Short-Term Associations of Road Density and Road Features with In-Vehicle PM 2.5 during Daily Trips in the Washington, D

: Increased daily exposure to fine particulate matter air pollution (PM 2.5 ) is associated with increased morbidity, yet high exposures over shorter timeframes (e.g., hourly) may also play a role. Transportation is a milieu for increased transient pollution exposures. Both the road traveled and nearby roadways (i.e., surrounding road density) may play a role in increased PM 2.5 exposure during commutes. For 2311 min of commutes, corresponding to 25 participants, we obtained in-vehicle PM 2.5 exposures using personal monitors and, through GPS data, road features, including road density and road type (e.g., highway vs. local roads). We considered the density of both the surrounding highways and the local roads at 500 m and 1000 m resolutions. We estimated associations of road features with minute-averaged in-vehicle PM 2.5 by applying linear mixed-effects models with random intercepts and autoregressive errors. The difference in log PM 2.5 , comparing the highest vs. lowest quartile of highway road density at 1 km resolution, was 0.09 log µ g/m 3 (95% confidence interval: 0, 0.19), which was similar to the difference between driving on highways vs. local roads (0.07 log µ g/m 3 (95% confidence interval: 0.00, 0.14)). Estimated differences were attenuated for local road density and 500 m resolution. The results were robust to adjustment for meteorology and ambient PM 2.5 . Unlike road features such as speed and road type, the surrounding road density is less modifiable during transportation. Therefore, road choice may not have a large impact on personal PM 2.5 exposures.


Introduction
Decades of research support the health effects associated with exposure to particulate matter air pollution less than 2.5 µm in aerodynamic diameter (PM 2.5 ) [1-6].Most epidemiologic studies of short-term clinical health associations have focused on daily or multiday PM 2.5 exposure, driven by the availability of pollution data at the daily timescale [1,2,7,8].However, associations have also been demonstrated between transient pollution exposures (e.g., hourly exposures) and sub-clinical health outcomes, including biomarkers of oxidative stress and airway inflammation [9][10][11].For workday exposures (~8 h), truck drivers, but not office workers, had decreased lung function following their shifts [12].Increased workday exposure to PM 2.5 was also associated with inflammation and cardiovascular markers among highway-patrol troopers [13] and increased inflammatory biomarkers among highway maintenance workers [14].Studies assessing even shorter exposures corresponding to ~2 h scripted roadway commutes [9,11,15,16], two hours of exposure in a transit hub [17], and two-hour walking exposure [18] have found associations with increased airway inflammation, decreased lung function, and changes in cardiovascular markers.These subclinical health outcomes are situated on the hypothesized biologic pathways by which air pollution leads to cardiorespiratory ED visits and hospitalizations [1].Therefore, modifiable factors associated with high transient exposures may serve as interventions to decrease both pollution exposures and subsequent clinical health impacts.
A daily activity associated with increased transient exposure to pollution is transportation, e.g., commuting to work, which includes travel via automobile [19][20][21][22][23]. Koehler et al. [22] reported that, compared to home levels, personal exposure to black carbon (BC), a traffic pollutant, was 129% (95% confidence interval (CI): 101,162%) higher during transit.Jung et al. [24] also found children had a higher peak exposure to BC during commutes compared to home environments.Higher exposure to pollution occurs in high-traffic settings and on highway or urban streets compared to local roads [25][26][27][28].Using scripted commutes in Birmingham, UK, in-vehicle PM 2.5 exposures were higher on urban roads compared to suburban roads [29].In our previous study of n = 46 women commuters in Washington, DC [30], rush-hour commuting was associated with increased in-vehicle PM 2.5 exposures relative to commuting outside of rush hour.Higher pollution exposures during peak travel times (i.e., rush hour) may be due, in part, to traveling via highway compared to other road types [20,29,31] or due to increased traffic [25,32].A recent study also suggested modifying trip routes (e.g., using lower-traffic roads) can lead to lower BC exposures [33], though they did not find an effect on PM 2.5 exposures.
These previous studies imply that understanding how specific road features impact transient in-vehicle PM 2.5 can guide route modifications that lower personal exposure to PM 2.5 .The impact of mobile sources on PM 2.5 concentrations may extend to 100-400 m from a roadway [34,35], though some studies did not find a spatial gradient for PM 2.5 [36,37].The density of the roads in the surrounding area, i.e., road density, is positively associated with increased traffic-related air pollutants in ambient air, including elemental carbon and nitrogen dioxide [38].Yet, road density may be less modifiable than other trip characteristics, such as speed and road type, because it is spatially coarser, meaning that it may not vary substantially between different routes to the same location.
Our study comprised 2311 min of observation from 69 separate trips taken by 25 participants in 2018-2019.The goals of this study were as follows.First, we identified road-location features, including road density, road type (e.g., highway, local road), and speed, for each observation in our study using GPS data recorded by in-vehicle on-board diagnostic (OBDII) data loggers.Second, we quantified minute-averaged in-vehicle PM 2.5 exposures during daily trips, capturing transient PM 2.5 over short time periods.Last, we estimated associations between road features and minute-averaged in-vehicle PM 2.5 using regression modeling.We focus on PM 2.5 because of its well-established associations with morbidity and mortality [1].Because in-vehicle exposures influence overall PM 2.5 exposures, this work helps to advance our understanding of whether vehicle trips can be modified to lower cumulative personal PM 2.5 exposures.

