Effects of Road Traffic on the Accuracy and Bias of Low-Cost Particulate Matter Sensor Measurements in Houston, Texas

Although PM2.5 measurements of low-cost particulate matter sensors (LCPMS) generally show moderate and strong correlations with those from research-grade air monitors, the data quality of LCPMS has not been fully assessed in urban environments with different road traffic conditions. We examined the linear relationships between PM2.5 measurements taken by an LCPMS (Dylos DC1700) and two research grade monitors, a personal environmental monitor (PEM) and the GRIMM 11R, in three different urban environments, and compared the accuracy (slope) and bias of these environments. PM2.5 measurements were carried out at three locations in Houston, Texas (Clinton Drive largely with diesel trucks, US-59 mostly with gasoline vehicles, and a residential home with no major sources of traffic emissions nearby). The slopes of the regressions of the PEM on Dylos and Grimm measurements varied by location (e.g., PEM/Dylos slope at Clinton Drive = 0.98 (R2 = 0.77), at US-59 = 0.63 (R2 = 0.42), and at the residence = 0.29 (R2 = 0.31)). Although the regression slopes and coefficients differed across the three urban environments, the mean percent bias was not significantly different. Using the correct slope for LCPMS measurements is key for accurately estimating ambient PM2.5 mass in urban environments.


Introduction
Exposure to fine particulate matter (PM 2.5 ) has been associated with several adverse health outcomes [1][2][3]. These health outcomes include increased hospitalizations, morbidity, and mortality from respiratory effects (infections, exacerbations of asthma, and chronic obstructive pulmonary disease) [4][5][6][7][8], and cardiovascular effects (ischemic heart disease, stroke) [4][5][6]8,9]. Due to the severity of these adverse health outcomes associated with PM 2.5 exposures, monitoring PM 2.5 for exposure assessment within communities has become a subject of interest among environmental scientists, governmental organizations, and interested private citizens. Recently, it has been suggested that the application of low-cost air sensors could improve the understanding of current indoor and ambient exposures to air pollutants, including PM 2.5 [10][11][12][13]. Low-cost air sensors are portable and lightweight [11,13]. Furthermore, the direct reading capabilities of these sensors allow for instantaneous readings to obtain fine temporal resolution of the data that can be used for the evaluation of short-term exposure to air pollutants. For example, in the city of Houston (land area: 1651 km 2 ), Texas (TX), only seven continuous air monitoring stations (CAMS) The Dylos DC1700 air quality monitor (Dylos) (Dylos Corporation, Riverside, CA, USA) is a relatively inexpensive (<USD 500) and portable air monitoring sensor. The instrument uses a light scattering method to measure the number of particles in two size bins: >0.5 µm and >2.5 µm. The Dylos was selected due to the availability of research data from previous studies for comparison with this study.
The Grimm Mini Laser Aerosol Spectrometer 11R (Grimm) (GRIMM 11R) (Grimm Aerosol Technik GmbH & Co. KG, Ainring, Germany) was selected to compare with the Dylos measurements. The GRIMM 11R counts particles in the size range from 0.25 µm to 32 µm using a light scattering method. PM 2.5 mass concentrations were obtained using Grimm Software 1178 V8-1 Rev IV. The cost of the Grimm 11R (USD 20,000) was about 40 times greater than the Dylos DC1700.
The Personal Environmental Monitor (PEM) Impactor (SKC Inc., Eighty-Four, PA, USA) was also used to compare with the Dylos data. The 37 mm polytetrafluoroethylene (PTFE) filter was placed in the PEM to collect PM 2.5 . The SKC Leland Legacy pump (SKC Inc., Eighty-Four, PA, USA) was used to pull air at 10 L/min. Unlike the Dylos and Grimm, PEM is a gravimetric method to measure time-integrated PM 2.5 mass. In this study, the sampling time was 3 h at each location. The total cost of the PEM and the pump was about USD 1900.
The Thermo Orion Cahn-35 Microbalance (Thermo Electron Corporation, Beverly, MA, USA) was used to weigh the filters from the PEM. The accuracy of the gravimetric analysis as 0.0012% with a weight range and sensitivity of 250 mg/1.0.
The HOBO Data Logger U12-012 (Onset Computer Corporation, Bourne, MA, USA) was used to monitor ambient air temperature. The HOBO measures temperatures are in the range from −20 to 70 • C. The accuracy of HOBO temperature measurements is documented by the manufacturers as ±0.35 • C from 0 to 50 • C.

