Modeling Traffic-Related Air Pollution Burden of Disease using High Spatial Resolution Data

Abstract

Introduction

Traffic-related air pollution (TRAP), particularly fine particulate matter (PM_2.5), is linked to adverse health effects, including premature mortality, particularly in dense urban areas. Modeling the chain from traffic to air quality, exposure, and burden of disease (BoD) enables health impact assessments of PM_2.5 for scenario planning. However, the influence of spatial resolution in population and mortality data on TRAP-attributable BoD estimates remains underexplored.

Methods

This study evaluated the BoD attributable to PM_2.5 exposure from traffic-related activities in the Dallas, Texas, USA region and the impact of incorporating varying spatial resolutions of population and mortality data. The air quality modeling was conducted using travel demand modeling, on-road emissions modeling and dispersion modeling. This was followed by spatial analysis incorporating population data for exposure assessment and BoD quantification. The study compared high-resolution population mapping (30-meter) with traditional census tract-level data to quantify differences in PM_2.5 exposure and resulting mortality estimates. The study also incorporated mortality rate data at county level and census tract levels.

Results

The use of high-resolution population data resulted in lower average exposure concentrations and mortality estimates compared to tract-level PM_2.5 exposure estimation. Higher mortality estimates occurred when using finer mortality rate data compared with coarser county-level data.

Conclusions

Spatial resolution in population and mortality data substantially affects TRAP-related BoD estimates. In the study region, high-resolution population data, which accounts for land use and excludes non-residential areas near major roads, led to lower estimates of population exposure and lower BoD estimates compared to coarser resolution data. In settings with substantial populations near high-traffic corridors, high-resolution population data, which captures actual population concentrations unlike the even distribution assumed in coarser census tract data, combined with localized mortality data, may increase BoD estimates by accurately mapping exposure and health variations.

INTRODUCTION

Air pollution is a global phenomenon that poses a major threat to the environment and public health (WHO, 2023a). Ambient air pollution, primarily derived from the transportation, industry, energy and power generation, and buildings sectors (WHO, 2022), is a leading cause of non-communicable diseases (WHO, 2018). Air pollution adversely affects human health including increasing the risk of all-cause and cause-specific premature mortality and cardiovascular, respiratory, metabolic and neurodegenerative diseases (HEI, 2022; USEPA, 2023a; Wilker et al., 2023). Particulate matter (PM) is a key component of ambient air pollution which adversely impacts human health at every stage of life (ALA, 2023). Traffic-related air pollution, particularly PM_2.5, is linked to adverse health effects including premature mortality in urban areas. Modeling of exposure assessment and burden of disease are affected by the spatial resolution of data. This study evaluates how different spatial resolutions of population and mortality data affect estimates of premature mortality attributable to traffic-related PM_2.5 in the Dallas-Fort Worth region.

PM is one of the six criteria air pollutants for which the United States Environmental Protection Agency (U.S. EPA) must set ambient air quality standards at the national level, as required by the Clean Air Act (USEPA, 2022). PM is composed of numerous types of chemicals; some of which are directly emitted from sources, such as construction sites, paved and unpaved roads, wildfires, and smokestacks (USEPA, 2023d), where the concentrations of primary PM are elevated in proximity to these sources (Gilbert et al., 2003). On the other hand, most PM is formed in the atmosphere through a series of complex reactions between pollutants, namely sulfur dioxide and nitrogen oxides (USEPA, 2023d). Secondary PM is then transported across regions and adds to the background concentration (Shapero et al., 2023).

PM is typically categorized based on particle size (UCS, 2023; ALA, 2023). Coarse particles, also known as PM₁₀, have a diameter of less than 10 μm while fine particles or PM_2.5 have a diameter of less than 2.5 μm. Their small size means that they can be easily inhaled and deposited through the breathing passageways. Exposure to high PM concentrations can pose increased risks for cardiovascular and respiratory health complications (USEPA, 2022; ALA, 2023). Certain population groups are more vulnerable including older adults with chronic cardiovascular or respiratory disease, pregnant women, infants and children (Bekkar et al., 2020; USEPA, 2023b). Research has also highlighted the unequal distribution of particle pollution, with marginalized communities consisting of people of color and low-income populations experiencing a higher burden (Tessum et al., 2021; Rentschler and Leonova, 2023).

Exposure to particle pollution is estimated to be responsible for approximately 4.2 million premature deaths globally (WHO, 2022). The Health Effects Institute estimated that exposure to PM_2.5 accounts for 48,000 premature deaths in the U.S. annually (HEI, 2020). Tailpipe and non-tailpipe emissions from gasoline- and diesel-fueled vehicles are major contributors to urban PM_2.5 (Boogaard et al., 2022; Li and Managi, 2021). Even though there has been a continuous decrease in traffic-related air pollution (TRAP) over the last decade due to air quality regulations and improvements in vehicle technology (Boogaard et al., 2022), PM remains a major health concern (Wang et al., 2023). The transportation sector accounts for less than 10 percent of both PM_2.5 and PM₁₀ in the U.S. (USEPA, 2023c), yet it contributes to at least 20 percent of total PM_2.5 attributable mortality (Lelieveld et al., 2015; Davidson et al., 2020). A study by Davidson et al. (2020) evaluated the recent and future health burden attributable to the transportation sector by source type in the U.S. They found that, among mobile sources, emissions from on-road sources are the largest contributor of PM_2.5- and ozone-attributable premature deaths (Davidson et al., 2020). This is primarily due to the intersection of high traffic activity, roadway densities, and population congregations resulting from rapid increase in urbanization and economic activity (Hajat et al., 2013; Han et al., 2014; Mainka et al., 2015; Larkin et al., 2016; Hao and Liu, 2016; UN, 2018; Yang et al., 2018; Wang et al., 2019; Zhu et al., 2019; Boogaard et al., 2022; She et al., 2022). Almost 79.5 million U.S. residents lived within 500 meters of major and high-volume roads in 2020 (Antonczak et al., 2023) increasing the risk of exposure since air pollution concentrations are highest within 0–300 meters from the point of release (Gilbert et al., 2003; Brugge et al., 2007; Rioux et al., 2010). In addition to the proximity to the point of release, urban PM spatially varies (Pinto et al., 2004) due to other influencing characteristics such as meteorology (Calkins et al., 2016; Manju et al., 2018), traffic volume (Boehmer et al., 2013), terrain and patterns of land development (Lu et al., 2018), and urban morphology (Li et al., 2021a).

