Assessing the Suitability of Historical PM2.5 ... - ACS Publications

Article pubs.acs.org/est

Assessing the Suitability of Historical PM2.5 Element Measurements for Trend Analysis Nicole P. Hyslop,* Krystyna Trzepla, and Warren H. White Crocker Nuclear Laboratory, University of California, Davis, California 95616, United States

Environ. Sci. Technol. 2015.49:9247-9255. Downloaded from pubs.acs.org by UNIV OF CAMBRIDGE on 09/05/15. For personal use only.

S Supporting Information *

ABSTRACT: The IMPROVE (Interagency Monitoring of Protected Visual Environments) network has characterized fine particulate matter composition at locations throughout the United States since 1988. A main objective of the network is to evaluate long-term trends in aerosol concentrations. Measurements inevitably advance over time, but changes in measurement technique have the potential to confound the interpretation of long-term trends. Problems of interpretation typically arise from changing biases, and changes in bias can be difficult to identify without comparison data that are consistent throughout the measurement series, which rarely exist. We created a consistent measurement series for exactly this purpose by reanalyzing the 15-year archives (1995−2009) of aerosol samples from three sites − Great Smoky Mountains National Park, Mount Rainier National Park, and Point Reyes National Seashore−as single batches using consistent analytical methods. In most cases, trend estimates based on the original and reanalysis measurements are statistically different for elements that were not measured above the detection limit consistently over the years (e.g., Na, Cl, Si, Ti, V, Mn). The original trends are more reliable for elements consistently measured above the detection limit. All but one of the 23 site-element series with detection rates >80% had statistically indistinguishable original and reanalysis trends (overlapping 95% confidence intervals).

■

INTRODUCTION The IMPROVE network (Interagency Monitoring of Protected Visual Environments) is distinguished by the length and stability of its air quality measurements. IMPROVE has collected 24 h samples of airborne particulate matter (PM) continuously since 1988.1,2 The network today includes about 160 sites, including 69 sites that have operated continuously at the same locations since 1994. Every third day, PM samples with aerodynamic diameters less than 2.5 μm (PM2.5) are collected on polytetrafluoroethylene (PTFE) filter membranes. These samples are analyzed for elemental content by nondestructive X-ray fluorescence (XRF) spectroscopy. The sampling equipment and techniques have remained fundamentally unchanged throughout IMPROVE’s history. The element analyses have always been performed by the same laboratory, but the analytical techniques have evolved over the years. New analytical techniques are introduced to improve measurement quality−decrease measurement error, lower detection limits (DLs), increase reliability, increase measurement stability, and improve measurement accuracy. The impact of these changes in measurement quality depends on both the type of change and how the data are used.3−5 Errors can be broadly categorized as either random (noise) or systematic (bias). Random errors can be effectively addressed by averaging multiple concentrations to reduce the error, but systematic errors are harder to address. Attempts are routinely made to correct measurements for known © 2015 American Chemical Society

systematic errors before they are reported (e.g., subtraction of sampling media blank or analytical background), but there is often not a way to evaluate the efficacy of these adjustments. The absolute error associated with a particular method can be quantified only by comparing it against a reference method known or defined to be completely accurate.6 No such reference method exists for PM speciation measurements, so the total error of these measurements is expected to include some uncorrected bias. Errors are not necessarily constant across the range of measured concentrations; they may be proportional to the measured concentration, such as multiplicative calibration errors, or arithmetic at lower concentrations where the blank variability drives the error.7−9 It is difficult to accurately estimate the uncertainty in a measurement method, particularly the systematic component of the uncertainty. In their historical review of the evolution of fundamental physical constants, Henrion and Fischhoff demonstrated that reported uncertainties consistently underestimated actual errors.10 Biases are often brought to light only by the arrival of improved measurement methods. Trend estimates are particularly sensitive to bias that can change over time.11,12 The reported uncertainties only account Received: Revised: Accepted: Published: 9247