Sampling Locations
To determine the effect of traffic volumes on the linear relationship between PM 2.5 measurements taken by the Dylos DC1700 (Dylos) and other research-grade instruments (PEM and GRIMM 11R), we simultaneously deployed a Dylos, a PEM, a Grimm 11R, and a HOBO at each sampling location ( Figure 1a).

1.
Clinton Drive, Houston, Texas (major road with a higher percentage of traffic emissions from heavy-duty diesel vehicles (HDDV)). Clinton Drive is located in the eastern part of the Houston metropolis in Texas ( Figure 1b). In this study, the traffic on this road was made up of a higher number of HDDV (28%), emitting diesel particles. All samplers were deployed near the fence line of the Clinton CAMS, about twenty meters from Clinton Drive.

2.
US-59 Highway, Houston, Texas (major highway with traffic emissions mainly from gasoline vehicles (GV)). The US-59 highway in Houston, Texas runs from southwest to northeast in Houston ( Figure 1c). The proportion of HDDVs among total traffic counts was 3% in this study. All samplers were deployed about 50 m from the road on a side street (Eastside Street).

3.
Residential home (residential location with no major sources of PM), which was located in a suburban area of Houston, Texas ( Figure 1d). There were no significant sources of PM near the sample location. The closest major roadway from the residence was about 6400 m away, and there were no factories or industrial facilities close to the residence. All samplers were deployed in the backyard of the residence.
The Personal Environmental Monitor (PEM) Impactor (SKC Inc., Eighty-Four, PA, USA) was also used to compare with the Dylos data. The 37 mm polytetrafluoroethylene (PTFE) filter was placed in the PEM to collect PM2.5. The SKC Leland Legacy pump (SKC Inc., Eighty-Four, PA, USA) was used to pull air at 10 L/min. Unlike the Dylos and Grimm, PEM is a gravimetric method to measure time-integrated PM2.5 mass. In this study, the sampling time was 3 h at each location. The total cost of the PEM and the pump was about USD 1900.
The Thermo Orion Cahn-35 Microbalance (Thermo Electron Corporation, Beverly, MA, USA) was used to weigh the filters from the PEM. The accuracy of the gravimetric analysis as 0.0012% with a weight range and sensitivity of 250 mg/1.0.
The HOBO Data Logger U12-012 (Onset Computer Corporation, Bourne, MA, USA) was used to monitor ambient air temperature. The HOBO measures temperatures are in the range from −20 to 70 °C. The accuracy of HOBO temperature measurements is documented by the manufacturers as ±0.35 °C from 0 to 50 °C.

Sampling Locations
To determine the effect of traffic volumes on the linear relationship between PM2.5 measurements taken by the Dylos DC1700 (Dylos) and other research-grade instruments (PEM and GRIMM 11R), we simultaneously deployed a Dylos, a PEM, a Grimm 11R, and a HOBO at each sampling location ( Figure 1a).
1. Clinton Drive, Houston, Texas (major road with a higher percentage of traffic emissions from heavy-duty diesel vehicles (HDDV)). Clinton Drive is located in the eastern part of the Houston metropolis in Texas ( Figure 1b). In this study, the traffic on this road was made up of a higher number of HDDV (28%), emitting diesel particles. All samplers were deployed near the fence line of the Clinton CAMS, about twenty meters from Clinton Drive. 2. US-59 Highway, Houston, Texas (major highway with traffic emissions mainly from gasoline vehicles (GV)). The US-59 highway in Houston, Texas runs from southwest to northeast in Houston ( Figure 1c). The proportion of HDDVs among total traffic counts was 3% in this study. All samplers were deployed about 50 m from the road on a side street (Eastside Street). 3. Residential home (residential location with no major sources of PM), which was located in a suburban area of Houston, Texas ( Figure 1d). There were no significant sources of PM near the sample location. The closest major roadway from the residence was about 6400 m away, and there were no factories or industrial facilities close to the residence. All samplers were deployed in the backyard of the residence. (a)