Quantifying the impacts of PM_2.5 exposure on health outcomes is critical and is supported by the strong scientific evidence from epidemiological and toxicological studies of harmful acute and chronic effects in humans (Xu et al., 2022; Bhat et al., 2021; Askariyeh et al., 2019). It is also critical to assess the spatial heterogeneity of PM and its attributable health impacts in urban settings (Zhang and Cheng, 2022; Li et al., 2021b). By estimating health impacts attributable to PM_2.5 and their distribution, decision-makers can acquire a deeper understanding of the health and equity impacts of PM, and accordingly devise evidence-informed mitigation strategies (Roser et al., 2021; Bhat et al., 2021; Southerland et al., 2021).

Quantifying the health impacts attributable to PM_2.5 exposure is feasible due to the availability of required input data and standard methods. The estimation process typically involves the following procedures: the exposure to air pollution is assessed and assigned to people in the study area, after estimating emissions and the resulting ambient air pollution at a selected spatial resolution. Then, the attributable health impacts are assessed using a comparative risk assessment approach often utilizing well-established concentration-response functions (CRFs), sometimes called exposure-response functions (ERFs), from critically appraised epidemiological evidence. Finally, the uncertainty in the estimates is calculated and results communicated. Researchers have developed integrated models that incorporate traffic data from travel demand models, emissions, and air quality modeling with the burden of disease assessment (Wu and Song, 2019). There are also widely used tools that provide information on air pollution exposure and its population health impacts including but not limited to COBRA, BenMAP-CE, Air Q+, InMAP, ITHIM, GAINS, SIM-air, HAPIT, Ecosense, and TM5-FASST (Bhat et al., 2021; Tessum et al., 2017; Anenberg et al., 2016; Woodcock et al., 2014). These tools apply mathematical and statistical models and common data inputs for population, air quality, and baseline mortality rates, to evaluate population exposure to air pollution and estimate selected health impacts. Yet these tools differ in their methodological approaches, intended use, spatial resolution of data, geographical coverage and their selection of the CRF/ERF.

The level of detail or granularity of input data plays an important role in quantifying health impacts since mobility patterns, exposure to air pollution, distribution of population, and socioeconomic and sociodemographic factors vary widely within urbanized areas (deSouza et al., 2023; Montgomery et al., 2023; Rowangould et al., 2019). This underpins the need to understand how the burden of disease changes at small spatial scales which would allow policymakers and city officials to identify localized intervention strategies that target the reduction of air pollution exposure and risk disparities (Nawaz et al., 2023; Montgomery et al., 2023; Southerland et al., 2021; Castillo et al., 2021; Rowangould et al., 2019). Mortality attributable to ambient air pollution, particularly PM_2.5, has been quantified in several studies at the global, national, regional, or city scale (Camilleri et al., 2023a; Nawaz et al., 2023; Pozzer et al., 2023; McDuffie et al., 2021; Southerland et al., 2021; Bhat et al., 2021; Sohrabi et al., 2020; Baxter et al., 2017). The variability in the burden of disease, however, may be obscured when estimating health impacts at aggregate levels, i.e., country, state, or county level, compared to finer spatial scales (Southerland et al., 2021). There has been an increasing use of disaggregate high-resolution datasets. This includes analyzing the burden of disease at different resolutions for one or multiple of the following input datasets: exposure estimates, population density, and baseline mortality rates (Camilleri et al., 2023a; Camilleri et al., 2023b; deSouza et al., 2023; Visa et al., 2023; Valencia et al., 2023; Agibayeva et al., 2022; Xu et al., 2022; Castillo et al., 2021; Southerland et al., 2021; Mohegh et al., 2020; Alexeeff et al., 2018; Morelli et al., 2016; Shandas et al., 2016; Chart-asa and Gibson, 2015). Researchers have also explored the implications of using emerging race- and ethnicity-specific CRFs along with race- and ethnicity-specific mortality rates to estimate health impacts of air pollution (deSouza et al., 2023; Spiller et al., 2021). The overall inference from these studies is that the resolution of data influences the burden of disease estimates and that this influence may vary by air pollutant.

This study adds to the existing and expanding literature by documenting the application of high-resolution data inputs to estimate the burden of mortality attributable to traffic-related PM_2.5. The methodological aim is to evaluate the influence of using population and mortality data inputs with varying spatial granularity on the burden attributable to PM_2.5 exposure. This burden of disease assessment is conducted for the Dallas-Fort Worth (DFW) region in the U.S., where we are the first to utilize the U.S. EPA Dasymetric mapping geospatial dataset containing detailed population density estimates (Baynes et al., 2022). Although some studies have employed the Dasymetric mapping technique to estimate exposure (albeit not at the fine resolution of 30m used in the current study), they did not perform a burden of disease assessment (Requia et al., 2018; Shandas et al., 2016; Hu et al., 2007).

MATERIALS AND METHODS

This section summarizes the methodology and data sources used. Figure 1 provides a visual representation of the process, starting with study area selection. This is followed by air quality modeling which involved a multilateral modeling framework consisting of travel demand, emissions, and high-resolution dispersion modeling of PM_2.5. The framework then integrates the high-resolution dispersion modeling results with advanced spatial analysis to capture the distribution of population densities and assign exposures. Then, the burden of disease is assessed to quantify premature mortality attributable to traffic-related PM_2.5 exposures. Finally, the design of different input data combination scenarios to evaluate the impact of data spatial resolution on the burden of disease estimates are described.