March 27, 2015 June 24, 2015 June 30, 2015 June 30, 2015 DOI: 10.1021/acs.est.5b01572 Environ. Sci. Technol. 2015, 49, 9247−9255

Article


Environmental Science & Technology for the repeatability (or precision) of the measurement using the existing technique and do not account for bias, particularly biases that change over long periods of time.8 Changes in DL can dramatically change a measurement distribution, particularly if the measurements are truncated (reported as zero) below the DL as they were in IMPROVE prior to 2011. Problems of interpretation typically arise from changing biases, and changes in bias can often only be identified if comparison data are available whose biases can be assumed consistent throughout the measurement series. We created a consistent measurement series such as this by reanalyzing archived samples with current techniques.13 Although the IMPROVE analysis methods changed over the years, the samples themselves preserve a long-term record of the ambient aerosols; because the original analyses were nondestructive, these samples can now be reanalyzed to provide methodologically consistent concentration time-series. We generated consistent element concentration data at three sitesMount Rainier National Park (MORA) in Washington state, Point Reyes National Seashore (PORE) in California, and Great Smoky Mountains National Park (GRSM) in Tennesseeby reanalyzing the 15-year sample archive for each site with a single analytical method under the same calibration and measurement techniques. In this paper, we compare the trend estimates based on the original concentrations with trend estimates based on the reanalysis concentrations. We also explore whether biases are consistent across the three sites and can be extended across the entire IMPROVE network. The results of this reanalysis inform historical explorations of the unique and irreplaceable IMPROVE data record and provide insight into observed trends that are now confounded by measurement issues. Our interest in the following analyses lies in the implications of evolving measurement methods for interpreting trend analyses, not in trends and their analysis per se. We accordingly focus on differences (or similarities) between results from the same trend analysis procedure performed on heterogeneous (original) and homogeneous (reanalyzed) measurement series. To limit the proliferation of possible comparisons, we do not explore the many alternative statistical approaches available for dealing with nondetects,14 seasonality,15 and fitting.16

Figure 1. Timeline of analytical methods for IMPROVE PTFE filter samples. The bars show the sequence of methods for elements lighter than Fe (top), Fe, and elements heavier than Fe (bottom). Samples were archived beginning in March 1995.

energy-dispersive X-ray fluorescence (EDXRF) system using a Mo-anode excitation tube (Mo XRF) was introduced in June 1992 to obtain better sensitivity for the heavier elements, Fe−Pb.19 A second, custom EDXRF system with a Cu-anode excitation tube replaced PIXE for the lighter elements, and was used to report Na−Fe starting in December 2001. The analysis chamber of the initial Cu XRF system was flushed with He (Cu XRF, He) to displace atmospheric Ar, which contributes unwanted spectral background. Operational difficulties caused by the He flush led to the installation of a vacuum chamber in January 2005 (Cu XRF, vac). From 1988 through 2010, the same spectral interpretation software was used to determine mass loadings, detection limits, and uncertainties from the XRF spectra generated by the various instruments. This software performed baseline correction using a laboratory blank PTFE filter; no further blank correction was performed during these years. Beginning in 2005, blank filters began to be regularly analyzed for IMPROVE. PANalytical Epsilon 5 EDXRF instruments have been used to perform the elemental analyses for all IMPROVE samples collected after 1 January 2011, but no data from the Epsilon 5 instruments are included in this analysis. In addition to these changes in the underlying systems, operational factors such as detector performance and calibration procedures introduced other variations in the measurements. Some shifts in reported concentrations and measurement interferences are documented on the IMPROVE Web site (http://vista.cira. colostate.edu/improve/Data/QA_QC/Advisory.htm). The analytical methods used for element determinations were nondestructive and left the PTFE sample filters intact. All PTFE filter samples collected since March 1995 have been archived in a storage building on the UCD campus. For this study, the 1995−2009 archived filter samples from GRSM, PORE, and MORA were retrieved, inspected, and assembled into batches for analysis on the same instruments used to process the 2010 samples.13 Cu-anode XRF in vacuum was used for the light elements, and Mo-anode XRF in air for the heavier elements. On each instrument the entire sample archive for a site was reanalyzed within a month. XRF stability was monitored with weekly calibration checks and by analyzing a designated set of filters before and after each reanalysis batch. Reinspection revealed some new damage invalidating additional filters after the handling and stress of reanalysis. Of the filters with valid original measurements reported, the survival rates for retrieval and reanalysis were 89%, 87%, and 90% (1441 of 1626, 1503 of 1736, and 1342 of 1494) at GRSM, MORA, and PORE, respectively. For a fair comparison, a sample day was included only if it had both valid original and reanalysis concentrations. Previous analyses of the GRSM reanalysis data indicated the

