Exposure assessment for air pollution epidemiology: a scoping review of emerging monitoring platforms and designs

Abstract

Background:

Both exposure monitoring and exposure prediction have played key roles in assessing individual-level long-term exposure to air pollutants and their associations with human health. While there have been notable advances in exposure prediction methods, improvements in monitoring designs are also necessary, particularly given new monitoring paradigms leveraging low-cost sensors and mobile platforms.

Objectives:

We aim to provide a conceptual summary of novel monitoring designs for air pollution cohort studies that leverage new paradigms and technologies, to investigate their characteristics in real-world examples, and to offer practical guidance to future studies.

Methods:

We propose a conceptual summary that focuses on two overarching types of monitoring designs, mobile and non-mobile, as well as their subtypes. We define mobile designs as monitoring from a moving platform, and non-mobile designs as stationary monitoring from permanent or temporary locations. We only consider non-mobile studies with cost-effective sampling devices. Then we discuss similarities and differences across previous studies with respect to spatial and temporal representation, data comparability between design classes, and the data leveraged for model development. Finally, we provide specific suggestions for future monitoring designs.

Results:

Most mobile and non-mobile monitoring studies selected monitoring sites based on land use instead of residential locations, and deployed monitors over limited time periods. Some studies applied multiple design and/or sub-design classes to the same area, time period, or instrumentation, to allow comparison. Even fewer studies leveraged monitoring data from different designs to improve exposure assessment by capitalizing on different strengths. In order to maximize the benefit of new monitoring technologies, future studies should adopt monitoring designs that prioritize residence-based site selection with comprehensive temporal coverage and leverage data from different designs for model development in the presence of good data compatibility.

Discussion:

Our conceptual overview provides practical guidance on novel exposure assessment monitoring for epidemiological applications.

1. Introduction

There have been steady improvements in approaches to obtaining individual-level exposure estimates to ambient air pollution for application to cohort studies (Hoek 2017; Hoek et al. 2013). These advances have strived to overcome a basic limitation in exposure assessment, namely that individual-level air pollution measurements are nearly always unavailable at the spatial locations and time scales of interest. Investigators have paid particular attention to approaches to adequately characterize spatial variability of air pollution exposures across individuals, as the spatial contrast in exposure is the primary contrast of interest in cohort studies where the scientific question centers on whether those exposed to higher levels of air pollution are likely to be at higher risks of death, disease occurrence, or progression compared to those exposed to lower levels (Eeftens et al. 2012b; Kaufman et al. 2016b). Advances in both air pollution monitoring and exposure prediction modeling have played major roles in these efforts. For instance, various prediction models have been developed, typically based on statistical, chemical transport, and/or physics-based models (Di et al. 2016; Eeftens et al. 2012a; Keller et al. 2015; van Donkelaar et al. 2016). These models have allowed investigators to quantify air pollution concentrations at residential or workplace locations, and characterize spatial variability of air pollution across individuals. In the realm of exposure monitoring, recent attention has focused on new and emerging monitoring technology along with novel monitoring designs to capture spatially varying pollutants.

New air pollution monitoring paradigms that rely on low cost sensors (LCSs) and/or a mobile monitoring framework have dramatically improved our ability to obtain air pollution measurements at finer spatial scales than have been available previously. The lower cost and ease of implementation of LCS networks have allowed investigators to expand the spatial coverage of monitoring and represent increasingly larger numbers of study participants (Clements et al. 2017; Cromar et al. 2019; Morawska et al. 2018; Zhao et al. 2021). Similarly, mobile monitoring platforms have become a much more popular monitoring option given extensive and readily accessible road networks as well as the high cost of instruments needed to measure pollutants such as ultrafine particles (Gozzi et al. 2016). Many previous mobile monitoring studies were motivated by project-specific research questions that focused on emissions from specific sources such as traffic-related air pollution (TRAP) and residential wood burning (Austin et al. 2021; Boanini et al. 2021; Larson et al. 2007; Loeppky et al. 2013; Pirjola et al. 2012; Su et al. 2007; Wagstaff et al. 2022). In contrast, the aims of recent mobile monitoring studies have often been expanded to exposure mapping and residential exposure assessment (Kerckhoffs et al. 2016; Messier et al. 2018). Mobile monitoring is even more appealing for studies focused on vehicular exposures, as TRAP has been implicated as an important exposure responsible for adverse health effects (Health Effects Institute 2010). And perhaps most important, repeat short-term measurements over space, such as obtained from fixed routes of mobile monitoring campaigns, have been shown to meaningfully capture the spatial contrast in residential exposure to air pollutants (Klompmaker et al. 2015; Larson et al. 2009; Levy et al. 2014; Montagne et al. 2015; Riley et al. 2016; van Nunen et al. 2017; Weichenthal et al. 2016). Furthermore, the measurements obtained from these LCS and mobile monitoring studies have been used for developing exposure prediction models; these have shown good performance for obtaining individual-level air pollution exposure estimates (Apte et al. 2017; Kerckhoffs et al. 2016; Saha et al. 2019a; Weichenthal et al. 2016; Xu et al. 2017).

Since these new monitoring paradigms provide numerous temporal and relatively dense spatial measurements, which are typically also temporally and spatially imbalanced, the design of the monitoring campaign can play an important role in achieving the inferential goal of the cohort study. For example, systematic exclusion of sampling during selected hours, days, or months or at some locations may lead to poor representation of individual long-term average exposure, which in turn may affect the accuracy and precision of the inference in subsequent health effect analyses (Blanco et al. 2022a; Blanco et al. 2022b; Szpiro et al. 2009, 2011). However, although increasing numbers of exposure assessment studies have employed new monitoring technologies, we have not observed the same attention to providing guidance regarding the best monitoring designs. Thus, an important step is to conduct a review of these emerging methodologies for application to epidemiologic cohort studies in order to provide a foundation for proposing optimal monitoring designs.