Figure 1.

Visual Representation of Study Methodology.

Site Selection

The North Central Texas Council of Governments (NCTCOG) region which consists of 16 counties and serves as the Metropolitan Planning Organization (MPO) for the DFW metroplex was selected as the study location. The vast and rapidly expanding area (Tsiaperas, 2023; Lee, 2021), significant and growing economic activity (City of Dallas, 2023a), and diverse socioeconomic and sociodemographic profile (City of Dallas, 2023b) make this region a suitable testbed for evaluating the health burden attributable to TRAP. In the DFW region, the transportation sector is a major contributor to air pollution (Lawrence, 2019; IQAir, 2023). The number of vehicles and population in the region are increasing (Lawrence, 2019).

Additionally, the availability of detailed high-resolution spatial data in the study area enhances the assessment. The sociodemographic characteristics of study area are described in Supplementary Material in Figure S1. Figure 2a shows that populations are highly distributed in the central and northcentral tracts and the roadway density map in Figure 2b. displays similar trends. This can be a possible indication of regions with higher human exposure to TRAP (Batterman et al., 2014; Alruwaili, 2020; Yu et al., 2023).

Figure 2.

(a) Population Density, and (b) Roadway Density, by Census Tract.

We employed a full chain modeling framework that integrates traffic, emissions, dispersion modeling with exposure and BoD assessment. The process is done using a Dispersion Modeling Framework (DMF) web application developed in R shiny (Jaikumar et al., 2024). The DMF has built-in functions to process the traffic data from the NCTCOG travel demand model by extracting and calculating vehicle miles traveled (VMT) by vehicle type, hour, average speed, and road type. The characteristics of each roadway link are then matched to the MOVES emissions rates to calculate hourly emissions for each link. Finally, the DMF prepares AERMOD input files based on the hourly emissions per link. The following sections describe each component of the DMF.

Travel Demand Modeling

Traffic data was sourced from the 2019 NCTCOG travel demand model (TDM) which consists of 5,352 traffic analysis zones (TAZs) covering 13 out of the 16 counties in the region. The network has a total of 51,715 roadway links including interstates, freeways/expressways, arterials, collectors, and approach roads (ramps) (Figure 3). The TDM uses the four-step sequential process to forecast VMT and other travel demand characteristics (NCTCOG, 2023). To account for temporal variations in traffic volume and vehicular speed, the link traffic data is modeled for three time periods of a typical weekday: morning peak, evening peak, and off-peak hours. Hourly activity factors were utilized to further disaggregate the traffic data to generate hourly results including VMT, speed, and volume at the link level. These factors were estimated using 10 years of activity data aggregated using the Texas Department of Transportation (TxDOT) Automatic Traffic Recorder (ATR) stations in each county (TxDOT, 2023).

Figure 3.

Roadway Network in NCTCOG’s Travel Demand model.

Vehicle Emission Rates Modeling

Emission rates for PM_2.5 in grams per mile were estimated using the third version of EPA’s Motor Vehicle Emission Simulator (MOVES3), by the Texas A&M Transportation Institute for environmental applications for the Texas Department of Transportation’s air quality toolkit (TxDOT, 2023). Composite rates applicable to the NCTCOG region were used to estimate total emissions for each link based on the roadway type and average link speed. The composite rates were generated by considering the different vehicle and fuel type combinations in MOVES and included the following processes: running exhaust, crankcase running exhaust, brake wear, tire wear, and running evaporative losses but did not account for the resuspension of road dust. These composite rates do not consider the vehicle classes in the TDM but reflect fleet mixes developed for the study region based on local vehicle registration data and traffic types specific to roadway types.

Dispersion Modeling

AERMOD was selected to allow for fine-scale dispersion analysis over larger areas. Meteorological data for 2019 processed by the Texas Commission on Environmental Quality (TCEQ) was used and considered all 365 days to ensure the comprehensive coverage of varying atmospheric conditions. To capture near-road concentration gradients, we adhered to a receptor placement strategy underpinned by a literature review which showed that traffic-related air pollutant concentrations and associated health effects fall to background levels at around 300–500 meters, with a steep decay with increasing distance from the roads (Liu et al., 2019; Weaver and Gauderman 2018; Volk et al., 2011; Kan et al., 2008; McConnell et al., 2006). Receptors were placed at intervals of 50m, 150m, and 300m from the midpoint of each roadway link on both sides, with additional gridded receptors every 500m across the entire study area. As a result, a total of 88,783 receptors were placed across the study area and the model estimated PM_2.5 concentration in μg/m³ at each of these receptors. Receptors were placed at ground level where human exposure is most relevant. Vertical dispersion, such as overpasses, was not modeled due to computational constraints, as traffic emissions are primarily ground-level sources. The AERMOD modeling parameters are summarized in Table 1.

Table 1.

AERMOD Modeling Parameters.

Modeling Parameters	Description
Geographical Region	NCTCOG
Analysis Year	2019
Pollutant	PM2.5
Month and Day	All 365 days of the year
Averaging Period	Annual
AERMOD Source Type	LINE
Variable Emission Rate	Emission rates vary by hour of day
Release Height	2.6 meters
Initial Vertical Dispersion Coefficient	2.18
Meteorological Data	Texas Commission on Environmental Quality (TCEQ) preprocessed 2019 data for all 365 days
Receptor Placement	50, 150, and 300 meters from the roadway centerline/midpoint, in addition to a mesh of gridded receptors at 500 meters × 500 meters beyond
Total Number of Receptors	88,783
Model Output	Pollutant concentration in μg/m3 at each receptor