■

MATERIALS AND METHODS Since sampling began in 1988, IMPROVE has continuously reported concentrations of the elements Na, Mg, Al, Si, P, S, Cl, K, Ca, Ti, V, Cr, Mn, Fe, Ni, Cu, Zn, As, Se, Br, Rb, Sr, Zr, and Pb. Samples have always been collected on 25 mm-diameter stretched PTFE membrane filters drawing ambient air at ∼23 lpm through a cyclone with a 50% cut at 2.5 μm aerodynamic particle diameter.17 In the early years, the sample filters were masked to concentrate the samples into a smaller deposit area of 2.2 cm2 effectively providing lower DL, but the smaller deposit areas sometimes resulted in clogging, particularly at sites with higher PM concentrations, and the masks were removed from the network over several years. The masks were removed from GRSM on 18 April 1995 (prior to the archiving of filters), PORE on 9 March 2004, and MORA on 23 October 2007. The change in sample area affects the DLs and may affect the measured concentrations. All elemental analyses have been performed by Crocker Nuclear Laboratory (CNL) at the University of California in Davis (UCD), using the analytical systems summarized in Figure 1. Proton induced X-ray emission (PIXE) analysis with the CNL cyclotron was initially used for all the elements.18 A custom 9248

DOI: 10.1021/acs.est.5b01572 Environ. Sci. Technol. 2015, 49, 9247−9255

Article


Environmental Science & Technology

indistinguishable from (i.e., confidence intervals overlapped with) the raw concentration trends for the majority of the element-site series (33 out of 36 in reanalysis set). Since deseasonalization did not prove necessary and to keep the analysis as simple as possible, the remainder of this paper only discusses trends based on the raw concentration data. We also explored the effects of the masks being removed from the filters during the reanalysis period at PORE and MORA. When the masks were removed, the DLs shifted because the deposit area used to convert the areal densities (mass cm−2) measured by the XRF instruments into loadings (mass filter−1) changed. Removing masks may also have affected the measured concentrations, particularly if the concentrations were close to the DLs. To explore whether removing the masks is an important factor in the trends, multivariate regression was performed with the filter masking as a factor. Theil-Sen regression cannot be used for multivariate regression, so instead multivariate regressions were performed using another robust regression technique (MASS package in R, http://cran.r-project.org/web/packages/MASS/index.html); eq 3 shows the model for the concentration as a function of both time and filter masking.