Study Design
One Dylos DC1700 (Dylos) was collocated with one PM2.5 PEM, one GRIMM 11-R (Grimm), and a HOBO data logger (HOBO) at all three locations. The Dylos, PEM, and Grimm measured ambient PM2.5 for 3 h over 20 days at each location. Sampling was conducted at one location each day, alternating between the three sampling locations. Samples were collected from October 2019 through January 2020. The HOBO was also deployed to record temperature during the 3 h sampling periods. The HOBO, Dylos, and Grimm were programmed to record data at 1 min intervals. Samples were collected in the mornings (8 a.m.-1 p.m.) on weekdays (Monday to Friday) and weekends (Saturday and Sunday). Samples were not collected when weather conditions were not conducive for sampling (e.g., on rainy days). A total of 53 valid sampling campaigns, 40 sample days on weekdays, and 13 sample days on weekends were obtained.
To estimate the ratio of HDDV to gasoline vehicles on the two major roads (US-59 and Clinton Drive), the number of each type of vehicle passing was counted using video recording during the sampling periods. Vehicles passing by were counted for a 10 min period in every hour. The total number of vehicles counted per hour was then multiplied by 6. Prior to this study, the Dylos, Grimm, and HOBO were previously calibrated by the manufacturer. The PTFE filters were pre-weighed using the Cahn-35 microbalance (Thermo Electron Corporation, Beverly, MA, USA) in a weighing room at the University of Texas Health Science Center at the Houston (UTHealth) School of Public Health. The batteries for the SKC Leland pumps (SKC Inc., Eighty-Four, PA, USA), the Dylos DC1700, and the Grimm 11R were fully charged before the deployment. On each day of sampling, the Leland legacy pump was calibrated at 10 L/min using the BIOS Dry-Cal DC-lite (BIOS International Corp., Butler, NJ, USA) at the sampling locations. The clocks on the Dylos, Grimm, and HOBO were all synchronized with the internet clock on a laptop. The information on the sampling date, time, location, sample ID, pump ID, filters' pre-weight, and pump pre-flow rates was recorded. The instruments were placed side by side on a portable folding table, which was about 1 m above the ground level ( Figure 2). All the sampling instruments were turned on simultaneously, and the start time was recorded. After three hours, all the instruments were turned off, and the stop time was recorded. After sampling, the post-flow rate for PEM was measured using the Dry Cal DC-Lite. All the

Study Design
One Dylos DC1700 (Dylos) was collocated with one PM 2.5 PEM, one GRIMM 11-R (Grimm), and a HOBO data logger (HOBO) at all three locations. The Dylos, PEM, and Grimm measured ambient PM 2.5 for 3 h over 20 days at each location. Sampling was conducted at one location each day, alternating between the three sampling locations. Samples were collected from October 2019 through January 2020. The HOBO was also deployed to record temperature during the 3 h sampling periods. The HOBO, Dylos, and Grimm were programmed to record data at 1 min intervals. Samples were collected in the mornings (8 a.m.-1 p.m.) on weekdays (Monday to Friday) and weekends (Saturday and Sunday). Samples were not collected when weather conditions were not conducive for sampling (e.g., on rainy days). A total of 53 valid sampling campaigns, 40 sample days on weekdays, and 13 sample days on weekends were obtained.
To estimate the ratio of HDDV to gasoline vehicles on the two major roads (US-59 and Clinton Drive), the number of each type of vehicle passing was counted using video recording during the sampling periods. Vehicles passing by were counted for a 10 min period in every hour. The total number of vehicles counted per hour was then multiplied by 6. Prior to this study, the Dylos, Grimm, and HOBO were previously calibrated by the manufacturer. The PTFE filters were pre-weighed using the Cahn-35 microbalance (Thermo Electron Corporation, Beverly, MA, USA) in a weighing room at the University of Texas Health Science Center at the Houston (UTHealth) School of Public Health. The batteries for the SKC Leland pumps (SKC Inc., Eighty-Four, PA, USA), the Dylos DC1700, and the Grimm 11R were fully charged before the deployment. On each day of sampling, the Leland legacy pump was calibrated at 10 L/min using the BIOS Dry-Cal DC-lite (BIOS International Corp., Butler, NJ, USA) at the sampling locations. The clocks on the Dylos, Grimm, and HOBO were all synchronized with the internet clock on a laptop. The information on the sampling date, time, location, sample ID, pump ID, filters' pre-weight, and pump pre-flow rates was recorded. The instruments were placed side by side on a portable folding table, which was about 1 m above the ground level ( Figure 2). All the sampling instruments were turned on simultaneously, and the start time was recorded. After three hours, all the instruments were turned off, and the stop time was recorded. After sampling, the post-flow rate for PEM was measured using the Dry Cal DC-Lite. All the collected filters were transferred to the UTHealth School of Public Health for gravimetric analysis. The post-weight and the post-flow rate were recorded. The data logged on the Dylos, the Grimm 11R, and the HOBO data logger were downloaded onto a laptop computer.
collected filters were transferred to the UTHealth School of Public Health for gravimetric analysis. The post-weight and the post-flow rate were recorded. The data logged on the Dylos, the Grimm 11R, and the HOBO data logger were downloaded onto a laptop computer.