Exposure Assessment

For exposure assessment, high-resolution gridded population density estimates for the year 2020 were obtained from the U.S. EPA’s Intelligent Dasymetric Mapping toolbox (Baynes et al., 2022). This dataset was developed using the Dasymetric Toolbox from the USEPA and employed the 30 m, 2019 land cover classification from the 2019 National Land Cover Dataset (NLCD) as the ancillary raster (U.S. Geological Survey, 2019). The source units were the 2020 United States Census Blocks with corresponding populations. Numerous geospatial datasets related to land use, such as roads, protected lands, and steep slopes, were utilized to identify uninhabited areas. These uninhabited areas, such as primary roads, wind turbines, and energy production sites, were assigned a new “uninhabited” class. Additionally, classes like open water, perennial ice/snow, and emergent wetlands were preset at zero population for every state. Blocks that were irregularly shaped or small were merged with neighboring blocks within the same census group or tract, and populations were reassigned at the pixel level. Representative population densities (RPDs) were determined for 17 land cover classes using sampling or intelligent area weighting (IAW). This refinement, along with updated data sources like ESRI StreetMap Premium, improved the precision of population allocation by ensuring alignment with habitable areas while filtering out uninhabited or mixed-use regions. The data was accessed from EPA’s EnviroAtlas website in a raster format with each pixel corresponding to a 30-meter grid (EnviroAtlas, 2020). The raster was converted to a points layer to extract pixel values corresponding at each grid to the center point of the 30-meter grid.

After the PM_2.5 concentration at each receptor was extracted from AERMOD’s outputs, a uniform air quality surface at 30 m resolution was generated using Inverse Distance Weighting (IDW) method to match the Dasymetric population layer. Although Kriging and spline interpolation methods were also tested, IDW was ultimately selected due to its ability to produce more variable results. Figure 4 illustrates the interpolation raster layer for the study area, with a zoomed-in portion of I-20 to show the receptor placement along the roadway. The zoomed-in portion depicts receptors positioned at the pre-specified intervals to capture detailed concentration gradients near roadways and the 500 m grid receptors. The figure also includes the Dasymetric raster layer, highlighting population distribution at this resolution.

Figure 4.

Exposure Assessment - AERMOD Receptor Placement and Interpolation Raster/Layers.

Two exposure estimates were determined for comparison:

1. Dasymetric Population Exposure Estimate: With the interpolated concentration data and Dasymetric population data both available at 30-meter intervals, a population-weighted average concentration was calculated for each census tract. This was achieved using the formula:

   $${PopWeightedAverageExposureConcentration}_{tract}(ug/m^3)=\frac{\sum {(C}_i\times P_i)}{\sum P_i}$$

   Where C_i is the interpolated concentration at each 30-meter grid point within the given tract and P_i is the population at each corresponding grid point from Dasymetric mapping data.
2. Tract Level Exposure Estimate: For the census tract-level exposure assessment, instead of population weighted average concentration, a simple average of the interpolated raster within each tract was computed.

   $${AverageExposureConcentration}_{tract}(ug/m^3)=\frac{\sum {(C}_i)}{N}$$

   Where C_i is the interpolated concentration at each 30-meter grid point within the given tract and N is the number of grid cells within each census tract

Burden of Disease Assessment

The burden of disease assessment quantifies all-cause mortality attributable to PM_2.5 within the NCTCOG. To conduct the assessment, we synthesized the-weighted average PM_2.5 concentrations from the exposure assessment phase with baseline all-cause mortality rates from three different data sources with different spatial resolutions, as described next.

CDC WONDER County-Level Mortality Rates

The 2010–2015 crude all-cause mortality rate per 100,000 people at the county level was obtained from the Center for Disease Control (CDC) Wide-ranging Online Data for Epidemiologic Research (WONDER) database (CDC, 2023). These years were selected to match the USALEEP data described next. Census tracts within counties were assigned the corresponding county mortality rates.

USALEEP Census-Tract Level Mortality Rates

For higher resolution all-cause mortality rates at the census tract level, we utilized the U.S. Small-Area Life Expectancy Estimates Project (USALEEP) abridged life tables published by the CDC. The data estimates age-stratified life expectancy at birth for 2010–2015 (NCHS, 2018). Age-specific baseline mortality rates for the same period were then estimated by Industrial Economic Incorporated by adjusting the USALEEP abridged life tables to allow users to apply the data in health benefit analyses (Raich et al., 2020). It should be noted that numerous tracts (10%) were missing from the dataset.

The age-stratified mortality rates for each census tract were translated to a crude rate per 100,000 people by initially multiplying the USALEEP mortality rate by the five-year population estimate of each age group. The U.S. population estimates by age groups between 2010 and 2015 were extracted from the National Cancer Institute (NCI) Surveillance, Epidemiology, and End Results (SEER) program dataset (NCI, 2022). The estimates were developed using a hybrid approach consisting of regression, demographic, and proportional models. Finally, tracts with missing USALEEP rates were assigned the county-level crude mortality rate from CDC WONDER. The total estimated number of mortalities within each census tract was then divided by the total population to compute the mortality rate.

Table 2 summarizes the different data sources and their spatial and temporal resolutions used in the current study.

Table 2.

Summary of Different Data Sources used in the Study

Dataset Name	Source	Original Spatial Resolution	Temporal Coverage/Year	Brief Description/Use in Study
NCTCOG Travel Demand Model	North Central Texas Council of Governments	Roadway link level	2019	Traffic volumes, speeds for emissions modeling.
Intelligent Dasymetric Population	U.S. EPA EnviroAtlas	30-meter grid	2020	High-resolution gridded population data for exposure assessment.
Meteorological Data	Texas Commission on Environmental Quality	Weather station point data	2019	Hourly data for AERMOD dispersion modeling.
MOVES3 Emission Rates	Texas A&M Transportation Institute (TTI) / Texas Department of Transportation (TxDOT) Air Quality Toolkit	Lookup tables (region-specific)	For 2019 conditions	PM2.5​ emission rates (g/mile) by roadway type, speed, vehicle mix.
Center for Disease Control - Wide-ranging ONline Data for Epidemiologic Research (CDC) WONDER Mortality Data	CDC WONDER database	County-level	2010–2015	Crude all-cause baseline mortality rates.
United States Small-Area Life Expectancy Estimates Project (USALEEP) Mortality Data	National Center for Health Statistics (NCHS) / Industrial Economics Inc.	Census-tract level	2010–2015	Age-specific and crude all-cause baseline mortality rates.
Texas Dept. of State Health Services DSHS Mortality Data	Texas DSHS	Census-tract level	2010–2015	All-cause mortality counts for baseline mortality rate calculation.