This paper aims to provide a conceptual overview of novel monitoring designs for ambient air pollution cohort studies that either leverage mobile platforms or incorporate LCS monitoring instruments. We complement this conceptual overview with a characterization of the design features employed in recent exposure assessment studies that have adopted these new monitoring paradigms and were conducted as part of or intended for future application to epidemiologic cohorts. Our overview can be defined as scoping review, as we seek to identify key attributes and concepts which have not been completely characterized, as opposed to a systematic review that includes exhaustive literature search and quality control to reach conclusive answers to specific questions (Grant and Booth 2009; Munn et al. 2018). Further, since this work focuses on studies that characterize high-resolution spatial contrasts in long-term exposure to ambient air pollution, we do not consider studies that focused on between-city variation, micro-scale variation within one season, characterization of temporal variability, instrument properties such as LCS degradation, or calibration using a few LCSs in a single city (Gao et al. 2015; Hofman et al. 2016; Weissert et al. 2019; Zusman et al. 2020). We also did not include passive samplers in our review, due to the paucity of epidemiological studies that leverage PM passive samplers.

2. Monitoring design concepts

Table 1 provides a conceptual overview of monitoring designs that leverage either mobile platforms or LCSs. We first divide monitoring designs into two classes: mobile and non-mobile. These classes are primarily distinguished by whether the monitoring equipment travels during some or all of its sampling periods. In the mobile class, most studies have loaded high-quality samplers onto a motor vehicle, driven along fixed routes multiple times within a defined urban area, and measured air pollution concentrations every second or at the finest instrument-level time scale. These campaigns have lasted a few days or weeks, and many include replicate sampling in multiple seasons. As measurements obtained from this mobile class often come from equipment loaded onto motor vehicles driving on roads, many studies have measured traffic-related air pollutants such as oxides of nitrogen (NO_X), ultrafine particles (UFP) often characterized as particle number concentration (PNC), and black carbon (BC). In contrast, the non-mobile class of designs has been used to deploy LCSs at many fixed locations, simultaneously or in rotation, typically over multiple seasons and years. These have been deployed in single or multiple cities, and have measured concentrations of various pollutants primarily including criteria pollutants such as particulate matter (PM), nitrogen dioxide (NO₂), carbon monoxide (CO), sulfur dioxide (SO₂), and ozone.

Table 1.

Typical characteristics of the mobile and non-mobile design classes of air pollution monitoring

Characteristics	Mobile design class	Non-mobile design class
Mounting platform	Motor vehicle, bicycle, or pedestrian	Trailer, telephone pole, rooftop, or tripod
Monitoring equipment	High-quality samplers	Costly and/or cost-effective devices including low cost sensors
Primary pollutants	UFP, BC, and gaseous pollutants	PM2.5 and gaseous pollutants
Spatial domain	One city at a time	Multiple cities at a time
Temporal domain	A few weeks to a few years	Multiple years
Spatio-temporal resolution of measurements	One or a few seconds on thousands of road segments or at hundreds of stops typically including repeat visits	More than ten fixed or rotating sites at the same time
Concurrent sampling	No	Yes
Prediction model domain	Spatial	Spatial or spatio-temporal

2.1. Major differences between mobile and non-mobile design classes

In addition to the mounting platform and monitoring equipment, the major characteristics that distinguish the mobile and non-mobile design classes include primary pollutants, spatial and temporal domain, concurrent sampling, and prediction model type (Table 1). There is an inter-relationship between the distinguishing characteristics for these two design classes, driven by logistical and financial feasibility.

Mobile monitoring campaigns are particularly useful for measuring pollutants not commonly measured by regulatory agencies such as UFP and BC (Klompmaker et al. 2015; Montagne et al. 2015). In particular, UFP has been typically monitored by using a mobile platform because of its high-cost instrumentation; low-cost UFP instruments that achieve high accuracy are not available (Liu et al. 2020; Morawska et al. 2018; Presto et al. 2021). Due to the high costs of both UFP sampling equipment and campaign operations, along with the bulk and weight of the instruments, most mobile monitoring studies have chosen to drive a single platform at a time in a highly urbanized study area (Kerckhoffs et al. 2021; Messier et al. 2018; Weichenthal et al. 2015). Mobile platforms have been expanded to include NO₂ and possibly other gases (Blanco et al. 2022c; Hatzopoulou et al. 2017; Li et al. 2019). Alternative mounting platforms have included bikes and public transportation (Hankey and Marshall 2015a, 2015b; Mueller et al. 2016a). Within the mobile design class, the spatial and temporal coverage of sampling can be maximized, although samples at individual locations are brief and multiple repeat visits are necessary to obtain stable location-specific estimates of long-term averages (Apte et al. 2017; Klompmaker et al. 2015; Messier et al. 2018). Typical mobile campaigns sample over defined area(s) and/or by following pre-specified routes. The brief repeat visit data from mobile monitoring campaigns are often averaged over time, with or without temporal adjustment, intended to account for varying background pollution levels depending on different hours, days, or seasons. These spatial averages are used for developing spatial prediction models (Kerckhoffs et al. 2019; Weichenthal et al. 2016). In addition, there have been a few studies that used the non-mobile design class where a few dozen UFP instruments were deployed simultaneously at multiple fixed sites for a relatively long time period. For example, in order to overcome the limitation of short-term sampling of PNCs for the purpose of characterizing long-term exposures, Saha et al. (2019) collected 3–6 week average samples at three groups of 12 rotating sites in Pittsburgh, U.S. (Saha et al. 2019b).