Data Analysis
To obtain PM2.5 particle number concentrations from the Dylos, the number counts for particles > 2.5 µm were subtracted from the number counts for particles > 0.5 µm. The 3 h mean particle count concentrations for the Dylos and 3 h mean mass concentration for the Grimm for each sampling day were calculated using 1 min raw data from both instruments. To obtain the 3 h integrated PEM PM2.5 mass concentration for each sampling day, the total mass (post weight-pre weight) of PM2.5 on each filter was divided by the total volume of air (average flow rate × sampling duration, approximately 3 h) on the same sample day. The mean of the 3 h temperature was also calculated from the 1 min interval data from the HOBO. All the data were analyzed using STATA15 (StataCorp LLC., College Station, TX, USA).
A descriptive analysis was performed using 3 h mean PM2.5 concentrations and ambient temperature across three sampling locations. Simple linear regression analysis was conducted using overall data and location-specific data, separately. The effect of sampling location (as a proxy of different traffic emission sources) on the linear relationship between the Dylos DC1700 and PEM measurements was assessed with the linear regression model from Equation (1) below. Before the regression analysis was carried out, a natural log transformation of the data was carried out to fit the data close to a Gaussian distribution.
where Y = natural log of the 3 h PM2.5 mass concentration measured by the PEM or Grimm; X1 = natural log of the 3 h PM0.5-2.5 particle number concentration measured by the Dylos; X2 = binary dummy variable coded as 1 for US-59 and zero (0) for the other two locations (Clinton Drive and the residence); and

Data Analysis
To obtain PM 2.5 particle number concentrations from the Dylos, the number counts for particles > 2.5 µm were subtracted from the number counts for particles > 0.5 µm. The 3 h mean particle count concentrations for the Dylos and 3 h mean mass concentration for the Grimm for each sampling day were calculated using 1 min raw data from both instruments. To obtain the 3 h integrated PEM PM 2.5 mass concentration for each sampling day, the total mass (post weight-pre weight) of PM 2.5 on each filter was divided by the total volume of air (average flow rate × sampling duration, approximately 3 h) on the same sample day. The mean of the 3 h temperature was also calculated from the 1 min interval data from the HOBO. All the data were analyzed using STATA15 (StataCorp LLC., College Station, TX, USA).
A descriptive analysis was performed using 3 h mean PM 2.5 concentrations and ambient temperature across three sampling locations. Simple linear regression analysis was conducted using overall data and location-specific data, separately. The effect of sampling location (as a proxy of different traffic emission sources) on the linear relationship between the Dylos DC1700 and PEM measurements was assessed with the linear regression model from Equation (1) below. Before the regression analysis was carried out, a natural log transformation of the data was carried out to fit the data close to a Gaussian distribution.
where Y = natural log of the 3 h PM 2.5 mass concentration measured by the PEM or Grimm; X 1 = natural log of the 3 h PM 0.5-2.5 particle number concentration measured by the Dylos; X 2 = binary dummy variable coded as 1 for US-59 and zero (0) for the other two locations (Clinton Drive and the residence); and X 3 = binary dummy variable coded as 1 for the residence and zero (0) for all other locations (Clinton Drive and US-59).
In Equation (1), the Clinton Drive location, coded as zero for both X 2 and X 3 , is the reference sampling location to which all the other locations (US-59 and Residence) were compared.
Two interaction terms (X 1 × X 2 and X 1 × X 3 ) were introduced into a linear regression to test for the difference in slopes across sampling locations. The null hypothesis for the model that the difference between the slope at Clinton Drive and the slope at US-59 (represented by the β 4 coefficient) equal to zero was tested. Similarly, a null hypothesis was tested whereby the difference between the slope at Clinton Drive and the slope at the residence (represented by the β 5 coefficient) was equal to zero.
To test for the difference between US-59 and the residence, the coding for the variables in the regression analysis from Equation (1) was slightly modified. The residence location was recoded as zero for both X 2 and X 3 (the reference sampling location). The X 3 was recoded as (1) for Clinton and (0) for all other locations, while the X 2 variable remained coded as (1) for US-59 and (0) for all other locations. The X 1 variable remained the same. The β 4 coefficient for this modified model was tested for the difference between the slopes of US-59 and the residence.
Further analysis was carried out to determine the effect of other covariates such as temperature and the ratio of heavy-duty trucks to gasoline cars (HDDV%) on the linear relationship between the Dylos and the each of the research grade instruments.
The agreement (or bias) between the Dylos DC1700 measurements and the researchgrade instruments across sampling locations was assessed by calculating the mean absolute relative error. To estimate the absolute relative error, the Dylos number concentration measurements were first converted to mass concentration measurements using fitted regression line equations from a linear regression of (i) the PEM on the Dylos measurements and (ii) the Grimm on the Dylos measurements. This conversion was carried out using two methods.