Texas DSHS Census-Tract Level Mortality Rates

Vital statistics data of the number of mortalities between 2010 and 2015 for Texas residents at the census tract level was requested from the Texas Department of State Health Services (DSHS) (Texas DSHS, 2023). Texas DSHS use mortality certificates to compile mortality data. To protect the identity of individuals, tracts with mortality counts between 1 and 9 were suppressed in the dataset, resulting in 7.35% missing data for all-cause mortality counts. Consequently, these tracts were assigned a count of 4.5 which is the midpoint of the values suppressed by Texas DSHS, as applied by Castillo et al. (2021). The methodology used to calculate the USALEEP mortality rates was also utilized to estimate mortality rates using data from Texas DSHS.

USALEEP and Texas DSHS data show higher variability with broader ranges compared to CDC WONDER, indicating differences in mortality rate distributions across the three sources (Table S1).

In the BoD assessment, the baseline mortality rates were combined with exposure estimates to calculate the relative risk differentials (RR_diff) and the population-attributable fraction (PAF). RR_diff which associates all-cause mortality to PM_2.5 exposures in the study area was estimated based on a published exposure-response functions (ERFs) from the systematic review and meta-analysis by Chen and Hoek (2020) as:

$${RR}_{diff}=\mathrm{e}\mathrm{x}\mathrm{p}[\left(\frac{\ln \left(RR\right)}{{RR}_{unit}}\right)\times \left(C\right)]$$

Where RR is the relative risk extracted from Chen and Hoek’s systematic review and meta-analysis and is equal to 1.08 (95% confidence interval (CI) 1.06 – 1.09 per 10 μg/m³ of PM_2.5). C represents the exposure concentration and RR_unit is the exposure unit RR obtained from the original ERFs which is equivalent to 10 ug/m³.¹

PAF, representing the proportion of all-cause mortality attributable to traffic-related PM_2.5 exposures is expressed as:

$$PAF=\frac{{RR}_{diff}-1}{{RR}_{diff}}$$

The final step involved estimating the attributable all-cause mortalities at the census tract level, using the population data, mortality rates, and estimated PAF as:

$$Deathsattributableto\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{f}\mathrm{f}\mathrm{i}\mathrm{c}-\mathrm{r}\mathrm{e}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{e}\mathrm{d}{PM}_{2.5}=PAF\times Mortalityrate\times Populationcounts$$

Model Input Combinations Development

To evaluate the impact of varying spatial data resolution on the TRAP burden of disease, we combined six alterations of input data with different population and baseline mortality datasets, as shown in Table 2.

RESULTS AND DISCUSSION

Exposure Estimates

PM_2.5 exposure estimated for the NCTCOG region using EPA’s Dasymetric mapping (estimated by calculating a population-weighted average of 30-meter resolution concentrations within each census tract) and using census tract-level population data (calculated by finding the spatial average of concentrations at all the receptor within each tract) are compared in Figure 5. We excluded the outliers, i.e. data points outside the whiskers or fences of each box plot, from Figure 5. to improve the legibility of data variability. However, the outliers were retained in the analysis.

Figure 5.

Box Plots comparing Census-Tract Level Mean PM_2.5 Exposure Concentration and Population-Weighted Mean PM_2.5 Exposure Concentration Using Data from Dasymetric Mapping.

Compared against the census-tract level population plot (1.26 μg/m³), Dasymetric mapping data resulted in a lower average exposure concentration (1.08 μg/m³). Furthermore, as evident from Figure 6, the rasterized Dasymetric population layer shows that, within the NCTCOG region, a smaller number of people reside near roadways in contrast to areas further away from roadway centerlines. These findings emphasize (1) an influence of population data resolution selection on pollutant exposure in a geographic region which can potentially impact the burden of disease estimates and (2) that this influence in context specific as it depends on residential and roadway configuration and proximity in the study area.

Figure 6.

Population and PM_2.5 Concentrations Distribution at 30 × 30 Meters Resolution at a Densely Populated Census Tract.

Burden of Disease Estimates

The mean and median attributable mortality to traffic-related PM_2.5 per census-tract as well as the total estimated mortalities in the study area are summarized in Table 3 for all different data combinations (i.e. scenarios developed and tested). The percent difference was also calculated by comparing the summary to the base case scenario, i.e., TR_CDC. The total mortality estimates for the TR_USALEEP combination is 16.44% higher than the base case, highlighting the higher burden of disease estimate when using modeled USALEEP data. Similarly, the TR_TxDSHS combination results in a total mortality estimate that is 16.06% higher than the base case, underscoring the impact of using actual mortality counts from Texas DSHS. The DY_CDC combination, which uses Dasymetric mapping with CDC WONDER data, results in a 6.89% lower estimate compared to the base case. This indicates that using the more refined population data tends to lower the total burden of disease estimate in the study area. In contrast, the DY_USALEEP and DY_DSHS scenarios show modest increases in total estimated mortality, 4.33% and 5.67% higher than the base case, respectively. These increases suggest that even with finer-scale population data, the baseline mortality rates from USALEEP and DSHS still result in higher estimates of PM_2.5-attributable mortality.

Table 3.

Scenarios Developed and Tested.