The non-mobile monitoring design class is the most common for studies that use LCS. Because these devices are cost-effective and easy to set up, it is straightforward to deploy multiple sensors simultaneously at many locations. In many exposure assessment studies for epidemiology, sensors measure multiple air pollutants concurrently, most often focusing on PM ≤ 2.5 μm in aerodynamic diameter (PM_2.5) and NO₂ (Morawska et al. 2018; Snyder et al. 2013). Typically, these studies calibrate their measurements to better align the LCS data to federal reference method standards (Bi et al. 2020, 2021, 2022; Castell et al. 2017; Zuidema et al. 2021; Zusman et al. 2020). There are some exceptions where LCSs have been deployed using the mobile design class, including loading them on cars (Suriano et al. 2015; Zhao et al. 2021), bikes (Wesseling et al. 2021), or trash trucks (deSouza et al. 2020) or carried by people (Lim et al. 2019). And because of their low cost and ease of use, there is an expanding role for the use of LCS by community members (Commodore et al. 2017; Wesseling et al. 2019). This includes the ever-expanding and easily accessible PurpleAir network (https://www2.purpleair.com/). Widespread deployment of PurpleAir accompanied by its powerful data sharing platform provides an excellent opportunity to calibrate measurements, develop models, and leverage these sensors in applications that have not been possible with prior LCS applications (Bi et al. 2020, 2022; Lu et al. 2021; Wallace et al. 2021). For example, as opposed to the more traditional calibration approach involving co-locating LCSs with few regulatory monitors over different seasons, Bi et al. (2020) developed an alternative calibration approach for the PurpleAir measurements based on more than 54 PurpleAir locations that were placed within 500 meters from 26 regulatory monitors in California, U.S.(Bi et al. 2020). This study also developed a down-weighting approach that lowered the weight of LCS measurements in order to reduce the impact of PurpleAir sensor uncertainty in prediction models.

2.2. Sub-designs of mobile and non-mobile designs

Based on the spatial and temporal features of monitoring designs described in previous studies, the mobile and non-mobile design classes can be further categorized into sub-classes (Figures 1 and 2). The mobile class includes “stationary” and “non-stationary” components to distinguish whether the vehicle parks during the mobile campaign. Whereas the vehicles in most mobile design studies collect data while in motion (the “non-stationary” condition) (Kerckhoffs et al. 2021; Larson et al. 2017), some studies also collect data while parked at fixed locations (the “stationary” condition) (Blanco et al. 2022c; Montagne et al. 2015; van Nunen et al. 2017). The stationary mobile design can be further subdivided: “short-term” stationary design studies stop their vehicles for a few minutes to hours, whereas the “long-term” stationary design studies park their vehicles for longer durations (Kim et al. 2016).

Figure 1.

Classes and sub-classes of monitoring campaign designs intended to ensure adequate spatial representation of air pollution exposure assessment for cohort studies

Figure 2.

Temporal and spatial scales of mobile and non-mobile design classes of air pollution monitoring campaigns

The non-mobile class includes “long-term” and “short-term” sub-classes, distinguished by the duration of sampling at fixed sites. Typical sampling durations range over one to several years in studies employing the long-term sub-class (Bi et al. 2020; Tanzer et al. 2019), while sampling durations are a few days to weeks in studies using the short-term sub-class. The short-term sub-class includes at least two additional sub-sub-classes: “rotating” where a small number of monitors are rotated among sites, typically each site is visited multiple times across seasons (Saha et al. 2019b; Shaffer et al. 2021), and a “snapshot” class where all sites are sampled at once, with the entire snapshot typically repeated across several seasons (Cohen et al. 2009).

We note that these various sub-classes have been called by different names in previous studies (Table S1). For example, the terms “central-site”, “fixed continuous”, “fixed-site”, “long-term fixed-site”, or “reference” in some studies are identical to our “non-mobile long-term” design class (Abernethy et al. 2013; Kerckhoffs et al. 2016; Saha et al. 2021; Saraswat et al. 2013; Simon et al. 2017). Our “non-mobile short-term snapshot” class has been called “home outdoor”, “short-term fixed”, or “snapshot” depending on the emphasis that other studies have placed on features such as sampling location, contrast with a mobile campaign, and concurrent sampling, respectively (Keller et al. 2015; Kerckhoffs et al. 2019; Saha et al. 2019b).

3. Characterization of the new monitoring paradigms and design classes in previous studies

Table 2 summarizes selected published papers that used mobile or LCS non-mobile designs. We have only included exposure monitoring studies that stated one of their ultimate goals was application to epidemiologic cohorts in their papers. Although increasing numbers of exposure monitoring studies have adopted new monitoring technologies, most of these were focused on characterizing exposure distributions or validating monitoring equipment and did not focus on assessing long-term exposure for epidemiological applications. The aims of the mobile or non-mobile monitoring studies summarized in Table 2 may have also included validation of sampling devices, exploration of spatial variation of pollutant concentrations, and/or development of exposure prediction models. We discuss the similarities and differences across our selected studies with respect to their approaches to spatial and temporal representation, data compatibility between design classes, and the data they leveraged from different design classes for model development.

Table 2.

Summary of selected recent studies using mobile and non-mobile monitoring design classes to characterize long-term exposure to air pollution for epidemiologic application.