Study Design and Trip Data
Data on personal PM 2.5 exposures during vehicle trips were obtained from a study of commuters in 2018-2019 from September through May [30].Participants were recruited from Northern Virginia (Virginia suburbs of the Washington, DC, metro area) through flyers, online message boards, and word-of-mouth.Inclusion criteria comprised being a woman aged 18-55 who commutes by a personal vehicle at least 15 min to work at least three days a week and lives in Northern Virginia.Only women were recruited because the parent study was focused on exposures relevant to reproductive health and because of established differences in the commute behaviors between women and men [39][40][41].Each participant was given a personal air pollution monitor to collect personal exposure to PM 2.5 and a vehicle monitor to collect trip information for 48 h starting between 6 and 10 a.m. on a Monday, Tuesday, or Wednesday morning, excluding federal holidays, to capture weekday trips.All data collection started at George Mason University's Fairfax campus.Personal exposure to PM 2.5 was obtained for ten-second intervals using the MicroPEM™ (RTI International, Research Triangle Park, NC, USA) and then averaged to one-minute exposures to remove the impact of short-lived, extreme observations.These exposures were corrected by their filter concentrations, which assumes that temporal fluctuations in continuous data are proportional to the actual airborne concentration collected on the filter during that time period.Trip information, including time and GPS location, was obtained using OttoView-CVS44 data loggers by Persen Technologies Inc. (Winnipeg, MB, Canada) connected to the OBDII port within a participant's personal vehicle.Surveys were used to collect demographic information about the participants.Participants were also given trip diaries and instructed to record their commute times (for verification against the data loggers) and to note any driving activities that might influence air quality, including window use, smoking, or visible smog.The study design is also described elsewhere [30].The George Mason University Institutional Review Board approved this study.Some trips were missing GPS location data (e.g., the OBDII data loggers did not report location) and were removed from the analysis.Locations were restricted to those that had corresponding personal PM 2.5 exposures during that minute.We linked each GPS location with personal PM 2.5 exposure data to road features, including road density, road type, and speed.
Of primary interest in this study was road density, which is the density of roads in the surrounding area of each trip.Our previous study showed higher traffic-related air pollution concentrations in areas with higher road density (or roadiness) [38].As in our previous study [38], we computed road density in each grid cell i in our study area using the sum of the length of roads (RoadLength j ) as: where j = 1 . . .J referenced all Northern Virginia grid cells, and d i,j denoted the distance between grid cells i and j.For the road density in grid cell i, we assigned maximum weight to roads within that grid cell (d 2 i,i = 1).Separately, we considered road density using the length of highways, which has been associated with increased in-vehicle pollution exposures [27,28], and local roads, which represent the majority of the road links in the dataset and are consistent with our previous work [38].For the grid-cell size, we considered both 500 m and 1 km resolution.Thus, our analysis consisted of four measures of road density: highway road density at 1 km resolution; local at 1 km; highway at 500 m; local at 500 m.We represented road density as higher or lower than average by standardizing road density to have a mean zero and a standard deviation of one within our study area.
For each GPS location with personal PM 2.5 exposure data, its corresponding road type was assigned as that of the road segment in the United States Geological Survey Transportation Dataset [42].Road types were categorized as controlled-access highway, secondary highway, or major connecting road (Highway), tunnel or ramp (Ramp/Tunnel), local connecting road (LocalConn), which consists of roads that connect neighborhoods to highways, and local road (Local).Speed was obtained by computing the distance between GPS locations over time (miles per hour, mph).To match road features (multiple GPS locations per minute from the data logger) to the frequency of personal PM 2.5 exposures (each minute), the road type was summarized as the mode for each minute whereas road density and speed were summarized using the mean.