1.
General equation (GE) method: A single fitted regression line equation from the linear regression of the total combined data was obtained and used to convert the Dylos PM 2.5 measurements from all three locations.

2.
Sampling location equation (SLE) method: A different regression line was constructed, stratified by each sampling location. Three fitted regression equations, one for each sampling location (Clinton, US-59, and the residence), were used to convert the Dylos PM 2.5 measurements.
After obtaining the converted Dylos mass concentrations, the absolute relative error was calculated for each day of sampling with Equation (2) below. The mean absolute relative error (MARE) was then estimated by calculating the average of all the absolute relative errors for all the sampling days. The data were then grouped by the sampling location, and the MARE for each location was calculated. With the equal variance structure, an analysis of variance (ANOVA) test with Bonferroni correction was used to compare the difference of the MAREs between the Dylos-Grimm at the three locations. For the Dylos-PEM comparison, because the equal variance between groups assumption was unmet, two-sample t-tests were used to compare the difference between the MAREs across the three locations.
where Dylos PM 2.5 = converted PM 2.5 mass concentrations from the 3 h mean Dylos count measurements collected over a single sample duration; and PEM or Grimm PM 2.5 = 3 h integrated PM 2.5 mass concentration collected by the PEM or Grimm over a single sample duration.

Statistical Summary
A total of 53 valid sampling days was obtained from the Dylos DC1700 after excluding seven outliers in this study. Outliers were identified as data points with residuals greater than two standard deviations. The mean and standard deviation (SD) of the particle count for the Dylos was 1439 ± 1053 particles/0.01 ft 3 (Range: 158-4394 particles/0.01 ft 3 ). The means and SDs of the PEM and the Grimm 11-R were 24.4 ± 24.4 µg/m 3 (Range: 5.1-137.8 µg/m 3 ) and 13.7 ± 10.7 µg/m 3 (Range: 1.9-47.6 µg/m 3 ), respectively. The mean ambient air temperature during the entire study period was 25.0 ± 6.5 • C (range: 10.9-37.3 • C).

Effect of Road Traffic as a Proxy of PM 2.5 Emission Source on the Linear Relationship between Dylos and Research Grade Instruments (PEM and Grimm 11R)
Dylos vs. PEM. The slope for the overall combined log-transformed data was 0.70 (R 2 = 0.52). The slope comparing the Dylos to the PEM measurements was 0.98 (R 2 = 0.77), 0.63 (R 2 = 0.42), and 0.29 (R 2 = 0.31) at the Clinton Drive, US-59, and the residence, respectively ( Figure 3).

Effect of Temperature and Truck Ratio (HDDV%) on the Linear Relationship between Dylos and Research Grade Instruments (PEM and Grimm 11R)
For the PEM, the coefficient for temperature (0.023) in the final model was significant, indicating an increase of 1.26 µg/m 3 in PEM mass concentration measurements with a 10 • C rise in temperature. The effect of truck ratio was not significant on the relationship between Dylos and PEM. A similar finding was obtained with the Grimm and Dylos. The coefficient for temperature (0.030) indicated an increase of 1.35 µg/m 3 in Grimm mass concentration measurements with a 10 • C rise in temperature. The slopes at three sampling locations and the difference in slopes by location between without temperature (Model 1) and with temperature (Model 2) are summarized in Table 2. For PEM, ambient temperature increased regression slope for US-59 (0.63 to 0.82), whereas it did not change the slopes for other locations. Similarly, we found that the regression slope for Grimm was incresed from 0.73 (Model 1) to 0.84 (Model 2) for US-59, while they were not significantly changed between Model 1 and Model 2 for other locations.