Mobile/non-mobile	Study areaa		Study	Aim	Pollutant	Monitoring design subclass	Site selection	Temporal coverage	Related health study
Mobile	North America	Oakland, US	Apte 2017	Data exploration	BC, NO, NO2	Non-stationary	Land use of study area	9 am-5 pm/weekdays/4 seasons
		Los Angeles and Baltimore, U.S.	Tessum 2021 b	Model developmentc	NOx, NO2	Non-stationary	Land use	2–7 pm/weekdays/2 seasons
		Boston and Chelsea, US	Simon 2017	Data exploration	PNC	Non-stationary	Residence of study participants	5 am-9 pm/weekday & weekend/4 seasons
		Pittsburgh, US	Li 2019	Data exploration	PNC, PM2.5, PM1, CO, NO2	Non-stationary	Land use	5 am-10 pm/weekdays (& dry condition)/2 seasons
		Puget Sound, US	Blanco 2021	Data exploration, model development	PNC, PM2.5, BC, CO, CO2, NO2	Non-stationary, Short-term stationary	Residence	4 am-11 pm/weekday & weekend/4 seasons
		3 cities, US	Saha 2021 d	Model development	PNC	Non-stationary	Land use	7–20 days/1 seasone
		3 cities, Canada	Weichenthal 2015	Data exploration, Model development	PNC	Non-stationary	Land use	7–10 am & 3–6pm/weekday/1 or 2 seasons	Weichenthal 2020
	Europe	Amsterdam and Rotterdam, Netherlands	Kerckhoffs 2016	Model development	PNC, BC	Non-stationary, Short-term stationary	Land use	9 am-4 pm/2 seasonsf
		3 cities, Netherlands	Kerckhoffs 2017 d	Model development	PNC, BC	Non-stationary, Short-term stationary	Land use	9 am-4 pm/2 seasonsf
		Netherlands	Kerckhoffs 2021 d	Model development	PNC	Non-stationary, Short-term stationary	Land use	9 am-4 pm/14 monthsf
		6 European areas	Van Nunen 2017	Model development	PNC	Short-term stationaryg	Land use	9 am-4 pm/weekday/3 seasons	Downward 2019
		Basel, Switzerland	Ragetth 2014	Model development	PNC	Short-term stationaryg	Land use	9 am-4 pm/3 seasonsf
Non-mobile	North America	Pittsburgh, US	Saha 2019	Data exploration	PNC, PM2.5, CO, NO2	Short-term rotating	Land use	3–6 weeks in winter
		Pittsburgh, US	Tanzer 2019	Data exploration	PM2.5, NO2, SO2	Long-term	Land use	14 months
		Pittsburgh, US	Li 2019	Data exploration	PNC	Long-term	Land use	2–8 months
		California, US	Bi 2020	Calibration, model development	PM2.5	Long-term	Residence	12 months
		Boston and Chelsea, US	Simon 2017	Data exploration	PNC	Short-term rotating	Residence	6 weeks between spring and fall
		Puget Sound, US	Bi 2022	Model development	PM2.5	Long-term	Residence	20 months
		Puget Sound, US	Shaffer 2021 h	Model development	PM2.5	Long-term Short-term rotating	Residence	> 2 years 2 weeks over 3 years	Shaffer 2021
	Europe	6 European areas	Van Nunen 2017	Model development	PNC	Short-term’	Land use	24 hours in each of the 3 seasons	Downward 2019
	Asia	New Delhi, India	Saraswat 2013	Model development	PNC, PM2.5, BC	Short-term rotating	Land use	1 −3 hours during the morning and/or afternoon over 4 months in spring and summer
	Africa	4 districts, Uganda	Coker 2021	Model development	PM2.5	Long-term	Land use	12 months
		Adama, Ethiopia	Abera 2020	Model development	PM2.5	Short-term rotating	Land use	1 month in each of dry and wet seasons

^a.The latest studies were presented when there were multiple studies performed in the same area.

^b.This paper also carried out non-mobile monitoring using passive samplers, but we did not include to our review because we did not consider passive samplers as LCS.

^c.Exposure prediction model for long-term average air pollution

^d.These papers also used non-mobile monitoring data published elsewhere for model evaluation or development.

^e.No information on time and weekday/weekend; one-season sampling in either summer or fall in each of the three cities f. No information on weekday/weekend

^f.No information on weekday/weekend.

^g.Although there is no specific information on whether these studies used mobile platforms, we classified as mobile monitoring based on their description of “30 min measurements taken between 9 am and 4 pm” and “20-min measurements on the sidewalk at 60 locations during non-rush hours” for van Nunen et al. 2017 and Ragettli et al. 2014, respectively.

^h.Paper focused on health effects that provides the documentation of the exposure assessment using LCS.

^i.Not clear about whether this non-mobile monitoring is rotating or snapshot campaign based on the description in van Nunen et al. 2017

3.1. Spatial and temporal representation

Most studies, regardless of design class, selected their monitoring sites based on the type of land use with the goal of representing diverse environments over a defined study area, including urban background, traffic, commercial, and industrial site locations (Table 2). Land use-based site selection may be an optimal choice for general population cohort studies such as studies that leverage a census-based cohort. For example, Canadian studies applied non-stationary mobile designs in Toronto and Montreal, selected their sites for PNC sampling based on variation in land use, and developed exposure prediction models for application to the Canadian Census Health and Environmental Cohort (CanCHEC) participants (Weichenthal et al. 2015, 2016). However, it is not clear whether exposure assessment for traditional cohort studies can be adequately characterized by locations selected based on variation in land use, because many of the sampled locations may not represent cohort residence locations. Spatial incompatibility between monitor and cohort locations can affect exposure measurement error, leading to inaccurate health effect estimates as shown in a simulation study (Szpiro and Paciorek 2013). Only a few studies have designed their exposure monitoring to specifically target the participants in subsequent epidemiological analyses. For instance, the Boston Puerto Rican Health Study (BPRHS) conducted mobile and non-mobile monitoring campaigns to collect PNC focusing on residences of BPRHS participants in Boston and Chelsea, U.S. (Simon et al. 2017). They located 20 non-mobile sites and 20–40 km mobile routes at or near BPRHS cohort participant homes. Likewise, in the Puget Sound area in U.S., the Adult Changes in Thought - Air Pollution (ACT-AP) study selected most of their more than 100 LCS sampling sites from participant homes for non-mobile monitoring (Shaffer et al. 2021). For their mobile campaign, they selected about 300 short-term stationary sampling locations at an average 611-meter distance from participant homes (Blanco et al. 2022c). They also explored land use characteristics to confirm that selected monitoring locations are well represented for diverse land use.

With the aim of characterizing long-term average concentrations, many mobile monitoring campaigns collected data on different days of the week in multiple seasons for a year or longer (Kerckhoffs et al. 2016, 2017; Simon et al. 2017). However, the temporal coverage often remains limited for accurate and precise representation of long-term exposure. Most of the studies in Table 2 operated their monitoring only during business hours, non-rush hours, weekdays, and/or dry conditions (Kerckhoffs et al. 2017; Li et al. 2019; Montagne et al. 2015; van Nunen et al. 2017; Weichenthal et al. 2015, 2016). The duration of sampling ranged between one and several weeks, and was distributed across two seasons over a few different years instead of four seasons in the same year. To account for this temporally imbalanced feature of measurements, some studies employed a central reference site within the study area over the entire time period of interest to allow correction of the measurements (Kerckhoffs et al. 2016, 2017). However, these central reference sites may not well characterize spatially-varying temporal patterns, particularly for UFPs, a feature that has been shown in a previous study (Klompmaker et al. 2015). Recently, there have been some efforts to improve the temporal representation in measurements. For instance, starting in 2019, the ACT-AP study sampled mobile monitoring data from 5 a.m. to 11 p.m. on 288 days over all four seasons, including weekdays and weekends (Blanco et al. 2022c).