Ambient PM 2.5 and Meteorological Data
To adjust for high daily or hourly ambient PM 2.5 during commutes, we obtained ambient 24 h and hourly PM 2.5 from outdoor monitors in the Northern Virginia area for each day of observation.To obtain the 24 h ambient PM 2.5 average across our study area, we used three ambient PM 2.5 monitors in Northern Virginia.We developed a daily 24 h ambient average for each study day by averaging PM 2.5 concentrations across monitors.For days with missing 24 h concentrations, the average of the preceding day and the following day were used.The one near-road hourly ambient PM 2.5 monitor in Fairfax County, VA, was used to obtain exposures for each hour of the day to capture diurnal trends in ambient PM 2.5 driven by meteorological and/or anthropogenic sources.All study days had corresponding hourly ambient PM 2.5 observations.We utilized hourly deviations from daily PM 2.5 concentrations at the hourly monitor to account for the differences in ambient PM 2.5 between the location of the hourly monitor and the study participants and to reduce the correlation between the ambient and hourly PM 2.5 measurements.
To account for the possible impacts of meteorology on ambient air quality, we obtained temperature, precipitation, and humidity data.Meteorological data for each day of observation were obtained from the National Oceanic and Atmospheric Association's (NOAA) Global Historical Climatology Network using the R package rnoaa [43].The daily variables extracted included minimum temperature ( • C), maximum temperature ( • C), precipitation (mm), snowfall (mm), average wind speed (m/s), and the direction of the fastest wind lasting 5 min (degrees).Because of the large number of days with no precipitation, days were categorized as having some precipitation (vs.none) and some snow (vs.none).To account for high correlations between minimum and maximum temperature, the mean temperature ( • C) was computed as the mean of the daily minimum and maximum temperatures.The wind direction in the study area was categorized.Hourly relative humidity (%) was averaged to create daily summaries and was obtained as the mean of two monitors in Fairfax County, VA, through the Local Climatological Data website of the NOAA National Centers for Environmental Information [44].

Statistical Analysis
We categorized each trip minute as occurring during morning rush hour (6-10 a.m.), evening rush hour (3-7 p.m.), or neither, as in our previous work [30].The final dataset included minute-averaged personal PM 2.5 exposures and the corresponding road features (road density, speed, and road type), hourly and daily ambient PM 2.5 , indicators of rush hour, and daily meteorology.To limit our analysis to those trips with sufficient roadfeature and personal PM 2.5 data, we restricted our data to any trip with at least 15 min of consecutive location and personal PM 2.5 data, as we have in our previous work [30].This removed the impact of both short trips and instances where the car was started but not driven.The number of trips (overall and per participant) and length of trip time observed were reported.Boxplots were used to assess the distribution of ambient PM 2.5 and in-vehicle PM 2.5 for each participant across trips.Daily ambient PM 2.5 , hourly ambient PM 2.5 deviations, wind speed, relative humidity, and minimum, maximum, and mean daily temperature were summarized using means and standard deviations.We computed proportions of demographic characteristics of commuters, road type, rush-hour indicators, precipitation days, snowfall days, and days with each categorized wind direction.Pearson correlations between road-density measures (highway at 1 km; local at 1 km; highway at 500 m; and local at 500 m) and speed, as well as correlations of road features with ambient PM 2.5 and meteorology, were computed to identify potential multicollinearity.To account for nonlinearity in associations with in-vehicle PM 2.5 , we categorized road density and speed by quartiles.The relationship between road density and road type was explored using conditional proportions.
The primary goal of this analysis was to estimate associations between road location features (road density, road type, speed) and minute-averaged personal PM 2.5 exposures, capturing transient PM 2.5 over short time periods.Linear mixed-effects models were used to estimate associations between road features and in-vehicle PM 2.5 exposures.For the models, personal PM 2.5 exposures were natural log transformed after adding 0.05 µg/m 3 to address skewness and zero concentrations.Road-density quartiles, speed quartiles, and road type were included in the models individually as fixed effects.The temporal correlation was controlled for using autoregressive errors on the residuals.Random intercepts for trip within participant were also incorporated.The main models were unadjusted for additional covariates.We also considered the lagged effects of road features for up to ten minutes to determine whether their associated PM 2.5 increases take time to infiltrate the vehicle cabin.The models were fitted using restricted maximum likelihood estimation.Coefficients for fixed effects of the road features were reported along with a 95% CI.
To assess whether associations were attributable to other factors, we also fitted models adjusted for additional covariates separately to avoid multicollinearity and overfitting.These separate models included (1) meteorology, including numeric variables of mean temperature, average wind speed, and relative humidity, and categorical variables including any precipitation, any snowfall, and wind direction; (2) ambient PM 2.5 , including daily PM 2.5 and hourly deviations from ambient PM 2.5 (e.g., is ambient PM 2.5 at 9 a.m. higher or lower than the daily ambient average) to reduce collinearity between ambient pollution measures; (3) rush hour (as in [30]), defined as a categorical variable with morning rush hour (6-10 a.m.), evening rush hour (3-7 p.m.), or non-rush hour; and (4) all road features, including road-density quartiles, speed quartiles, and road type simultaneously.We did not include minimum or maximum temperature in the adjusted models because of their high correlation (r = 0.8); however, in sensitivity analyses, we replaced mean temperature separately with minimum or maximum temperature.All analyses were conducted using R 4.3 [45], leveraging the tidyverse package [46].

