08:00 and 23:00 (average air pollution within a buffer around the home)
10,000 m
F/M
-
Commuter 2
M
-
Commuter 3
F
-
Commuter 4
F
High
Commuter 5
+
+
F
Low
Commuter 6
8 hours at the working location (average air pollution at the work location)
Remaining time at home (average air pollution at home location)
F
Middle
Commuter 7
M
High
Commuter 8
M
Low
Commuter 9
+
M
Middle
remaining time at home (average air pollution at home location)
a Commute mode was randomly drawn based on Euclidean distance to work location, with a chance of mode defined as: <1,000 m: 50% walk, 50% bike; 1,000–10,000 m: 50% bike, 50% car; 10,000+ m: 50% car, 50% public transport.
The path between the residence and work location was obtained by routing over OpenStreetMap traffic infrastructure data using the shortest route. Our ABM approach supports four different commute modes (by foot, bike, car, and public transport, including trains). An agent’s commute mode was selected depending on the distance to work. For network distances smaller than 1,000 m, we used 50% walk and 50% cycle mode; for distances between 1,000 and 10,000 m, we used 50% bike and 50% car; and for distances greater than 10,000 m, we used 50% car and 50% public transport. The duration of the activity was then the result of the commute distance over the traffic infrastructure and the estimated speed for the corresponding mode. A short description of the profiles can be found in S4.1.
An agent’s work location can be directly specified if known (as in the SAPALDIA subset and most tracking campaign participants). For unknown work locations (as in the SNC and EPIC-NL cohorts), we randomly assigned a work location from an origin–destination matrix (described below) on the basis of the postcode/municipality in which the participant lives. Work postcodes/municipalities were weighted by all visits derived from mobility survey data (see Figures S4.1A and S4.1B). Within a selected work postcode/municipality, we randomly selected a potential work location. If the commute time to a potential work location exceeded the maximum allowed duration, a new work location was selected. If, within 25 attempts, no suitable work location could be determined, a work location in the (postal code) area of the residential address was assigned. The diurnal profiles were combined to weekly exposure profiles (e.g., a profile representing a full-time worker is combined with 5 weekday commute profiles and 2 weekend homemaker profiles with the 10,000-m buffer).
The origin–destination matrices for the Netherlands were created from data from the National Travel Surveys from Statistics Netherlands (see S4.2), which are annual surveys of daily mobility behavior among ~40,000 randomly selected households in the country. Participants were asked to report all trips they made during one designated day, including information about start and end time, geographic location of the origin and destination, mode of transport, and individual sociodemographics (e.g., income). We combined the years 2011–2017 from the survey Onderzoek Verplaatsingen in Nederland and 2018/2019 from the subsequent Onderweg in the Nederland survey. We selected only the commute trips provided in the surveys and extracted the postcodes of residential and work location, sex, and income for each commute trip (181,489 in total). We used this information to generate an origin–destination matrix for all Dutch inhabitants, separately for male and female, and subdivided further into low-, middle-, and high-income groups, resulting in nine different commute profiles (overall, male/female, and male/female, and each combined with income).
The origin–destination matrix for Switzerland was created using data from the Structural Survey, which is a component of the population census. At least 200,000 people are interviewed each year as part of the survey, which collects individual-level data on travel behavior, including car ownership, total distance covered per day, time spent commuting, number of trips per day, means of transport, reasons for each trip, and socioeconomic characteristics, among other information.30 The surveys from 2010 to 2019 were compiled, including responses from over 2.8 million people. For the working population, information on work and home location was aggregated at the municipality level. The resulting origin–destination matrix was stratified by age, sex, and socioeconomic status.
The same set of profiles was used for both countries, but each was supplied with country-specific origin–destination matrices. The municipality and postcode for origin and destination correspond to the administrative level of municipality (for Switzerland and the Netherlands). We used cadastral datasets to obtain the coordinates for all buildings in each country (see S4.2). Buildings classified as residential and partly residential were used for residential locations, while those classified as nonresidential were taken as potential work locations. For Switzerland, this resulted in about 1.87 million residential address locations and about 1.33 million work address locations. For the Netherlands, about 8.3 million residential address locations and 1.1 million work address locations were extracted.
We used the Python programming language to implement the ABM framework with the Geospatial Data Abstraction Library, NumPy, and Campo modules for spatial operations and storing the exposure data. Python was also used to develop pre- and postprocessing scripts, such as converting travel survey data to origin–destination matrices, for plotting results and generating Slurm scripts for execution on high-performance computing systems. OpenStreetMap data snapshots for both countries were obtained on December 6, 2022. The Open Source Routing Machine was used to obtain commute routes. The Quantum Geographic Information System (QGIS) was used for visual inspection of GIS datasets. Web links to software packages and datasets used are given in the Additional Materials (see S4.2).
4.2 TRACKING CAMPAIGN
Time–activity data were gathered through a mobile app, while position data were obtained from both the app and a high-precision GPS tracking device. Time–activity information was assigned to the corresponding location points. In instances when participants stayed in a location for an extended period (e.g., overnight) with missing location points, we imputed one point per hour. Each location point was linked to NO2 and PM2.5 exposure values extracted from the hourly air pollution raster files corresponding to weekdays and weekends. The time window covered by each point was calculated. For every activity block, we computed the time-weighted exposure, then generated summaries per activity type for each participant. We performed this analysis for both the combined location points of the phone and GPS tracker, and separately for each device type.
We compared exposures based on this tracking method with those based solely on home location and with exposure derived from the ABM. For the comparison with the ABM, participants were assigned to the categories defined in the ABM (i.e., residential, homemaker, or commute profile) based on information gathered in the participant questionnaire. Participants who were in paid employment were assigned a commuter profile based on their sex and socioeconomic status. Participants who were not in paid employment were assigned a homemaker profile, depending on time spent at home as indicated in the baseline questionnaire.
We developed common R scripts that were used to process the data in both countries. In brief, the scripts resolved the following issues:
Combine GPS data from SODAQ and the mobile app
Combine data from the time–activity questionnaire with the GPS tracks
Impute missing data for home indoor activity if GPS points were missing (one point per hour)
Extract the respective hourly weekday or weekend exposure for NO2 and PM2.5
Exclude outliers (by distance and time to previous and following point)
Exclude travel activities if there were fewer than six points per hour
Calculate, for each point, the time period it represents and subsequently the time-weighted total exposures
4.3 AIR POLLUTION EXPOSURE ASSESSMENT
4.3.1 Air Pollution Exposure Model Development
Hourly surfaces of NO2 and PM2.5 for Switzerland and the Netherlands were not available at the start of this study, and thus were specifically developed for MOBI-AIR. Different approaches for the two countries were used for the exposure modeling.
In Switzerland, we used an approach developed by de Nazelle and colleagues31 where annual average NO2 surfaces were adjusted into hourly average weekday and weekend day surfaces using conversion factors obtained from hourly background monitoring data at multiple background monitoring stations. We applied this method to obtain the required hourly surfaces for NO2 and PM2.5. We used annual average PM2.5 and NO2 surfaces for the year 2016 from earlier published spatiotemporal geostatistical models32,33 and rescaled these to long-term hourly weekday and weekend day surfaces. Measured air pollution data from background monitoring stations in Switzerland were used to calculate ratios that were then applied to the long-term concentration surfaces using the formula:
(1)
where Ratioannual–t is the ratio of the air pollutant concentration measured at the background monitoring station (BS) at time t (hours 1–24 averaged over a year, for weekday and weekend); CBS-t is the measured air pollution concentration averaged for each given hour per year; and CBS–annual is the annual average air pollution concentration at all background stations in Switzerland.
These ratios were then applied to the annual average air pollution surfaces using the following formula:
(2)
where CP–t is the time-adjusted prediction at point P and time t; and CP–annual is the annual average pollutant concentration predicted by the spatiotemporal land use regression models at point P.
In the Netherlands, specifically for this study, maps of annual average hourly concentrations of PM2.5 and NO2 were developed using land use regression models based on air quality monitoring data in the Netherlands and in the neighboring countries Belgium and Germany. Details of the modeling are provided in a publication based on this project.34 Previous studies show that an increased number of observations provides more diverse monitoring data and allows the land use regression models to better capture the relationships between predictors and pollutants, leading to improved performance and generalizability.35 Thus, to increase the number of observations to be used in training the land use regression models in this study, we added stations from neighboring countries. We modeled the spatial variation of the average hourly concentration for the years 2016–2019 combined, for two seasons (cold and warm), and two weekday types (weekdays and weekend days). The total number of stations for NO2 and PM2.5 was 544 and 227, respectively.
Two modeling approaches were used in the Netherlands: supervised linear regression (SLR) and random forest (RF). The potential predictors included population, road, land use, satellite retrievals, and chemical transport model pollution estimates variables with different buffer sizes. We followed the ELAPSE protocol36 to train the SLR models. For each model, first the predictor variable with the highest R2 value, and thus the one explaining the most variation in the concentrations, was added. In the following steps, additional predictor variables were considered for inclusion in the model if they improved the adjusted R2 and had the predetermined direction of effect, and if their coefficient value was statistically significant (P < 0.1). For each hour, a different SLR was built, with possibly different predictor variables. RF is a tree-based machine learning method.37 The advantage of RF is that it can handle nonlinear relationships, interactions between variables, and highly correlated variables by randomly selecting a subset of variables in each split node of a tree. The ranger package version 0.12.1 in R was used to train RF. Detailed information on how the RF was trained can be found in Shen and colleagues.38
The mean and the standard deviation of the fivefold cross-validation R2s of the hourly models were used to evaluate the performance of the Dutch models. The cross-validation included new model development, where observations were grouped based on station location and then randomly subdivided into five groups of equal size. Four groups were used to train the model, and one was used to test it. The fivefold cross-validation R2 of the SLR and RF was calculated as a mean squared error-based R2. In the current study, we used the RF models to combine outdoor concentrations with time activity data, as this algorithm had slightly better performance; see the Results section.
4.3.2 Air Pollution Exposure Assignment to the Cohorts
For each participant in the SAPALDIA, EPIC-NL, and SNC cohorts, nine possible exposures were calculated. For all cohorts, all time–activity exposures were calculated using hourly concentration averages. For EPIC-NL, as a sensitivity analysis, time–activity exposures were also calculated using yearly concentration averages. In both cohorts, the first exposure estimate was the averaged air pollution concentration at the residential address (residential only, not counting participants who moved during follow-up), which assumes that the individual spends all their time at home. Then, three exposure estimates were calculated assuming that participants were not employed (i.e., the homemaker profile, which included those unemployed or retired) and spent 2 hours of the day outside the house (HOME1 to HOME3), in a 1,000-m, 5,000-m, or 10,000-m buffer around the residential address. Thus, exposure estimates were calculated by averaging concentrations at the home address and the possible locations within the buffers. Finally, nine exposure estimates (COM1 to COM9) assumed that the participants were employed and thus spent time at the workplace and commuted. In addition to calculating totals, each commuting exposure was tailored to sex and socioeconomic status (see Table 3 and S4.1 for a more detailed description of the profiles).
Descriptive analyses were conducted to examine the correlations between the exposure scenarios for NO2 and PM2.5. First, we assigned the residential exposure estimates to the participants. Second, we assigned the best-fitting commuting-type exposure, according to employment status, economic status (educational level in the Netherlands, socioeconomic position–score in Switzerland), and sex. Because it was not possible in the SNC and EPIC-NL to decide how far from home those who were not employed (i.e., considered homemakers) would go, we assigned the exposure estimates assuming they would most often stay within a 5,000-m buffer around the home (HOME2). In the SNC, the largest of the three cohorts, exposures from the three homemaker profiles were very highly correlated (R > 0.99). For the tracking campaigns, we used HOME1, HOME2, or HOME3 depending on the time the participants spend at home as indicated in the baseline questionnaire.
As the primary metric for the setting of an unknown work location, we assigned the mean of the 50 realizations of the ABM. We further documented the variability of the modeled exposure by providing the standard deviation of the mobility-enhanced exposure. As an exploratory analysis, we assigned one realization per person by drawing at random from the person-specific distribution. We selected a random number between the mean plus and minus two times the standard deviation. We assumed equal density in that interval. Drawing a single observation provides a “worst case” of deviation from the mean value.
4.4 EPIDEMIOLOGICAL ANALYSES
4.4.1 Comparison of Residential and Mobility-Enhanced Exposure Estimates in the Large Administrative Cohort (SNC) and the Adult Lifestyle-Rich Cohort (EPIC-NL)
For SNC and EPIC-NL, the protocol and statistical scripts followed those of the ELAPSE study administrative and pooled cohorts.9,14 Analyses were conducted separately for each cohort.
We analyzed natural mortality (SNC and EPIC-NL), cardiovascular mortality (SNC), and stroke and coronary events incidence (EPIC-NL) in adults; our main objective was to examine whether associations differed when using the mobility-enhanced air pollution exposure versus the residential exposure. In EPIC-NL, because of the limited number of events, we did not evaluate cardiovascular mortality. In SNC, stroke and coronary events data were not available.
To focus on the difference between mobility-enhanced and residential exposure, we analyzed NO2 and PM2.5 in single-pollutant models only and assumed a linear association. We used Cox proportional hazards models, with age (as timescale) and sex (as strata), adjusting for cohort-specific individual- and area-level covariates (including socioeconomic status at multiple scales). The populations included adults 20 years and older at baseline. Individuals were followed for up to 21 years, depending on cohort (Table 2) and censored on date of event, death, or loss to follow-up (e.g., emigration), whichever came first.
For the SNC, the fully adjusted model included age, sex, marital status (single, married, widowed, divorced), occupational status (employed, unemployed, homemaker, retired), origin (Swiss, non-Swiss), Swiss socioeconomic position score,39 language region as a proxy for cultural differences (German, French, Italian, Rhaeto-Romansch), Swiss physical region to account for topography and air sheds (n = 7), and area-level socioeconomic status by way of the Swiss socioeconomic position-score and percentage unemployed at the neighborhood (n = 2,583 municipality) and regional (n = 26 cantons) levels.
For EPIC-NL, the fully adjusted model included smoking status (never, former, current), smoking intensity (number of cigarettes/day), smoking duration (years of smoking), fruit and vegetable intake, alcohol consumption (low, medium, high), body mass index (BMI) (underweight: <18 kg/m2; normal weight: 18–25 kg/m2; overweight: 25–30 kg/m2; obese, class 1: > 30 kg/m2), educational level (low, medium, high), employment and occupational status (employed, unemployed, homemaker, retired), and marital status (single, married/with partner, divorced, widowed) at the individual level. The model further included mean income, ethnicity, and unemployment rate at the neighborhood level.
In an additional analysis, retired persons were excluded from the SNC due to their potentially high, but spatially unknown and less predictable, mobility compared with the working-age population. This further allowed for a potentially larger difference between the residential and enhanced, as all individuals have the potential to commute.
We expressed effect estimates for a fixed exposure increment, specifically 5 μg/m3 for PM2.5 and 10 μg/m3 for NO2. Additionally, we compared effect estimates expressed per interquartile range (IQR) for each method to investigate how much differences in exposure contrast contributed to differences in effect estimates.
4.4.2 Comparison of Residential and Mobility-Enhanced Exposure Estimates with Lung Function and Blood Pressure in SAPALDIA
We analyzed lung function and blood pressure in SAPALDIA participants with the following objectives:
To examine what difference the mobility-enhanced air pollution estimates make in the association with lung function and blood pressure, compared with the residential estimates
To examine what difference a randomized work location makes in the mobility-enhanced air pollution estimates, compared with the air pollution estimates based on true work locations
SAPALDIA participants were included in the analysis for the following:
They participated in both surveys SAPALDIA3 (2010/2011) and SAPALDIA4 (2017).
Geocoded residential address was available in SAPALDIA3 and SAPALDIA4.
Nonmissing information on all the covariates is available in SAPALDIA3 or 4.
For the second objective, the geocoded work address was available in SAPALDIA3.
Repeated cross-sectional analysis, combining SAPALDIA3 and SAPALDIA4 data, was conducted for the first objective, while only SAPALDIA3 was used for the second objective because the work address was unavailable in SAPALDIA4. Participants were assigned to ABM profiles based on their employment status, sex, and education.
Association Analysis of Lung Function
We applied linear regression to each of the three lung function parameters: forced expiratory volume in 1 second (FEV1); forced vital capacity (FVC); and FEV1/FVC from spirometry, after adjustment for a priori selected covariates: age, age2, height, sex, BMI, BMI2, smoking status (never/former/current), cigarette pack-years in lifetime (0 for never smokers), cigarettes/day (past or present, 0 for never smokers), doctor-diagnosed asthma, neighborhood socioeconomic index,39 seasonality using sine function of the date of the examination, skin-prick test (at SAPALDIA1), atopy, parental smoking, Swiss nationality, respiratory medicine within 12 hours before exam, currently exposed to gas/smoke/aerosols/fumes/vapors at workplace, early childhood respiratory infection, time of day (sine and cosine functions), and study area (eight areas). For repeated cross-sectional models, a random intercept for the individual was included.