3.2. Data compatibility between design classes

Among the papers based on mobile and non-mobile monitoring summarized in Table 2, a small subset used more than one design class or sub-class to monitor or model the same pollutants, in the same area, or for the same period. Tables 3 and S2 show the design characteristics of those papers that applied both the mobile and non-mobile classes or different sub-classes within one class, while Tables S3 and S4 summarize their major findings. Although comparable measurements of various monitoring designs can provide a unique opportunity to improve exposure characterization when applied to a single cohort study, studies using the mobile and non-mobile designs mostly measured different sets of pollutants over different times and spatial locations.

Table 3.

Design characteristics of exposure assessment studies that used measurements from multiple design classes of new technology monitoring for data exploration and model development to assess long-term exposure to air pollution for epidemiological application

N	Study	Area	Period	Aim	Pollutant		Monitoring designa
							Mobile		Non-mobile
							Non- stationary	Stationary short-term	Long-termb	Short-term rotating	Short-term snapshot
1	Simon 2017	Boston and Chelsea, US	2011–2015	Data exploration	PNC	Spatial	40 and 6 km2 in each of 2 cities		1 central site per city	20 homes
						Temporal	3–6 hr at 5:00–21:00 on 42–46 days, wkday/wknd, 2011–2015			2–11 sites for 6 wks in 2012–2014
2	Li 2019 c	Pittsburgh, US	2016–2017	Data exploration	PNC, PM2.5, PM1, CO, NO2	Spatial	All public roads within 8 1 -km2 areas		10 sites
						Temporal	45–60 min at 5:00–22:00, wkday in summer and winter, 2016–2017		2–8 mths in 2016–2017
3	Blanco 2021d	Puget Sound, US	2019–2020	Data exploration, model development	PNC, PM2.5,BC, CO, CO2, NO2	Spatial	9 fixed routes (1,069 total km)	309 2-min locations
						Temporal	4–8 hr at 5:00–23:00 on 288 days, wkday/wknd, 2019–2020
4	Saha 2021 c	Continental US	2016–2017	Model development	PNC	Spatial	Every street in multiple neighborhoods in 3 cities, summarized on a 1 km2 grid		15 rural, 19 urban, and 4 near-airport sites e
						Temporal	7–20 days in summer/fall, 2017–2019		1–12 mths for 2009–2019
5	Kerckhoffs 2016 c	Amsterdam and Rotterdam, Netherlands	2013	Model development	PNC, BC	Spatial	2,964 road segments	161 30-min locations	1 site in Utrecht
						Temporal	9:00–16:00 on 42 days in two seasons in 2013f
6	Kerckhoffs 2017 c,g	Amsterdam, Maastricht, and Utrecht, Netherlands	2014–2015	Model development	PNC, BC	Spatial	5,236 road segments in 3 cities	240 30-min parked locations	1 site in Utrecht	42 homes
						Temporal	9:00–16:00, 84 days in two seasons in 2014–2015f			24-hr3 seasonsin2014–2015
7	Kerckhoffs 2021 c,g	Netherlands	2016–2017	Model development	PNC	Spatial	14,392 road segments in 5 major cities and multiple towns	400 repeated 30-min parked locations		42 homes (H)e,h,i 20 background sites (B)e,h,i
						Temporal	9:15–16:00, for 14 mths in 2016–2017f			24-hr 3 seasons in 2014–2015 (H) 2-wk 3 seasons in 2016–2017 (B)

^a.Additional information is presented in Table S2

^b.

All or some of the long-term sites described in the papers were not considered as non-mobile long-term sites in our review.

• – One central site per city in Simon et al. 2017, Kerckhoffs et al. 2016, and Kerckhoffs et al. 2017 (gray cells in the table) was not considered as non-mobile long-term monitoring because the aim of this monitoring was limited to temporal adjustment.
• – Kerckhoffs et al 2021 defined 20 regional background sites as long-term sites, but we considered these as non-mobile short-term rotating or snapshot monitoring because they only collected three 2-wk samples during the study period.
• – 8 out of 10 sites in Li et al. 2019 had 2–3 month data which may not be considered as non-mobile long-term monitoring.
• – 16 out of 38 sites in Saha et al. 2021 had 1–3 month data which may not be considered as non-mobile long-term monitoring.

^c.Li et al. 2016 and Saha et al. 2021, as well as Kerckhoffs et al. 2016, Kerckhoffs et al. 2017, and Kerckhoffs et al. 2021 used overlapping data

^d.Although both stationary and non-stationary mobile monitoring campaign were conducted, stationary data were used for exploration and model development.

^e.Monitoring data were not obtained by the papers but obtained from other previous studies (dark gray cells in the table)

^f.

No description of weekday-only or weekend/weekend sampling

• – 38 sites in Saha et al. 2021 from the National Oceanic and Atmospheric Administration, Bay Area Air Quality Management District, and other field campaigns
• – 42 homes in Kerckhoofs et al. 2017 from van Nunen et al. 2017
• – 20 background sites in Kerckhoffs et al. 2021 from Beek et al. 2020

^g.Mobile non-stationary data used only for model development and all the other types of monitoring data used for model evaluation or temporal adjustment

^h.Not clear about whether these non-mobile monitoring is rotating or snapshot campaigns based on the description in the paper