Results
Of the n = 46 participants in our original commuting study [30], 14 participants were excluded from the present study because the OBDII data logger did not connect to their personal vehicle (generally model years before 2005).Of the remaining n = 32 participants, our final sample of n = 25 had at least one trip (>15 min) consisting of both consecutive GPS and personal PM 2.5 data.On average, the participants (n = 25) were 26 years (SD = 8) and had at least some college education (n = 19, 76%), worked full-time (n = 17, 68%), were students (n = 15, 60%), did not have children (n = 18, 72%), were not Hispanic or Latino (n = 23, 92%), and were of Asian (n = 11, 44%) or white race (n = 9, 36%) (Table 1).All participants drove gasoline vehicles.The demographics of this study were generally similar to the full commuter study (n = 46), though participants in the present study were more likely to be Asian (44% vs. 34%) and less likely to be Hispanic or Latino (8% vs. 15%).
The number of minutes of total trip time observed per participant ranged from 24 to 267 min, with a median of 54 min.We observed one trip for 10 participants (40%).We observed 2-4 trips for nine participants (36%) and 5-6 trips for six participants (24%).The average participant in-vehicle PM 2.5 exposure across all trips was 5.5 µg/m 3 (SD = 3.1).Visualizing the distributions of in-vehicle PM 2.5 across participants, there was variability both between participants and within participants (Supplementary Material, Figure S3).We did not observe clear trends in in-vehicle PM 2.5 across each trip, displayed as standard deviations from trip-specific means (Supplementary Material, Figure S4).
Road density was highly correlated between measures at 1 km and 500 m for both highway (r = 0.97) and local road lengths (r = 0.99) (Table 2).Therefore, we present results using 1 km grid cells in the main text, which captures broader spatial impacts and present most of the results using 500 m in the Supplementary Material.Across the study area, road density based on the length of highways had greater spatial variation around roadways compared to road density based on local roads for both the 1 km and 500 m grid cells (Figure 1; Supplementary Material, Figure S5).For each measure (highway at 1 km; local at 1 km; highway at 500 m; local at 500 m), road density was standardized across trips to have mean zero and standard deviation of one.Road density varied both across trips and between trips (Figure 2; Supplementary Material, Figure S6) but was not correlated with ambient PM 2.5 , wind speed, temperature, or humidity (Supplementary Material, Table S2).For road type, most observations (n = 1215, 52.6%) occurred on local roads.Of the remaining observations, 26% (n = 600) were on highways or secondary highways, and 17% (n = 399) were on local connecting roads.The remainder (n = 97, 4.2%) were on ramps or in tunnels.Some participants spent nearly all of their time traveling on local roads, while others distributed time across road types (Supplementary Material, Figures S7 and S8).Although all participants spent some time on local roads, four participants (16%) did not travel on a highway, and 10 participants (40%) did not travel on a local connecting road.
Road type was associated with road-density quartiles for both highways and local roads; for example, for each increasing quartile of highway road density, there was an increased proportion of trips occurring on highways (Figure 3).Despite this association, 45% of trips in the highest quartile of local road density were on highways and 32% of trips in the highest quartile of highway road density were on local roads.This implies that highway road density captures trips occurring on local roads in the vicinity of highways and that local road density captures trips occurring on highways in the vicinity of local roads.The average speed across trips was 24 miles per hour (mph), and the speed ranged from 0 to 94 mph.Speed was not strongly correlated with either highway or local road density (Table 2).We did not observe any clear patterns of speed across each trip (Supplementary Material, Figure S8).For road type, most observations (n = 1215, 52.6%) occurred on local roads.Of the remaining observations, 26% (n = 600) were on highways or secondary highways, and 