Agreement between the Dylos and the Reference Grade Instruments (PEM and Grimm 11R)
The bias between the Dylos and the PEM PM 2.5 measurements was about two times greater than the bias between the Dylos and the Grimm (Table 3). Using the GE method, for instance, the MARE between the Dylos and PEM measurements for the combined data was 42 ± 35%, whereas the MARE between the Dylos and Grimm measurements was 22 ± 16%. Table 3 also summarizes the MARE using the SLE method. For Dylos to PEM PM 2.5 measurements, the MAREs were greater at Clinton (37%) and US-59 (37%) than at the residence (27%). However, the MAREs across sampling locations were not statistically different from each other (p-values = 0.89 (Clinton vs. US-59), 0.22 (Clinton vs. the residence), and 0.39 (US-59 vs. the residence). The MAREs using the SLE method between the Dylos and Grimm at all locations were similar (Clinton: 14%, US-59: 19%, and the residence: 19%), and the p values for all comparisons were 1.0.  Figure 5 shows temporal trends between the instruments on weekdays (Monday through Friday) and weekends (Saturday and Sunday). The 3 h mean PM 2.5 concentrations from all three instruments show similar temporal trends across different sampling days. The peak 3 h mean PM 2.5 concentrations were much larger on weekdays compared to weekends. The MARE (estimated using the SLE) between the Dylos and PEM PM 2.5 was similar on weekdays (32 ± 32%) and weekends (36 ± 30%). The MARE between the Dylos and the Grimm was also similar on weekdays (18 ± 14%) and weekends (15 ± 12%). weekends. The MARE (estimated using the SLE) between the Dylos and PEM PM2.5 was similar on weekdays (32 ± 32%) and weekends (36 ± 30%). The MARE between the Dylos and the Grimm was also similar on weekdays (18 ± 14%) and weekends (15 ± 12%).