^i.Monitoring data used for model evaluation only and not for model development

Among the seven studies documented in Table 3, two studies in the U.S. measured PNC or other pollutants using both mobile and non-mobile monitoring campaigns in the same study areas and periods. These studies also declared that one of their aims was to compare monitoring data across different designs. The first, the BPRHS study conducted in Boston and Chelsea, used the non-mobile short-term rotating design class with 6-week sampling at 2–11 rotating home sites and a maximum of 20 sites to measure PNC between 2012 and 2014 (Simon et al. 2017). Using the same instrument type, they also used the mobile non-stationary design to collect 3- to 6-hour samples between 5 a.m. and 9 p.m. from 20 and 40 km routes of driving on 42 and 46 days in Boston and Chelsea, respectively, both over four seasons between 2011 and 2015. In the comparison using these two different design class datasets, they found similar hourly and monthly patterns, but different spatial patterns (Table S3). On-road measurements from mobile monitoring, even those sampled near participant homes, gave significantly higher concentrations compared to home-site measurements from non-mobile monitoring. They also compared to measurements obtained from the two long-term sites deployed for more than a year. However, because this long-term sampling was carried out only at a single site in each of the cities, we did not classify this sampling as non-mobile long-term monitoring. The other U.S. study in Pittsburgh aimed to explore the fine-scale spatial variation in eight 1-km urban neighborhoods, defined by different land use characteristics, with combined mobile and non-mobile monitoring (Li et al. 2019). In the non-mobile long-term monitoring design, they operated 10 fixed sites and measured PNC, PM_2.5, CO, and NO₂ for 2–8 months in 2016–2017. Their mobile non-stationary monitoring measured BC and PM₁ components in addition to PNC, CO, and NO₂, using different instrumentation for the same pollutant from their non-mobile sampling. Drive routes covered all public roads in the same study neighborhoods during the same overall time period; each of the 9–25 drives per neighborhood lasted 45- to 60- minutes and was carried out during dry conditions between 5 a.m. and 10 p.m. on weekdays in summer and winter. Their findings, using temporally corrected data, showed that within-neighborhood spatial variability was correlated with the contrast of land use in each neighborhood. However, despite their intention to combine both mobile and non-mobile data, their spatial analysis relied on mobile monitoring data. They compared mobile measurements from 50-meter grid cells to those in one central reference cell (pseudo monitor), selected as the most representative grid cell in each neighborhood. They did not use their long-term sites as the reference due to concerns about data comparability resulting from different sampling equipment and limited non-mobile data.

One U.S. study and three studies in the Netherlands used two mobile monitoring sub-designs with both non-stationary and short-term stationary sampling. The U.S. study, as a part of the ACT-AP study, conducted a mobile monitoring campaign in the Puget Sound area for a year starting in March 2019 and measured PNC, BC, NO₂, PM_2.5, CO, and CO₂ (Blanco et al. 2022c). They used the non-stationary sub-class along 9 fixed routes (1,069 total km) in conjunction with the stationary short-term sub-class at 309 2-minute parked locations in order to better represent residential exposure to traffic-related air pollutants than may be possible when only sampling on roads. Using about 29 visit-level measurements at each stop location, they reported larger temporal variability than spatial variability for all pollutants; only UFP had higher spatial variability than temporal variability. There was also a non-mobile monitoring campaign carried out by the ACT-AP study that included both short-term snapshot and rotating designs as well as long-term sampling (Bi et al. 2022; Shaffer et al. 2021); however, because this non-mobile monitoring using LCS measured criteria pollutants only without PNC, we did not consider two ACT-AP studies as a single study that used mobile and non-mobile design classes. Among three related papers focused on UFPs in the Netherlands (Kerckhoffs et al. 2016, 2017, 2021), the first study leveraged stationary and non-stationary monitoring campaigns in Amsterdam and Rotterdam during 2013 (Kerckhoffs et al. 2016). In their non-stationary design, a vehicle drove between 9 a.m. to 4 p.m. on 42 days during two seasons in two cities combined; it provided 1-minute measurements on nearly 3,000 road segments averaging 130 meters. In addition, as the short-term stationary sampling, the same vehicle stopped for 30 minutes at stationary sites along their routes each sampling day, revisiting each of the 161 sites in both seasons. The second study expanded this approach to include more than 5,000 road segments in three cities (Amsterdam, Maastricht, and Utrecht) during 2014–2015 (Kerckhoffs et al. 2017). The last study even further expanded to over 14,000 road segments in more than 5 cities (Kerckhoffs et al. 2021). All three studies showed higher non-stationary measurements than stationary measurements both at traffic and urban background locations. These studies further used their measurements to develop prediction models. Although they also deployed one reference site, we did not categorize this as non-mobile long-term monitoring, because it is only a single long-term site and measurements were specifically used for temporal correction of the temporally unbalanced mobile measurements (Klompmaker et al. 2015; Montagne et al. 2015).

3.3. Data leveraged for model development

Data from different design classes and sub-classes each provide unique strengths that, when integrated into the same model, can possibly contribute to improving the assessment of long-term exposure. However, three out of four papers that used different design classes and developed prediction models in Table S4 only applied monitoring data from the different design classes in separate models. Their model development relied on data from a single type of design classes with measurements from different classes used for model validation (Tables 3 and S4). As the only exception, one U.S. study combined the measurements from mobile and non-mobile monitoring to develop a national prediction model of UFPs (Saha et al. 2021).

