Impact of Chemical Speciation Network method changes on time series ion and carbon species concentrations

Abstract

Numerous studies rely on long-term PM_2.5 speciation data from the EPA’s Chemical Speciation Network (CSN), for example, to estimate health impacts or investigate the sources and transport of PM_2.5 pollution. These studies rely on consistent, long-term time series measurements of PM_2.5 species to draw conclusions about PM_2.5 emissions sources and their health impacts. However, changes in contractors and associated methodological changes in 2015 and 2018 led to disruptions in the consistency of the CSN data, specifically, concentration discontinuities in the CSN time series for ions and elemental carbon (EC) from November 2015 to September 2018 and from October 2018 onward, respectively. To address the impact of these changes on downstream air quality and health analyses, this study developed correction factors by comparing collocated CSN measurements to measurements from the Interagency Monitoring of Protected Visual Environment (IMPROVE) network, which used consistent instrumentation and contractors throughout the study period. These correction factors reduced the discontinuities in the ions and EC concentration time series data, which could be critical for time series source apportionment receptor modeling, air pollution policy and accountability investigations, and health effect studies.

Graphical Abstract

1. Introduction

In the United States, monitoring of fine particulate matter (PM_2.5) composition is conducted by two main networks: the Chemical Speciation Network (CSN) (EPA, 2023), overseen by the EPA, and the Interagency Monitoring of Protected Visual Environment (IMPROVE) network, managed through a steering committee comprising representatives from various agencies and organizations (IMPROVE 2024; Solomon et al. 2014; Malm et al. 1994). CSN, established in 1997, and IMPROVE, established in 1985, collect 24-hour filters for analysis of organic and elemental carbon, major ions, and various elements (Solomon et al. 2014), with different sampling schedules (every 3^rd or 6^th day for CSN and every 3^rd day for IMPROVE). These networks have similarities and differences, discussed in detail by Solomon et al. (2014) and Gorham et al. (2021) and presented briefly in Table S1 and Table S2 in the Supplemental Information.

The US EPA has used the combined data from these networks to develop effective emission control strategies for fine particles. Additionally, research studies rely on these publicly available measurements to understand the contribution of various sources to ambient PM_2.5 levels. To understand the impact of short and long-range pollutant transport, to determine the effectiveness of regulations, to study the effect of factors like economic changes, mitigation measures, changes in source signatures, and changing chemistry on air quality (Farahani et al. 2021; Park et al. 2020; Masiol et al. 2019; Hasheminassab et al. 2014) and to estimate health impacts of PM_2.5 constituents (Bi et al. 2023; Hao et al. 2016, 2023; Huang et al. 2019; Pennington et al. 2019; Krall et al. 2013, 2017a; Thurston et al. 2016; Xiao et al. 2016; Dominici et al. 2015; Peng et al. 2009; Sarnat et al. 2008) and sources (Krall et al. 2017a, 2017b; Hao et al. 2016). These studies all rely on consistent long-term time series concentrations of PM_2.5 species.

The measurement and analytical techniques in long-term networks like IMPROVE and CSN evolve, and these evolutions can introduce artificial shifts in the data, making it a challenge to maintain consistency over time. These changes in analytical methods are often necessary, despite the challenges they pose in maintaining consistent sampling methods, calibration, laboratory procedures, and equipment updates (Zhang et al. 2021). For instance, past CSN changes in sampler and analytical methods in 2007-2009, and contractors during the early 2000s had a significant impact on total PM and PM species concentrations (Malm et al. 2011). These shifts in the data affect the results of studies that rely on these data, including studies focusing on source apportionment, air quality trends, and health effects. Unlike CSN, IMPROVE contractors for the analysis of ions, elements, and carbon have remained consistent over time. There are several methodological changes implemented by IMPROVE over the years (some listed in Solomon et al. (2014)). For most changes, IMPROVE provided a detailed analysis of the changes made (presented under gray literature on their website), such as the transition from the DRI 2001 to the DRI 2015 thermal optical reflectance instruments, and supplied corrected data for user utilization.

More recently in 2015 and 2018 (Figure 1), the CSN network changed contractors for ion, elements, and carbon analysis (Hopke et al. 2023; Zhang et al. 2021), while IMPROVE did not change their contractors. Regarding CSN ions, Research Triangle Institute (RTI) handled the analysis before Nov 20, 2015, and from Oct 2018 onwards, and between these two periods, Desert Research Institute (DRI) handled the analysis (Young 2021). Also, starting 1 Jan 2016, the ions were blank corrected using median values of the field blanks corresponding to the sampling month, further discussed in the Methods section. For carbon analysis, from 2007-2009 (coincided with the carbon sampler transition period, i.e., transition to URG (University Research Glassware) 3000N between 2007-2009) through Oct 2018, DRI handled the carbon analysis by implementing IMPROVE-A protocol on the DRI-2001 instrument model, and later switched to DRI 2015 model starting 1 Jan 2016. However, in Oct 2018, the contract laboratory changed from DRI to the University of California (UC) Davis, with DRI using DRI-2001 and DRI-2015 model analyzers and UC Davis using Sunset laboratory model 5L analyzer for EC (elemental carbon) and OC (organic carbon) analysis (Figure 1) (Young 2021). Before Nov 2015, the contractor performing the analysis of elements through energy-dispersive X-ray fluorescence (XRF) changed from RTI to UC-Davis. Both contractors employed different techniques, leading to elevated detection limits for heavy metals during the UC-Davis period. The XRF protocol was consequently modified starting October 2018 to lower these detection limits.