measured elements, with the exception of Br, were stable on the filters over time.13 Br is excluded from the trends analyses for this reason. Cl was not consistently detected at GRSM, but it is at PORE, which is a coastal site, and Cl may also exhibit stability problems. Detection rates were previously shown to be generally predictive of precision, relative bias, and correlation between the original and reanalysis values.13 Trend results are sensitive to the treatment of below-detection measurements, particularly for elements whose nondetects were unequally distributed through time in the original data set. To minimize complications related to DLs, only elements measured above their DL more than 60% percent of the time overall are included in the trend analyses that follow. This criterion eliminates Mg, P, and Zr, which had low detection rates throughout the period considered; it also eliminates Al and Cr because of their low detection rates in the early years. IMPROVE reports an estimate of the DL with each measurement, and for the trends analysis we substituted these sample-specific DLs for below-detection concentrations. DL estimates for the period considered were based on the estimated noise in a spectral baseline free of interfering peaks, and can substantially underestimate actual detection limits.7,20 Trends based on the 24 h concentrations from 1995 through 2010 were estimated using Thiel-Sen regression (zyp package in R, http://cran.r-project.org/web/packages/zyp/index.html).21−23 Thiel-Sen is a robust regression technique that is insensitive to outliers; the Thiel-Sen slope is the median slope of all the slopes calculated from all possible pairs of values. To take full advantage of this equal weighing, no averaging of the concentration data was performed; the reanalysis data sets contain at least 1342 concentrations per element and site, which means a minimum total of 899 811 slope estimates (1342 × 1341/2 unique pairs) for each data series. eq 1 shows the model for the concentrations as a function of time.

C time = C0 × Rtime

C time = C0 × maskM × Rtime ln(C time) = ln(C0) + M × ln(mask) + time × ln(R )

(4)

The binary value (0 or 1) of the indicator ln(mask) represents the masking status of the sample, and the fitted parameter M represents the effect of masking. The estimates of mask effect were highly variable across elements and sites; approximately equal numbers of positive and negative mask coefficients (M) were obtained, with approximately half of their 95% confidence levels including zero. The trends (time coefficients) estimated with multivariate regression were statistically indistinguishable from the Thiel-Sen trend estimates for the majority of the data series; 18 of 24 were indistinguishable in the original data set and 22 of 24 in the reanalysis data set. This analysis suggests that removing the masks did affect the measurements of some elements, but the effect is not consistent across all the elements and thus difficult to quantify. There are a few potential

(1)

where Ctime is the concentration at time = time, C0 is the concentration at time = 0, and R = Ctime+1/Ctime is the change in concentration over a unit time step, expressed as a ratio. This formulation assumes a constant proportional rate of change over all years, analogous to compounding interest. Time is expressed in years, and trends are expressed in %/year, as (R − 1)100%. Taking logarithms of eq 1 converts the relationship into linear terms eq 2 for the regression analysis. ln(C time) = ln(C0) + time × ln(R )

(3)

(2)

The regression fits for ln(Ctime) and ln(R) in eq 2 are then exponentiated to yield estimates for the model trend of eq 1. In addition to this basic approach to trend analysis, two additional complexities were explored: seasonality and the removal of filter masks. Some elements have seasonal cycles that overwhelm the long-term trends in the data; seasonal cycles add variability to the concentrations unrelated to trends and increase the uncertainty in the trend estimates. To assess the influence of seasonal trend on the long-term trends, we calculated the 15 year trends after removing the seasonality. The seasonality was removed by first calculating the monthly median element concentrations across all measurement years (e.g., monthly median S concentrations from 1995 through 2009) and dividing each 24 h concentration by the corresponding monthly median concentration. The regressions were then run on these deseasonalized values. The deseasonalized trends were statistically

Figure 2. Trends estimated from the original concentrations for the elements measured above the detection limit (DL) >60% of the time at each site independently. A black circle around a point indicates the trend is significant at the 95% confidence level. 9249


Article



Figure 3. Original and reanalysis concentrations of select elements measured at MORA, GRSM, or PORE. The colored data points (opposed to black) indicate that the concentration value was below the DL so the DL was substituted, and the different colors indicate the analytical method. The purple vertical lines indicate when the masks were removed from the sample filters. Trend lines are in red.

measured unsampled areas of the filter, or self-attenuation effects decreased when the masks were removed, thus affecting the reported concentrations. These speculations cannot be confirmed without further analysis. The objective of this paper

explanations for why concentrations may have shifted when the filters were unmasked: water-soluble elements deposited on the filter leached outside the masked area, the X-ray beam was not always centered on the masked deposit, and sometimes 9250