Three Dutch studies developed land use regression models using either stationary or non-stationary mobile monitoring data (Kerckhoffs et al. 2016, 2017, 2021). Kerckhoffs et al. (2016) developed two separate prediction models of two-season average concentrations from non-stationary and stationary mobile data, respectively, predicted at mobile stationary and non-stationary sites as well as residential locations, and compared the model performances and predictions. Although model R²s for PNC were low (0.12–0.28), the predictions from non-stationary and stationary mobile data at residential locations were highly correlated (R²=0.89). The second study, expanded to three cities, showed that models based on mobile non-stationary monitoring data over-estimated predictions despite their high correlation with predictions based on stationary monitoring data (Kerckhoffs et al. 2017). The model evaluation using non-mobile monitoring data at 42 homes obtained from a different study (van Nunen et al. 2017) gave moderate and similar performances with R²s of 0.46 and 0.47 for non-stationary and stationary data, respectively. This study used the measurements at residences only for model validation and not for model development. Although PNC instruments differed across the mobile and non-mobile campaigns, the authors justified this approach by noting good agreement found in previous literature (Meier et al. 2013). Their additional assessment that considered the difference in predictions between the two sub-designs and whether these varied by geographic characteristics did not provide any clear insights beyond the basic observation that predictions based on stationary data were generally 30% lower than predictions from non-stationary data. The most recent study (Kerckhoffs et al. 2021) combined mobile and non-mobile monitoring data that were obtained from the present or their previous studies (Kerckhoffs et al. 2016, 2017; van de Beek et al. 2021; van Nunen et al. 2017). Their mobile monitoring included non-stationary measurements collected over 14 months on 14,392 road segments in 5 major cities and multiple towns as well as stationary measurements at 400 repeated 30-min parked locations. The non-mobile monitoring included 24-hour samples at 42 homes across three seasons in 2014–2015 and 2-week samples at 20 background sites across three seasons in 2016–2017. They developed three different land use regression and machine learning models of 14-month average concentrations using the non-stationary mobile data as training data, after temporal correction based on a reference site, and evaluated model performance using stationary mobile and non-mobile home-site data as test data. While the mobile and non-mobile PNC measurements came from different instruments, co-located data at a background site measured on an every 3-week basis showed good agreement (correlation coefficient=0.81). External validation R²s ranged between 0.25 and 0.6 depending on the model.

Finally, the first paper that combined mobile and non-mobile monitoring data into a single model, Saha et al. (2021) reported a nationwide land use regression model for PNC. Their non-mobile long-term data included measurements from 19 urban, 15 rural, and 4 near-airport locations collected over 1–12 months between 2009 and 2019. In addition, they used spatially dense intraurban mobile monitoring data from three cities (Pittsburgh PA, Oakland CA, and Baltimore, MD) collected in summer and fall of 2017–2019. Their prediction model of annual average concentrations was based on data temporally corrected to represent 2016–2017. An external validation using 3–6-week average concentrations at 30 locations in Pittsburg gave R² of 0.54 with RMSE of 2,600 pt/cm³. Although the model performance was higher than that reported by many previous city-specific studies of UFP, as authors noted, large spatial variability resulting from the national scale of their model could have affected the performance. Further, although they investigated the consistency of some measurements between different design classes, this investigation was limited to specific seasons and/or areas. Their monitoring data came from different seasons, years, sampling durations, and studies, some of which used different sampling equipment; these discrepancies could have impacted their results.

4. Suggestions for application to epidemiologic cohorts

Our scoping review suggests that there are some overarching principles regarding spatial and temporal representation, data compatibility between design classes, and data leveraged for model development. All these design features will contribute to the characterization of spatial variability in long-term residential exposure to air pollution, which will affect the success of future epidemiological cohort studies. We recommend that first, given a target cohort, monitoring site selection should prioritize participants’ residences or locations with similar geographic features to those at participants’ residencies over a diverse land use representation. While heterogeneity of data is important for development of well-performing prediction models, locations represented by some land use characteristics may not necessarily represent where people live. Previous studies have shown that misaligned monitoring and prediction locations can result in biased and/or imprecise health effect estimates (Szpiro et al. 2011; Szpiro and Paciorek 2013). Second, temporal resolution of the monitoring campaign should cover as many hours, days of the week, and seasons as possible. Measurements obtained only during business or rush hours on weekdays, even for an extended period, will provide average exposure estimates that are dominated by some diurnal or weekly patterns that do not accurately represent long-term average exposure (Blanco et al. 2022a; Blanco et al. 2022b). Despite increasing attention by many researchers to ensure their designs have adequate spatial variability, many published designs have not also addressed temporal balance in their sampling designs. Although many studies operated a single central site for a relatively long period in order to allow temporal correction of the temporally imbalanced measurements, it has not been demonstrated that any single site can sufficiently represent temporal patterns over space, particularly for pollutants with fine-scale spatio-temporal variability such as UFP. Third, research is needed to investigate whether and how measurements from different design classes can be compared and then combined into a single exposure prediction model. Ideally the combined model will overcome the limitations of each separate design class and maximize benefit from each to capture adequate spatial variation affected by different emission sources or pollutant features. While a few studies have examined data compatibility between different design classes, they relied on limited data over small areas and/or short time periods and mostly compared measurements or predictions from two different design or sub-design classes (Kerckhoffs et al. 2016, 2017; Li et al. 2019; Saha et al. 2021; Simon et al. 2017). Regarding approaches to combine the data from different design classes into a single prediction model, an existing prediction model framework could be expanded to include multiple sources of data from different design classes. In this case, the monitoring data would be adjusted using different calibration weights that would allow the uncertainty of the measurements to vary by data sources (Bi et al. 2020, 2022; Saha et al. 2021). For example, Saha et al. 2021 applied different calibration weights (which they call correction factors) to their monitoring data collected from mobile and non-mobile design classes in the national UFP land use regression model. Another possible approach is to develop an ensemble model that is a weighted combination of predictions, rather than measurements, obtained from separate models developed using monitoring data from different designs (Di et al. 2019).

5. Discussion

We explored novel monitoring designs that leveraged on emerging monitoring technology and/or approaches to characterize long-term average spatial distributions of pollutants. Although the number of monitoring campaigns using mobile monitoring or LCS has dramatically increased, many studies were focused on equipment validation or exposure mapping rather than an epidemiological application. Only very recently have exposure assessment studies begun to focus on epidemiological applications with the aim of estimating long-term average exposures with substantial spatial variability. Our review draws particular attention to the monitoring designs of these studies in order to gain critical insight into the advantages and limitations of design characteristics for epidemiological applications and provide practical suggestions to benefit the design of future studies.