17% (n = 399) were on local connecting roads.The remainder (n = 97, 4.2%) were on ramps or in tunnels.Some participants spent nearly all of their time traveling on local roads, while others distributed time across road types (Supplementary Material, Figures S7 and S8).Although all participants spent some time on local roads, four participants (16%) did not travel on a highway, and 10 participants (40%) did not travel on a local connecting road.highway road density captures trips occurring on local roads in the vicinity of highways and that local road density captures trips occurring on highways in the vicinity of local roads.The average speed across trips was 24 miles per hour (mph), and the speed ranged from 0 to 94 mph.Speed was not strongly correlated with either highway or local road density (Table 2).We did not observe any clear patterns of speed across each trip (Supplementary Material, Figure S8).The quartiles of road density and speed are shown in the Supplementary Material, Table S3.Using linear mixed-effects models, in-vehicle PM2.5 exposures were, on average, higher for increasing quartiles of highway road density for 1 km grid cells (Figure 4A; Supplementary Material, Table S4), although the magnitudes of associations were generally modest.The difference in in-vehicle log PM2.5 comparing the top to the lowest quartile of highway road density was 0.09 log µg/m 3 (95% CI: 0.00, 0.19), or a 9% increase in PM2.5.Compared to 1 km, associations of in-vehicle PM2.5 with highway road density were attenuated for 500 m grid cells (Supplementary Material, Figure S9A).There was little evidence of associations between in-vehicle PM2.5 exposures and local road density (Figure 4B; Supplementary Material, Figure S9B).Our results did not change substantially between our main unadjusted models and additional models that separately adjusted for (1) meteorology, (2) ambient PM2.5, (3) rush-hour trips, and (4) all road features simultaneously (Figure 4; Supplementary Material, Figure S9).Because the mean temperature was highly correlated with both minimum temperature (r = 0.93) and maximum temperature (r = 0.96) (Supplementary Material, Table S2), the results from the models adjusting for minimum or maximum temperature were similar to the results for models adjusting for The quartiles of road density and speed are shown in the Supplementary Material, Table S3.Using linear mixed-effects models, in-vehicle PM 2.5 exposures were, on average, higher for increasing quartiles of highway road density for 1 km grid cells (Figure 4A; Supplementary Material, Table S4), although the magnitudes of associations were generally modest.The difference in in-vehicle log PM 2.5 comparing the top to the lowest quartile of highway road density was 0.09 log µg/m 3 (95% CI: 0.00, 0.19), or a 9% increase in PM 2.5 .Compared to 1 km, associations of in-vehicle PM 2.5 with highway road density were attenuated for 500 m grid cells (Supplementary Material, Figure S9A).There was little evidence of associations between in-vehicle PM 2.5 exposures and local road density (Figure 4B; Supplementary Material, Figure S9B).Our results did not change substantially between our main unadjusted models and additional models that separately adjusted for (1) meteorology, (2) ambient PM 2.5 , (3) rush-hour trips, and (4) all road features simultaneously (Figure 4; Supplementary Material, Figure S9).Because the mean temperature was highly correlated with both minimum temperature (r = 0.93) and maximum temperature (r = 0.96) (Supplementary Material, Table S2), the results from the models adjusting for minimum or maximum temperature were similar to the results for models adjusting for mean temperature.There was some indication of stronger associations of road density with in-vehicle PM 2.5 for lags ≥ 1 min for local road density and road density at 500 m (Figure 5B; Supplementary Material, Figure S10).
Environments 2024, 11, x FOR PEER REVIEW 10 of 17 Figure 5. Associations of road density with in-vehicle PM2.5 concentrations from linear mixed models for main models (lag 0) and lags up to 10 min, reported as the change in log µg/m 3 (95% confidence intervals) comparing each quartile of road density (Q2, Q3, Q4) to the lowest quartile (Q1) for (A) highways and (B) local roads at 1 km resolution.