Association Analysis of Blood Pressure
We applied tobit regression to take account of hypertension medication, assuming that for those who are on hypertension medication, their actual blood pressure is at least as high as their observed blood pressure. Systolic or diastolic blood pressure was regressed on exposure estimate (residential or mobility enhanced), after adjustment for a priori–selected covariates: age, age2, sex, BMI, BMI2, smoking status, cigarette pack-years in lifetime (0 for never smokers), education level, neighborhood socioeconomic index, alcohol consumption, physical activity, seasonality using sine and cosine function of the date of the examination, occupational status, and study area (eight areas). For repeated cross-sectional models, a random intercept for the individual was included.
5. RESULTS
5.1 MOBILITY (ABM)
We ran simulations for Switzerland and the Netherlands using the 13 different activity profiles (Table 3). For each residential address and all profiles, we calculated NO2 and PM2.5 weekly average exposure estimates using 50 realizations for the SNC and EPIC-NL cohorts and 1,000 realizations for the SAPALDIA and tracking cohorts. Outputs are the ensemble mean and standard deviation per exposure for each agent and each profile. Values can be used by look-up for further statistical, spatial, and epidemiological analysis by matching on sex, socioeconomic status, and employment status. Comparisons between exposures of the different commuter profiles (i.e., male/female and socioeconomic status) revealed high correlations. Figures S5.1A and S5.1B show the high correlations (R > 0.97) between the different profiles linked to participants in the EPIC-NL cohort for both NO2 and PM2.5. The high correlations are partly explained by the small differences in the origin–destination matrix between the profiles. Figures S5.2A and S5.2B visualize the origin–destination matrix used for the probabilistic selection of work locations with examples of persons with a residential address in Basel and Utrecht, respectively. The thickness of the solid lines represents the number of persons commuting to the surrounding destination areas; the thickness of the transparent vertical bar represents the number of persons commuting to a work location in the same areas as their residential address. In the ABM, work locations were randomly drawn with a probability weight proportional to the number of persons commuting to the areas. The example shows very little difference between the mobility patterns of the different commuter profiles; however, commuters with low socioeconomic status tend to work closer to their residential area compared with those with intermediate or high socioeconomic status.
The calculated values can be used to evaluate different activity patterns and their influence on exposure estimates. Different profiles can be compared, as shown for the SNC and EPIC-NL cohorts; for example, the homemaker and full-time worker profiles of NO2 (Figures 3 [A and B] and 4 [A and B], respectively). The figures show that among those with low (<10 μg/m3) estimated NO2 exposures, exposures are more likely to be underestimated for a homemaker compared with a working person; however, among those with high (>40 μg/m3) estimated exposures, homemaker exposures are likely to be overestimated compared with a working person. This can be because people living in low-pollution, urban locations work in more polluted areas. On the contrary, people living in the more polluted areas tend to work in less polluted locations and therefore overall have less exposure. Because our ABM framework is capable of propagating model input uncertainties to model outputs, exposure estimates come with uncertainties in these estimates. For the commuter profiles with randomly sampled work locations, the uncertainty at the standard deviation level is approximately 1–2 μg/m3 for the SNC and EPIC NL (Figure S5.3 [A and C]) and 0.0–0.6 μg/m3 for SAPALDIA and the Dutch tracking cohort (Figure S5.3 [B and D]). This uncertainty is mainly caused by the uncertainty of the exposure during work and the commute trip, both due to the uncertainty in the work location within the destination regions of commuters.
(A and B) Comparison of mean NO2 values (50 realizations for 1.8 million Swiss addresses) of the homemaker and commuter profiles using a scatterplot (A) and histogram (B); (C and D) Scatterplots comparing mean NO2 values of Swiss tracking campaign participants, using the known work location (C) and randomly sampled work locations (1,000 realizations) (D).
(A and B) Comparison of mean NO2 values (50 realizations for EPIC-NL) of the homemaker and commuter profiles using a scatterplot (A) and histogram (B); (C and D) Mean NO2 values of Dutch tracking campaign participants, using the known work location (C) and randomly sampled work locations (1,000 realizations) (D).
Figures 3 (C and D) and 4 (C and D) show a comparison of NO2 exposure estimates for commuters between known and randomly sampled work locations in Switzerland and the Netherlands, respectively. For randomly sampled work locations, the correlation with homemaker exposure is high; this is because when averaging out exposures across randomly sampled work locations, as applied here, we cannot distinguish between commuters regarding their exposure at work. For model runs with a known work location, the exposures for commuters become more different from homemaker exposures, because ABM uses commute and work-time exposures specific to each person. The known work location exposures for NO2 also result in considerably smaller uncertainties in exposure estimates, compared with those calculated with randomly sampled work locations, approximately ranging between 0.01 μg/m3 and 0.2 μg/m3 (Figure S5.3 [B and D]).
Figure 5A shows the diurnal variation of NO2 exposure for a set of randomly selected agents. Clearly visible is the trend following the hourly changing air pollution concentrations, short-term changes due to the commute trip occurring around 08:00 and 17:00, and exposure differences between residential and work locations. Figure 5B shows the difference between the homemaker and commuter profiles, with randomly sampled work locations for residential locations in the city of Bern. Residents living along busy roads (e.g., the motorway) have reduced NO2 exposure estimates (in green), while those living in locations with low pollution have increased NO2 exposure estimates (in red). Under- or overestimation can be up to 8 μg/m3 NO2 over short distances. In an additional analysis, we compared different activity profiles showing small differences between the residential exposure and the homemaker profile (see Figure S5.4) and between the residential exposure and the homemaker profile using varying buffers of 1, 5, and 10 km to estimate exposure during leisure time (see Figure S5.5). Socioeconomic status–based results for commuters differed from results for homemakers (see Figures S5.6 and S5.7, for SNC and EPIC-NL, respectively). The full-time worker profile (Figure 3) showed the largest difference from the homemaker profile because the work location and commute trip were included. Part-time worker profiles align more closely per day with the homemaker profile (Figure S5.8).
(A) Example showing the diurnal exposure estimates for 40 SAPALDIA participants; (B) Differences in NO2 exposure (in μg/m3) between homemaker and commuter shown for residential addresses in the city of Bern.
5.2 TRACKING CAMPAIGN
5.2.1 Dutch and Swiss Tracking Campaigns
The COVID-19 pandemic, causing partial lockdowns in the Netherlands and Switzerland well into 2022, led to delays in starting the tracking campaigns. The lifting of all restrictions in the Netherlands (on March 23, 2022) and in Switzerland (from April 1, 2022) returned mobility patterns to almost prepandemic levels. This gave us the confidence to start both tracking campaigns in September 2022. The DPIA procedure led to further delays in the Netherlands.
In Basel, a total of 489 people participated in the tracking campaign during two waves (September–December 2022 and January–May 2023). The tracking campaign benefited from the COVCO-Basel cohort implemented during the COVID-19 pandemic. For recruitment, 1,475 people in the COVCO-Basel study were invited by email, 502 signed the informed consent form, and 489 participated (33% recruitment rate). The high recruitment rate builds on the vast cohort experience available in COVCO-Basel and active and labor-intensive communication with COVCO-Basel participants by phone and email. COVCO-Basel also implemented personalized feedback so that each participant received an email summarizing their exposure after completion of the tracking campaign (see template letter in S3.5a).
In the Netherlands, we started with a random population survey of 3,000 letters in September 2022, to which 41 individuals responded. Recruitment through social media and the distribution of 5,500 leaflets in the city and province of Utrecht in May 2023 were performed to increase the number of participants. In total, 189 participated fully in the tracking campaign, which was finished in August 2023. We lost people subsequently because of the complicated DPIA procedure. In total, 380 people showed an interest, of whom 248 signed an informed consent letter and 208 completed the baseline questionnaire. Some participants who had already agreed to participate did not complete tracking, even after receiving the equipment. Despite a large personnel and financial effort (25 Euro incentive), we did not manage to achieve the targeted 1,000 participants. The Dutch enrollment is lower than the Swiss campaign, likely because of the lack of a recent existing cohort from which to start recruitment. Building on the Swiss experience, we developed personalized feedback for the Dutch participants, so that each received an email summarizing their exposure distributed over different activities after completion of the tracking campaign (see template letter in S3.5b). This allowed participants to understand where they experienced the highest exposures and how their exposure compared with that of the other Dutch participants. In the Netherlands, tracking took place in seven waves, between November 2022 and August 2023, thus covering winter and summer seasons. Most participants were measured between May and August 2023.
Table 4 shows details and demographics of the tracking campaign participants in Switzerland and the Netherlands. To evaluate how representative the tracking campaigns were, national demographic data on the Swiss and Dutch populations were added. In both campaigns, females were overrepresented. Most participants were employed, with a small fraction who were retired. Participants with a higher education and higher income were also generally overrepresented in the tracking campaign. Compared with the national data, it is clear that the tracking campaigns are not representative of the full Swiss and Dutch populations. The 40- to 60-year-old, high-income, higher-education, and full-time employment groups were overrepresented in both Switzerland and the Netherlands. However, in the tracking populations, there was a reasonable representation of the other age groups, low- to medium-education and low- to medium-income groups, and nonemployed groups. The larger share of employed participants and 40- to 60-year-old participants likely resulted in greater differences between residential and mobility-enhanced exposures compared with the full adult population.
Table 4.
Demographics of Participants in the Swiss and Dutch Tracking Campaigns Compared with National Population Dataa
Table 5 shows the comparison in the percentage of time spent during different models of travel between the tracking campaign and national statistics in Switzerland and the Netherlands. While the share of transport modes in the Dutch tracking campaign matches the national figures quite well, this is not the case for Switzerland. The participants of the tracking campaign conducted in Basel traveled half as much by car and walking (21% vs. 41% nationally) and more by public transport (32% vs. 11%) and by bicycle (26% vs. 7%). The differences are likely explained by the urban character of the Basel study area, with a well-connected public transport system, including an extensive bus, tram, and train network, compared with the whole of Switzerland. In the Dutch tracking campaign, participants cycled modestly more and used the car modestly less than the national average, possibly related to differences between the Utrecht province study area and the entire country (majority of participants) and the higher education/income of the tracking population.
Table 5.
Comparison of Percentage of Time per Commuter Mode Obtained from the Tracking Campaigns with National Figures for Switzerland and the Netherlands
b Hoeveel reisden inwoners van Nederland en hoe? | CBS, 2022.
5.2.2 Comparison of Exposures Based on GPS Readings from the App and SODAQ Device
In both the Swiss and Dutch campaigns, all participants were asked to use both the mobile phone app and the SODAQ device simultaneously to track their location. As explained in Section 3.3, the two tracking methods had different intervals and behaviors. For the final exposure assessment, we used the combined data from both devices to have the most complete dataset possible. Here, we present a comparison between the two tracking methods by splitting the datasets and calculating exposures for each device separately.
Figure S5.9 shows an example of the GPS readings from the mobile phone app and the SODAQ device recorded during piloting. It shows a commute to and from work (Figure S5.9A, top left) by bicycle and the difference in sampling rate, with approximately nine to ten SODAQ GPS readings per mobile app GPS reading. Figure S5.9B shows an example with missing GPS readings from the SODAQ tracking device during a portion of a cycle ride in Basel.
Imputation of missing GPS readings was performed only during the time when participants were at home. As only one tracking point per hour spent at home was needed, only a small number of tracking points needed to be imputed (2% of the total tracking points were imputed in Switzerland, and 1% were imputed in the Netherlands).
We performed a sensitivity analysis using GPS readings from the mobile phone app, the SODAQ devices, and the combination of both in the two countries. We compared the computed time-weighted air pollution exposures using both sets of GPS readings (Figure 6). A good agreement was found between NO2 exposures from the two different GPS capture methods by participants for Switzerland (R2 = 0.82), but less so for the Netherlands (R2 = 0.57) when looking at all activities together (Total). Correlations by activity showed a similar pattern between the two countries, with moderate correlations for Socializing, Travel, and Work/School (R2 = 0.43–0.63), and high correlations for Home/Other (R2 = 0.86 in Switzerland and R2 = 0.71 in the Netherlands). For PM2.5, the agreement between the two GPS capture methods was substantially smaller, especially in the Netherlands (Total: R2 = 0.80 in Switzerland; R2 = 0.28 in the Netherlands). Correlations by activity followed a similar pattern as with NO2, with higher correlations for Travel than the other three activities. Figure S5.10 shows a similar comparison as in Figure 6, but here the data were aggregated by activity. A good agreement was found between NO2 and PM2.5 exposures from the two different GPS capture methods, with R2 between 0.60 and 0.96 when aggregated by activity (Figure S5.10). For a small number of individual participants, large differences were found.
Comparison of NO2 (two left columns) and PM2.5 exposures (two right columns) based on GPS readings from the mobile phone app and the SODAQ device. NO2 and PM2.5 exposures were calculated and compared over the whole 2-week period (Total) and by activity (Home, Socializing, Work/School, and Travel). Each point in the scatterplots is a participant of the Swiss (top row) or Dutch (bottom row) tracking campaign. (Reprinted from Kyriakou et al. 2025; Creative Commons license CC BY 4.0)
In a few cases, either the SODAQ device or the mobile phone app recorded only very few GPS data points. Figure 6 shows the outliers in the graph of the Total exposure. The outliers represent participants who had incomplete measurement data from one device. This can result in a very different Total exposure, as one dataset represents a smaller time window, while the other dataset covers more points (and therefore more time and space). The plots by different activities for both countries show higher correlations between the two GPS measurements for Home/Other compared with the other activities — the only exception being Work/School PM2.5 exposure in the Netherlands, which had a higher R2. Thus, outlier points that had lots of missing data might not be evenly represented in all activities. When looking into single activity blocks, this is even more obvious because only activity blocks with data points from both devices are represented.
5.2.3 Comparison of ABM and Tracking Campaign Exposures
For the evaluation, we compared exposures derived from the tracking campaign and the ABM, as well as those based on the residential location only. Figure 7 shows the Bland-Altman plots for the three NO2 and PM2.5 exposure comparisons, where each point represents a participant of the tracking campaign. Figure 8 shows the corresponding scatterplots, including the correlation. As mentioned earlier, this evaluated the time–activity component of the ABM. Because the same ambient air pollution surface was used, it does not address the full uncertainty.
Bland-Altman plots showing comparisons between NO2 and PM2.5 (μg/m3) exposures for the Swiss and Dutch tracking campaign participants, based on the tracking data (combination of mobile phone app and SODAQ device GPS data), the ABM-enhanced mobility modeling, and the residential location only. (Adapted from de Hoogh et al. In review.)
The Bland-Altman plots (Figure 7) show that the majority of the points lie within ± 2 times the standard deviation, meaning that overall, there was good agreement between the three comparisons. However, there were differences between the three comparisons regarding the slope of the plots. The comparison of residential versus ABM for both NO2 and PM2.5 shows a clear positive slope in both Basel and the Netherlands. At low concentrations, the ABM-estimated concentrations were higher than the residential exposure, and vice versa at high concentrations. This can be explained by the fact that people who live in a less polluted area tend to get a higher overall exposure when including mobility (i.e., ABM), as they often work in a highly polluted area. This effect was the opposite at the higher end, where people who live in a highly polluted area tend to work in a less polluted area (and their ABM estimates will be lower). The same pattern, although weaker, can be observed in the comparison between residential and tracking, especially in Switzerland; in the Netherlands, this pattern was less pronounced. The Bland-Altman plots for the comparison between ABM and tracking were much more stable (i.e., no real patterns were observed, meaning that no clear pattern was detected, and that tracking and ABM-estimated concentrations were similar).
High correlations (Figure 8) were found between the tracking and ABM exposures for NO2 (R2 = 0.77 in Switzerland; R2 = 0.81 in the Netherlands) and moderate to high for PM2.5 (R2 = 0.80 in Switzerland; R2 = 0.52 in the Netherlands). Similar correlations were found in the comparison between tracking and residential-based exposures (NO2: R2 = 0.78–0.80; PM2.5: R2 = 0.55–0.76). The comparison between ABM and residential location yielded the highest correlations (R2 > 0.97 for both NO2 and PM2.5). Therefore, the high correlation between residential and ABM exposure was supported by the relationship with the tracking-based estimated exposure, though with more variability. In the sensitivity analysis, with assignment of a single realization instead of the mean ABM estimate from 50 realizations, we also found correlations between the tracking and mean ABM exposure comparison (Figure S5.11).