Figure 1:

Timeline showing changes in the CSN contractor/instrumentation for carbon and ions analysis (more detail in Table S3). For ions, RTI handled the analysis before Nov 20, 2015, and from Oct 2018 onwards. Between these two periods, DRI handled the analysis. The ion comparison included three time periods: RTI-pre (1 Dec 2012 to 19 Nov 2015), DRI (20 Nov 2015 to 30 September 2018), and RTI-post (1 Oct 2018 to end date (2/24/2021 – 3/29/2021, depending on the site). The end date represents the end date of the downloaded data for this study. The RTI-pre period began in Dec 2012 to provide a consistent duration for both the DRI and RTI-post intervals. For carbon, this study included two time periods: 1 Jan 2015 to 30 September 2018 (DRI-2015) and 1 Oct 2018 to end date (2/24/2021 – 3/29/2021, depending on the site) (Sunset). For carbon, the cross-hatched area indicates the sampler transition period (i.e., 2007-2009), and the hatched area indicates when DRI used model DRI 2001 (until 31 Dec 2015). This study did not focus on the hatched or the cross-hatched time periods.

The 2018 contractor change impacted carbon analysis, as reported in CSN advisories and other studies (Hopke et al. 2023; Zhang et al. 2021). Although both DRI-2015 and Sunset 5L follow the IMPROVE-A protocol (Chow et al. 2007). Zhang et al. (2021) analyzed filters from 2007 and 2017 using three carbon instruments, including DRI-2015 and Sunset 5L, reporting an inter-instrument mean bias of around 5% for TC (total carbon) and OC (higher OC for DRI-2015) and 12% for EC (higher EC for Sunset). They also identified a mean bias ranging from 5% to 140% for thermal subfractions of OC and EC. Additionally, they highlighted issues related to optical measurement saturation due to strong-absorbing material (e.g., EC), leading to a substantial inter-instrument EC difference (mean bias 21.7% +/− 12.2%) for samples with high EC content (about 5% of samples). However, their study did not include filters collected during winter months. Regions of the western US experience winter temperature inversions and accompanying elevated EC levels (Hand et al. 2013), making these instrument differences potentially noteworthy. Zhang et al. (2021) suggested conducting receptor modeling separately for the periods before and after October 2018. Hopke et al. (2023) reported on impacts from the EC/OC instrument change on EC and OC subfractions, resulting in significant changes to diesel and gasoline source contributions using positive matrix factorization (PMF). The OC and EC fractions are critical to resolving contributions from gasoline, wood-burning, diesel, and secondary SOAs (Hopke et al. 2023; Kotchenruther 2021; Squizzato et al. 2018; Kelly et al. 2013). Hopke et al. (2023) suggested reevaluation of the samples ollected after October using the original and temporally consistent instrument designed to implement the IMPROVE_A protocol as the only solution to enable long-term studies.

A challenge with temporal interruptions including contractor changes is they can impact any downstream analysis that relies on short-term variation, including source apportionment and epidemiologic analysis. For example, in time series studies of short-term pollutant health effects, the estimated health effect relies on the variation in pollutant concentrations as well as the temporal correlation between pollutant concentrations and the health outcome, both of which may be impacted by a contractor change. Epidemiologic models could directly estimate different health effects for different contractor periods, but this approach has several disadvantages. First, such a model would inherently assume a different true health effect for each contractor period, though a contractor change in pollution measurement would not impact the toxicity of pollution. Second, epidemiologic time series models generally contain multiple confounding variables that may have different impacts on health depending on time, complicating the model. Last, even if separate health effects are estimated during different contractor periods, it is unclear how to weigh them to provide an overall estimate of the pollutant health effect, the target of the analysis.

For receptor modeling, prior studies addressing contractor-related changes in carbon concentrations proposed either separate models for each contractor (Zhang et al. 2021) or filter reanalysis (Hopke et al. 2023). Yet, filter reanalysis remains impractical for most CSN data users who lack access or resources for it. Filter reanalysis would be particularly infeasible (due to cost and access) for long-term, multi-city studies, like those used to relate particle composition to health impacts. While the use of OC and EC fractions for the period after Oct 2018 is infeasible for source apportionment studies (Hopke et al. 2023), using total OC and EC fractions could be informative, although it may combine certain sources, such as gasoline and diesel. This study evaluates the impact of contractor changes on total OC and EC and devises correction measures using collocated IMPROVE and CSN measurements. This approach mirrors previous efforts (e.g., Malm et al. (2011)), who utilized collocated IMPROVE and CSN monitors to address changes in EC/OC samplers and analysis protocols.

Although the effect of contractor changes on carbon and carbon fractions has been studied, the effect on ion concentrations has not been studied. This study further aims to evaluate the effect of this contractor change on ion concentrations and to develop a correction method to address the associated measurement discontinuities. This correction method also relies on collocated CSN/IMPROVE measurements. As for elements, this study did not attempt to develop a correction for element concentrations because of existing biases between IMPROVE and CSN that are likely associated with the difference in the samplers (flow rates, flow control methods, and cyclones), sampling, and analytical protocols (Gorham et al. 2021).

Given the potential impact of contractor changes on species concentration, species contributions, and downstream analyses, including source apportionment and epidemiologic studies, it is critical to develop correction methods to minimize the effect of contractor changes in the time series species concentrations. Therefore, this technical note aims to: 1) evaluate the effect of contractor changes on carbon and ion concentrations; and 2) develop factors to correct the CSN measurements by using collocated IMPROVE and CSN measurements.

2. Methods

2.1. Data collection

This study considered sites with collocated CSN and IMPROVE monitors and with data availability for the years 2015 and later. These sites (CSN site code, abbreviation code, state) included Fresno (06-019-0011; FRES, CA), Phoenix (04-013-9997; PHOE, AZ), Puget Sound (53-033-0080, PUSO, WA), North Birmingham (01-073-0023, BIRM, AL), Detroit (26-163-0001; DETR, MI), Atlanta (13-089-0002; ATLA, GA), and Pittsburgh (42-003-0008; PITT, PA) (Table 1, and Figure S1, online Supplemental Information). Table 1 also lists the IMPROVE site code. We accessed CSN and IMPROVE measurements for these sites from the Federal Land Manager Environmental Database (FED 2023). Carbon measurements for both networks were available at all seven sites. Ion measurements were available for all CSN sites but only for four IMPROVE sites (FRES, PUSO, PHOE, and BIRM) for 2012 and onwards (Table S4). Table S4 summarizes the raw downloaded data in terms of date and number of measurement days. Both networks collect 24-hour integrated PM_2.5 filters, which are then analyzed by contract labs for total PM_2.5 mass (CSN network stopped reporting the total PM_2.5 mass starting Nov 2014), elemental composition by energy dispersive X-ray fluorescence, organic and elemental carbon (OC, EC) by thermal evolution in eight temperature fractions (OC1, OC2, OC3, OC4, OP, EC1, EC2, EC3), and anions/cations by ion chromatography.