Article



Figure 4. Trend estimates based on the reanalysis data versus the original data for all measured elements (excluding Br); elements with >60% detection rate are identified by color excluding Ti, V, and Mn which are black but circled in blue. The thick black line indicates the 1:1 relationship, and the thin black lines indicate 1% above and below the 1:1 relationship.

is to compare trends estimates based on the original and reanalysis data set, and not to estimate trends themselves. Therefore, since the majority of the regressions proved to be robust to the changes in masks, the remainder of this paper discusses only the Thiel-Sen regression results.

Cu-XRF in He atmosphere instrument was employed, most Na concentrations were below the DL at PORE. The shifts in the Cl concentrations are more subtle with the analytical changes and may not affect the trend estimates; the shift in the Cl trend from negative to positive suggests that Cl may not be stable on the filters over time, similar to Br.13 The trend estimates for Rb, Na, and Cl are questionable based on the original time-series graphs alone, but reanalyzing the filters with a consistent analysis technique provides new insight. Figure 4 compares the trend estimates based on the original and reanalysis concentrations for all elements with the elements measured above the DL > 60% of the time identified by distinct colors, excluding Ti, V, and Mn which are colored black but circled in blue. The thick black line indicates the 1:1 relationship, and surrounding thin black lines indicate 1% above and 1% below the 1:1 relationship. This graph illustrates that the DL > 60% criterion alone is not adequate to eliminate unstable measurements although most well-detected elements lie within the 1% deviation lines. Na, Cl, and Rb are three of the exceptions, and as suggested by the graphs in Figure 3, the original trends were questionable because the originally measured concentrations for these elements were affected by changes in detection limits over the years; the reanalysis revealed that concentrations measured close to the detection limits were often biased high. Si slopes at all three sites lie outside the 1% lines; a documented spectral interference between Si and S is likely causing the unstable trend estimates for Si.24,25 The 1% deviation lines in Figure 3 provide a simple visual evaluation of the agreement between the trends. Table 1 summarizes the statistics including the detection rates. Another way to evaluate the agreement between the trends is to look at the confidence intervals (CI). The original and reanalysis trends are considered statistically indistinguishable if their CIs overlap, and in most cases, they do. In Table 1, the 7 (out of 36) estimates that do not have overlapping CI for the original and reanalysis trends are highlighted in pink; with the exception of Pb and Si, these elements have relatively high levels of nondetects. By simply increasing the detection rate criterion from 60% to 80% and eliminating Si from the comparison, the outlying points are largely eliminated; Figure 5 plots the trend estimates

■

RESULTS AND DISCUSSION Trend estimates along with the 95% confidence intervals (CI) for the original concentrations using Thiel-Sen regression based on the eq 2 model are shown in Figure 2. The CI are used to determine if the slope is significant; if the CI spanned zero it was considered insignificant. Most trend estimates are between −5 and +1% yr−1. Trends are mostly negative at GRSM but mixed at PORE and MORA. Three elements (Ti, V, and Mn) show anomalously large negative trends that invite closer inspection. Figure 3 time-series plots for Ti, V, and Mn at MORA illustrate that the change from PIXE to Mo-XRF lowered both the DL and measured concentrations of these elements and that this method change drives the observed trends. Ti, V, and Mn concentrations show discontinuities at other IMPROVE sites also; these elements are not suitable for trends analysis and are therefore excluded from subsequent discussions. Figure 3 illustrates the danger in analyzing data that are consistently measured near the DL as the data may just be noise.14 More samples were below the DL with PIXE than with Cu-XRF for all three elements, and the DL is not a reliable bound or indicator for the nondetected concentrations; previous work has shown that reported DL were unrealistically low for several IMPROVE XRF measurements.20 The trends are sensitive to the treatment of below DL measurements (e.g., excluding from trends, substituting DL, substituting one-half DL), particularly when nondetects in the original analyses were not equally distributed throughout the time period, as was often the case. The GRSM Rb time-series in Figure 3 shows that the DL for the heavier elements measured by MoXRF throughout the time period were more stable although the DL did decrease over the years, possibly from changes to the detectors, electronics, or calibration procedures. The PORE Na time-series shows that the DL did not improve for all light elements with the switch from PIXE to Cu-XRF in He atmosphere; while the 9251