Scatterplots showing comparisons (plus R2s) between NO2 and PM2.5 exposures for the Swiss (left column) and Dutch (right column) campaign participants, based on the tracking data (combination of mobile phone app and SODAQ device GPS data), the ABM-enhanced mobility modeling, and the residential location only. (Adapted from de Hoogh et al. In review.)
5.3 EXPOSURE MODELING
5.3.1 Development of Long-Term Hourly NO2 and PM2.5 for Switzerland
We converted previously developed annual NO2 and PM2.5 surfaces (100 m × 100 m)32,33 into long-term hourly NO2 and PM2.5 concentration surfaces for the whole week (Mon–Sun), weekday (Mon–Fri), and weekend day (Sat–Sun) in Switzerland for 2016. The conversion was based on analysis of daily NO2 and PM2.5 monitoring data across Switzerland at 63 and 9 monitoring stations, respectively, representing urban, suburban, and rural background locations. Figures S5.12 and S5.13 show the diurnal variation of NO2 and PM2.5 at the Swiss monitoring sites for weekdays and weekend days. Diurnal patterns for NO2 showed more pronounced morning and evening rush hour peaks than for PM2.5.
Diurnal patterns of the ratios (based on formulas 1 and 2 presented in Section 4.3.1) were extracted for the long-term week (7-day), weekday (5-day), and weekend (2-day) periods and applied to the annual NO2 and PM2.5 surfaces (Figure 9 and Table S5.1). Figure S5.14 shows maps of predicted long-term hourly PM2.5 and NO2 concentrations across Switzerland for selected hours and averaged for weekday and weekend days.
Diurnal patterns of ratios for NO2 and PM2.5 for all days, weekdays, and weekend days for 2016, based on 63 NO2 and 9 PM2.5 background air pollution monitoring stations in Switzerland.
5.3.2 Development of Long-Term Hourly NO2 and PM2.5 for the Netherlands
Figures S5.15 and S5.16 show the diurnal variation of NO2 and PM2.5 at the monitoring sites, combining data from the Netherlands, Belgium, and Germany. Patterns were similar across the three countries (data not shown). Diurnal patterns for NO2 showed more pronounced morning and evening rush hour peaks than PM2.5. Patterns in the cold season and weekdays were more pronounced than in the warm season.
Figure 10 and Table 6 document the performance of the hourly models for NO2 and PM2.5. NO2 SLR models performed quite well overall (fivefold cross-validation mean R2 = 0.61–0.71), while the PM2.5 performed moderately well (fivefold cross-validation R2 = 0.39–0.50). For both NO2 and PM2.5, the warm-season models performed worse than the cold-season models. The models for the weekends performed worse than the weekdays. The performance of the RF models was slightly better than that of the SLR for both pollutants (Table 6). For both SLR and RF, variables with larger buffer sizes (representing variation in background concentrations) were selected more often in the weekend models than in the weekdays, and in the warm season than in the cold season. The difference in model performance and selection of variables across hours, seasons, and type of day (weekday or weekend day) highlights the benefit of developing independent hourly models when matching modeled exposure to hourly time–activity data.
Table 6.
Fivefold Cross Validation R2 of the NO2 and PM2.5 Dutch Models Developed Through Supervised Linear Regression (SLR) and Random Forest (RF)a
a The Mean SLR and Mean RF columns show the mean R2 of the hourly nitrogen dioxide (NO2) and fine particulate matter (PM2.5) models averaged by season (Warm and Cold), weekdays (0), and weekends (1). The columns SD, SLR, and SD RF show the standard deviation (SD) of the hourly R2 values. The column Improvement shows the percentage increase in R2 in RF compared with SLR. Note that the fivefold cross-validation R2 of the SLR and RF was calculated as a mean squared error (MSE)-based R2.
R2of the fivefold cross-validation hourly models developed through supervised linear regression (r2_slr) and random forest (r2_rf) for NO2and PM2.5, for cold and warm seasons, weekdays, and weekends. The NO2 (N = 544 sites) and PM2.5 (N = 227 sites) models were built from monitoring data for the Netherlands, Belgium, and Germany from 2016 to 2019. Note: Scales of NO2 and PM2.5 are different to illustrate differences between methods and season/weekday per pollutant. The fivefold cross-validation R2 of the SLR and RF was calculated as a mean squared error-based R2.
Figure S5.17 illustrates the predicted spatial patterns for selected hours, showing more pronounced spatial contrasts during daytime and the cold season for the city of Amsterdam. Figure S5.18 shows maps of predicted long-term hourly PM2.5 and NO2 concentrations across the Netherlands for selected hours and averaged for weekend and weekday.
5.4 LINKING EXPOSURES TO COHORTS
We extracted two different exposures to the cohorts: (1) the residential-only exposure (the exposure at the home address of the cohort participants), and (2) the mobility-enhanced exposure (the exposure resulting from ABM, taking into account commuting and work address). For a subset of SAPALDIA, the work address was known, and we also calculated an exposure based on the true workplace address. Table 7 shows the summary statistics for the three cohorts for these main exposures (residential only, mobility enhanced, and true workplace address) for both NO2 and PM2.5. In all cohorts, we documented slightly lower air pollution exposures for the mobility-enhanced exposures compared with the residential exposures. The difference was less than 5% in the cohorts. In SAPALDIA, using the actual work location showed slightly higher enhanced exposures. We also generally found slightly lower exposure contrasts for the mobility-enhanced exposure method, as reflected by the lower IQR. Differences between residential and mobility-enhanced exposure were larger for NO2 than for PM2.5, consistent with the larger spatial variation of this pollutant. Table S5.2 shows similar trends when we evaluated (more readily available) yearly pollution surfaces instead of the hourly surfaces.
Table 7.
Exposure Distributions (μg/m3) for Each of the Cohorts for Residential (RES) and Mobility-Enhanced (ENH) Exposuresa
a This includes Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults (SAPALDIA) (SAPALDIA3 and SAPALDIA4 follow-ups) and the subset of individuals with known workplace address; the full Swiss National Cohort (SNC) cohort and subset excluding those retired at baseline; and European Prospective Investigation into Cancer study, Netherlands cohort (EPIC-NL). n for SAPALDIA3 & SAPALDIA4 refers to the blood pressure outcomes.
Table 8 shows correlations between residential and mobility-enhanced exposure in each of the three cohorts. In all three, we reported very high correlations (R > 0.98) between residential exposure and mobility-enhanced exposures. The correlation was also very high (R > 0.95) when using the known work address (SAPALDIA3 subset only).
Table 8.
Pearson Correlations Between Residential (RES) and Mobility-Enhanced (ENH) NO2 and PM2.5 Exposures for the Three Cohorts, and at the Known Work Address (WP) for SAPALDIA
SAPALDIA = Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults; SNC = Swiss National Cohort; EPIC-NL = European Prospective Investigation into Cancer study, Netherlands cohort.
a For both repeated cross-sectional samples (SAPALDIA3 & SAPALDIA4).
b For only the cross-sectional sample (SAPALDIA3 subset).
For EPIC-NL (Table S5.3) and SNC (Figure S5.19), the standard deviation of the modeled exposure for the different profiles was 1–2 μg/m3 for NO2 and about 0.1–0.2 μg/m3 for PM2.5. When we randomly selected a single realization from the distribution, we still found very high correlations with residential exposure, though not as high as with the mean (Figures S5.20 and S5.21). Specifically, in the Netherlands, we found R2 values of 0.77 and 0.84 for NO2 and PM2.5, respectively, assuming the full cohort was a commuter. This represents an overestimate, as 64% of the cohort was employed at recruitment. For the SNC, the R2 decreased from 0.98 to 0.95 (for both pollutants) when using a randomly selected single realization versus the mean of the 50 realizations in the full cohort that combined commuters and homemakers, as applicable on an individual basis. The decrease in R2 was more obvious when focusing only on the commuters (R2 = 0.89 for NO2 and 0.90 for PM2.5).
5.5 RESULTS OF EPIDEMIOLOGICAL ANALYSIS FOR SNC, EPIC-NL
Study Population (SNC)
At baseline (January 1, 2011), there were 7.9 million individuals in Switzerland (thus in the SNC), with 5.2 million over the age of 30 years and eligible for this analysis. A complete case approach was used to form the analytical dataset. Sequential exclusions were made as follows: 1% had invalid home coordinates, 17.7% had missing employment status impacting the assignment of the mobility-enhanced exposures, and 0.7% had missing PM2.5 exposure due to gaps in the exposure surfaces. Finally, 2.9% and 14.6% were excluded on the basis of not being able to assign mobility-enhanced exposure for NO2 and PM2.5, respectively. In total, 33.4% of eligible adults were dropped, for a final study population of 3,488,705. During follow-up (until 31.12.2018; 26.1 million person-years), there were a total of 375,180 natural deaths and 126,385 cardiovascular disease–related deaths. An additional analysis where we dropped those who were retirement age at baseline, which is 65 for men and 64 for women, included 2,213,377 individuals, 17.3 million person-years, and 42,867 natural deaths and 8,242 cardiovascular deaths.
The mean (standard deviation) age of individuals in the SNC was 57.5 (15.6) years in the main cohort and 47.7 (9.4) in the subset not retired. The main cohort had slightly more women than men, and well over half of all individuals were married and employed. The majority, as expected, were Swiss nationals (84.1%), with the largest cultural group being German (74.3%) and residing mainly in the regions of Northwestern, Eastern, and Central Switzerland, and Zurich (Table S5.4).
Study Population (EPIC-NL)
During the follow-up period, 3,291 participants died of natural causes, of which 739 were from cardiovascular disease and 330 from lung cancer. We observed 3,137 incident stroke cases and 1,446 coronary events. Table S5.5 presents the population characteristics of the EPIC-NL study. Participants were on average 50 years old at baseline. Most participants were women (because one of the cohorts consisted of women only), were married/living together, had a medium educational level, and were employed.
Associations Between Air Pollution and Mortality (SNC)
Figures 11A and 11B show the association between residential exposure and mobility-enhanced exposure of PM2.5 and NO2 and natural and cardiovascular mortality in the SNC, respectively, for the full cohort and subset excluding those retired at baseline. The associations in tabular form are presented in Table S5.6A (expressed for fixed increments and per IQR).
Association for NO2 and PM2.5 and natural mortality and cardiovascular (CVD) mortality for the two different exposures (residential vs. mobility-enhanced) for the full SNC cohort per 5 μg/m3 for PM2.5, 10 μg/m3 for NO2. (Adapted from Niaye at al. 2025; Creative Commons license CC BY 4.0.)
Association for NO2 and PM2.5 and natural mortality and cardiovascular (CVD) mortality for the two different exposures (residential vs. mobility-enhanced) for the SNC subset excluding those retired at baseline per 5 μg/m3 for PM2.5, 10 μg/m3 for NO2. (Adapted from Niaye at al. 2025; Creative Commons license CC BY 4.0.)
Both NO2 and PM2.5 were positively associated with natural mortality, with only small differences in hazard ratio (HR) and confidence interval between residential exposure and mobility-enhanced exposure. NO2 was also associated with cardiovascular mortality, but PM2.5 was not (Figure 11A). In the additional analysis excluding those retired at baseline, the associations for natural mortality in particular were stronger and, again, were similar for residential only versus mobility enhanced, by pollutant (Figure 11B).
Associations Between Air Pollution and Mortality and Morbidity (EPIC-NL)
Figure 12 and Table S5.6A show the association between residential exposure and mobility-enhanced exposure of PM2.5 and NO2 and natural mortality and incidence of coronary events and stroke (expressed for fixed increments and per IQR in the table). Figures S5.22–S5.24 and Table S5.7 show these associations for different representations of the air pollution surface (hourly vs. annual) and the commuter profile to reflect the uncertainty of assignment of profile to cohort members. Additional adjustment for subcohort (as performed in ELAPSE) did not result in changes in the HRs.
Association for NO2 and PM2.5 and natural mortality, stroke incidence, and coronary events (CE) incidence for the two different exposures (residential vs. mobility-enhanced) for EPIC-NL per 5 μg/m3 for PM2.5, 10 μg/m3 for NO2. (Adapted from Niaye at al. 2025; Creative Commons license CC BY 4.0.)
PM2.5 was positively associated with natural mortality, with only small differences in the HR and confidence interval between residential exposure and mobility-enhanced exposure. The large effect estimate and wide confidence interval were due to the use of an increment that is much larger than the IQR (Table S5.6a). NO2 was negatively associated with natural mortality, with only small differences in the HR and confidence interval between residential exposure and mobility-enhanced exposure (Figures 12 and S5.22).
PM2.5 and NO2 were negatively associated with stroke incidence, with only small differences in the HR and confidence interval between residential exposure and mobility-enhanced exposure (Figures 12 and S5.23). Confidence intervals were wider than for natural mortality, related to a smaller number of cases.
PM2.5 was negatively associated with coronary events, with only small differences in the HR and confidence interval between residential exposure and mobility-enhanced exposure. NO2 was positively associated with coronary events, with only small differences in the HR and confidence interval between residential exposure and mobility-enhanced exposure (Figures 12 and S5.24).
5.6 RESULTS OF EPIDEMIOLOGICAL ANALYSIS FOR SAPALDIA
Repeated Cross-Sectional Analysis of Blood Pressure and Lung Function at SAPALDIA3 and SAPALDIA4
PM2.5, but not NO2, was associated with increased blood pressure, albeit with wide confidence intervals. PM2.5 but not NO2 was associated with lower FEV1 and FVC, with wide confidence intervals. Both PM2.5 and NO2 showed a small positive association with FEV1/FVC, with wide confidence intervals. Little difference was documented between residential and mobility-enhanced effect estimates (Figure 13).
Association for NO2 and PM2.5 and blood pressure and lung function for the two different exposures (residential vs. mobility-enhanced) for SAPALDIA3 and SAPALDIA4 per 5 μg/m3 for PM2.5, 10 μg/m3 for NO2.
Cross-Sectional Analysis of Blood Pressure and Lung Function at SAPALDIA3
No association was observed for blood pressure and lung function, regardless of the exposure estimate (Figure 14). No consistent difference in association was found for the known versus the estimated work location.
Association for NO2 and PM2.5 and blood pressure and lung function for the three different exposures (residential, mobility-enhanced, true work location) for SAPALDIA3 per 5 μg/m3 for PM2.5, 10 μg/m3 for NO2.
The validation of exposures derived from ABM using exposures derived from observed time–activity patterns is shown in Section 5.2. Good agreement was found, supporting the activity-based modeling approach. The exposure error evaluation using the EXPOsOMICS data is reported in Section S5.1.
No correlations existed between the personal PM2.5 measurement data and the ABM mobility-enhanced and ABM residential exposures for either the Basel or Netherlands sample (see Figure S5.25 in S5.1). The poor correlation can be attributed to a couple of factors. First, we are comparing an average of three 24-hour averages (from EXPOsOMICS) with an annual average (from MOBI-AIR), which is unlikely to correlate. Second, in the Netherlands, the range in modeled PM2.5 exposure was very small, limiting the possibilities to evaluate the correlation with external measures of exposure. Third, a review by Hoek40 points out that it is almost impossible to use personal measurements for long-term air pollution exposure assessment. The interpretation of the data comparison was difficult because of the insufficient number of measurements per person to assess long-term exposures with any certainty. The review also pointed out that personal measurements include indoor and outdoor air pollution concentrations that are difficult to separate. All the issues raised in the review applied to our comparison.
6. DISCUSSION AND CONCLUSIONS
The vast majority of studies investigating the associations between air pollution and health have done so by assuming a residential-based exposure, thereby ignoring the possible misclassification due to time–activity. In this study, we investigated whether a more sophisticated mobility-enhanced exposure assessment would lead to different results compared with using a residential-based exposure. Using ABM, we simulated a mobility-enhanced exposure to air pollution (NO2 and PM2.5), informed by travel survey and census data, for two large cohorts in Switzerland (SAPALDIA and SNC) and one (EPIC-NL) in the Netherlands. We evaluated the ABMs with GPS and time–activity data collected independently in tracking campaigns that included almost 700 participants from selected areas in the two countries. For these participants, the exposures based on the tracking data versus those derived from ABM showed a moderate to good agreement (R2= 0.52–0.81). Within the three cohorts, the mobility-enhanced exposure compared with exposure based only on the residential location showed very high correlations (R > 0.95). Finally, the epidemiological analyses revealed very small differences in the associations across health outcomes within the three cohorts for the different exposure estimates. In SAPALDIA, where the work address was known for a subset of individuals, we were further able to conduct a comparison with the ABM-estimated work address, indicating little difference in mobility-enhanced exposures.