Data Combination	Population Data Resolution	Baseline Mortality Rate Resolution	Description of Combination
TR_CDC (Base Case Scenario)	Census-tract level	County-level (CDC WONDER)	Population is assumed to be uniformly distributed within a census tract. Baseline mortality rates at the county level are applied to the census tract.
TR_USALEEP	Census-tract level	Census-tract level (USALEEP)	Population is assumed to be uniformly distributed within a census tract. Estimated baseline mortality rates are applied at the census-tract level resolution.
TR_TxDSHS	Census-tract level	Census-tract level (Texas DSHS)	Population is assumed to be uniformly distributed within a census tract. Observed baseline mortality rates are applied at the census-tract level resolution.
DY_CDC	Dasymetric mapping (30 × 30 meters)	County-level (CDC WONDER)	Population distribution is adjusted using high-resolution Dasymetric mapping. Baseline mortality rates at the county level are applied to the census tract.
DY_USALEEP	Dasymetric mapping (30 × 30 meters)	Census-tract level (USALEEP)	Population distribution is adjusted using high-resolution Dasymetric mapping. Estimated baseline mortality rates are applied at the census-tract level resolution.
DY_TxDSHS	Dasymetric mapping (30 × 30 meters)	Census-tract level (Texas DSHS)	Population distribution is adjusted using high-resolution Dasymetric mapping. Observed baseline mortality rates are applied at the census-tract level resolution.

When holding the population data constant at the census-tract level, the USALEEP mortality rates produce the highest estimated total mortalities (460 [95% CI 349, 514]), followed closely by the Texas DSHS (458 [95% CI 348,513]), and the CDC WONDER mortality rates (395 [95% CI 300, 442]). This suggests that using tract level mortality data from Texas DSHS and USALEEP remains higher than the CDC WONDER baseline, which aggregates over larger geographic regions. When comparing mortality data using Dasymetric population mapping, Texas DSHS rates again yield the highest estimate (417 [95% CI 317,467]), followed by USALEEP (412 [95% CI 313,461]), and CDC WONDER (368 [95% CI 279,411]).

Overall, using baseline mortality rates at the higher census-tract level resolution results in higher total burden of disease estimates in this study area when compared to using rates at the county-level resolution. However, using population data at the higher 30 × 30 meters resolution results in lower total burden of disease estimates in this study area, which are at least partly explained by trends in the exposure estimates summarized in the previous section.

The box plots in Figure 7 demonstrate the variability in the estimated number of mortalities attributable to traffic-related PM_2.5 across all scenarios. Generally, scenarios that use tract-level population data tend to have higher estimated all-cause mortality in comparison to scenarios that use the Dasymetric mapping technique. Moreover, the tract-level population data generated more variable/heterogeneous mortality estimates.

Figure 7.

Box Plots Comparing the Estimated All-Cause Mortality Attributable to Traffic-Related PM_2.5 Across the Different Data Combinations in the Study Area.

Figure 8 displays a matrix of scatterplots (on the left) and a correlation matrix (on the right) to illustrate the relationship between pairs of input data combinations, while the diagonal histograms show the distribution of deaths attributable to Traffic-Related PM_2.5 within each data combinations.

Figure 8.

Scatterplots and Correlation Matrix of Deaths Attributable to Traffic-Related PM_2.5 Across the Different Data Combinations in the Study Area.

Scenarios which share the same population data resolution tend to have stronger correlations with each other. The correlation matrix reveals high correlations between scenarios with similar methodologies, particularly those using Dasymetric mapping and those using census-tract level population data. The scenarios using census-tract level mortality rates (USALEEP and TxDSHS) tend to have broader distributions and higher mortality estimates as seen in the scatterplots and histograms, compared to the narrower and slightly lower mortality range seen with scenarios using county level data (CDC). Using coarser spatial data, such as county-level rates, can mask significant local differences by averaging data over larger areas, reducing the visibility of specific variations in individual census tracts. In contrast, finer resolution data, such as census tract-level mortality rates, better capture these localized differences in mortality rates, leading to higher estimates of the attributable mortality to PM_2.5 exposure.

To provide a more detailed analysis of the relationship between population and road characteristics and the health impacts of traffic-related PM_2.5 exposure, we computed the mean and standard deviation of mortalities attributable to PM_2.5 for each data combination. This analysis disaggregates results into population density deciles and roadway density deciles, allowing us to examine how different levels of population and roadway densities influence mortality estimates across the study area.

Figure 9 illustrate the means of mortalities attributable to traffic-related PM_2.5, along with their standard deviations as bars against population density deciles, across the different data combinations.

Figure 9.

All-Cause Mortality Attributable to Traffic-Related PM_2.5 Across the Different Data Combinations in the Study Area by Population Density Deciles

The analysis reveals the mortalities attributable to traffic-related PM_2.5 generally increase with rising population density in the lower population density bins up to 3000 residents per square mile. This trend is visible particularly in scenarios using census-tract level population data (TR_CDC, TR_USALEEP, TR_TxDSHS). In mid-range population density bins (between 3,000 and 10,000 residents per square mile), estimates generally begin to converge across all data combinations. This suggests that population density in these tracts is more evenly distributed, leading to more consistent mortality estimates. In tracts with more than 10,000 residents per square mile, which are limited in number, lower estimated mortality is observed when using Dasymetric population data. This trend occurs because high population density tracts are typically smaller. When using Dasymetric data, greater weight is assigned to grid cells with high population densities, which, in the NCTCOG region, are often located slightly away from major roadways. Consequently, this approach results in lower exposure estimates for these highly dense areas compared to methods that do not account for fine-scale population distribution.

Figure 10 illustrate the means of mortalities attributable to traffic-related PM_2.5, along with their standard deviations as bars against roadway density deciles, across the different data combinations.

Figure 10.