Article


Table 1. Statistics−trends with Confidence Intervals (CI), Total Concentration Change over the 15 Year Period, And Percent of Records below the DL, for the Original Concentrations, Reanalysis Concentrations, And Ratios of the Original to Reanalysis Concentrationsa


original

reanalysis

element

site

trend (%/yr)

CI

15 yr change

below DL

trend (%/yr)

Na Na Si Si

MORA1 PORE1 PORE1 MORA1

−3.8* −2.8* −2.6* −1.2

(−5.2, −2.5) (−4.6, −0.95) (−4.2, −1.1) (−2.7, 0.32)

−44% −35% −33% −17%

37% 20% 13% 9%

−0.89 −0.19 −0.51 0.05

(−2.2, (−1.8, (−1.8, (−1.5,

Si S S S

GRSM1 PORE1 GRSM1 MORA1

0.89 0.64 −2.9* −2.7*

(−0.2, 2) (−0.25, 1.5) (−3.9, −1.9) (−3.9, −1.5)

14% 10% −36% −34%

6% 1% 0% 0%

−2* 0.06 −2.7* −2.3*

Cl K K

PORE1 PORE1 GRSM1

−3.1* −0.82 −1.5*

(−5.5, −0.69) (−1.7, 0.06) (−2.1, −0.84)

−38% −12% −20%

24% 0% 0%

K Ca

MORA1 MORA1

−1.8* −0.53

(−2.6, −1) (−1.6, 0.52)

−24% −8%

Ca

GRSM1

−1.2*

(−2, −0.31)

Ca Fe Fe Fe Ni Cu Cu Zn Zn Zn As As Se Se Se Rb Sr Sr Sr Pb Pb Pb

PORE1 PORE1 MORA1 GRSM1 PORE1 GRSM1 MORA1 MORA1 PORE1 GRSM1 GRSM1 MORA1 MORA1 GRSM1 PORE1 GRSM1 MORA1 PORE1 GRSM1 GRSM1 PORE1 MORA1

−0.8 −1.5* 1.2 −1.3* 0.81 −2.6* 0.86 −1.8* −0.92 −3* 0.5 1.2 −3.2* −3.7* −0.02 −4.8* 0.49 0.4 −0.53 −5.2* −0.95 −2.4*

(−1.9, 0.27) (−2.9, −0.14) (−0.56, 2.9) (−2.2, −0.44) (−0.61, 2.2) (−3.4, −1.8) (−0.5, 2.3) (−2.9, −0.66) (−2.3, 0.51) (−3.7, −2.4) (−0.76, 1.8) (−0.19, 2.7) (−4.2, −2.1) (−4.4, −2.9) (−1.1, 1.1) (−5.5, −4.1) (−0.44, 1.4) (−0.79, 1.6) (−1.4, 0.35) (−5.9, −4.6) (−2.4, 0.55) (−3.5, −1.1)

ratio 15-yr change

below DL

trend (%/yr)

−13% −3% −7% 1%

10% 1% 6% 2%

−2.9* −1.4* −0.83* −0.41*

(−3, −1) (−0.79, 0.91) (−3.7, −1.7) (−3.5, −1.1)

−26% 1% −34% −29%

1% 0% 0% 0%

3.2* 0.41* −0.07 −0.28*

1.6 −1.2* −1.1*

(−0.47, 3.5) (−2.1, −0.34) (−1.7, −0.41)