Discussion
Using linear mixed models, small increases in in-vehicle PM2.5 were observed for increasing highway road density at 1 km (Figure 4A).The results indicate that highway road density is a different attribute than highway road type, as shown in the substantial proportion of trips (32%) occurring on local roads for the highest quartile of highway road density (Figure 3A).This indicates, further, that the impact of highways on in-vehicle PM2.5 exposures may extend beyond the road being traveled.We found the difference in invehicle PM2.5 between traveling on highways vs. local roads was comparable to the differ-

Discussion
Using linear mixed models, small increases in in-vehicle PM 2.5 were observed for increasing highway road density at 1 km (Figure 4A).The results indicate that highway road density is a different attribute than highway road type, as shown in the substantial proportion of trips (32%) occurring on local roads for the highest quartile of highway road density (Figure 3A).This indicates, further, that the impact of highways on in-vehicle PM 2.5 exposures may extend beyond the road being traveled.We found the difference in in-vehicle PM 2.5 between traveling on highways vs. local roads was comparable to the difference between the top vs. lowest quartile of highway road density (0.07 log µg/m 3 (95% CI: 0.00, 0.14) vs. 0.09 log µg/m 3 (95% CI: 0.00, 0.19)).The results were similar in models controlling for meteorology, ambient PM 2.5 , rush-hour trips, and other road features.
Our previous work demonstrated that an increased local road density was associated with an increased annual average ambient pollution, including elemental carbon, a trafficrelated component of PM 2.5 [38].However, in this present study, we did not find that increased local road density was associated with increased short-term in-vehicle PM 2.5 .From Figure 2B, local road density varied slowly across trips, with greater variation between trips.It is possible that local road density did not vary sufficiently to identify an association with in-vehicle PM 2.5 in our study.The association between in-vehicle PM 2.5 and highway road density was attenuated at finer resolutions (500 m grid cells) compared to 1 km grid cells (Supplementary Material, Figure S9A).The 500 m resolution may be too fine to sufficiently capture the impact of nearby roadways (which are frequently sited more than 500 m away) on in-vehicle pollution.
Similar to our findings, the length of highway roads within a 500 m buffer was associated with increased in-vehicle PM 2.5 [47].Road length was also associated with increased personal exposure to PM 2.5 [48].We used road lengths to define road density in our study of in-vehicle PM 2.5 ; road lengths are also associated with ambient traffic pollution [49][50][51][52].A previous study accounted for differences in traffic intensity in roadlength measures using an approach to standardize road length [53].Road-density measures have also been extended to incorporate other relevant road features, such as traffic counts, that were unavailable in the present study [54].
In our study, in-vehicle PM 2.5 was higher for highways compared to local roads, similar to previous studies [27].A previous scripted commute study found higher PM 2.5 on the New Jersey turnpike compared to local roads [28].A study of highway-patrol officers identified higher in-vehicle PM 2.5 near major roadways [55].Further, Dons et al. [25] examined BC, a marker of traffic PM, and found that 5 min concentrations were higher for highways compared to local roads.There were strong associations of in-vehicle pollution with the number of diesel vehicles [47], a factor we were unable to examine in our study.A study in Beijing, China, identified diesel vehicles as contributing to increased on-road BC concentrations [56].Our study did not find travel speed to be associated with in-vehicle PM 2.5 concentrations.A previous study found slightly higher BC concentrations at both low and high speeds [25].Speed and proximity to highways may be related to one another [47] because of increased speed limits on highways; however, in our dataset, road density and speed were not positively correlated, possibly due to the spatial resolution of our road-density measures.
Previous studies have compared pollution exposures across modes of transportation, including personal vehicle use, walking, biking, and public transit [20,23,33,57,58].Our study focused on pollution exposures during personal vehicle trips, which are the most common commute mode in the US [59] and, therefore, represent a possible target of interventions to reduce pollution exposures.A previous study of direct (i.e., arterial and larger collector roads) vs. alternate (including collector roads) car routes did not find differences in mean PM 2.5 exposures [33].Our present study might help explain these results; if the alternate routes selected are in the vicinity of larger roads, the impact of road choice might be limited.In our previous study of n = 46 commuters, we found that rush-hour trips were associated with greater in-vehicle PM 2.5 exposures [30].This finding could have been due, in part, to participants traveling on or near highways more frequently during rush hour.In this present study, controlling for rush hour trips did not substantially change the associations between road density and in-vehicle PM 2.5 (Figure 4).
We utilized in-vehicle monitoring of both air pollution concentrations and trip information using OBDII data loggers, which have the advantage of limiting GPS information to when the car is being operated.However, there were several limitations to our approach.The OBDII data loggers were not compatible with older vehicles (<2005), and therefore not all participants in our original study (n = 46) could be included in the present study.To confirm our findings, future research should determine associations between road features and in-vehicle PM 2.5 using a larger number of participants.Additionally, we found that the OBDII GPS connectivity was occasionally disrupted, leading to some missing data.While our study was an observational study of all vehicle trips taken over 48 h, this approach did not allow us to control for the types of trips taken (e.g., commuting to work, child care).We did not instruct participants on how to utilize their windows or in-vehicle ventilation, but we asked them to report window use on their trip diaries.Most participants in our full study (n = 23, 92%) did not report traveling with their windows down.Ventilation decisions, for example, window or air conditioning use, can impact in-vehicle pollution concentrations [16,21,29,[60][61][62][63][64].
We were unable to control for traffic in this analysis, which has been associated with increased pollution exposures [27].However, traffic and road features are difficult to disentangle because they are highly correlated [25].It is important to note that road features may impact PM 2.5 exposures through increased traffic.Our study only collected continuous measurements of PM 2.5 , which is highly relevant to human health but is not traffic specific.PM 2.5 is emitted by multiple natural and anthropogenic sources and represents only one size distribution of PM.MicroPEMs can only collect either PM 2.5 or PM 10 size distributions.We selected PM 2.5 , as it is most relevant for human health.Studies of road density and other traffic-related pollutants (e.g., PM 1 , BC, nitrogen dioxide) are needed.
Additionally, this study was unable to examine sources of traffic-related air pollution because continuous speciation data were unavailable.Using integrated, 48 h PM 2.5 data, our previous study found that major sources of PM 2.5 in this population were road salt, anti-icing/de-icing, secondary pollution, and a mixed mobile source representing tailpipe emissions and brake/tire wear, which were difficult to separate using integrated data [65].MicroPEMs do not report uncertainties for continuous data including PM 2.5 ; therefore, we have likely underestimated the uncertainty for our associations with road features.Future work should determine whether these results are consistent with more recent vehicle fleets, including the increased adoption of electric vehicles.However, the approaches we have utilized in this study could be applied to study new vehicle-fleet mixes.
Our study also had several strengths, including the ability to automatically determine when participants were traveling by personal vehicle, without relying on trip diaries, and our ability to match personal PM 2.5 exposures to GPS locations on roadways, since our GPS data only consisted of roadway locations.Previous studies have utilized personal GPS monitoring [19,55,66] or activity trackers [57,67] to classify time-activity patterns in studies of air pollution.Our study examined all real-world trips taken by participants during a 48 h period, which increases the breadth of trips included, such as trips to and from work, as well as household management trips (e.g., childcare, errands).Last, the Washington, DC, metropolitan study area encompasses a wide range of road types; our trip locations included those near the Capital Beltway as well as suburban local roads.