All-Cause Mortality Attributable to Traffic-Related PM_2.5 Across the Different Data Combinations in the Study Area by Roadway Density Deciles

For the scenarios using census-tract level population data (left side: TR_CDC, TR_USALEEP, TR_TxDSHS), there is a noticeable upward trend in the mean attributable mortalities as roadway density increases. This increase is expected, as areas with higher roadway density typically experience higher levels of traffic-related pollution, leading to greater health impacts. The denser the roads are in the area, the higher the exposure to traffic-related PM_2.5, and consequently, the higher mortality attributable to traffic-related PM_2.5. In contrast, the Dasymetric population data (right side: DY_CDC, DY_USALEEP, DY_TxDSHS) show a much flatter trend across roadway density deciles. The mean number of attributable mortalities remain relatively consistent across all deciles, with little variation between areas of low and high roadway density. This flattening effect indicates that Dasymetric mapping, which provides a finer spatial distribution of the population in relation to roadways, reduces the impact of roadway density on the estimated attributable mortality as more people is living away from the roadways in the NCTCOG region. It suggests that, when accounting for where people live more accurately, the effect of roadway density on mortality becomes less pronounced and is context specific (i.e., it may become more pronounced in areas which are denser in population near roadways). This trend could be very different in case of another area where a larger percent of population lives near the roadways.

Discussion

The results of this study contribute to a growing body of literature on the health impacts of PM_2.5 exposure, focusing on traffic-related PM_2.5, and emphasizing the importance of spatial resolution in health risk assessment as one factor affecting attributable mortality estimates. Several key themes emerge when comparing our findings with previous studies.

Spatial Resolution of Mortality Rates and Attributable Mortality Estimation

This study shows that the spatial resolution of baseline mortality rate data plays a critical role in estimating the burden of mortality attributable to traffic-related PM_2.5. Our study estimated approximately 16% higher PM_2.5-attributable mortality when using tract-level mortality rates (both USALEEP and TxDSHS data) compared to county-level rates from CDC WONDER. These findings align with broader research, which shows that using coarser spatial data (e.g., county-level mortality rates) tends to result in lower attributable mortality estimates when compared to finer resolutions (Southerland et al., 2023; deSouza et al., 2023).

For example, Southerland et al. (2023) found that using Census Block Group (CBG) level mortality rates resulted in a 13% higher PM_2.5-attributable mortality compared to county-level rates. Similarly, deSouza et al. (2023) reported 8.4% higher PM_2.5-attributable mortality when using Census Tract-level mortality rates and 8.5% higher when using CBG-level mortality rates, both compared to county-level rates. It is important to note that these other studies used ambient PM_2.5 concentrations, which include all sources, as opposed to the traffic-related PM_2.5 used in the current study.

Using even finer resolution baseline mortality rates at the Census Block Group or Census Block level may result in different and potentially higher estimates of mortality attributable to PM_2.5—a viable avenue for future research.

Importance of Near Road Population

Another important finding from this study is that using Dasymetric population data resulted in lower exposure and hence lower traffic-related PM_2.5-attributable mortality compared to exposure estimations based on Census Tract-level population data. Our results align with those of Shandas et al. (2016), who adapted population data with varying spatial resolutions to assess air quality and evaluate exposure to NO₂ in the Portland metropolitan area. Their findings suggested considerable variations in exposure estimates near major highways when contrasting the cadastral-informed Dasymetric system to the standard uniform method of allocating populations. Similarly, in that study, exposure was higher when using Census Tract and Census Block Group population datasets compared to Dasymetric mapping (Shandas et al., 2016).

It is important to note that the trend—where Dasymetric data produces lower mortality estimates—applies specifically to the current study area, where the population tends to reside away from high-density roadway networks. If the study area were different—particularly if large segments of the population lived near major roadways with high PM_2.5 concentrations—these results might be reversed. In such a scenario, using high-resolution Dasymetric data could result in higher mortality estimates by more accurately capturing the exposure of populations living near high-pollution areas.

For example, in a study conducted in the city of Pamplona, Spain, NO₂ exposure estimates using population data at 100m resolution yielded 34% higher results than when using city-level population data (Santiago et al., 2022). The study found that a significant portion of the population lived closer to roadways, which was not captured when using city-level averages (Santiago et al., 2022). This underscores the importance of context when applying Dasymetric methods. The impact of population distribution on exposure can vary significantly between study areas, depending on the proximity of large population segments to major traffic sources. In areas where people live near busy roads, using fine-grained population data may reveal greater health risks than those estimated using coarser, tract-level data. Additionally, as the Dasymetric method is based on land use data and satellite imagery, improvements in algorithms for determining habitable areas could significantly affect exposure estimates.

Methodological Considerations and Limitations

One strength of this study lies in its methodological approach, particularly the use of varying spatial resolutions for population and mortality data. This allowed for a more nuanced understanding of how exposure and mortality estimates can vary depending on the granularity of the input data. However, it is important to acknowledge potential sources of exposure misclassification that may arise for several reasons.

First, movement patterns are typically not captured in traditional exposure assessments, as epidemiological studies generally focus on residential exposure to ambient air to estimate health risks. Similarly, in this study, we did not incorporate location-based-services mobility data, which has been gaining visiblity in recent research (Chen et al., 2024; Dėdelė et al., 2020; Han et al. 2025; Younkin et al., 2024) to overcome the limitation on assuming static residential locations for the population under study.

Future studies could benefit from incorporating mobility data to further refine exposure estimates. Including time-activity patterns and mobility data—capturing where people are, how they travel, and how long they spend in different environments—could enhance exposure assessments and provide a more dynamic and accurate representation of exposure. This integration would likely result in more precise estimates of exposure.

Another limitation of our study lies in the placement strategy of the receptors. Receptors were placed at intervals of 50m, 150m, and 300m from the midpoint of each roadway link on both sides, with additional gridded receptors every 500m across the entire study area. While we recognize that a uniform horizontal receptor placement along the roadway, rather than orthogonally around roadway centerlines, could have been beneficial—especially for the Inverse Distance Weighting (IDW) interpolation—resource constraints prevented us from implementing this computationally intensive adjustment. We acknowledge that this receptor placement method may have introduced artifacts in the IDW surface, such as “bubbles” around the placed receptors due to gaps in horizontal receptor placement. However, given the low population density near roadways in our study area, the impact on exposure assignment is expected to be minimal. Also, the study used IDW for interpolation due to its simplicity and suitability when using dense receptor data (Idir et al., 2021). While kriging could account for spatial correlation, it was not used due to high computational demands and given the high receptor density and considering that the methodological aim of this study, where the relative differences between the scenarios is more informative than the absolute values. Future studies could explore kriging to assess potential accuracy improvements.