Table 1:

Study-specific summary of sites, time periods, and measurements for nitrate (NO₃) ion (AQS code: 88306), sulfate ion (SO₄) (AQS code: 88403), ammonium ion (NH₄) (AQS code: 88301, only CSN), OC (AQS code: 88320 IMPROVE only), EC (AQS code: 88321 IMPROVE only). For the CSN network, the OC and EC were calculated using the sum of thermal fractions (OC = OC1 + OC2 + OC3 + OC4 + OP and EC = EC1 + EC2 + EC3 − OP). The time interval for the OC and EC ranged from 1 Jan 2016 until the listed end date (end date of the downloaded data). The time interval for NH₄, NO₃, and SO₄ ranged from 1 Dec 2012 (specific start date differs depending on species and site) until the listed study end date. The number of days for each species represents days when both CSN and IMPROVE measurements are available and measurements were greater than 3*MDL. For NH₄, the number of days represents days when CSN reported NH₄ was greater than 3*MDL, and CSN reported NO₃ and SO₄ were also available.

Site (abbreviation)	Improve site code	CSN site code	Species	# of days	Start date	End date
Fresno (FRES)	FRES1	060190011	OC	530	1/1/2016	2/15/2021
			EC	579	1/1/2016	2/15/2021
			NO3	891	12/5/2012	2/24/2021
			SO4	880	12/5/2012	2/24/2021
			NH4	910	12/5/2012	3/29/2021
Phoenix (PHOE)	PHOE1	040139997	OC	501	1/1/2016	2/27/2021
			EC	575	1/1/2016	2/27/2021
			NO3	904	12/2/2012	2/27/2021
			SO4	918	12/2/2012	2/27/2021
			NH4	838	12/2/2012	3/29/2021
Seattle (PUSO)	PUSO1	530330080	OC	417	4/12/2016	2/24/2021
			EC	504	4/12/2016	2/27/2021
			NO3	788	12/2/2012	2/27/2021
			SO4	786	12/2/2012	2/27/2021
			NH4	561	12/11/2012	3/26/2021
North Birmingham (BIRM)	BIRM1	010730023	OC	436	1/1/2016	3/26/2021
			EC	494	1/1/2016	3/29/2021
			NO3	769	12/5/2012	3/29/2021
			SO4	826	12/5/2012	3/29/2021
			NH4	873	12/2/2012	3/29/2021
Atlanta (ATLA)	ATLA1	130890002	OC	493	1/1/2016	3/26/2021
			EC	574	1/1/2016	3/29/2021
			NH4	785	12/2/2012	3/29/2021
Detroit (DETR)	DETR1	261630001	OC	450	1/1/2016	3/23/2021
			EC	580	1/1/2016	3/29/2021
			NH4	937	12/2/2012	3/29/2021
Pittsburgh (PITT)	PITT1	420030008	OC	476	1/1/2016	3/26/2021
			EC	578	1/1/2016	3/29/2021
			NH4	840	12/2/2012	3/29/2021

This study utilized the following species: nitrate (NO₃) ion (AQS code: 88306), sulfate ion (SO₄) (AQS code: 88403), ammonium ion (NH₄) (AQS code: 88301), OC (AQS code: 88320), and EC (AQS code: 88321), OC1 (AQS code: 88324), OC2 (AQS code: 88325), OC3 (AQS code: 88326), OC4 (AQS code: 88327), OP (AQS code: 88328), EC1 (AQS code: 88329), EC2 (AQS code: 88330), and EC3 (AQS code: 88331). IMPROVE network calculates the total OC and EC by summing up fractions (OC = OC1 + OC2 + OC3 + OC4 + OP and EC = EC1 + EC2 + EC3 − OP) after adjusting each fraction for blanks. CSN determines the total OC blank and EC blank, which are subsequently employed to blank-correct the total OC and total EC measurements. To be consistent with the IMPROVE network, the total OC and EC for the CSN network were calculated from the sum of fractions (reported after blank correction starting 20 Nov 2015). The difference between the OC and EC calculated from fractions and CSN reported total OC and EC, respectively, were rather small (less than 5%, Figure S2, online Supplemental Information). For simplicity, from here onward, the OC and EC calculated from the fractions for CSN would be referred to as CSN OC and EC.

For NO₃, SO₄, OC, and EC, the measurements from the CSN were compared with the IMPROVE measurements using scaled residual difference. The IMPROVE network does not report ammonium (NH₄) ion concentrations. Therefore, the NH₄ concentrations were corrected solely based on the CSN data (details in Section Comparisons between the different time intervals). The potassium ion (K+, AQS code: 88303) and sodium ions (Na+, AQS code: 88302) were reported by CSN but not by the IMPROVE network, and therefore were not considered. For chloride ions (AQS code: 88203), the CSN reported the measurements starting Feb 2017, but only 13% of the reported values were greater than 3*MDL, therefore, chloride was not considered.

2.2. Blank correction, minimum detection limit, and excluded measurements

The OC and EC fraction measurements from the CSN (starting 1 Jan 2016) and IMPROVE network were field blank corrected by their respective laboratories. The ions reported by CSN were not blank corrected before 1 Jan 2016 but were field-blank corrected for the remaining period. This study did not blank-correct ions for the period before Jan 2016, because the blanks available before Jan 2016 were filter-lot background blanks and not the field-blanks. Also, the CSN advisory (Young 2021) reported a minimal effect of field blank correction on the NH₄, NO₃, and SO₄, i.e., less than 5% difference between the raw and blank-corrected concentrations.