Conclusions
This study estimated associations of road density, along with road type and speed, with minute-averaged in-vehicle PM 2.5 using real-world personal vehicle trips in the Washington, DC, metropolitan area.We utilized GPS data collected from vehicle OBDII data loggers to characterize road features, personal monitors to determine in-vehicle PM 2.5 exposures, and regression modeling to assess associations between road features and PM 2.5 .Our study found that most commutes took place on local roads.In-vehicle PM 2.5 exposures differed both between individuals and within individuals over time, and invehicle PM 2.5 was greater for the highest vs. lowest quartile of highway road density, with an estimated difference similar to the difference between traveling on highways vs. local roads.These findings align with previous studies that have concluded that PM 2.5 exposures are higher on highways compared to local roads.This study was limited in its ability to examine the impact of traffic, in-vehicle ventilation, or vehicle type on in-vehicle PM 2.5 exposures.Additionally, our study included only observational data collected from 25 participants.However, the strengths of our study include examining road-density impacts on PM 2.5 among real-world commutes using GPS to objectively measure travel information.Because personal vehicle use in the US allows for individual choices in route planning and timing, our work contributes to how individuals can modify their trip routes to decrease their personal PM 2.5 exposures.Our work implies that traveling by a smaller, local road may not result in lower in-vehicle PM 2.5 if the local road is in the vicinity of a highway.To confirm our findings, future work to examine road density should utilize scripted commutes that can directly compare high and low road-density commutes without relying on observational data.

Supplementary Materials:
The following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/environments11070135/s1:S1: Meteorological characteristics (n (%)) for n = 45 person-days of observation; Table S2: Correlations of road features with ambient PM 2.5 and daily meteorology; Table S3: Quartiles (Q1, Q2, Q3, Q4) for road density and speed; Table S4: Unadjusted differences in in-vehicle PM 2.5 (log µg/m 3 ) comparing quartiles of road density for highways and local roads at 1 km and 500 m resolution.The supplementary material contains supplemental tables and figures that provide further support for the results discussed in this article, including greater exploration of the demographic, meteorological, ambient PM 2.5 , in-vehicle PM 2.5 , road density at 500 m, road type, and speed data.
analysis, and interpretation of data; in the writing of the report; or in the decision to submit the article for publication.
Institutional Review Board Statement: Participants provided informed consent prior to study participation.The George Mason University Institutional Review Board approved this study.