6.1 TRACKING CAMPAIGN
Largely due to the COVID-19 pandemic that started at the onset of this project, we were unable to reach the initial goal of 1,000 participants each in Switzerland and the Netherlands. The Swiss tracking campaign, however, was more successful in terms of recruitment than the Dutch. This was facilitated by recruiting people from the COVCO-Basel cohort who stated in a COVCO survey that they were willing to participate in further environmental research, allowing us to achieve almost 500 participants and a high recruitment rate (33%). In the Netherlands, the tracking campaign participants could not be recruited from an existing cohort. Instead, we first attempted to recruit from a random sample of 3,000 people from the whole Dutch population, with a disappointing response rate of 1.4%. More targeted subsequent attempts using social media offered little gains, so we changed tactics and ultimately distributed 5,000 leaflets in the province of Utrecht, adding a voucher of 25 Euros as an incentive to participate. This was more successful, adding around 150 participants for a total of just under 200. We note that with 200–500 participants obtained in our campaigns, both can be considered large tracking studies when compared with previous tracking studies.20,41
Participants in the campaigns were tracked in two ways: (1) using a wearable tracking device and (2) using a mobile phone app with GPS functionality. As explained in Section 3.3, our initial plan to use an existing mobile phone app was not possible because of the severe battery drain from the GPS location functionality in both iOS and Android. This led to a change in protocol, using a wearable SODAQ device instead, with the added advantage of obtaining accurate GPS readings at a high temporal frequency (20 seconds). We also arranged for a purpose-designed mobile phone app to extract time–activity information from our participants. In the meantime, both iOS and Android changed the architecture of their software such that it was possible to also use the new mobile app for GPS measurements, albeit with a low temporal frequency (3–4 minutes). Ultimately, we decided to use both SODAQ and the app in our study and combine information from both sources. The study allowed us to compare the two methods. We found high correlations within the two study areas between total NO2 exposures based on the app and the SODAQ device. The correlations were smaller when compared by activity (i.e., Socializing, Work/School, Travel), except for when staying at home. For PM2.5 correlations, there are substantially smaller differences between the two data capture methods. The smaller correlation was mainly due to the difference in the frequency of the GPS readings between the two methods (e.g., while traveling). PM2.5 showed lower correlations, in part due to the smaller spatial contract in PM2.5 concentrations. Overall, we conclude that with the app — the more cost-effective approach — it is possible to obtain accurate time–activity data. This is important for scaling the approach to large populations, where deployment of a SODAQ-type device would likely not be feasible, both for practical and cost reasons.
6.2 AGENT-BASED MODELING
Modeling continuous movements for millions of agents is currently not feasible for two main reasons. First, a detailed parameterization for each agent is not available; second, such simulations would require a significant amount of computing time. By using our activity-based modeling approach to aggregate the space–time paths of agents, we were able to perform an exposure assessment for large numbers of agents using high-resolution air pollution datasets. Integrating the different components in the modeling framework provided the flexibility to execute runs to feed the model with different activity profiles, cohorts of various sizes, and multiple air pollutants for two different countries.
The results showed that only using the residential location for exposure assessment marginally estimates larger contrasts between individuals, as it only considered the home-based exposures. In addition, using the residential exposure for commuting persons overestimated the total exposure for those with a relatively low exposure at home, while it underestimated the total exposure for those living at locations with high air pollution. Under- or overestimation can be up to 8 μg/m3 NO2 over short distances. Exposure at the residential location thus did not fully capture time–activity, agent-based modeled exposure patterns, in particular for full-time workers who are commuting.
The Open Source Routing Machine routing engine used in the ABM supports various commute modes, including car, bike, walking, train, and other public transport, covering the most common commute modes. Due to a lack of consistent data for both countries and large population groups, commute modes could not be estimated directly. Instead, we used distance ranges for each commute mode and empirically set maximum commute distances for walking and biking, and a total maximum time for commuting. We decided to calculate the Euclidean distance to a work location as the first step to determining the corresponding distance range for commuting, to avoid a more computationally intensive evaluation across different transport modes over the traffic infrastructure. As the commute mode within a selected distance range was still unknown, we used a probabilistic assignment to account for the uncertainty of commute mode choices. This contributed to the total uncertainty in exposures calculated through the Monte Carlo simulation. The uncertainty introduced depends on the distance traveled and a randomly chosen commute mode. The largest uncertainties in the estimated exposures are expected when the distance to the work location exceeds 10 km and car or train modes are randomly assigned. Monte Carlo realizations that select a commute by car will likely result in higher exposure values because of the need to follow highways or national roads with high traffic intensities, while trains often pass through less densely populated and, therefore, less polluted areas and away from major roads. Differences in total exposure estimates between walking and biking as the commute mode are expected to be less pronounced because of the shorter durations of the commute activities.
We applied the same allocation scheme to all residential locations, neglecting spatial differences in commuting behavior, such as for urban or rural areas. The data scarcity did not allow us to account for other variables, such as age or socioeconomic status, in the commute mode allocation or to consider trips with different consecutive commute modes. Models incorporating more refined commute mode allocation may lead to somewhat different total exposure estimates.
We used 13 different activity profiles to cover the broad range of demographics in the cohorts and to understand how different assumptions would impact the resulting exposures. A smaller number of profiles could be used in exposure assessment to reduce computational demands. A residential profile could be omitted, unless comparison between residential and mobility-enhanced exposure is desired, and only one instead of three homemaker profiles could be used as the different buffer sizes only show minor variation in the resulting values. In addition, a known work location and the commute mode per participant would help to further reduce uncertainty in the commuter profiles.
Our study indicates that exposure values derived from our agent-based methodology differ from those estimated using the conventional approach, which typically relies on air pollution at the residential location — often represented by the front door. This holds for exposures to air pollution, as shown in this report. Similarly, a recent application of our agent-based approach to assessing exposure to traffic noise revealed comparable or greater differences between activity-based and residential exposures.42 These findings suggest that activity-based methods may offer valuable insights for exposure assessment and merit further consideration. A potential challenge in future applications of agent-based exposure assessments lies in the sensitivity of exposure estimates to the assumptions underpinning the agent-based model. In our study, the model’s design was guided by both an understanding of human activity and practical constraints specific to the study — in particular, the information available on individuals’ activities and the need for feasible model runtimes, given the requirement to process a large number of addresses and socioeconomic groups. While we have not explicitly quantified the variation in exposure values across different agent-based model configurations, our findings provide an initial basis for evaluating the stability of exposure values under varying modeling configurations.
In examining the agent-based model itself, differences emerge in the magnitude of the contribution of activities to the long-term average exposure. Commute trips lead to elevated exposure levels due to higher pollutant concentrations near roads and during commute hours. However, the relatively short duration of commuting (two commute trips, maximum commute time of 2 hours), compared with other activities, limits its overall contribution to the total exposure. As a result, different approaches than those followed in our model to calculate commute-related exposure are unlikely to affect the long-term average exposure estimates considerably. Similarly, the approach followed to calculate exposure during leisure activities may be of limited relevance, as these activities also tend to be of short duration. Our findings support this: varying the buffer sizes used to estimate exposure during leisure time — whether 1 km, 5 km, or 10 km — does not result in notable differences in long-term average exposure estimates. By contrast, long-term exposures are likely to be more sensitive to the methods used for estimating exposure at residential and work locations, given that these activities account for the majority of time spent: approximately 50%–100% of the total duration. To assess these exposures, our ABM averages air pollution over a small buffer, representing the size of a typical building, centered at the geographical coordinates of the residence or workplace. Models that calculate residential or workplace exposure based on air pollution at the front door or over a broader area (e.g., neighborhood, city) may yield different long-term exposure values. Infiltration factors, which we did not apply in our study, and different assumptions (e.g., specific values for different buildings) are expected to give different values.
Approaches used to estimate inputs of the ABM will also affect the values of the calculated air pollution exposures. Road network is an essential input for modeling commute trips, and in most regions, the data are expected to be sufficiently detailed and reliable given the extensive global coverage of OpenStreetMap. However, other inputs are likely to have a more pronounced impact on exposure estimates. One is the methodology employed to estimate the work location. In many cohorts, the geographical coordinates of the work location are unknown and must be inferred from available information. Our study employed origin–destination matrices to estimate the administrative region of the workplace. Although these matrices were stratified by socioeconomic status and sex, this stratification did not result in considerable differences in exposure estimates across strata. Instead, 50 workplace locations were randomly selected from buildings in the region, from which routes along the transport network were modeled; after linkage to air pollution surfaces, average air pollution exposures were calculated. Future studies may need to employ approaches to further constrain the work location. Such an approach would require geographical information, along with thematic information regarding the types of work carried out at specific locations. Another important input for working people is the number of days the person works at the workplace, as this determines the contribution of the commute exposure and workplace exposure to the overall exposure. Other essential inputs include the air pollution surface and the infiltration factors applied, both of which have been shown in previous studies18 to play an important role in the exposure.
As explained in Section 3.1, due to severe delays in starting the tracking campaigns as a result of the COVID-19 pandemic, we had to adapt the original plan of using information obtained from the tracking campaigns to inform choices in the ABM. Instead, we evaluated the performance of ABM with mobility patterns obtained from the tracking campaigns. Our team believes that this has not impacted the study because the travel survey data used in both countries are more representative of the total and cohort populations in terms of education, income, employment, etc. The same level of representativeness can never be achieved in a small tracking campaign; therefore, tracking campaigns tend to be biased from the onset.
Exposure Differences Between Residential and Enhanced Exposures (Unknown Work Address)
In all three cohorts, very high correlations were found between annual average residential exposure and mobility-enhanced exposure of NO2 and PM2.5. Furthermore, exposure levels were very similar between the two approaches. We did find a slightly smaller exposure contrast for mobility-enhanced exposure. This can be explained by the phenomenon that participants living in residential locations with high concentrations can only work in areas with lower concentrations and vice versa. Finally, we found that the residential exposure also correlated well with the mobility-enhanced exposure based on actual time–activity data in the tracking population of almost 700 participants. Correlations based on the mean exposure across the 50 ABM realizations were lower than observed and were more in line with those from a single, randomly selected realization from the 50.
As elaborated in our published review assessing bias in epidemiological studies due to the use of residential address-based exposure,43 our observations are in line with several previous studies comparing residential and mobility-enhanced exposures.10,11,44-46 In all these studies, ABM was conducted using time surveys. Setton and co-workers46 documented that the agreement between residential and mobility-enhanced exposure diminished with increasing time at work and with increasing distance between the home and work locations. A recent study using tracking data also reported high correlations for noise and PM2.5.41 Other studies have performed comparisons without ABM, based solely on exposure determined at the residential address (home) versus exposures including work address (work) and during commute. Two studies compared air pollution exposure at the residential address with exposure while commuting and at the Work/School address with the results from Basel10 and Vancouver and Southern California46; these indicate that, while there is room for improvement, it is reasonable to use exposure characterized at the residential address. A study in Montreal, Canada, showed that almost 90% of individuals had a lower estimated 24-hour daily average NO2 exposure at home compared with a mobility-enhanced NO2 exposure.47 Researchers in the Netherlands followed 269 adults with a GPS-enabled smartphone app for 7 days and compared the residential exposure with only a mobility-enhanced PM2.5 exposure, concluding that the residential exposure was a good proxy for overall exposure to outdoor air pollution.41 A study in Shenzhen, China, used cell phone data from more than 300,000 individuals to assess the impact of mobility on air pollution exposures. They concluded that mobility did not significantly impact exposures, in particular for large-scale population studies, but that it did impact exposure on the individual level.48 A study in the United Kingdom compared population-weighted NO2, PM2.5, and O3 exposures at the residential address with only a combined residential and work exposure. They used a chemistry transport model to estimate rush-hour-specific long-term averages and found only a small increase in population-weighted NO2 and PM2.5 exposures when including the work location (2% and 0.3%, respectively).49
Comparisons of Known and Unknown Work Address
In the SAPALDIA study subset, where the actual work address of the subject was known, we were able to compare exposures and associations using ABM mobility-enhanced exposures and exposures based on the actual work address. We observed very high correlations (R = 0.95–0.97) between exposure estimated for the known work address and the estimated work address, suggesting that for epidemiological studies, ABM with the estimated work address is useful. We found moderate to high correlations between ABM simulations when the work location was estimated from travel surveys versus known work locations in the tracking campaigns, supporting ABM as a reasonable approach. It can, therefore, be beneficial for an agent-based exposure assessment if the work location of cohort participants is known. Thus, we recommend that cohort studies collect information on the work address. Additionally, we found no consistent differences in associations for blood pressure and lung function when using the actual work address or the estimated, a logical result given the high correlation found between the two exposures.
Associations with Health
Consistent with the very high correlation and the modest difference in exposure level and contrast between residential and mobility-enhanced exposure, we found no consistent differences in effect estimates between the two exposures when applied in three cohort studies of adults. We found positive and generally statistically significant associations with natural and cardiovascular mortality in the large SNC cohort, of similar magnitude for both exposure methods. These associations are also in line with previous results from the ELAPSE study that followed individuals in the SNC from 2000 to 2014 and had slightly different data for confounder adjustment.9,50 For natural mortality, the hazard ratios were 1.05 (1.04; 1.06) versus 1.04 (1.03; 1.04) per 10 μg/m3 NO2 and 1.03 (1.02; 1.04) versus 1.02 (1.01; 1.03) per 5 μg/m3 PM2.5 in ELAPSE versus using the residential only exposure from the current study, respectively. For cardiovascular mortality, the comparisons were 1.03 (1.01; 1.04) versus 1.02 (1.01; 1.03) per 10 μg/m3 NO2 and 1.03 (1.01; 1.04) versus 1.00 (0.99; 1.01) per 5 μg/m3 PM2.5 in ELAPSE versus using the current study’s residential-only exposure.
We found both positive and negative, generally nonsignificant, associations between PM2.5 and NO2 and mortality and cardiovascular morbidity in the Dutch EPIC-NL study, with nearly identical effect estimates for residence and mobility-enhanced exposures. Our EPIC-NL effect estimates were mostly in line with previous papers in this cohort on these outcomes.14,27,51Table S6.1 presents effect estimates in the current EPIC-NL study and previous analyses in the same cohort. Consistently, none of the analyses showed positive associations with these outcomes. For mortality, we found the same pattern of larger effect estimates for PM2.5. For coronary events, we found nonsignificant positive associations with NO2 and negative associations with PM2.5 in all three papers. The health and covariate data of the current study are taken from the ELAPSE study, but the exposure model differs (ELAPSE model based on 2010 European monitoring vs. new models based on the hourly land use regressions developed using routine monitoring in the Netherlands, Belgium, and Germany in 2016–2019 in the current study). In the Downward study, follow-up and covariate data were different, and the exposure model differed (European Study of Cohorts for Air Pollution Effects [ESCAPE] models52,53). In the Downward study, significant associations with heart failure and myocardial infarction were found for NO2 but not PM2.5, but as these were based on a much smaller number of events, we preferred to use broader categories to compare associations between residential and mobility-enhanced exposures.
We also found positive and negative, generally nonsignificant associations between PM2.5 and NO2 and lung function and blood pressure in the Swiss SAPALDIA study, with nearly identical effect estimates for residential and mobility-enhanced exposures. There are no previously published results based on the same SAPALDIA follow-up as used in the current study. However, in a meta-analysis of 15 European cohorts in the ESCAPE project,54 of which SAPALDIA contributed with earlier survey data (SAPALDIA2), neither PM2.5 nor NO2 showed significant associations with blood pressure. Another ESCAPE meta-analysis of five European cohorts, including SAPALDIA (SAPALDIA1/2), reported that NO2 but not PM2.5 exposure was associated with lower levels of FEV1 and FVC, particularly in obese individuals.21
The epidemiological analyses were performed with the mean ABM exposure from 50 realizations, which may have resulted in a smaller difference in effect estimates between residential and mobility-enhanced exposures. However, because the actual work location was unknown (in all except the SAPALDIA subset, as discussed above), we consider the mean as the best estimate for an individual. The ABM approach did allow for potential systematic differences in the mean between individuals, specifically between profiles and different origin–destination matrices depending on the residential area. The very high correlation between a single, randomly selected ABM exposure and residential exposure suggests that the observed difference in effect estimates is likely representative.