Discussion
The effects of different types of road traffic on the linear relationship between the Dylos DC1700 and the two research-grade instruments (PEM and Grimm 11R) were assessed at three specific locations with varying traffic types. The agreement (or bias) between the Dylos and the research-grade instruments was also examined. Environments with different traffic characteristics significantly changed the regression slopes between the Dylos DC1700 and the other instruments. However, the agreement (or bias) between the instruments was not significantly affected by traffic conditions. The 3 h integrated PM 2.5 mass concentration measured by the PEM at the Clinton Drive location (HDDV% = 28 ± 11%) was at least two times higher than those at the US-59 highway (HDDV% = 3.3 ± 0.7%) and the residence (no HDDV). Although both the Clinton Drive and the US-59 locations were in close proximity to busy roadways, one possible explanation for the higher concentration of PM 2.5 at Clinton Drive than the US-59 may be associated with more resuspension of PM 2.5 at Clinton Drive. Askariyeh et al. (2020) observed that PM 2.5 previously deposited on roadways was heavily resuspended due to the movement of heavy-duty vehicles (i.e., trucks) over the road. Estimated traffic-related PM 2.5 increased up to 208% on roadways, as resuspended particles were included in the model [28]. A larger proportion of heavy vehicles on a road is also a factor increasing the rate of resuspended particles [28]. Hence, the elevated PM 2.5 concentration at the Clinton Drive may be explained by the higher percentage of heavy-duty trucks in this study.
The regression of the log-transformed PEM over the log-transformed Dylos PM 2.5 concentrations for the overall combined data showed positive and moderate to strong correlation (r = 0.72). Previous studies showed a similar correlation between a DRI and gravimetric samplers (Kim et al., 2004: r = 0.68 and Zhu et al., 2011: r = 0.63) [29,30]. We also found that the correlation (r = 0.95) between the Dylos and the Grimm 11-R was stronger than that between the Dylos and the PEM. The higher correlation between the Dylos and Grimm is probably due to the similar light scattering method used by both samplers as opposed to the gravimetric method used by the PEM. Previous studies observed a strong correlation between an LCPMS and other direct reading instruments (Han et [19,23]. Slopes from the linear regression equations obtained from collocated measurements of the Dylos and research-grade instruments are often used to estimate correction factors in the conversion of Dylos particle count concentration (particles count/0.01 ft 3 ) to mass concentration (µg/m 3 ). Converting the Dylos counts to mass concentration enables equivalent comparison between the Dylos measurements, other research-grade instruments, and the national air quality standards. The slope of simple linear regression of the log transformed PEM and Dylos data for the overall data was 0.70, whereas the slope of the log transformed Grimm11R and Dylos was 0.91. Using the linear regression equations in this study, the estimated Dylos PM 2.5 for a particle count of 10,000 particles/0.01 ft 3 , for instance, was 82.9 µg/m 3 using the PEM model and 76.8 µg/m 3 using the Grimm 11R model. The results of this study overestimated PM 2.5 approximately 20-30 percent compared to the PM 2.5 for a 10,000 particle/0.01 ft 3 in previous studies: 62.1 µg/m 3 [22] or 67.8 µg/m 3 [23].
Additionally, the slope of the Dylos-PEM varied by location. The discrepancies in regression slopes may be explained by the heterogeneous characteristics of aerosols at the different sampling locations. The PM 2.5 at Clinton Drive was composed of a relatively higher percent of diesel particle emissions (HDDV = 28%) compared to US-59 (HDDV = 3%) and the residential location (no HDDV). The PM 2.5 at the residence consisted of particles from suburban background sources with no significant traffic sources nearby. Given the different slopes by locations, sampling location specific conversion factors (SLE slopes) can be helpful in improving the accuracy of the estimated PM 2.5 mass concentration from the Dylos PM 2.5 count measurements (and other LCPMS) at locations where the characteristics of PM 2.5 differ.
The effects of temperature and truck ratio on the linear relationship between the Dylos and the other research grade instruments were examined using the regression model (Model 2). The effect of truck ratio (ratio of heavy-duty trucks to regular gas vehicles) was not significant. Although temperature was a significant factor in both the Dylos-PEM and Dylos-Grimm linear relationship, the effect size of temperature was minimal compared to the location variable. The inclusion of temperature in the regression model modified less than 10 percent of the overall slope and the slopes at each specific sampling location. Only one exception was the difference between the US-59 and residence slopes which became statistically significant for the Dylos-PEM relationship.
The mean percentage bias as MARE between the Dylos and the PEM (42%) with the combined data in this study was similar to a previous study (40%) [29]. Further analysis of the MARE by sampling location showed that the mean percent bias did not significantly differ by sampling location. The results suggest that, in urban environments similar to Houston, TX, the bias (or precision) was consistent between the converted Dylos PM 2.5 mass and the PM 2.5 measured from the research-grade instruments.
The limitations of this study include the ambient PM 2.5 at the sampling locations not being fully characterized for their composition aerosols. Chemical characterization of the PM 2.5 at the sampling locations would help show how chemical constituents in PM 2.5 affect the relationship between the LCPMS and research-grade instruments. Further, a longer sampling period (i.e., >30 days) at each location would be preferable to the 18 days of sampling in this study, as this would have provided a larger sample size to make stronger statistical conclusions. This knowledge would enable better conclusions about the results of the study.

Conclusions
The Dylos had a moderate to strong correlation with both the PEM and the Grimm. The slopes of the linear regression comparing the Dylos to both the PEM and the Grimm varied by about 30 percent across sampling locations. This finding suggests that calibration at each specific sampling location with different traffic characteristics is key to determine the PM 2.5 mass concentrations using an LCPMS. On average, the bias between the converted Dylos measurements and the PEM measurements was about two times higher than the bias between the Dylos and the Grimm measurements. The biases between the Dylos and both research-grade instruments remained similar across different sampling locations and PM concentration levels. Overall, the LCPMS, after proper calibration, could provide supplementary PM 2.5 data to evaluate air pollution in urban environments with different traffic characteristics.
Author Contributions: T.O. designed the study, collected, and analyzed the data, wrote manuscript. L.W., E.S., C.B. and A.C. reviewed the data and revised the manuscript. I.H. designed the study, analyzed, and reviewed the data, and revised the manuscript. All authors have read and agreed to the published version of the manuscript.
Funding: This research was partially funded by the National Institute of Environmental Health Sciences (NIEHS) (Award #R21ES031795 and #P30ES030285) and the Centers for Disease Control and Prevention/National Institute of Occupational Safety and Health (CDC/NIOSH) (Award #T42OH008421). The content is solely the responsibility of the authors and does not necessarily represent the official views of NIEHS or CDC/NIOSH. The APC was funded by I.H.

Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.