The study uses 2019 traffic data, reflecting pre-COVID conditions, alongside 2020 population data. The COVID-19 pandemic significantly reduced traffic volumes in 2020, which may lead to overestimation of exposure levels if our findings are interpreted as representing 2020 conditions. As our focus is on methodological comparisons, the relative differences between scenarios are expected to hold and be informative, but absolute estimates should be contextualized with the data years used.

Another aspect to consider is the potential impact on ERFs at low PM_2.5 concentrations. The study’s TRAP-attributable PM_2.5 increments (1.08–1.26 μg/m³) coupled with the background concentrations in the region of 2.4 to 7.4 μg/m³ (Shapero et al., 2023) fall into the range reported by the primary studies underlying the Chen and Hoek (2020) meta-analysis. We applied a linear ERF from Chen and Hoek (2020), consistent with guidance from public health advisory bodies which recommends assuming a continued linear (no-threshold) relationship when quantifying mortality attributable to long-term PM_2.5 exposure (Committee on the Medical Effects of Air Pollutants, 2022), since current evidence does not robustly support a specific population-level threshold below which no adverse effects occur, nor is there a universally applicable non-linear CRF for PM_2.5-attributable mortality at these very low incremental exposure levels. Further, supra-linear effects at low exposure concentrations may underestimate risks (Health Effects Institute, 2021, 2023).

This study did not explicitly incorporate socioeconomic status (SES) or sociodemographic variables in the burden of disease estimation. While our analysis of varying mortality rate resolutions (e.g., tract-level USALEEP vs. county-level CDC) may indirectly capture some SES-related health outcome disparities, a dedicated analysis stratifying by SES indicators was beyond the scope of this study.

Lastly, we reiterate that the current study focuses on leveraging Dasymetric and USALEEP data. The same IDW surface was used throughout, ensuring a consistent exposure assessment method, which allowed us to focus on investigating the impact of different dataset combinations on the final outputs. Future studies could shed more light on how further refining the exposure assessment methods might impact the final results.

CONCLUSIONS

This study highlights the critical role that spatial resolution plays in the assessment of traffic-related PM_2.5 exposure and the attributable mortality impacts in a large metropolitan area. By comparing the attributable mortality estimates utilizing population at different resolutions in the Dallas-Fort Worth region, we demonstrate how the use of finer-scale Dasymetric population data produced lower estimates of traffic-related PM_2.5 attributable mortality when compared to using coarser population data across the same base mortality rates. This finding is particularly important in regions like Dallas-Fort Worth, where much of the population lives away from major roadways, thus reducing exposure to traffic-related air pollution. However, this trend may not hold universally—if a study area has a significant population living near major traffic corridors, the use of high-resolution Dasymetric data could lead to higher attributable mortality estimates, more accurately reflecting the localized exposure.

Our analysis also underscores the need to consider not only the spatial distribution of population but also the resolution of the baseline mortality rates data. Scenarios using county-level mortality data, tended to result in lower attributable traffic related PM_2.5 mortality when compared to using census-tract level mortality data.

Future studies should aim to incorporate dynamic population mobility data, which would account for movement patterns throughout the day, and improve the receptors’ placement strategy to cover larger areas, which could further improve the exposure assessment exercise. Improved algorithms for Dasymetric mapping and further refinement of baseline mortality data resolution—down to census block or block group levels—could further enhance the precision of these health impact assessments. Future research could integrate fine-grained SES and sociodemographic data with the high-resolution exposure and population data explored here to explicitly quantify disparities in traffic-related air pollution burden across different socioeconomic population subgroups. Moreover, emerging data sources, such as location-based services (LBS), offer promising avenues to refine exposure assessment by capturing dynamic mobility patterns.

Our findings can also have applications for urban planning and public health policy. The high-resolution exposure maps generated by dasymetric population data can be used to support land use planning or other interventions to mitigate exposures. We have leveraged advances in computing power to utilize high-resolution datasets (including an innovative application of dasymetric data) and large-scale modeling to provide insights into how traffic related BoD can be modeled to assess health benefits and support policy analyses.

In conclusion, this study contributes to the growing literature on air pollution exposure and health risk assessment in rapidly growing urban areas by demonstrating the importance of spatial resolution in both population and baseline mortality rate data. The findings suggest that policymakers and researchers must carefully select and apply population and mortality data resolutions that best capture the localized characteristics of their study areas to inform more effective air quality and public health interventions.

Table 4.

Descriptive Statistics of All-Cause Mortality Attributable to PM_2.5 Across the Different Data Combinations in the Study Area (% Differences were Calculated in Comparison to TR_CDC).

Data Combination	Mean [95%CI]	Median [95%CI]	Total [95%CI]	Mean (% Difference)	Median (% Difference)	Total (% Difference)
TR_CDC	0.228[0.173,0.255]	0.187 [0.142,0.209]	395 [300, 442]
TR_USALEEP	0.266[0.202,0.297]	0.197 [0.150,0.221]	460 [349, 514]	16.44%	5.70%	16.44%
TR_TxDSHS	0.265[0.201,0.296]	0.189 [0.143,0.211]	458[348,513]	16.06%	1.13%	16.06%
DY_CDC	0.213[0.161,0.238]	0.146[0.110,0.163]	368[279,411]	−6.89%	−22.03%	−6.89%
DY_USALEEP	0.238[0.181,0.266]	0.147[0.111,0.164]	412[313,461]	4.33%	−21.48%	4.33%
DY_TxDSHS	0.241[0.183,0.270]	0.149[0.113,0.167]	417[317,467]	5.67%	−19.97%	5.67%