The calculation of the minimum detection limit (MDL) for the CSN network also changed starting on 20 Nov 2015 (Young 2021). Before November 20, 2015, MDLs were calculated using three times the standard deviation of replicate measurements of reagent, matrix, or calibration standards, characterizing detection limits solely in the laboratory. Between November 2015 and February 2017, MDLs for ion and carbon species were calculated from monthly field blanks (three times the standard deviation of field blanks using 50 field blanks for each filter type collected both during and closest to the sampling month). Starting in March 2017, the MDLs were calculated using the percentile calculation method (95^th percentile – median) on 50 field blanks collected during or closest to the sampling month. This study uses the MDLs to remove the carbon and ion measurements below 3*MDL. The CSN-reported MDLs were used as reported, regardless of the different methods of MDL calculation. This resulted in a lower percentage of measurements above 3*MDL for OC and NH₄. Between 2015-11-23 and 2017-01-31, the MDL values of OC were highest (~0.17 μg m⁻³ as compared to ~0.07 μg m⁻³ for the remaining period, (Young 2021) resulting in a lower percentage of OC > 3*MDL (77.3 %, 79.4 %, 81.9 %, 82.9%, 85.3%, 86.3%, and 90.1 for DETR, PUSO, BIRM, PITT, PHOE, ATLA, and FRES, respectively, Table S4). This study did not develop factors for OC (discussed in Results), so the relatively low percentage of measurements should not affect the results in this study. The MDL of NH₄ was high during the period between 2015-01-01 and 2015-11-22 (~0.027 μg m⁻³ as compared to 0.005 μg m⁻³ during the remaining period). A much lower percentage of concentrations greater than 3*MDL was observed for NH₄ at sites that do not experience inversions (61.6 %, 84.8, and 86.4% for PUSO, PHOE, and ATLA, respectively), and low NH₄ concentrations were expected for these sites. Our approach assumes that the correction factors for the lowest concentration bin are the same regardless of whether the concentrations are above or below 3*MDL. This approach allows the concentrations >3*MDL to guide the associations, thus reducing the impact of low, imprecise, concentrations.

2.3. Time intervals

Table 1 and Table S4 provide the time intervals and species evaluated in this study. For OC and EC, the comparison included 1 Jan 2016 – 30 September 2018 (referred to as DRI-2015; when DRI was the contractor and used the DRI 2015 model for analysis) and 1 Oct 2018 – end date (Table 1) (referred to as Sunset, when UC Davis used Sunset 5L model for analysis). For ions, the comparison included three time periods based on the contractors: RTI-pre (1 Dec 2012 to 20 Nov 2015), DRI (21 Nov 2015 to 30 September 2018), and RTI-post (1 Oct 2018 to end date, shown in Table 1). The end date represents the end date of the downloaded data used in this study, and this end date varies for each species and each site (Table 1). For all the sites, if either IMPROVE or CSN measurements were missing (all species except NH₄), these days were excluded from further analysis. In addition, to reduce the effect of measurements near the detection limit, for all sites and all species, the days with measurements less than 3 times the MDL from either IMPROVE or CSN network were eliminated. Table 1 provides a summary of the number of days included in the analysis. The number of days for OC, EC, NO₃, and SO₄ represents days when both CSN and IMPROVE measurements are available and measurements were greater than 3*MDL. For NH4, the number of days represented when CSN reported NH₄ was greater than 3*MDL, and CSN reported NO₃ and SO₄ were also available.

2.4. Comparisons between the different time intervals

We compared the concentration differences between IMPROVE and CSN using the scaled residual difference (SRD), as used in previous studies (Gorham et al. 2021; Zhang et al. 2021). SRD was calculated for each species, each day using the following formula:

$$\mathit{SRD}=100*\frac{\frac{(\mathit{IMPROVE}-\mathit{CSN})}{\sqrt{2}}}{\frac{(\mathit{IMPROVE}+\mathit{CSN})}{2}}$$ (1)

The study focuses on the following evaluations:

2.4.1. Effect of contractor change on SRD

Each site underwent a comparison of daily SRD across specific time intervals: DRI-2015 versus Sunset for OC and EC, and RTI-pre, DRI, and RTI-post for NO₃ and SO₄.

2.4.2. Effect of concentration range on SRD

Zhang et al. (2021) revealed that discrepancies in EC measurements between the DRI 2015 model and Sunset 5L model were most pronounced in samples with elevated EC levels, suggesting that differences between CSN and IMPROVE measurements might vary depending on concentration. Thus, we investigated whether the variations observed in species concentrations between IMPROVE and CSN were linked to concentration levels. Data points from all sites were aggregated, and for each time interval, species measurements were categorized into three or four bins based on CSN-network concentration thresholds. For instance, EC bins were delineated as <0.5 μg m⁻³, 0.5 – 1 μg m⁻³, and >1 μg m⁻³. OC bins were categorized as <1.5 μg m⁻³, 1.5 – 3 μg m⁻³, and >3 μg m⁻³. NO₃ bins were established as <0.5 μg m⁻³, 0.5 – 1 μg m⁻³, 1 – 5 μg m⁻³, and >5 μg m⁻³. SO₄ bins comprised <0.5 μg m⁻³, 0.5 – 1 μg m⁻³, >1 μg m⁻³ (Table 2). These bins were chosen based on visual inspection of CSN concentration vs SRD plots (Figure S3, online Supplemental Information) with a fitted smooth line. CSN concentrations with noticeable changes in the smooth line curvature were taken as concentration bins.