Few previous studies have compared health effects directly in epidemiological studies. Our observations are in line with a study in schoolchildren, showing nearly identical effect estimates for lung function comparing residential and mobility-enhanced exposures.11 In that study, little difference in the effect estimate was found using the estimated and actual school locations. Our results are also in line with studies evaluating the potential bias in epidemiological studies related to assessing exposure only at the residential address.10,46 Generally, these studies showed only moderate bias. Specifically, in a comparison between residential and mobility-enhanced exposure in Vancouver and Southern California, Setton and colleagues showed that the bias in assessing exposure at the residential address remained below 30%.46 A similar study in Basel, Switzerland, estimated the bias to be lower, at 12%.10 Researchers in Canada assigned estimates of home and workplace PM2.5 exposures to the 2001 Canadian Census Health and Environment Cohort (CanCHEC) and found no differences in nonaccidental mortality when comparing the combined exposures with home exposure only.55 A recent US-wide study investigated the influence of using both the home and work address in determining PM2.5 exposures in populations of different socioeconomic status. In addition to disparities in exposure by racial and ethnic groups, other groups, including those living in urban areas and the younger populations, showed a higher bias in health impact by using home-only exposures compared with a time-weighted home and work exposure.56
In two (SNC and EPIC-NL) of the three cohorts used in this study, no information was available about the actual work address, and in all three, the actual commuting route was unknown. Instead, we used modeled routes and work addresses based on statistics from national surveys, resulting in random error in the enhanced exposure estimates. Generally, random error in exposure tends to bias health effects downward. However, the high correlation between the different exposures (and subsequent very good agreement in the health effects observed for adults in Switzerland and a large part of the Netherlands) suggests that the random error component is small and likely not very important. Given the findings in previous studies, we suspect that this conclusion will apply in more settings in developed countries. If time–activity patterns differ substantially from those analyzed in our study, the differences between residential and activity-enhanced exposures may be larger.
6.3 STRENGTHS AND LIMITATIONS
6.3.1 Strengths
For this study, we developed a flexible framework for ABM and used it to simulate mobility patterns for a large number of people, including the entire adult population in Switzerland in the nationwide SNC (3.5 million), EPIC-NL (33,000), and SAPALDIA (5,000). ABM was performed for different profiles reflecting specific demographics in the cohorts studied. We applied all 13 profiles and exposure estimates for each to every home location (i.e., building geocodes) in Switzerland, meaning that these can be easily applied to derive mobility-enhanced exposures for participants in any other Swiss cohort.
We conducted tracking campaigns in two countries with almost 700 participants. Two weeks of tracking detailed time–activity data allowed for robust comparisons to evaluate the ABM estimates, and provided a rich resource for future research. The mobile phone app developed by Games for Health was a success, and it has the potential to remain a viable application going forward.
The three cohorts included in this study are large and representative of the general population in each country; they were used in previous (influential) air pollution studies, such as ESCAPE and ELAPSE. The cohorts were complementary in that they could be used to investigate specific issues: SAPALDIA was used to compare known versus estimated work location; EPIC-NL looked deeply into the 13 profiles and compared ABM-derived air pollution estimates from hourly and yearly air pollution surfaces; SNC offered full national coverage and enabled analysis of the working-age population by excluding individuals retired at baseline.
6.3.2 Limitations
The time between the baseline of the cohort studies (1993–2013) does not align with the air pollution surfaces used in the exposure assessment (2016 for Switzerland, 2016–2019 for the Netherlands). However, as our main goal was to compare residential with mobility-enhanced exposure, we did not consider this a serious limitation, especially given the previously documented spatial stability of air pollution exposure contrasts.36
A potential limitation is that the tracking campaigns started shortly after returning to normal following the COVID-19 pandemic lockdown and other restrictions. However, a detailed evaluation of travel trends using Google and TomTom was conducted in both study areas to ensure travel had returned to a stable and near-prepandemic level of activity.
The participants of the tracking campaigns in both countries were not representative of the full Swiss and Dutch population, with ages 40–60, higher-education, high-income, and full-time workers groups being overrepresented. Unfortunately, we were unable to influence the recruitment and better match the study population of our tracking campaigns with the overall population. The overrepresentation of highly educated participants agrees with most exposure and epidemiological studies in which participants are invited.
Our findings apply to long-term air pollution exposure, and we do not know the impact of mobility on short-term exposure estimates. Furthermore, our findings apply to the studied pollutants, NO2 and PM2.5. We cannot completely rule out different findings for other pollutants. However, Bouma and colleagues57 reported high spatial correlation in the Netherlands between NO2 and UFPs and BC, and Saha and colleagues16 reported a high correlation between modeled UFPs and NO2. In Switzerland, Vienneau and colleagues50 also showed a high correlation between NO2 and BC (0.91) and Cu and Fe (>0.88) within the SNC. Given these high correlations and the fact that NO2 is often considered a surrogate for traffic-related pollutants, we suspect that qualitatively, our findings apply to more traffic-related pollutants, such as UFPs, BC, and PM2.5 Fe and Cu. However, if the spatial pattern of pollutants differs substantially, results may differ. For future research needs to increase the monitoring of other pollutants of interest, including nontraffic-related pollutants.
7. IMPLICATIONS OF FINDINGS
Our results suggest that the assessment of air pollution exposure at the residential address in epidemiological studies generally does not lead to substantial bias in health effects estimates. This conclusion likely applies to many studies conducted in Europe and possibly North America. If time–activity patterns differ substantially from the patterns analyzed in our study, differences between residential and activity-enhanced exposures may be larger.
Despite the good agreement between residential and work location, exposure research should continue to strive toward improving exposure assessment in large-scale epidemiological studies to minimize exposure misclassification.
Tracking campaigns can be conducted successfully using either a mobile phone app or a wearable GPS device. The mobile phone app’s low frequency (every 3–4 minutes) of GPS reading turned out to be sufficient for our purpose. For studies requiring finer-scale readings, a purpose-built GPS device is recommended.
The use of ABM in this study has proven successful in simulating time–activity patterns for large study populations and allowing the calculation of mobility-enhanced air pollution exposures. It is important to base these simulations on information extracted from travel surveys from the studied region/country and to apply these to different target groups in terms of age, sex, and socioeconomic position. ABM is useful for both health impact assessment and epidemiological studies, and even more useful when the work address is available.
Our study findings should be understood in the broad spectrum of measurement error in long-term exposure to air pollution. Among the various sources of bias in the air pollution exposure assessment, such as spatiotemporal resolution, composition of the air pollutants, building penetration rates that determine indoor-outdoor contrasts, and individual- and activity-driven respiratory rate, our study only addresses the people’s mobility. Our study deployed approaches scalable to large cohorts without relying on costly personalized information.
Epidemiological analysis for EPIC-NL was conducted on a secure IT environment at Utrecht University; no individual data can be transferred. Likewise, an analysis of SNC was conducted on a secure IT environment at the Swiss Tropical and Public Health Institute, and no individual data can be transferred. COVCO Basel and SAPALDIA data are available upon reasonable request to Principal Investigator N. Probst-Hensch, at nicole.probst@swisstph.ch. The MOBI-AIR app has been discontinued and is no longer available on the App Store.
ACKNOWLEDGMENTS
Access to the Dutch national supercomputer facilities was granted by SURF (project EINF-4618).
SAPALDIA was supported by the Swiss National Science Foundation, SNF-SAPALDIA (grant numbers 33CS30-148470/1, 33CSCO-134276/1, 33CSCO-108796, 324730_135673, 3247BO-104283, 3247BO-104288, 3247BO-104284, 3247-065896, 3100-059302, 3200-052720, 3200-042532, 4026-028099, PMPDP3_129021/1, PMPDP3_141671/1), SNF-SiRENE (grant number CRSII3_147635), and the Swiss Federal Office for the Environment. SAPALDIA is also supported by the Federal Office of Public Health; the Federal Office of Roads and Transport; the cantonal governments of Aargau, Basel-Stadt, Basel-Land, Geneva, Luzern, Ticino, Valais, and Zürich; the Swiss Lung League; the cantons’ Lung League of Basel Stadt/Basel Landschaft, Geneva, Ticino, Valais, Graubünden, and Zurich; Stiftung ehemals Bündner Heilstätten; SUVA; Freiwillige Akademische Gesellschaft; UBS Wealth Foundation; Talecris Biotherapeutics GmbH Abbott Diagnostics; European Commission 018996 (GABRIEL); and Wellcome Trust (WT 084703MA).
SAPALDIA could not have been done without the help of the study participants, technical and administrative support, and the medical teams and field workers at the local study sites.
Study directorate: NM Probst-Hensch (PI; e/g); T Rochat (p), C Schindler (s), N Künzli (e/exp), JM Gaspoz (c).
Scientific team: JC Barthélémy (c), W Berger (g), R Bettschart (p), A Bircher (a), C Brombach (n), PO Bridevaux (p), L Burdet (p), Felber Dietrich D (e), M Frey (p), U Frey (pd), MW Gerbase (p), D Gold (e), E de Groot (c), W Karrer (p), F Kronenberg (g), B Martin (pa), A Mehta (e), D Miedinger (o), M Pons (p), F Roche (c), T Rothe (p), P Schmid-Grendelmeyer (a), D Stolz (p), A Schmidt-Trucksäss (pa), J Schwartz (e), A Turk (p), A von Eckardstein (cc), E Zemp Stutz (e).
Scientific team at coordinating centers: M Adam (e), I Aguilera (exp), S Brunner (s), D Carballo (c), S Caviezel (pa), I Curjuric (e), A Di Pascale (s), J Dratva (e), R Ducret (s), E Dupuis Lozeron (s), M Eeftens (exp), I Eze (e), E Fischer (g), M Foraster (e), M Germond (s), L Grize (s), S Hansen (e), A Hensel (s), M Imboden (g), A Ineichen (exp), A Jeong (g), D Keidel (s), A Kumar (g), N Maire (s), A Mehta (e), R Meier (exp), E Schaffner (s), T Schikowski (e), M Tsai (exp) (a) allergology, (c) cardiology, (cc) clinical chemistry, (e) epidemiology, (exp) exposure, (g) genetic and molecular biology, (m) meteorology, (n) nutrition, (o) occupational health, (p) pneumology, (pa) physical activity, (pd) pediatrics, (s) statistics.
Footnotes
* A list of abbreviations and other terms appears at the end of this volume.
HEI QUALITY ASSURANCE STATEMENT
The conduct of the study “Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects” was subjected to an independent audit by David Bush and Scott Adamson of Trinity Consultants, Inc. Mr. Bush and Mr. Adamson are experts in quality assurance for air quality monitoring studies and data management.
The audit included a review of data quality for conformance to the study protocol as detailed in the final report and the study’s quality assurance plan, reviewing data quality for each of the study components. In January 2025, an off-site audit was conducted via a teleconferencing platform with primary study personnel. The audit concentrated on the study’s quality assurance and data management activities and included a review of the overall process utilized to collect new data and to manage and combine the exposure, air quality, epidemiological, and modeling data. Also evaluated were the procedures and measures undertaken to ensure quality and consistency in the processed databases and modeling results. Examples of datasets for the different types of data and modeling files were displayed by study personnel and reviewed for consistency, clarity, and completeness.
A written report of the audit was provided to the HEI project manager, who transmitted the findings to the principal investigator. The quality assurance audit demonstrated that the study was conducted by an experienced team with a high concern for data quality. Study personnel were responsive to audit questions and recommendations, providing formal responses that adequately addressed all issues. The report appears to be an accurate representation of the study.
The Additional Materials contain figures, tables, and other items for Sections 3, 4, 5, and 6, which are not included in the main report. They are available on the HEI website at www.healtheffects.org/publications.
Kees de Hoogh has a PhD from the University of Leicester, United Kingdom. He is currently a group leader of the Environmental Exposure Modelling group at the Swiss Tropical and Public Health Institute, Allschwil, Switzerland. He has coordinated exposure assessment to air pollution and other environmental stressors in national and international projects.
Benjamin Flückiger has an MA in Geography from the University of Basel, Switzerland. He is working as a scientific collaborator at the Swiss Tropical and Public Health Institute, Allschwil, Switzerland. He specializes in spatial data processing and environmental modeling.
Nicole Probst-Hensch has a PhD from the University of Basel, Switzerland, and the University of California, Los Angeles, USA. She is head of the Department of Epidemiology and Public Health and of the Exposome Science Group at the Swiss Tropical and Public Health Institute, Allschwil, Switzerland. She conducts exposome research as principal investigator of cohorts and biobanks in high-income, as well as low- and middle-income, countries.
Danielle Vienneau has a PhD from Imperial College London, United Kingdom. She is currently an associate professor in the Julius Center for Health Sciences and Primary Care at the University Medical Center Utrecht, the Netherlands, and previously co-led the Environmental Exposure Modelling group at the Swiss Tropical and Public Health Institute, Allschwil, Switzerland. Her research focuses on spatial modeling of environmental pollution and understanding the health consequences of exposure.
Ayoung Jeong has a PhD from the University of Basel, Switzerland. She is currently a senior scientist at the Swiss Tropical and Public Health Institute, Allschwil, Switzerland. She conducts exposome research in cohorts and biobanks established in high-income, as well as low- and middle-income, countries as part of the institute’s Exposome Science Group.
Medea Imboden has a PhD from the University of Pierre et Marie Curie, Paris, France. She is currently a senior scientist at the Swiss Tropical and Public Health Institute, Allschwil, Switzerland. She conducts exposome research in cohorts and biobanks established in high-income, as well as low- and middle-income, countries as part of the institute’s Exposome Science Group.
Aletta Karsies has a PhD from the University of Basel, Switzerland. She is experienced in air pollution modeling and big environmental data processing.
Sophie Baruth has an MSc in Environmental Sciences, majoring in ecology and evolution, at the Swiss Federal Institute of Technology Zurich, Switzerland.
Désirée De Ferrars has an MSc in Environmental Sciences from the Swiss Federal Institute of Technology Zurich, Switzerland.
Oliver Schmitz received his MSc degrees in computer science (2003) and environmental management (2007) from Kiel University, Germany, and a PhD (2014) in Geosciences from Utrecht University, the Netherlands. He is working as a research software engineer in the Department of Physical Geography at Utrecht University, with an emphasis on the development of generic software applications for constructing spatiotemporal models simulating environmental systems. In this project, he mainly developed the modeling framework for agent-based exposure assessment and ran all simulations for the Netherlands and Switzerland.
Meng Lu has a PhD in Geoinformatics from the University of Muenster, Germany. She is a junior professor in geoinformatics in the Department of Geography, University of Bayreuth, Germany, specializing in the theory and application of statistics, machine learning, agent-based modeling, remote sensing, and big computational techniques to the quantification and understanding of the spatiotemporal phenomena of our environment and society.
Roel Vermeulen has a PhD from Wageningen University, the Netherlands. He is a distinguished university professor at Utrecht University, the Netherlands, with a focus on developing new methods for quantifying exposures and their biological impacts. He is the chair of the Planetary Health Program at Utrecht University. He leads the Dutch program on the Exposome and the International Human Exposome Network. He coordinates the EU-funded, multicenter study EXPANSE EXposome Powered tools for healthy living in urbAN Settings, a 5-year project on the impact of the urban exposome on cardiometabolic disease.
Kalliopi Kyriakou has a PhD from the Aristotle University of Thessaloniki, Greece, where she now holds an assistant professor position. Her research interests include developing methods for environmental exposure assessment, with a focus on geospatial data, to address urban challenges.
Aisha Ndiaye is a PhD candidate at the Vrije Universiteit Amsterdam, the Netherlands. She currently works at the Dutch National Institute for Public Health and the Environment (RIVM). Her research focuses on air pollution and temperature epidemiology.
Youchen Shen is a PhD candidate at Utrecht University, the Netherlands. Her research focuses on using geospatial data and statistical methods to estimate environmental factors.
Derek Karssenberg has a PhD from Utrecht University, the Netherlands. He is a professor of computational geography at Utrecht University, focusing on spatiotemporal modeling to understand physical geographical and socioecological systems.
Gerard Hoek has a PhD from Wageningen University, the Netherlands. He is an associate professor at the Institute for Risk Assessment Sciences, Utrecht University, the Netherlands. His research is focused on air pollution exposure assessment and air pollution epidemiology.
OTHER PUBLICATIONS RESULTING FROM THIS RESEARCH
de Hoogh K, Flückiger B, Probst-Hensch N, Jeong A, Imboden M, Karsies A, et al. In review. Comparison of residential and mobility-integrated air pollution exposures from tracking campaigns and agent-based modelling in Switzerland and the Netherlands. J Expo Sci Environ Epidemiol.
Hoek G, Vienneau D, de Hoogh K. 2024. Does residential address-based exposure assessment for outdoor air pollution lead to bias in epidemiological studies? Environ Health 23:75, https://doi.org/10.1186/s12940-024-01111-0.
Lu M, Schmitz O, de Hoogh K, Hoek G, Li Q, Karssenberg D. 2022. Integrating statistical and agent-based modelling for activity-based ambient air pollution exposure assessment. Environ Model Softw 158:105555, https://doi.org/10.1016/j.envsoft.2022.105555.
Kyriakou K, Flückiger B, Vienneau D, Probst-Hensch N, Jeong A, Imboden M, et al. 2025. GPS tracking methods for spatiotemporal air pollution exposure assessment: comparison and challenges in study implementation. Int J Health Geogr 24:17, https://doi.org/10.1186/s12942-025-00405-x.
Ndiaye A, Shen Y, Kyriakou K, Karssenberg D, Schmitz O, Flückiger B, et al. 2024. Hourly land-use regression modeling for NO2 and PM2.5 in the Netherlands. Environ Res 256:119233, https://doi.org/10.1016/j.envres.2024.119233.