27% −17% −15%

23% 0% 0%

−1.8* 0.31* −0.42*

0% 3%

−0.72 0.51

(−1.6, 0.15) (−0.55, 1.6)

−10% 8%

0% 0%

−0.82* −0.32*

−17%

1%

−0.35

(−1.2, 0.5)

−5%

0%

−0.42*

−11% −20% 20% −18% 13% −33% 14% −24% −13% −37% 8% 20% −39% −43% 0% −52% 8% 6% −8% −55% −13% −31%

0% 0% 0% 0% 22% 6% 22% 1% 4% 0% 29% 31% 37% 1% 9% 34% 20% 4% 17% 1% 17% 7%

−1.1 −1.4 1.3 −1.2* −0.82 −1.7* 0.64 −0.8 0.09 −2.5* −1.8* 1.4* −0.33 −2.4* 0.27 0.77* 1.3* −1.1 −0.51 −3.7* −2.3* −2.2*

(−2.2, 0.02) (−2.8, 0.06) (−0.45, 3) (−2.1, −0.33) (−2, 0.37) (−2.3, −0.97) (−0.61, 1.9) (−1.9, 0.37) (−1.3, 1.5) (−3.2, −1.9) (−2.5, −1) (0.33, 2.4) (−1.2, 0.6) (−3.2, −1.6) (−0.7, 1.3) (0.11, 1.5) (0.29, 2.3) (−2.2, 0.1) (−1.3, 0.29) (−4.3, −3.2) (−3.6, −0.93) (−3.2, −1)

−15% −19% 21% −17% −12% −23% 10% −11% 1% −32% −24% 23% −5% −31% 4% 12% 21% −15% −7% −43% −29% −28%

0% 0% 0% 0% 8% 0% 5% 0% 1% 0% 29% 37% 28% 0% 3% 15% 8% 1% 7% 0% 9% 1%

0.33* 0 0.01 −0.02 2.1* −1.1* −0.14 −0.93* −0.75* −0.5* 2* −0.58 −2.2* −1.2* 0.1 −5.5* −0.72* 1.5* 0.04 −1.4* 1.4* 0.16

CI 0.45) 1.4) 0.75) 1.6)

CI (−3.8, −2) (−2.1, −0.75) (−1.4, −0.24) (−0.78, −0.04) (2.7, 3.6) (0.29, 0.54) (−0.14, 0.01) (−0.38, −0.17) (−2.2, −1.5) (0.19, 0.43) (−0.51, −0.33) (−0.94, −0.7) (−0.53, −0.11) (−0.57, −0.27) (0.21, 0.46) (−0.15, 0.13) (−0.1, 0.13) (−0.09, 0.05) (1.5, 2.7) (−1.4, −0.82) (−0.67, 0.4) (−1.1, −0.77) (−1.1, −0.45) (−0.6, −0.4) (1.1, 2.9) (−1.3, 0.12) (−3, −1.4) (−1.4, −1) (−0.35, 0.56) (−6.3, −4.8) (−1.4, −0.09) (1.1, 1.9) (−0.67, 0.76) (−1.7, −1.1) (0.61, 2.2) (−0.21, 0.52)

a Only elements with detection rates >60% are included. An asterisk indicates a trend is significant at the 95% confidence level. Original and reanalysis trends without overlapping CI are highlighted in pink.