Table 2:

Concentration bin specific $m3$ and $b3$ for use with equations Eq 2 and Eq 5 and spline-fit parameters ${\beta }_0,{\beta }_1,{\beta }_2$, and ${\beta }_3$ for use with Eq 8 and Eq 9, to correct the EC CSN concentration for the Sunset period and ions for the DRI time interval. Either the concentration-bin-specific coefficients or the spline-fit parameters can be utilized for these corrections. However, it is important to note that NO₃ concentrations exceeding 5 μg m⁻³ were not corrected, and consequently, no correction factors were provided. The decision to omit correction factors for NO₃ concentrations exceeding 5 μg m⁻³ also led to the removal of the knot corresponding to 5 μg m⁻³ for NO₃. The values within square brackets represent confidence interval [2.5%, 97.5%]. The confidence interval for $m3$ and $b3$ were estimated using confidence intervals of $m1,b1,m2$ and $b2$ (Section S2, Supplementary material).

Species	Concentration bin (μg m−3)	Concentration bin specific coefficient		Spline-fit parameters
		$m3$	$b3$	$k_1$	$k_2$	${\beta }_0$	${\beta }_1$	${\beta }_2$	${\beta }_3$
EC	<0.5	0.985 [0.92, 1.05]	−0.034 [−0.052, −0.015]	0.5	1	−0.061 [−0.063, −0.059]	1.081 [1.075, 1.087]	−0.181 [−0.19, −0.173]	−0.149 [−0.153, −0.146]
	0.5-1	0.818 [0.73, 0.90]	0.093 [0.047, 0.139]
	>1	0.749 [0.70, 0.80]	0.183 [0.088, 0.278]
NO3	<0.5	1.202 [1.12, 1.28]	0.013 [−0.006, 0.031]	0.5	1	0.037 [0.031, 0.043]	1.097 [1.079, 1.115]	0.152 [0.123, 0.18]	−0.168 [−0.185, −0.152]
	0.5-1	1.610 [1.41, 1.81]	−0.292 [−0.392, −0.193]
	1-5	1.103 [1.05, 1.16]	0.062 [−0.057, 0.182]
SO4	<0.5	1.254 [1.20, 1.31]	−0.027 [−0.044, −0.009]	0.5	1	−0.010 [−0.013, −0.008]	1.192 [1.186, 1.199]	−0.146 [−0.155, −0.138]	−0.042 [−0.046, −0.038]
	0.5-1	1.116 [1.05, 1.18]	0.012 [−0.032, 0.055]
	>1	1.010 [0.98, 1.04]	0.092 [0.044, 0.140]
NH4	<0.5	1.106 [1.03, 1.18]	0.014 [−0.007, 0.035]	0.5		0.0058 [0.005, 0.007]	1.171 [1.168, 1.173]	−0.174 [−0.177, −0.171]
	>0.5	0.992 [0.96, 1.02]	0.11 [0.047, 0.173]

2.5. Statistical analysis

The mean, minimum, maximum, 5^th percentile, and 95^th percentile of the daily SRD data were calculated for each time interval for each site and each concentration bin. The average SRDs between the different contractors were compared before and after the correction to evaluate the effect of the correction factors.

3. Results and Discussion

3.1. Effect of contractor change on carbon concentrations

Figure 2 (a & b) illustrates the effect of the contractor change on EC and OC daily SRDs for each site. Both the CSN and IMPROVE networks use the IMPROVE-A method for the OC and EC and carbon fractions, but these networks have subtle differences (Table S1 and Table S2). For example, sample preparation, sample storage, and shipping procedures (CSN ships the filter with ice packs, whereas IMPROVE ships at ambient conditions). These differences can create bias in carbon concentrations and thermal subfractions (Solomon et al. 2014).

Figure 2:

Boxplots showing the effect of contractor and instrument change (as SRD %) on OC and EC concentrations. (a & b) SRD of each site; (c & d) Concentration-binned SRD of all-site aggregated measurements. The box represents the interquartile range (IQR). The line in the box represents the median. The dot (red) with a border (black) represents the average SRD. The whiskers represent the 1.5*IQR range. The dots (black) represent the outliers.

In Figure 2 (b), during the DRI-2015 time interval, the average SRDs of the sites for EC ranged from −10.6% to −3.3%. These negative daily SRDs for OC and EC for all sites could be caused by the subtle differences between the networks, discussed in the previous paragraph. However, during the Sunset analyzer time interval, the average SRD for EC further decreased compared to the DRI-2015 period, ranging from −35.7% to −6.5% (Figure 2 (b)). This indicates that the Sunset analyzer contributes to the increased bias between the EC of two networks. Previous studies have also reported on biases between the DRI 2015 and Sunset 5L model analyzer (Zhang et al. 2021). Figure 2 (a) illustrates that the OC SRD of the two-time intervals had similar ranges with an average SRD between −10.8% to −3.6% for DRI-2015 compared to −12.1% to 1.3% for Sunset).

The North Birmingham site reported higher EC concentrations during 2019, causing a high negative SRD with a larger interquartile range (IQR) compared to the other locations (Figure 2(b)). The decline in the EC SRD appears to be more pronounced at locations where over 20% of the days recorded EC concentrations surpass 1 μg m⁻³. Supplemental Figure S4 illustrates this pattern: as the number of days with EC concentrations exceeding 1 μg m⁻³ increases, the slope’s magnitude (Figure S4, indicating the relationship between the percentage of days with EC greater than 1 μg m⁻³ and average SRD %) seems to increase notably during the Sunset time interval. This indicates that the difference between IMPROVE and CSN-reported EC concentrations during the Sunset analyzer time interval increased as EC concentrations increased. This is consistent with Zhang et al. (2021) who noted that the Sunset 5L overestimation increases with higher EC concentrations, resulting in more negative SRD values. For sites like North Birmingham, Fresno, Atlanta, and Phoenix, over 35% of the Sunset time interval recorded EC concentrations exceeding 1 μg m⁻³. In contrast, Detroit and Pittsburgh had EC concentrations surpassing 1 μg m⁻³ less than 20% of the time during both the DRI-2015 and Sunset analyzer time intervals.