Environments 2024 , 17 Figure 1 .
Figure 1.Map of standardized road density in the Washington, DC, metropolitan area for (A) highways and (B) local roads at 1 km resolution.

Figure 2 .
Figure 2. Variation in road density for (A) highways and (B) local roads at 1 km resolution (in standard deviations (SD)) and over each observed trip (n = 69).

Figure 1 .Figure 1 .
Figure 1.Map of standardized road density in the Washington, DC, metropolitan area for (A) highways and (B) local roads at 1 km resolution.

Figure 2 .Figure 2 .
Figure 2. Variation in road density for (A) highways and (B) local roads at 1 km resolution (in ard deviations (SD)) and over each observed trip (n = 69).For road type, most observations (n = 1215, 52.6%) occurred on local roads.remaining observations, 26% (n = 600) were on highways or secondary highways, an (n = 399) were on local connecting roads.The remainder (n = 97, 4.2%) were on ram in tunnels.Some participants spent nearly all of their time traveling on local roads others distributed time across road types (Supplementary Material, Figures S7 an Although all participants spent some time on local roads, four participants (16%) d travel on a highway, and 10 participants (40%) did not travel on a local connecting

Figure 3 .
Figure 3. Proportion of road type by quartiles of road density (for highways at 1 km (A), highways at 500 m (B), local roads at 1 km (C), and local roads at 500 m (D) resolution).

Figure 3 .
Figure 3. Proportion of road type by quartiles of road density (for highways at 1 km (A), highways at 500 m (B), local roads at 1 km (C), and local roads at 500 m (D) resolution).

Figure 4 .
Figure 4. Associations of road density with in-vehicle PM2.5 concentrations from linear mixed models reported as the change in log µg/m 3 (95% confidence intervals) comparing each quartile of road density (Q2, Q3, Q4) to the lowest quartile (Q1) for (A) highways and (B) local roads at 1 km resolution.Results include main (unadjusted) models as well as models that separately adjusted for meteorology, ambient PM2.5, rush-hour trips, and all road features.

Figure 4 .
Figure 4. Associations of road density with in-vehicle PM 2.5 concentrations from linear mixed models reported as the change in log µg/m 3 (95% confidence intervals) comparing each quartile of road density (Q2, Q3, Q4) to the lowest quartile (Q1) for (A) highways and (B) local roads at 1 km resolution.Results include main (unadjusted) models as well as models that separately adjusted for meteorology, ambient PM 2.5 , rush-hour trips, and all road features.

Figure 5 .
Figure5.Associations of road density with in-vehicle PM 2.5 concentrations from linear mixed models for main models (lag 0) and lags up to 10 min, reported as the change in log µg/m 3 (95% confidence intervals) comparing each quartile of road density (Q2, Q3, Q4) to the lowest quartile (Q1) for (A) highways and (B) local roads at 1 km resolution.
Figure S1: Distributions of daily ambient PM 2.5 (µg/m 3 ) and hourly PM 2.5 (µg/m 3 ) for n = 45 person-days of observation; Figure S2: Wind rose plot showing direction of the fastest wind (degrees) sustained over five minutes; Figure S3: Distributions of in-vehicle PM 2.5 exposures (µg/m 3 ) for each participant (n = 25); Figure S4: Variation in PM 2.5 (in standard deviations (SD) from trip-specific means) over each observed trip (n = 69); Figure S5: Map of standardized road density in the Washington, DC metropolitan area; Figure S6: Variation in road density for A. highways and B. local roads at 500 m resolution; Figure S7: Percentage of observations on each road type per participant (n = 25); Figure S8: Variation in A. road type and B. speed (mph) over each observed trip (n = 69); Figure S9: Associations of road density with in-vehicle PM 2.5 concentrations for A. highways and B. local roads at 500 m resolution; Figure S10: Associations of road density with in-vehicle PM 2.5 concentrations for A. highways and B. local roads at 500 m resolution; Figure S11: Associations of road type with in-vehicle PM 2.5 concentrations comparing travelling on highways, local connecting roads, and ramps/tunnels to local roads; Figure S12: Associations of road type with in-vehicle PM 2.5 for main models (lag 0) and lags up to 10 min; Figure S13: Associations of speed with in-vehicle PM 2.5 concentrations; Figure S14: Associations of speed with in-vehicle PM 2.5 concentrations for main models (lag 0) and lags up to 10 min; Table

Table 2 .
Correlations between road features, including standardized road density (for highways and local roads at 1 km and 500 m resolution) and speed (MPH).