Ndiaye A, Vienneau D, Flückiger B, Probst-Hensch N, Jeong A, Imboden M, et al. 2025. Associations between long-term air pollution exposure and mortality and cardiovascular morbidity: A comparison of mobility-integrated and residential-only exposure assessment. Environ Int. 198:109387, https://doi.org/10.1016/j.envint.2025.109387.
Schmitz O, de Hoogh K, Probst-Hensch N, Jeong A, Flückiger B, Lu M, et al. 2025. A computational framework for agent-based assessment of multiple environmental exposures. J Expo Sci Environ Epidemiol Aug 2, https://doi.org/10.1038/s41370-025-00799-7.
3.GBD 2021 Risk Factors Collaborators. 2024. Global burden and strength of evidence for 88 risk factors in 204 countries and 811 subnational locations, 1990–2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet
403:2162–2203, https://doi.org/10.1016/s0140-6736(24)00933-4.
[DOI] [PMC free article] [PubMed] [Google Scholar]
4.Health Effects Institute (HEI). 2022. Systematic Review and Meta-analysis of Selected Health Effects of Long-Term Exposure to Traffic-Related Air Pollution. Special Report 23. Boston, MA: Health Effects Institute. [Google Scholar]
5.Boogaard H, Patton AP, Atkinson RW, Brook JR, Chang HH, Crouse DL, et al. 2022. Long-term exposure to traffic-related air pollution and selected health outcomes: a systematic review and meta-analysis. Environ Int
164:107262, https://doi.org/10.1016/j.envint.2022.107262.
[DOI] [PubMed] [Google Scholar]
7.Ott W, Wallace L, Mage D, Akland G, Lewis R, Sauls H, et al. 1986. The Environmental Protection Agency's research program on total human exposure. Environ Int
12:475–494, https://doi.org/10.1016/0160-4120(86)90064-4. [Google Scholar]
9.Stafoggia M, Oftedal B, Chen J, Rodopoulou S, Renzi M, Atkinson RW, et al. 2022. Long-term exposure to low ambient air pollution concentrations and mortality among 28 million people: results from seven large European cohorts within the ELAPSE project. Lancet Planet Health
6:e9–e18, https://doi.org/10.1016/S2542-5196(21)00277-1.
[DOI] [PubMed] [Google Scholar]
11.Ntarladima AM, Karssenberg D, Vaartjes I, Grobbee DE, Schmitz O, Lu M, et al. 2021. A comparison of associations with childhood lung function between air pollution exposure assessment methods with and without accounting for time-activity patterns. Environ Res
202:111710, https://doi.org/10.1016/j.envres.2021.111710.
[DOI] [PubMed] [Google Scholar]
13.Steinle S, Reis S, Sabel CE. 2013. Quantifying human exposure to air pollution — moving from static monitoring to spatio-temporally resolved personal exposure assessment. Sci Total Environ
443:184–193, https://doi.org/10.1016/j.scitotenv.2012.10.098.
[DOI] [PubMed] [Google Scholar]
14.Brunekreef B, Strak M, Chen J, Andersen ZJ, Atkinson R, Bauwelinck M, et al. 2021. Mortality and Morbidity Effects of Long-Term Exposure to Low-Level PM2.5, BC, NO2, and O3: An Analysis of European Cohorts in the ELAPSE Project. Research Report 208. Boston, MA: Health Effects Institute. [PMC free article] [PubMed] [Google Scholar]
15.Hoek G, Beelen R, Kos G, Dijkema M, van der Zee SC, Fischer PH, et al. 2011. Land use regression model for ultrafine particles in Amsterdam. Environ Sci Technol
45:622–628, https://doi.org/10.1021/es1023042.
[DOI] [PubMed] [Google Scholar]
16.Saha PK, Hankey S, Marshall JD, Robinson AL, Presto AA. 2021. High-spatial-resolution estimates of ultrafine particle concentrations across the continental United States. Environ Sci Technol
55:10320–10331, https://doi.org/10.1021/acs.est.1c03237.
[DOI] [PubMed] [Google Scholar]
18.Lu M, Schmitz O, de Hoogh K, Hoek G, Li Q, Karssenberg D. 2022. Integrating statistical and agent-based modelling for activity-based ambient air pollution exposure assessment. Environ Model Softw
158:105555, https://doi.org/10.1016/j.envsoft.2022.105555. [Google Scholar]
19.Jeong A, Galliker F, Imboden M, Keidel D, de Hoogh K, Vienneau D, et al. 2022. Residential exposure to greenspace and life satisfaction in times of COVID-19: a cross-sectional analysis of 9444 participants from a population-based study in Basel-Stadt and Basel-Landschaft. Swiss Med Wkly
152:w30204, https://doi.org/10.4414/smw.2022.w30204.
[DOI] [PubMed] [Google Scholar]
20.Donaire-Gonzalez D, Valentín A, van Nunen E, Curto A, Rodriguez A, Fernandez-Nieto M, et al. 2019. ExpoApp: an integrated system to assess multiple personal environmental exposures. Environ Int
126:494–503. https://doi.org/10.1016/j.envint.2019.02.054.
[DOI] [PubMed] [Google Scholar]
22.Martin BW, Ackermann-Liebrich U, Leuenberger P, Künzli N, Stutz EZ, Keller R, et al. 1997. SAPALDIA: methods and participation in the cross-sectional part of the Swiss study on air pollution and lung diseases in adults. Soz Praventivmed,
42:67–84, https://doi.org/10.1007/bf01318136.
[DOI] [PubMed] [Google Scholar]
23.Ackermann-Liebrich U. 1993. Schweizer Studie “Luftverschmutzung und Atemwegserkrankungen bei Erwachsenen: SAPALDIA.”
Atemw Lungenkr
5:190–194. [Google Scholar]
24.Cerletti P, Eze IC, Schaffner E, Foraster M, Viennau D, Cajochen C, et al. 2020. The independent association of source-specific transportation noise exposure, noise annoyance, and noise sensitivity with health-related quality of life. Environ Int
143:105960, https://doi.org/10.1016/j.envint.2020.105960.
[DOI] [PubMed] [Google Scholar]
26.Strak M, Weinmayr G, Rodopoulou S, Chen J, de Hoogh K, Andersen ZJ, et al. 2021. Long-term exposure to low-level air pollution and mortality in eight European cohorts within the ELAPSE project: pooled analysis. BMJ
374:n1904, https://doi.org/10.1136/bmj.n1904.
[DOI] [PMC free article] [PubMed] [Google Scholar]
27.Wolf K, Hoffmann B, Andersen ZJ, Atkinson RW, Bauwelinck M, Bellander T, et al. 2021. Long-term exposure to low-level ambient air pollution and incidence of stroke and coronary heart disease: a pooled analysis of six European cohorts within the ELAPSE project. Lancet Planet Health
5:e620–e632, https://doi.org/10.1016/S2542-5196(21)00195-9.
[DOI] [PubMed] [Google Scholar]
28.Bopp M, Spoerri A, Zwahlen M, et al. 2008. Cohort Profile: The Swiss National Cohort — a longitudinal study of 6.8 million people. Int J Epidemiol
38:379–384, https://doi.org/10.1093/ije/dyn042.
[DOI] [PubMed] [Google Scholar]
33.de Hoogh K, Saucy A, Shtein A, Schwartz J, West EA, Strassmann A, et al. 2019. Predicting fine-scale daily NO2 for 2005–2016, incorporating OMI satellite data across Switzerland. Environ Sci Technol
53:10279–10287, https://doi.org/10.1021/acs.est.9b03107.
[DOI] [PubMed] [Google Scholar]
35.Basagaña X, Rivera M, Aguilera I, Agis D, Bousa L, Elosua R, et al. 2012. Effect of the number of measurement sites on land use regression models in estimating local air pollution. Atmos Environ
54:634–642, https://doi.org/10.1016/j.atmosenv.2012.01.064. [Google Scholar]
36.de Hoogh K, Chen J, Gulliver J, Hoffmann B, Hertel O, Ketzel M, et al. 2018. Spatial PM2.5, NO2, O3, and BC models for Western Europe – evaluation of spatiotemporal stability. Environ Int
120:81–92, https://doi.org/10.1016/j.envint.2018.07.036.
[DOI] [PubMed] [Google Scholar]
38.Shen Y, de Hoogh K, Schmitz O, Clinton N, Tuxen-Bettman K, Brandt J, et al. 2022. Europe-wide air pollution modeling from 2000 to 2019 using geographically weighted regression. Environ Int
168:107485, https://doi.org/10.1016/j.envint.2022.107485.
[DOI] [PubMed] [Google Scholar]
42.Schmitz O, de Hoogh K, Probst-Hensch N, Jeong A, Flückiger B, Lu M, et al. 2025. A computational framework for agent-based assessment of multiple environmental exposures. J Expo Sci Environ Epidemiol
Aug
2, https://doi.org/10.1038/s41370-025-00799-7. [DOI] [PubMed]
45.Ntarladima AM, Vaartjes I, Grobbee DE, Dijst M, Schmitz O, Uiterwaal C, et al. 2019. Relations between air pollution and vascular development in 5-year-old children: a cross-sectional study in the Netherlands. Environ Health
18:50, https://doi.org/10.1186/s12940-019-0487-1.
[DOI] [PMC free article] [PubMed] [Google Scholar]
46.Setton E, Marshall JD, Brauer M, Lundquist KR, Hystad P, Keller P, et al. 2011. The impact of daily mobility on exposure to traffic-related air pollution and health effect estimates. Article. J Expo Sci Environ Epidemiol
21:42–48, https://doi.org/10.1038/jes.2010.14.
[DOI] [PubMed] [Google Scholar]
47.Shekarrizfard M, Faghih-Imani A, Hatzopoulou M. 2016. An examination of population exposure to traffic-related air pollution: comparing spatially and temporally resolved estimates against long-term average exposures at the home location. Environ Res
147:435–444, https://doi.org/10.1016/j.envres.2016.02.039.
[DOI] [PubMed] [Google Scholar]
48.Yoo E-h, Pu Q, Eum Y, Jiang X. 2021. The impact of individual mobility on long-term exposure to ambient PM2.5: assessing effect modification by travel patterns and spatial variability of PM2.5. Int J Environ Res Public Health
18:2194, https://doi.org/10.3390/ijerph18042194.
[DOI] [PMC free article] [PubMed] [Google Scholar]
49.Reis S, Liška T, Vieno M, Carnell EJ, Beck R, Clemens T, et al. 2018. The influence of residential and workday population mobility on exposure to air pollution in the UK. Environ Int
121:803–813, https://doi.org/10.1016/j.envint.2018.10.005.
[DOI] [PubMed] [Google Scholar]
50.Vienneau D, Stafoggia M, Rodopoulou S, Chen J, Atkinson RW, Bauwelinck M, et al. 2023. Association between exposure to multiple air pollutants, transportation noise, and cause-specific mortality in adults in Switzerland. Environ Health
22:29, https://doi.org/10.1186/s12940-023-00983-y.
[DOI] [PMC free article] [PubMed] [Google Scholar]
51.Downward GS, van Nunen EJHM, Kerckhoffs J, Vineis P, Brunekreef B, Boer JMA, et al. 2018. Long-term exposure to ultrafine particles and incidence of cardiovascular and cerebrovascular disease in a prospective study of a Dutch cohort. Environ Health Perspect
126:127007, https://doi.org/10.1289/ehp3047.
[DOI] [PMC free article] [PubMed] [Google Scholar]
52.Beelen R, Hoek G, Vienneau D, Eeftens M, Dimakopoulou K, Pedeli X, et al. 2013. Development of NO2 and NOx land use regression models for estimating air pollution exposure in 36 study areas in Europe: the ESCAPE project. Atmos Environ
72:10–23, https://doi.org/10.1016/j.atmosenv.2013.02.037. [Google Scholar]
53.Eeftens M, Beelen R, de Hoogh K, Bellander T, Cesaroni G, Cirach M, et al. 2012. Development of Land use regression models for PM2.5, PM2.5 absorbance, PM10, and PMcoarse in 20 European study areas; results of the ESCAPE project. Environ Sci Tech
46:11195–11205, https://doi.org/10.1021/es301948k. [DOI] [PubMed] [Google Scholar]
54.Fuks KB, Weinmayr G, Foraster M, Dratva J, Hampel R, Houthuijs D, et al. 2014. Arterial blood pressure and long-term exposure to traffic-related air pollution: an analysis in the European Study of Cohorts for Air Pollution Effects (ESCAPE). Environ Health Perspect
122:896–905, https://doi.org/10.1289/ehp.1307725.
[DOI] [PMC free article] [PubMed] [Google Scholar]
56.de Souza P, Anenberg S, Makarewicz C, Shirgaokar M, Duarte F, Ratti C, et al. 2023. Quantifying disparities in air pollution exposures across the United States using home and work addresses. Environ Sci Technol
58:280–290, https://doi.org/10.1021/acs.est.3c07926.
[DOI] [PubMed] [Google Scholar]
Commentary by HEI Improved Exposure Assessment Studies Review Panel
Research Report 229, Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects, by K. de Hoogh et al.
Dr. Kees de Hoogh’s 3-year study, “Accounting for Mobility in Air Pollution Exposure Estimates in Studies on Long-Term Health Effects,” began in June 2020. Total expenditures were $915,597. The draft Investigators’ Report from de Hoogh and colleagues was received for review in March 2024. A revised report, received in September 2024, was accepted for publication in October 2024. During the review process, the HEI Improved Exposure Assessment Studies Review Panel and the investigators had the opportunity to exchange comments and clarify issues in the Investigators’ Report and the Panel’s Commentary. Note: Review Committee member Ulrike Gehring was not involved in the review of this report due to a conflict of interest.
This report has not been reviewed by public or private party institutions, including those that support the Health Effects Institute, and may not reflect the views of these parties; thus, no endorsements by them should be inferred.
Outdoor air pollution is a major global public health concern. There is now a broad consensus among experts that exposure to ambient air pollution causes a range of adverse health effects, based on evidence from a large body of scientific literature that has grown exponentially since the mid-1990s.1–5
The assessment of long-term exposure to ambient air pollution for epidemiological studies, however, remains challenging. Early cohort studies characterized participants’ exposure by assigning the average concentration measured at one or a few central monitoring sites within a city to each participant therein.6,7 Fixed-site networks — even those in North America and Western Europe — continue to have relatively limited spatial coverage in many areas, particularly in suburban and rural locations, and insufficient density to capture small-scale (within-city) variation in air pollution.
Recent developments in measurement technologies and modeling approaches have increasingly been used to estimate long-term air pollution exposure at fine spatial scales for epidemiological studies of large populations. Advances include novel air pollution sensors, mobile monitoring, satellite data, and machine learning approaches.8 Even with these advances, important limitations and challenges remain in estimating long-term air pollution exposure, particularly for pollutants that vary widely across space and time.
In 2019, HEI issued Request for Applications (RFA*) 19-1, Applying Novel Approaches to Improve Long-Term Exposure Assessment of Outdoor Air Pollution for Health Studies (see Preface). The goal of the RFA was to develop and apply scalable, novel approaches to improve assessments of long-term exposures to outdoor air pollutants that vary highly across space and time — such as ultrafine particles (UFP) black carbon (BC), and nitrogen dioxide (NO2). Studies were intended to evaluate exposure measurement error quantitatively and to determine how exposure assessment approaches might ultimately affect the estimated health effects.
de Hoogh and colleagues proposed to investigate whether exposure estimates that account for mobility (time–activity throughout the day) would improve exposure assessment and potentially reduce bias in health studies. The HEI Research Committee recommended the study for funding because they appreciated its leveraging of a wealth of data, including three cohort studies, and because they thought the study would address an important outstanding issue. Most health studies to date have assessed long-term air pollution exposure estimated as outdoor concentrations at participants’ residential locations only.
This Commentary provides the HEI Improved Exposure Assessment Studies Review Panel’s evaluation of the study. It is intended to aid the sponsors of HEI and the public by highlighting the study’s strengths and limitations and by placing the results presented in the Investigators’ Report into a broader scientific and regulatory context.
SCIENTIFIC AND REGULATORY BACKGROUND
Traffic-related air pollution is an important risk factor for poor health across the globe, with the highest exposures occurring in urban areas and at residences near busy roadways.9 Traffic-related air pollution is a complex mixture of gases and particles resulting from the use of motor vehicles. Motor vehicles emit a variety of pollutants, including NO2 and fine particulate matter (PM2.5).9 Because traffic-related air pollution remains an important public health concern, governments may use several approaches to reduce exposure, including those focusing on traffic itself (e.g., by implementing low-emission zones) or on individual pollutants (e.g., by setting targets for ambient concentrations).10 Regulatory bodies in the United States and Europe have recently adopted more stringent PM2.5 annual standards — 9 μg/m3 and 10 μg/m3, respectively — that align more closely with the 2021 World Health Organization (WHO) Air Quality Guidelines.11,12 A more stringent annual standard was also set in Europe for NO2 (Commentary Table 1).