Figure 2 (d) also illustrates the impact of EC concentration on the SRD; it indicates a decrease in SRD with increasing concentrations during the Sunset period. The average SRD for concentration bins <0.5, 0.5 – 1, and >2 μg m⁻³ during the DRI-2015 period interval were −3.09%, −8.52%, and −10.32%, respectively. These averages become more negative during the Sunset period to −15.1%, −17.7%, and −23.3%, respectively. This implies that employing a concentration-based correction could potentially mitigate EC concentration discontinuities during the Sunset period. Across all OC concentration bins, the average SRD remained consistent between the DRI-2015 and Sunset time intervals (Figure 2(c)). Specifically, the average SRD for OC for concentration bins <1.5, 1.5 – 3, and >3 μg m⁻³ during the DRI-2015 period interval were −4.89%, −8.60%, and −8.68%, respectively, while for the Sunset period, they were −7.06%, −7.94%, and −5.03%, respectively. Consequently, OC concentrations were not adjusted, and no correction factors were developed.

3.2. Effect of contractor changes on ion concentrations

Figure 3 (a & b) demonstrates an increase in the average SRD for NO₃ and SO₄ during the DRI time interval. An increase in SRD during the DRI period occurred at all four collocated monitoring sites. For instance, the average SRD for NO₃ across all sites ranged from −16.2% to −2.7% during the RTI-pre period, increased to 6.7% to 12.4% during the DRI period, and returned to −11% to −2.4% during the RTI-post period. The average SRD for SO₄ exhibited a similar trend. The average SRD for ions during the RTI-pre and RTI-post intervals were comparable for most sites. For instance, for NO₃, for sites BIRM, FRES, PHOE, and PUSO, the average SRD for the RTI-pre were −16.1%, −2.7%, −10.0%, and −6.8%, respectively, and for RTI-post, the average SRD was −10.9%, −2.4%, −6.6%, and −3.9%, respectively.

Figure 3:

Boxplots showing the effect of contractor/instrument change (as SRD %) NO₃ and SO₄ concentrations. (a & b) SRD of each site (c & d). Concentration-binned SRD of all-site aggregate measurements. The box represents the interquartile range (IQR). The line in the box represents the median. The dot (red) with a border (black) represents the average SRD. The whiskers represent the 1.5*IQR range. The dots (black) represent the outliers.

When RTI analyzed the filters for both networks (CSN and IMPROVE), the CSN NO₃ concentrations were higher than IMPROVE concentrations (negative SRD, Figure 3 (a & c)). This disparity in NO₃ concentrations between CSN and IMPROVE may be due to differing filter face velocities (43.2 cm s⁻¹ for IMPROVE vs. 9.5 cm s⁻¹ for CSN), as well as variations in filter handling and extraction protocols between the networks (Table S1 and Table S2). On the other hand, SO₄ concentrations were similar between the networks (SRD between −2.2% to −0.2%, Figure 3 (b & d)) during the RTI contract period.

Figure 3 (c) depicts the fluctuation in SRD of NO₃ concentration bins during different time intervals. During the RTI-pre period, the average SRD ranged from −14.3% to 0.7%, whereas it increased to −1.8% to 12.3% during the DRI period before decreasing again to −7.2% to −1.9% in the RTI-post interval. Similarly, for SO₄ concentration bins (Figure 3 (d)), the average SRD varied from −1.4% to 1.1% in the RTI-pre period, rose to 3.2% to 12.2% during the DRI period, and then decreased to −2.2% to −1.3% in the RTI-post interval. Gorham et al. (2021) also observed higher NO₃ concentrations reported by IMPROVE compared to CSN during the DRI period when NO₃ concentrations were less than 3 μg m⁻³. Additionally, they noted higher SO₄ concentrations reported by the IMPROVE network compared to the CSN network during the DRI period, which aligns with our study’s findings.

The analysis in Figure 3 (c & d) reveals that the transition from RTI-pre to DRI had the most significant impact on the smallest concentrations of NO₃ and SO₄. As concentration increased, the disparity between the DRI and RTI-post intervals decreased, particularly evident in NO₃ bins exceeding 5 μg m⁻³. This concentration-dependent pattern underscores the necessity for a concentration bin correction factor. Because the average SRD for NO₃ bins exceeding 5 μg m⁻³ was similar between DRI (average SRD of −1.8%) and RTI-post periods (average SRD of −1.9%), no correction was applied to NO₃ concentrations exceeding 5 μg m⁻³ during the DRI period.

3.3. Development of correction factors

The study aimed to reduce discontinuities in the CSN concentrations by maintaining the differences between the CSN-measured concentrations throughout the different time intervals, rather than to decrease the differences in the measurements between the CSN and IMPROVE networks (as indicated by SRDs). This was accomplished by developing correction factors for each concentration bin and each target species measured by CSN. For example, the EC factors corrected the EC concentration during the Sunset time interval measurements such that the SRD between DRI-2015 and the Sunset time interval became more consistent. Similarly, for ions, the factors developed for the DRI time interval kept the SRD consistent with the RTI-post time interval.

3.3.1. Development of a correction factor for EC, NO₃, and SO₄

Linear regression analysis was conducted between all-sites, aggregated IMPROVE and CSN measurements for each species (EC, NO₃, and SO₄) and each concentration bin during different time intervals: DRI-2015 and Sunset time intervals for EC, and the DRI and RTI-post time intervals for NO₃ and SO₄ (Figure S5, S6, and S7, online Supplemental Information). Before applying the linear regression, the outliers in the EC, NO₃, and SO₄ SRD series were removed using the interquartile range (IQR) method using first quartile (Q1) and third quartile (Q3) (values smaller than Q1 – 1.5*IQR, and values greater than Q3+1.5*IQR were removed). Less than 4% of data points were rejected due to the IQR method. The adjusted CSN concentration during the Sunset period for EC and DRI time intervals for NO₃ and SO₄ were calculated as follows:

$${\mathit{CSN}}^{\mathit{adj}}=\mathit{CSN}*m3+b3$$ (2)

with,

$$b3=b2-(b1*m3)$$ (3)

$$m3=\frac{m2}{m1}$$ (4)

where, $m1$ and $m2$ represent the slopes, while $b1$ and $b2$ denote the intercepts of the linear regression between IMPROVE and CSN concentrations. Specifically, $m1$ and $b1$ pertain to the Sunset time interval, while $m2$ and $b2$ elate to the DRI-2015 time interval for EC. Regarding ions, $m1$ and $b1$ align with the DRI time interval, whereas $m2$ and $b2$ aligns with the RTI-post time interval. The derivation to obtain the equations (2) and (3) can be found in section S1 of the online Supplemental Information.