When exposure assessment studies focus on specific cohorts at the design stage, it is straightforward for them to address spatial representation of sampled locations, data compatibility, and the data leveraged for model development. To our knowledge, to date only three cohort studies have both reported their novel exposure assessment campaigns and applied predictions from these campaigns to inference about health effects (Table 2), although many studies stated the eventual goal of epidemiological application. The Canadian mobile monitoring studies collected PNC in Toronto and Montreal, developed land use regression prediction models, and examined the association with brain tumor incidence from CanCHEC (Weichenthal et al. 2020). In addition, ACT-AP carried out LCS non-mobile monitoring for PM_2.5 and NO₂, developed spatio-temporal prediction models, and investigated the association with Alzheimer’s disease pathology and dementia incidence (Shaffer et al. 2021). In Europe, predictions from mobile monitoring campaigns of PNC, PM₁₀, PM_2.5, NO_X, and NO₂ in three Dutch cities were used to estimate associations with incidence of cardiovascular and cerebrovascular diseases in two Dutch cohorts of the European Prospective Investigation into Cancer and Nutrition (EPIC-NL) study (Downward et al., 2018). Their affiliated exposure assessment studies adopted new features in their monitoring designs in order to characterize individual-level long-term average exposure, possibly because of their original intended cohort applications. For example, exposure assessment in the Netherlands relied on multiple monitoring campaigns that expanded the spatial and temporal resolutions and investigated data comparability from mobile stationary and non-stationary designs. Exposure assessment in the ACT-AP study paid careful attention to design at both spatial and temporal scales for their mobile and non-mobile monitoring campaigns in order to obtain representative long-term average exposure estimates for cohort participants. However, none of these studies provided health effect estimates using the predictions combined from or compared between different design classes.

Previous cohort studies that employed monitoring designs with multiple design classes can also provide future guidance. Over a decade ago, two large cohort studies were among the first to include cohort-specific monitoring using some non-mobile design classes. For example, the Multi-ethnic Study of Atherosclerosis and Air Pollution (MESA Air) and its ancillary studies adopted non-mobile long-term, short-term rotating, and short-term snapshot sampling to estimate individual-level air pollution concentrations for more than 7,000 MESA participants in six U.S. cities (Kaufman et al. 2012). They measured 2-week samples of PM_2.5, PM_2.5 components, and gaseous pollutants at 26 fixed sites and over 300 MESA participant homes in their long-term and short-term rotating campaigns, respectively, along with a subset of the pollutants at urban background and street sites in their snapshot campaign in 6 U.S. cities mostly during 2005–2009 (Cohen et al. 2009; Keller et al. 2015; Kim et al. 2016). The European Study of Cohorts for Air Pollution Effects (ESCAPE) also used non-mobile long-term and short-term snapshot sampling to assess individual exposures in more than 20 European cohorts (Beelen et al. 2014). Their snapshot campaign provided 2-week samples of PM_2.5, PM_2.5 components, and gaseous pollutants at 20 background and street sites in 20 cities, each in three seasons over one year during 2008–2011 (Eeftens et al. 2012b). These were among the first large-scale efforts to adopt new paradigm monitoring for cohort studies; they improved their characterization of spatial variability by adopting exposure assessment designs that substantially increased the number of monitoring locations near roads and at residential locations. Their improved monitoring data subsequently contributed to the development of exposure prediction models and the investigation of the association with mortality and disease incidence (Beelen et al. 2014; Kaufman et al. 2016a; Raaschou-Nielsen et al. 2013; Wang et al. 2019). More recent exposure assessment campaigns have built upon this early work and expanded to LCS and mobile sampling.

Several other large air pollution cohort studies have adopted an alternative exposure assessment strategy that leverages existing regulatory monitoring data in combination with other exposure data sources not discussed in this scoping review. For example, the exposure surfaces developed for the U.S. Medicare cohort studies were based on exposure prediction models that combined regulatory monitoring data with satellite imagery data (Di et al. 2019, 2020; Requia et al. 2020). These studies relied on satellite-based air pollution estimates that allowed the exposure assessment in the areas with few regulatory monitoring sites and/or the improvement of exposure prediction when complemented with ground monitoring data. However, the accuracy of satellite-based estimates is unknown and likely inadequate at locations where regulatory monitors are sparse (Bi et al. 2020, 2022). Furthermore, these satellite observations are not available for UFP. These limitations provide additional justification for alternative novel monitoring designs.

There are some new approaches to exposure assessment that take advantages of features of emerging monitoring technology and these have the potential to be widely adopted in future monitoring designs that are intended to better characterize spatial variability of air pollution over a long-term period. First, mobile monitoring has been extended to different travel modes such as biking, walking, or modes of public transportation including bus and tram (Farrell et al. 2016; Hankey and Marshall 2015b; Mueller et al. 2016b). Alternative travel options allow mobile monitoring to characterize study areas where sampling from motorized vehicles is less efficient or impossible due to limited road networks, complex road conditions, and/or traffic congestion. In addition, global positioning systems (GPS) allow accurate individual characterization of space-time-resolved location information. This has the potential to broaden exposure assessment to capture exposures based on individuals’ time activities outside of their homes. Lastly, LCS monitoring has engaged communities and allowed participant involvement; examples include PurpleAir and citizen science studies that address environmental justice and identify air pollution sources (Bi et al. 2020; Jerrett et al. 2017). Community involvement can improve understanding of population exposure characteristics that are not well represented by regulatory monitoring networks. This community engagement can be expanded to citizen science initiatives where scientific projects are initiated and run by citizen scientists; a recent narrative review of these initiatives summarized their exposure assessment successes and possible expansion to epidemiology, noting that these initiatives need to overcome their challenges with e.g., scientific expertise and data quality, particularly for epidemiological applications (Froeling et al. 2021).

New air pollution monitoring paradigms made possible by advanced technology based on mobile platforms and low-cost sensors have opened up novel opportunities for characterizing air pollution exposures and improving our ability to make inference about the impact of long-term exposure on health outcomes. This in turn allows us to answer scientific questions that we were not feasible to address in previous air pollution cohort studies. Although exposure assessment design plays an important role in the quality of epidemiological inference, there has been little guidance in the literature on this topic. This review provides specific suggestions help future studies improve their exposure assessment designs, namely to prioritize residence-based site selection, provide comprehensive temporal coverage, and encourage leveraging data from different monitoring design classes for model development.