Commentary Table 1.
Annual NO2 and PM2.5 Standards in the US, EU, and WHO Guidelines
NAAQS = National Ambient Air Quality Standards; US EPA = United States Environmental Protection Agency; WHO = World Health Organization.
Exposure models applied in epidemiological studies underpin the air quality standards and guidelines. However, exposure assessment of traffic-related air pollutants is challenging because they are characterized by high spatial and temporal variability. Epidemiological studies have used various tools to address these challenges, including land use regression (LUR) models based on fixed-site routine monitoring, low-cost sensor networks, mobile monitoring, and dispersion models.
Most health studies to date have assessed participants’ long-term air pollution exposure based on estimates of outdoor concentrations at their residential locations only. The justification for this approach is that a substantial proportion of time is spent at home, and information about the time spent outside the home is typically not collected in epidemiological studies and not available in administrative databases.13 Similarly, the HEI traffic review relied almost exclusively on studies assessing outdoor concentrations at participants’ residential addresses only, although a few studies in children incorporated long-term exposures at their school addresses.9,14 Even fewer studies in adults have incorporated workplace address in the exposure assessment.15,16
Transport microenvironments (e.g., within different modes of transportation) tend to have higher air pollutant concentrations than other settings that most people encounter in their daily lives. Although commutes are typically brief, they might substantially contribute to total daily exposure and inhalation of air pollutants.17,18 It is unclear, however, how important it is to account for mobility in exposure assessments and how doing so would influence epidemiological studies.
Two main approaches to assessing time–activity patterns have been applied in air pollution exposure assessments and epidemiological studies.13 First, agent-based modeling (ABM) has been applied to simulate time–activity, typically based on existing survey data. ABM is a computer modeling approach that simulates actions and interactions among people, things, places, and time. It takes a bottom-up approach and allows for the integration of individual and population behaviors.19 In most applications, the individual workplace or school addresses are not known and are estimated. This approach can be applied to large populations.
Second, tracking studies have been performed, in which volunteers carry a GPS tracker or use a smartphone app so researchers can determine time–activity patterns empirically. Tracking studies typically follow small population groups (several hundred people at most) and cover a limited amount of time, often 1–2 weeks, because of the demand on participants’ time.13
The current study used ABM to estimate the influence of mobility on long-term exposures to NO2 and PM2.5 in Switzerland and the Netherlands, and used tracking study results to evaluate the ABM. The study then examined the influence of the mobility enhancement on exposure estimate quality using health data from three cohort studies.
SPECIFIC AIMS AND APPROACH
The overarching goal of de Hoogh’s study was to investigate whether accounting for mobility (time–activity throughout the day) would improve exposure assessment and potentially reduce bias in health studies. The investigators specified the following study aims:
Quantify long-term individual exposure to NO2 and PM2.5 by incorporating spatiotemporal mobility patterns and hourly air pollution maps.
Derive spatiotemporal mobility patterns for samples of people in Switzerland and the Netherlands that can be scaled to the general population.
Determine the influence of the mobility-enhanced air pollution exposure estimates on health effects in cohort studies.
Evaluate the measurement error of the mobility-enhanced air pollution exposure estimates.
To simulate mobility-enhanced exposure to NO2 and PM2.5, de Hoogh and colleagues applied an ABM that was informed by existing travel survey and census data in Switzerland and the Netherlands. They developed nationwide hourly maps of the two pollutants and linked those to the time–activity data to estimate exposure. To evaluate the ABM, they collected GPS location data over 2 weeks in tracking campaigns that included nearly 700 participants in the two countries. The investigators conducted epidemiological analyses in three adult cohort studies to compare health effects estimates of residential and mobility-enhanced exposure.
SUMMARY OF METHODS AND STUDY DESIGN
STUDY POPULATION AND HEALTH OUTCOMES
de Hoogh and colleagues selected three population-based cohorts that differ in size, population demographics, location, health outcomes, and study period (Commentary Table 2). The Swiss National Cohort (SNC) is a large administrative cohort (3.5 million individuals) that includes all Swiss adult citizens 30 years or older, followed from 2011 until 2018. The SNC was formed by linking census data, population registries, and death registries, but it contains less detailed covariate data than the other selected cohorts. In addition, the investigators selected two smaller epidemiological cohorts with detailed information available on lifestyle factors from the European Prospective Investigation into Cancer, Netherlands (EPIC-NL), and the Swiss Cohort Study on Air Pollution and Lung and Heart Diseases in Adults (SAPALDIA). EPIC-NL is a Dutch adult cohort of mostly women (75%) recruited between 1993 and 1997, with up to 20 years of follow-up data. SAPALDIA is a data-rich cohort established in 1991 with several waves of follow-up. For the current study, analyses were conducted using SAPALDIA3 (follow-up in 2010/2011) and SAPALDIA4 (follow-up in 2017). Both the SNC and EPIC-NL were also included in the HEI-funded Effects of Low-Level Air Pollution: A Study in Europe (ELAPSE) that investigated the health effects of low levels of air pollution.20 Earlier data from SAPALDIA were included in the European Study of Cohorts for Air Pollution Effects (ESCAPE).21
Commentary Table 2.
Key Study Characteristics of the Three Health Studies
EPIC-NL = European Prospective Investigation into Cancer and Nutrition: Netherlands cohort; SAPALDIA = Study on Air Pollution and Lung Disease in Adults; SNC = Swiss National Cohort.
EXPOSURE ASSESSMENT
Agent-Based Modeling
de Hoogh and colleagues applied ABM to approximate the spatiotemporal mobility patterns of the cohort participants, using input data from existing travel surveys and census data in Switzerland and the Netherlands. The ABM simulated space–time trajectories of individuals based on their activity profiles, which distinguished between commuters and noncommuters (i.e., those not in active employment outside the home, such as retired individuals). In the study, simulated routes for the commute between home and the workplace, together with the home and work locations, were overlaid on hourly air pollution maps to calculate mobility-enhanced exposures.
The investigators developed 13 activity profiles to represent different time–activity patterns using characteristics reported for cohort members (e.g., employment status). Specifically, various exposure estimates for noncommuters were calculated assuming that people spent 2 hours of the day between 8:00 and 23:00 outside the home, within a 1 km, 5 km, or 10 km buffer around their residential locations. Thus, exposures were calculated as the time-weighted average concentration at the home and at the possible activity locations within the buffers. The investigators distinguished between weekdays and weekend days in the activity profiles.
Likewise, various mobility-enhanced activity profiles for commuters assumed that people spent time at a workplace and on their commute. These commuter profiles included two commute trips (to work in the morning and from work in the afternoon), an 8-hour period at the work location, and the remaining time at home on weekdays. The investigators simulated four different commute modes (by foot, bike, car, and public transportation [i.e., train or tram]) for a maximum of 2 hours per day. The commute mode was probabilistically assigned based on distance to the work location.
Work locations were unknown in the three cohorts except for a subset of SAPALDIA participants. Hence, those unknown locations and the commute routes were randomly assigned, informed by country-specific travel survey data and each participant’s residential location. Annual census data on travel behavior for more than 200,000 residents from 2010 to 2019 were used for Switzerland. Annual surveys of daily mobility conducted from 2011 to 2019 among 40,000 randomly selected households were used in the Netherlands. To account for the uncertainty in the work location and commute route, the ABM was run between 50 and 1,000 times, with mean values used for subsequent analysis (Commentary Figure 1).
The investigators developed nationwide hourly maps of NO2 and PM2.5 and linked those to the time–activity modeling from ABM. For Switzerland, the investigators used annual average maps for 2016 from previously published spatiotemporal models22,23 and applied a temporal adjustment approach described in de Nazelle and colleagues.18 The annual exposure estimates were rescaled into hourly average weekday and weekend day maps by a fixed ratio, based on hourly background monitoring data from multiple background monitoring stations.
For the Netherlands, hourly LUR models for NO2 and PM2.5 were developed specifically for this study using regulatory monitoring data from 2016 to 2019 and stratified by season and weekday type (weekday or weekend day).24 The total number of monitoring stations for NO2 and PM2.5 was 544 and 227, respectively, and included stations from neighboring countries. The predictors available for inclusion were derived from the Europe-wide EXposome Powered tools for healthy living in urbAN Settings (EXPANSE) model25 and included population, road, and land use variables; satellite retrievals; and chemical transport model pollution estimates within different buffer sizes. To develop the LUR models, the investigators used both supervised linear regression and random forest to accommodate nonlinearity and complex data interactions better. They assessed model performance in terms of explained variance (R2) using fivefold cross-validation. The investigators used random forest models to combine outdoor concentrations with time–activity data.
For the health analyses in the three cohorts, the investigators estimated exposures for the cohort participants in two ways: annual average residential exposure at the baseline home address, and mobility-enhanced exposure estimated using ABM, which takes into account commuting, work locations, and nonwork activity outside the home. The mobility-enhanced exposure was assigned based on the most appropriate activity profile per participant, depending on their sex, employment status, and socioeconomic status. Additionally, mobility-enhanced exposure was estimated for a subset of SAPALDIA participants using their known workplace addresses.
Tracking Campaign
To evaluate the ABM, the investigators collected GPS location data over 2 weeks in tracking campaigns that included nearly 700 participants in the two countries. The tracking campaign began in 2022 following the lifting of COVID-19 restrictions. The goal was to recruit a representative sample that could be scaled to the general population. The recruitment strategy differed across the two countries. In Switzerland, the investigators built on an existing cohort (COVCO-Basel)26 and were able to recruit 489 participants (33% response rate). In contrast, in the Netherlands, the investigators had to set up a new recruitment campaign and, after several attempts, were able to include only 189 participants (<1 % response rate). In total, nearly 700 participants were recruited. To evaluate the representativeness of the tracking campaigns, they compared the demographic characteristics of the participants to national population data for the two countries.
The investigators gathered time–activity data through a mobile phone app and position data from both the app (every 3–4 minutes) and a high-precision wearable GPS tracking device (every 20 seconds). They compared the position data from both devices and combined them for the ABM evaluation to increase accuracy and precision. No exposure or health outcomes were measured in the tracking campaigns.
EVALUATION OF EXPOSURE MEASUREMENT ERROR
de Hoogh and colleagues used two approaches to evaluate measurement error in the ABM. First, for the tracking participants, the investigators compared exposure estimates based on the tracking data with those based solely on home location and with the mobility-enhanced estimates from ABM. Second, the investigators compared the annual average residential-only and mobility-enhanced exposures with existing personal measurement data from the European EXPOsOMICS project from 2013 to 2015. For this evaluation, they used data from the 35 participants from SAPALDIA and 41 participants from EPIC-NL who had complete data with three 24-hour measurements of PM2.5 over three seasons in 1 year. The investigators compared the different exposure approaches by preparing scatterplots and Bland-Altman plots and then calculating Pearson correlations between the different exposure estimates.
EPIDEMIOLOGICAL ANALYSES
The investigators applied Cox proportional hazards models to assess the association between each estimate of exposure to NO2 and PM2.5 and various health outcomes in the SNC and EPIC-NL. For SAPALDIA, they conducted repeated cross-sectional linear regression analyses to investigate the associations between the different exposure estimates and lung function and blood pressure. They censored the blood pressure data for the participants on hypertension medication to ensure the modeled value was at least as high as the measured value.
The main results for the SNC were adjusted for age, sex, and individual and area-level socioeconomic status. In addition, results were adjusted for various lifestyle factors, such as smoking and diet, for EPIC-NL and SAPALDIA. The SNC and SAPALDIA results were also adjusted for region and accounted for repeated measurements in the data (in SAPALDIA3 and 4).
The main analyses focused on single-pollutant models, and effect estimates were calculated for fixed exposure increments (10 μg/m3 for NO2 and 5 μg/m3 for PM2.5). Additional cross-sectional analyses were conducted in the subset of SAPALDIA with known workplace addresses to investigate the importance of known versus estimated workplace location.
SUMMARY OF RESULTS
COMPARISON OF DIFFERENT EXPOSURES IN THE COHORTS
The newly developed hourly LUR in the Netherlands performed slightly better for NO2 (cross-validation R2 = 0.6–0.7) than for PM2.5 (cross-validation R2 = 0.4–0.5). The cold season and weekday models performed better than the warm season and weekend models for both air pollutants. The Random forest models performed slightly (up to 4%) better than the supervised linear regression models for both pollutants. In both countries, diurnal patterns for NO2 showed more pronounced morning and evening rush hour peaks than did patterns for PM2.5, and patterns in the cold season and weekdays were more pronounced than patterns in the warm season and weekend days.
In all cohorts, slightly lower mean air pollution exposures and contrasts (interquartile range) in individuals’ exposures were estimated for the mobility-enhanced exposures compared to the residential exposures. For both exposures, the contrast was higher for NO2 than for PM2.5, consistent with the larger spatial variation of NO2. The investigators observed little contrast in PM2.5 estimates in the EPIC-NL study, hampering the interpretation of the analyses (Commentary Table 3). The mobility-enhanced exposure estimates were highly correlated (correlation > 0.95) with residential-only estimates for NO2 or PM2.5 within each cohort.
Commentary Table 3.
Mean and Interquartile Range of the Residential and Mobility-Enhanced Exposure Estimates (μg/m3) for the Three Health Studies
EPIC-NL = European Prospective Investigation into Cancer and Nutrition: Netherlands cohort; SAPALDIA = Study on Air Pollution and Lung Disease in Adults; SNC = Swiss National Cohort.
EVALUATION OF EXPOSURE MEASUREMENT ERROR
For the tracking campaign participants, the NO2 exposures based on the tracking data were highly correlated with those based solely on home location and those using ABM mobility estimates (R2 > 0.8). The correlation was lower for PM2.5 when comparing tracking exposure to residential and mobility-enhanced exposure, especially for the Netherlands (R2 > 0.50).
For the tracking campaign participants, a clear pattern emerged such that at low concentrations of NO2 and PM2.5, the mobility-enhanced exposure was higher than the residential exposure, and vice versa at high concentrations. One reason for this observation might be that participants living in less polluted locations work in more polluted areas, while people living in more polluted areas work in less polluted locations. The same pattern was observed in the cohorts.
The tracking campaign participants were not a representative sample of the general adult population. The investigators reported that the participants in both countries were more likely to be female and 40 years or older, have a high income and level of education, and work full-time. Moreover, the participants drove less and bicycled more than the general population.
No correlation was reported between the PM2.5 residential and mobility-enhanced exposure estimates and existing (EXPOsOMICS) personal measurements from this small sample in both countries.
COMPARISON OF HEALTH EFFECTS
The application of mobility-enhanced exposure estimates in the health studies yielded similar findings to those obtained from the application of residential-only estimates. The results were consistent with the high correlations of the different exposure estimates. Hence, the investigators concluded that using mobility-enhanced exposure estimates compared to residential-only estimates did not reduce bias in epidemiological studies.
Positive (adverse) associations were observed most clearly in the SNC (Commentary Figure 2). The researchers found that exposure to NO2 or PM2.5 was associated with an increased risk of natural mortality and that NO2 but not PM2.5 exposure was associated with an increased risk of cardiovascular mortality. They reported a lack of association and, in some instances, even negative associations for the various health outcomes in the EPIC-NL cohort and SAPALDIA. No consistent difference in the association was found with exposures using the known versus estimated workplace locations in the subset of SAPALDIA participants for which workplace locations were available.
Association between NO2 and PM2.5 and mortality in the Swiss National Cohort.
HEI IMPROVED EXPOSURE ASSESSMENT STUDIES REVIEW PANEL’S EVALUATION
In its independent review, the Panel considered the study to be well motivated and determined that it effectively leveraged a wealth of air pollution and health data across two countries. The Panel thought the study contributed new knowledge to exposure assessment for epidemiological research, with thorough analyses and findings of broad interest and value to a wide audience. The study found slight differences in residential and enhanced-mobility exposure estimates for NO2 and PM2.5, resulting in similar associations between each exposure estimate and participant health outcomes. The results suggest that the exposure measurement bias in epidemiological studies based on outdoor concentrations at residential locations might be small and that accounting for mobility might not be an important consideration. The Panel agreed with the overall findings but noted that further research is needed for other pollutants and in other locations and populations to examine fully the added value of collecting mobility information on a large scale in air pollution cohort studies.