3.3.2. Development of a correction factor for NH₄

In the absence of NH₄ concentration measurements from the IMPROVE network, a method was devised solely utilizing the CSN-reported NH₄, NO₃, and SO₄ concentrations. Correcting NH₄ is crucial, particularly during wintertime in the western and north-central US, due to the large contributions of NO₃ and SO₄ as ammonium nitrate and ammonium sulfate in the particulate phase, to PM_2.5 (Hand et al. 2023, 2024). The contractor changes are expected to influence NH₄ concentrations, along with NO₃ and SO₄. Since NO₃ and SO₄ concentrations for the DRI time interval were being adjusted, it was necessary to correct NH₄ concentrations during the DRI interval as well. This correction process assumes that the stoichiometrically calculated NH₄ from NO₃ and SO₄ should maintain a similar ratio to CSN-measured NH₄ before and after the correction of NO₃ and SO₄ (discussed in step 3). The correction process was as follows:

$${\mathit{NH}4}^{\mathit{adj}}=\mathit{NH}4*m3+b3$$ (5)

with,

$$b3=b2-(b1*m3)$$ (6)

$$m3=\frac{m2}{m1}$$ (7)

where, $m1$ and $m2$ represent the slopes, while $b1$ and $b2$ denote the intercepts of the linear regression between stoichiometrically calculated NH₄ and CSN-reported NH₄ concentrations, specifically for the DRI time interval (20 Nov 2015 – 30 September 2018). The $m1$ and $b1$ correspond to the stoichiometric NH₄ concentration from corrected NO₃ and SO₄ versus the CSN-reported NH₄, while $m2$ and $b2$ aligns with the stoichiometric NH₄ from uncorrected NO₃ and SO₄ versus CSN-reported NH₄ (Figure S8, online Supplemental Information). Similar to other species, this study developed two different corrections for NH₄, one for CSN-reported concentrations less than 0.5 μg m⁻³ and one for CSN-reported concentrations greater than 0.5 μg m⁻³. Basically, after NH₄ concentrations were divided into the two concentration bins based on the CSN-reported concentrations, a linear regression was performed differently for each concentration bin.

3.3.3. Linear spline fit

To mitigate abrupt discontinuities occurring at the boundaries of concentration bins when employing the equations obtained in step 3 and 4, a linear spline approach was employed. This involved plotting the adjusted concentrations (EC for the Sunset period and ions for the DRI period) obtained using the concentration bins equation 2 or equation 5 on the y-axis against the uncorrected concentrations on the x-axis. The concentration bin divisions were utilized as knot points (e.g., 0.5 μg m⁻³ and 1 μg m⁻³ for EC). Separately for each pollutant, the linear spline model was formulated as follows:

$${\mathit{CSN}}^{\mathit{adj\_spline}}={\beta }_0+{\beta }_1(x)+{\beta }_2{\left(x-k_1\right)}_{+}+{\beta }_3{\left(x-k_2\right)}_{+}$$ (8)

with, ${\beta }_0,{\beta }_1,{\beta }_2$, and ${\beta }_3$ as the coefficients of the spline fit, and $k_1$ and $k_2$ denote the knots. The $x$ denoted the uncorrected concentration. The spline terms are defined as:

$${\left(x-k_i\right)}_{+}=\{\begin{array}{ll}0& \mathit{if}x<k_i\\ x-k_i& \mathit{if}x\ge k_i\end{array}$$ (9)

In case of NH₄ only one knot (at 0.5 μg m⁻³) was used, which limited the coefficients to ${\beta }_0,{\beta }_1$, and ${\beta }_2$.

Table S5 presents the linear regression parameters used to formulate correction factors for EC, NO₃, SO₄, and NH₄. Additionally, Table 2 lists the parameters $m3$ and $b3$ for each species and concentration bin, while spline-fit parameters ${\beta }_0,{\beta }_1,{\beta }_2$, and ${\beta }_3$, along with knots $k_1$ and $k_2$, can also be found in Table 2.

The slope $m3$ for EC (Table 2) varied between 0.749-0.985, depending on the concentration bin. Zhang et al. (2021) reported that the Sunset reported EC was ~12% higher than the DRI-reported EC concentrations, which corresponds to a DRI/Sunset ratio of ~0.893, similar to this study. The $m3$ value decreases as concentrations increase across all species. The declining $m3$ values for EC suggest that corrected concentrations decrease more as the uncorrected concentrations increase. Conversely, concerning ions, $m3$ values decrease as concentration-bins increase suggesting that at higher concentrations, ions undergo relatively minimal changes (Figure S9, online Supplemental Information).

3.4. Comparison of uncorrected and corrected SRDs

Figure 4 and Table 3 provide a comparison of the daily SRDs between original and corrected measurements of EC, NO₃, and SO₄. Figure 4 illustrates the discontinuity in the EC and ions concentrations. If someone were evaluating the trends in the EC and ions, they might draw the incorrect conclusion about emission changes (e.g. accountability studies that rely on these concentrations to investigate policy implications). The corrected measurements correspond to the concentrations obtained from the linear spline fits. After correction, the disparity in EC between the DRI-2015 and Sunset time intervals decreased, as evidenced by the average SRD of −11% for the corrected versus uncorrected Sunset readings (−18.5%). Additionally, following corrections for the DRI period, SRDs for NO₃ and SO₄ tended to converge with those of the RTI periods, as depicted in Figure 4 (b & c) and Table 3. Figures S10 and S11 provide visual representations of the impact of corrections on site-specific and concentration-bin-specific SRD distributions, showcasing a reduction in differences across sites post-correction. In general, after correction, there was a notable reduction in the discrepancy between average SRDs, resulting in a more continuous time series, as illustrated in Figure 4 and Figure S12 (online Supplemental Information).