STRENGTHS OF THE STUDY
The Panel noted several strengths of the research. First, the use of ABM was considered a useful approach to accounting for mobility in estimating exposure when time–activity data on individuals is unavailable; such data are often unavailable in large-scale epidemiological studies. The investigators defined 13 mobility-enhanced profiles and made reasonable assumptions related to those profiles and subsequent modeling. They used existing travel surveys and census data in Switzerland and the Netherlands, where possible. The Panel appreciated the extensive modeling that went into estimating the workplace locations and the commuting routes because workplace locations were only known for a subset of SAPALDIA participants. The Panel also appreciated the repeated sampling of the ABM to estimate some of the uncertainties. The use of ABM in air pollution and health studies is relatively novel as only a handful of studies have used this approach to date.13
Second, the Panel found the comparison of the two different tracking devices in the tracking campaigns to be useful. For tracking time–activity information, a mobile phone is preferred over a dedicated GPS device based on cost and availability, but the wearable tracking device used in this study was expected to provide greater precision and temporal resolution. The investigators combined data from the two devices for their analyses, but concluded that the mobile phone app provided adequate location information for estimating mobility-enhanced exposure.
Third, linking time–activity data to hourly air pollution models enabled the capture of diurnal variations in air pollution. The investigators developed new hourly LUR models in the Netherlands specifically for this study. The Panel particularly appreciated the inclusion of the seasonal, diurnal, and weekly variability in LUR models built with two different algorithms. The Panel also noted the large number of monitoring stations used to train the model, including some from neighboring countries.
Fourth, the Panel recognized the value of the application of the various exposure estimates in relation to health outcomes. The investigators applied the various exposure estimates to three population-based cohorts in Switzerland and the Netherlands that differ in size, population demographics, location, health outcomes, and study period. In particular, the health analysis for a large population (3.5 million participants) that included all Swiss adult citizens 30 years or older was considered informative. The inclusion of both mortality and morbidity outcomes and the availability of lifestyle factors (e.g., smoking and diet) in the two smaller cohorts were also noted as strengths.
Although the Panel broadly agreed with the investigators’ conclusions, some limitations should inform the interpretation of the results, as explained next.
REVISED SCOPE OF THE STUDY
Partly due to the COVID-19 pandemic, several setbacks led to substantial changes to the study compared to the original scope. The tracking campaigns had to be postponed until COVID-19 restrictions were lifted. Several recruitment strategies were tried due to a low response rate in the Netherlands. Eventually, a small convenience sample was obtained, but the approach fell short of the original plan to recruit a representative, population-based sample of 1,000 participants in each country. Hence, the investigators decided to use the tracking data only to evaluate the ABM, rather than to inform the model directly. Instead, they used national census and travel survey data from multiple years before the pandemic in the ABM. The Panel concurred with this decision because the routinely collected census and survey data are designed to be representative.
The evaluation of the ABM using tracking data was inherently limited because of the small convenience sample obtained after the pandemic, which was not as representative as intended of the general population, cohort participants, and prepandemic activity patterns. The ABM evaluation was also limited because it focused on overall exposures, with the same hourly air pollution maps used to generate exposures based on the tracking data and the time–activity modeling from the ABM. It would be of interest to examine how the underlying time–activity patterns compared between ABM model estimates and tracking data before they were combined with the air pollution maps. Cross-validation revealed that the air pollution maps explained 40% to 70% of the variation in pollutant concentrations observed across the two countries; this means that a substantial portion of exposure variation remained unexplained.
The study originally sought to evaluate mobility-enhanced exposure for UFPs, BC, and PM2.5 composition, in addition to NO2 and PM2.5. However, the investigators were not able to develop daily (let alone hourly) air pollution maps due to the sparse routine monitoring data for those pollutants. This issue could not be resolved without more detailed monitoring in place, which was beyond the scope of the project. Hence, the investigators decided to focus entirely on NO2 and PM2.5. The Panel noted that the reduced number of pollutants examined was a limitation of the study. It remains unknown whether the findings of this study hold for other traffic-related air pollutants, particularly UFPs and BC, although those pollutants are generally highly correlated with NO2.
INTRODUCING NEW SOURCES OF UNCERTAINTY
The use of ABM was considered novel and a strength by the Panel, but also introduced new sources of uncertainty, including the assigned workplace locations (rather than known workplace locations) and the limited number of activity profiles. The uncertainty related to the unknown workplace addresses was partly captured by the repeated sampling of the model. In addition, the investigators reported similar results in the subset of SAPALDIA where the workplace addresses were known; this analysis alleviates the concern to some extent. For future studies, collecting more detailed information on workplace locations in cohort studies could reduce uncertainty in activity-based exposure assessments.
Furthermore, relationships between indoor and outdoor concentrations of pollutants were not taken into account, nor was the variability in exposure induced by transportation mode. These factors add uncertainty and measurement error to the mobility-enhanced exposure estimates, sometimes in unpredictable ways. For example, an HEI study in Hong Kong found that when pollutant infiltration indoors was taken into account, mobility-enhanced exposures were on average about 20% lower than residential exposures because infiltration of outdoor air pollution was higher in residences than in commercial buildings.27 Depending on the transportation mode used (e.g., car, bus, bicycle, or walking), commuting can increase exposures anywhere from 40% to more than 400% compared to typical urban background concentrations.18 Moreover, the exposure can differ by transportation mode and by pollutant, as motorized modes (cars, motorcycles, buses, tram, rail, and subway) tend to have higher air pollutant concentrations of gaseous pollutants — such as NO2 — compared to active modes (cycling and walking). For particulate matter, the results varied by particle size and transportation mode.28
One might expect that including mobility would increase variability, but slightly smaller contrasts were estimated for the mobility-enhanced exposures compared to the residential exposures in the current study. The investigators explained, and the Panel concurred, that one reason for this finding is that participants living in less polluted locations might work in more polluted areas and vice versa. Moreover, ABM makes use of spatial averaging, and a smoothing effect of integrating multiple microenvironments in the exposure assessment might have led to smaller exposure contrasts.
COMPARISON TO PERSONAL EXPOSURES
Accounting for time–activity patterns, including commuting and time spent at nonresidential locations, rather than only residential addresses, is an important step toward assessing personal exposure to outdoor air pollution. Lack of consideration of time–activity data and infiltration rates adds exposure measurement error, which is often assumed to bias the estimated outdoor air pollution and health estimates toward finding no association, although the nature of the potential bias cannot be fully known.29,30
Methodological complexity and lack of data prevented the investigators from propagating exposure measurement error into the health effects estimations. How to propagate exposure measurement error into health effects estimation in long-term air pollution and health studies remains an area of future research.31
The investigators compared the mobility-enhanced exposure estimates to personal exposure measurements for PM2.5 obtained from the EXPOsOMICS project. The Panel considered the comparison limited for several reasons. The short duration of the measurements, the nonsimultaneous sampling, the small number of repeated measurements for the different seasons, and the different years of sampling hampered the utility of the personal measurements for validating the residential and mobility-enhanced exposure estimates. Moreover, the personal measurements included both indoor and outdoor contributions. Because of those limitations, the investigators agreed with the Panel to de-emphasize the comparison and move those results to the Additional Materials.
Reconciling modeled exposures (whether by ABM, LUR, or another approach) with measured long-term personal exposure remains a challenging topic of research.
COMPARISON OF HEALTH EFFECTS
The application of various exposure estimates in relation to different health outcomes across three population-based cohorts was a strength noted by the Panel. In particular, the health analysis for a large population (3.5 million participants) that included all Swiss adult citizens 30 years or older was considered informative and confirmed that exposure to NO2 and PM2.5 was associated with an increased risk of natural mortality. The lack of associations in the EPIC-NL cohort and SAPALDIA hampered the assessment of the influence of mobility and known workplace location on health estimates. The Panel thought the discussion regarding the lack of associations or, in some instances, even negative associations in those cohorts could have been expanded.
One limitation of the EPIC-NL analyses the Panel noted was that the investigators assigned residential exposure at baseline addresses (1993–1997) but used air pollution models that were developed more recently (2016–2019). That is, the analyses did not consider the possibility of change in residence during the 20 years of follow-up and assumed that air pollution concentrations remained constant throughout that time, which is a large assumption that they did not test. The temporal mismatch between the period captured by the exposure model and the exposure window most relevant for epidemiological purposes was much less in the SNC and SAPALDIA. The EPIC-NL study was also limited by the low exposure contrast for PM2.5.
The study documented similar associations between residential and mobility-enhanced exposure estimates across various health outcomes within the cohorts. The results suggest that exposure measurement bias in epidemiological studies based on outdoor concentrations at residential locations might be small and that accounting for mobility might not be an important consideration. A recent review that was part of the current HEI study reported similar findings in five of six identified health studies.13 Although this is reassuring, the Panel determined that replication in other settings and for other pollutants is needed, given the potentially limited generalizability of the current study population and time–activity patterns in relatively well-off Western European countries.
Ultimately, the Panel thought the study provided important insights because it shows that adding complexity to an exposure model does not necessarily improve the estimation of health effects, likely because new sources of uncertainty are introduced at the same time. This finding has also been demonstrated in other applications.32–34
SUMMARY AND CONCLUSIONS
de Hoogh and colleagues applied ABM to simulate a mobility-enhanced exposure to NO2 and PM2.5 in Switzerland and the Netherlands. They then conducted epidemiological analyses in three adult cohort studies in the two countries to compare health effects estimates derived from residential and mobility-enhanced exposure estimates. To evaluate the mobility-enhanced exposure estimates, they collected GPS location data over 2 weeks in two small tracking campaigns.
In its independent review of the study, the HEI Review Panel found that the study had several strengths. These included the novel use of ABM to account for mobility in exposure estimates for large populations spanning entire countries in Europe and the linking of time–activity data to hourly air pollution models to capture the diurnal variation of air pollution. The application of the various exposure estimates in the health analysis for a large population (3.5 million participants) that included all Swiss adult citizens 30 years or older was considered especially informative to assess the influence of mobility on long-term exposures.
Although the Panel broadly agreed with the investigators’ conclusions, some limitations should be considered when one interprets the results. The evaluation of the ABM using tracking data was limited because of the small convenience sample. Moreover, the time–activity estimates from the ABM were not separately evaluated before they were combined with the air pollution models. Furthermore, workplace locations were estimated rather than known, and the study did not incorporate infiltration rates of pollutants into homes or other buildings or account for exposure induced by different modes of transportation.
Overall, the study contributed new knowledge to exposure assessment for epidemiological research and generated findings that will be of broad interest and value to a wide audience. The study documented that the mobility-enhanced exposure estimates and the residential-only estimates were highly correlated for NO2 and PM2.5. The associations across various health outcomes within the cohorts were similar between residential and mobility-enhanced exposure estimates. The results suggest that exposure measurement bias in epidemiological studies based on outdoor concentrations at residential locations might be small and that accounting for mobility might not be an important consideration. Although this is reassuring, further research is needed for other pollutants and in other locations and populations to examine fully the added value of collecting mobility information on a large scale in air pollution cohort studies.
ACKNOWLEDGMENTS
The HEI Review Committee is grateful to the Improved Exposure Assessment Review Panel for their thorough review of the study. The Committee is also grateful to Dan Crouse for oversight of the study, to Elise Elliott for assistance with review of the Investigators’ Report, to Hanna Boogaard for preparing its Commentary, to Allison Bowman for editing the Investigators’ Report and its Commentary, and to Kristin Eckles for her role in preparing this Research Report for publication.
Footnotes
* A list of abbreviations and other terms appears at the end of this volume.
REFERENCES
1.IARC (International Agency for Research on Cancer); Working Group on the Evaluation of Carcinogenic Risks to Humans. 2016. Outdoor air pollution. IARC Monogr Eval Carcinog Risks Hum
109:9–444, PMID: 29905447; PMCID: PMC7682275.
[PMC free article] [PubMed] [Google Scholar]
2.US EPA (United States Environmental Protection Agency). 2016. Integrated Science Assessment for Oxides of Nitrogen – Health Criteria. EPA/600/R-15/068. Washington, DC: US EPA. [Google Scholar]
3.US EPA (United States Environmental Protection Agency). 2019. Integrated Science Assessment for Particulate Matter. EPA/600/R-19/188. Washington, DC: US EPA. [PubMed] [Google Scholar]
4.US EPA (United States Environmental Protection Agency). 2022. Supplement to the 2019 Integrated Science Assessment for Particulate Matter. EPA/600/R-22/028. Washington, DC: US EPA. [PubMed] [Google Scholar]
5.WHO (World Health Organization). 2021. WHO Global Air Quality Guidelines: Particulate Matter (PM2.5 and PM10), Ozone, Nitrogen Dioxide, Sulfur Dioxide, and Carbon Monoxide. Available: https://www.who.int/publications/i/item/9789240034228/. [PubMed]
9.HEI (Health Effects Institute). 2022. Systematic Review and Meta-Analysis of Selected Health Effects of Long-Term Exposure to Traffic-Related Air Pollution. Special Report 23. Health Effects Institute: Boston, MA. [Google Scholar]
10.Khreis H, Sanchez KA, Foster M, Burns J, Nieuwenhuijsen MJ, Jaikumar R, et al. 2023. Urban policy interventions to reduce traffic-related emissions and air pollution: a systematic evidence map. Environ Int
172:107805, https://doi.org/10.1016/j.envint.2023.107805.
[DOI] [PubMed] [Google Scholar]
11.EU (European Union). 2024. Directive (EU) 2024/2881 of the European Parliament and of the Council of 23 October 2024 on Ambient Air Quality and Cleaner Air for Europe. Available: https://eur-lex.europa.eu/eli/dir/2024/2881/oj/eng.
14.Patton AP, Boogaard H, Vienneau D, Brook JR, Smargiassi A, Kutlar Joss M, et al. 2025. Assessment of long-term exposure to traffic-related air pollution: an exposure framework. J Expo Sci Environ Epidemiol
35:493–501, https://doi.org/10.1038/s41370-024-00731-5.
[DOI] [PubMed] [Google Scholar]
16.Puett RC, Hart JE, Suh H, Mittleman M, Laden F. 2011. Particulate matter exposures, mortality, and cardiovascular disease in the health professionals follow-up study. Environ Health Perspect
119:1130–1135, https://doi.org/10.1289/ehp.1002921.
[DOI] [PMC free article] [PubMed] [Google Scholar]
19.Crooks AT, Heppenstall AJ. 2012. Introduction to agent-based modelling. In: Agent-Based Models of Geographical Systems. Springer, pp. 85–105. [Google Scholar]
20.Brunekreef B, Strak M, Chen J, Andersen ZJ, Atkinson R, Bauwelinck M, et al. 2021. Mortality and morbidity effects of long-term exposure to low-level PM2.5, BC, NO2, and O3: an analysis of European cohorts in the ELAPSE project. Research Report 208. Boston, MA: Health Effects Institute. [PMC free article] [PubMed] [Google Scholar]
23.de Hoogh K, Saucy A, Shtein A, Schwartz J, West EA, Strassmann A, et al. 2019. Predicting fine-scale daily NO2 for 2005–2016: incorporating OMI satellite data across Switzerland. Environ Sci Technol
53:10279–10287, https://doi.org/10.1021/acs.est.9b03107.
[DOI] [PubMed] [Google Scholar]
25.Shen Y, de Hoogh K, Schmitz O, Clinton N, Tuxen-Bettman K, Brandt J, et al. 2022. Europe-wide air pollution modeling from 2000 to 2019 using geographically weighted regression. Environ Int
168:107485, https://doi.org/10.1016/j.envint.2022.107485.
[DOI] [PubMed] [Google Scholar]
27.Barratt B, Lee M, Wong P, Tang R, Tsui TH, Cheng W, et al. 2018. A dynamic three-dimensional air pollution exposure model for Hong Kong. Research Report
194. Boston, MA: Health Effects Institute. [PMC free article] [PubMed] [Google Scholar]
28.Ramel-Delobel M, Heydari S, de Nazelle A, Praud D, Salizzoni P, Fervers B, et al. 2024. Air pollution exposure in active versus passive travel modes across five continents: a Bayesian random-effects meta-analysis. Environ Res
261:119666, https://doi.org/10.1016/j.envres.2024.119666.
[DOI] [PubMed] [Google Scholar]
30.Kioumourtzoglou MA, Spiegelman D, Szpiro AA, Sheppard L, Kaufman JD, Yanosky JD, et al. 2014. Exposure measurement error in PM2.5 health effects studies: a pooled analysis of eight personal exposure validation studies. Environ Health
13:2, https://doi.org/10.1186/1476-069x-13-2.
[DOI] [PMC free article] [PubMed] [Google Scholar]
Epidemiological analysis for EPIC-NL was conducted on a secure IT environment at Utrecht University; no individual data can be transferred. Likewise, an analysis of SNC was conducted on a secure IT environment at the Swiss Tropical and Public Health Institute, and no individual data can be transferred. COVCO Basel and SAPALDIA data are available upon reasonable request to Principal Investigator N. Probst-Hensch, at nicole.probst@swisstph.ch. The MOBI-AIR app has been discontinued and is no longer available on the App Store.