Figure 4:

Daily SRDs, aggregated from all the sites (7 for EC and 4 for ions). (empty circle (black)) SRD from uncorrected concentrations, (solid circles (red)) SRD from corrected concentrations (obtained after linear spline fit) for specific time intervals (EC for Sunset interval, ions for DRI interval). The dashed (black) vertical line represents the dates of contractor change, i.e., 20 Nov 2015 and 1 Oct 2018. The solid line (black) represents the time-interval specific average SRD, and the solid lines (white) represent the average SRD for the corrected data corresponding to the time interval (Sunset for EC and DRI for ions).

Table 3:

Statistics including average, 5^th percentile, and 95^th percentile of the daily SRD % for the different species and time intervals before and after correction. OC values were not adjusted.

	OC			NO3
	before correction		after correction	before correction			after correction
	DRI	Sunset		RTI-pre	DRI	RTI-post	DRI
Average	−7.3	−6.8		−8.9	9.6	−5.6	−3.3
5th percentile	−24.3	−24.1		−36	−13.4	−22.8	−27.4
95th percentile	10.9	9.4		10.9	31.5	8.9	18.5
	EC			SO4
	before correction		after correction	before correction			after correction
	DRI	Sunset	Sunset	RTI-pre	DRI	RTI-post	DRI
Average	−6.4	−18.5	−11	−0.9	7.7	−1.8	−0.4
5th percentile	−31.9	−49.4	−41.3	−12.2	−6	−9	−12.9
95th percentile	19.1	4.3	12.7	12.4	21.9	5.8	13.3

3.5. Correction of the CSN NH₄

In the absence of a direct comparison between CSN and IMPROVE NH₄ concentrations, NH₄ concentration during the DRI time interval was corrected based on the assumption that the relationship between CSN-reported NH₄ and stoichiometrically calculated NH₄ from NO₃ and SO₄ should be similar, whether the stoichiometrically calculated NH₄ was derived from reported CSN NO₃ and SO₄ concentrations or corrected NO₃ and SO₄ concentrations. Figure S8 (online Supplemental Information) shows the relationship between the CSN-reported NH₄ and stoichiometrically calculated NH₄ from uncorrected and corrected NO₃ and SO₄ concentrations for the DRI time interval. Table 2 provides the correction factor $m3$ and $b3$ for use with equations Eq 2 and Eq 5 and spline-fit parameter ${\beta }_0,{\beta }_1,{\beta }_2$, and ${\beta }_3$ for use with Eq 8 and Eq 9.

Figure 5 (a) illustrates the comparison between NH₄ concentrations before and after correction (obtained using linear-spline fit parameters) for the DRI time interval. There was an overall increase observed in NH₄ measurements across the entire concentration range. NH₄ concentrations below 0.017 μg m⁻³ experienced a change of more than 50%, albeit these measurements accounted for only approximately 4%. Comparing the corrected NH₄ and stoichiometrically calculated NH₄ from corrected NO₃ and SO₄, shown in Figure S13 (online Supplemental Information), reveals similar relationship and slope to the CSN reported NH₄ reported and stoichiometrically calculated NH₄ from CSN reported NO₃ and SO₄ (Figure S8 (top), online Supplemental Information).

Figure 5:

(a) Temporal distribution of original NH₄ concentration (empty circle (black)), and corrected NH₄ concentration (solid circles (red), obtained using linear-spline fit parameters) for the DRI interval. The dashed (black) vertical line represents the dates of the contractor change, i.e., 20 Nov 2015 and 1 Oct 2018. (b) CSN uncorrected NH₄ concentration versus the difference between the corrected (obtained using linear-spline fit parameters) and uncorrected NH₄ concentration. The dashed (black) vertical line represents the concentration bin at 0.5 μg m⁻³. The dashed-dot (black) horizontal line represents the % difference of 0%. The difference was calculated as: (uncorrected concentration – corrected concentration) *100/uncorrected concentration.

4. Conclusions

The investigation into the impact of contractor changes on EC, OC, and ion concentrations within the CSN network, beginning in November 2015 for ions and October 2018 for carbon and ions, revealed significant discontinuities in the time series SRDs (scaled relative difference) between IMPROVE and CSN collocated measurements. The comparison of SRDs revealed that the ion concentrations analyzed by DRI were markedly lower than those analyzed by RTI, while EC measurements from the Sunset 5L model consistently exceeded those from the DRI model across all concentration ranges, with the extent of increase varying depending on the concentration. To minimize these discontinuities, correction factors were developed based on linear regression between CSN and IMPROVE measurements for different contractor time intervals (DRI-2015 and Sunset for EC, and RTI-pre, DRI, and RTI-post for ions), followed by refinement through linear spline fitting. Additionally, a correction factor for NH₄ was derived from CSN measurements alone, assuming that the stoichiometrically calculated NH₄ from NO₃ and SO₄ should maintain a similar ratio to CSN-measured NH₄ before and after the correction of NO₃ and SO₄. Implementation of these correction factors led to a reduction in discontinuities within ion and EC concentration time series and minimized differences in SRDs across various time intervals. Reducing discontinuities is important for downstream analyses, including epidemiologic studies, although further research will determine the impact of these contractor changes on estimated health associations.

5. Data availability

All the raw data was downloaded from the Federal Land Manager Environmental Database (https://views.cira.colostate.edu/fed/), which is an open-source database.