with caution; one alternative approach for elements that do not meet these criteria is to determine trend estimates based on only the top percentiles of the concentration data (e.g., 80th percentiles). One objective of this study was to determine if biases in the measurements were consistent across the three sites and if adjustments could be extended to all IMPROVE sites to account for the biases. Ratios of the original to reanalysis concentrations simplify the comparisons and help identify biases in the measurements; Table 1 includes the trends in the daily ratios and CI. Figure 6 shows time-series plots of the original concentrations, reanalysis concentrations, and ratios for S at all three sites (plots like these for all the elements are included in the Supporting Information). Ideally, the ratios would be

based on the original and reanalysis concentrations for all elements measured above the DL at least 80% of the time. This criterion is somewhat crude and excludes some elements that had consistent original and reanalysis trends (Ni at PORE, Cu at MORA, and As at MORA). Nevertheless to obtain more reliable trends from IMPROVE data, we recommend performing trends analysis only on elements that are measured above the DL > 80% of the time, have consistent detection rates over time, and do not display discontinuities in the concentration time-series. These three criteria lead to statistically indistinguishable estimates for the original and reanalysis trends for all the site-element series except Pb at GRSM, where the trend estimates are similar but the CI do not overlap. Trend estimate for elements detected 60% and shows the ratio trends are not consistent across all three sites for any elements except Zn. Zn is a well-detected element but also a ubiquitous contaminant that was previously identified as a problem in the XRF systems.26 The Zn trends in the original analyses may result from additional cleaning measures performed after the contamination was discovered in 2008. The trends in the Zn ratios are negative at all three sites, which is consistent with the early original measurements possibly being biased high by contamination. The variability in the ratio trends across the three sites suggests that no adjustments based on reanalysis at these three sites can be extended to the entire network.

Figure 5. Trend estimates based on the reanalysis data versus original data for elements with >80% detection rate. The thick black line indicates the 1:1 relationship, and the thin black lines indicate 1% above and below the 1:1 relationship.

centered around one, display only random error, and have no temporal trend indicating that the measurements were stable over the entire time period. If the original measurements were consistently biased across the entire network, the ratios are expected to display similar patterns and trends at the three sites. These generalities are complicated by the fact that measurement behavior often varies depending on the measured

Figure 6. Original concentrations, reanalysis concentrations, and ratios of the original to reanalysis concentrations for S at all three sites. Trend estimates along with CI are printed at the bottom left of each graph. 9253


Article


analysis provided by Research Triangle Institute, and carbon analysis provided by Desert Research Institute.


■

(1) Malm, W. C.; Schichtel, B. A.; Pitchford, M. L.; Ashbaugh, L. L.; Eldred, R. A. Spatial and monthly trends in speciated fine particle concentration in the United States. J. Geophys. Res. 1994, 109, D03306. (2) Hand, J. L., Copeland, S. A., Day, D. E., Dillner, A. M., Indresand, H., Malm, W. C., McDade, C. E., Moore, C. T., Pitchford, M. L., Schichtel, B. A., Watson, J. G. Spatial and Seasonal Patterns and Temporal Variability of Haze and its Constituents in the United States: Report V, Colorado State, Fort Collins, CO, 2011. http://vista. cira.colostate.edu/improve/Publications/Reports/2011/2011.htm. (3) Rousseau, R. Detection limit and estimate of uncertainty of analytical XRF results. Rigaku Journal 2001, 18 (2), 33−47. (4) Padilla, R.; Alvarez, A.; Markowicz, D.; Wegrzynek, E.; Chinea Cano, S.; Bamford, A.; Hernandez Torres, D. Quality management and method validation in EDXRF analysis. X-Ray Spectrom. 2007, 36, 27− 34. (5) White, W. H.; Ashbaugh, L. L.; Hyslop, N. P.; McDade, C. E. Estimating measurement uncertainty in an ambient sulfate trend. Atmos. Environ. 2005, 39, 6857−6867. (6) Guide to the Expression of Uncertainty in Measurement; International Organization for Standardization (ISO): Switzerland, 1995. (7) White W. H., Eldred R. A., Feeney P. J., McDade C. E., Perley B. P., Shadoan D. J., Wakabayashi P. H. (2004) Behavior of fine-particle elemental data near the detection limit. Paper presented at Regional and Global Perspectives on Haze: Causes, Consequences and Controversies, Asheville NC. 2004. http://vista.cira.colostate.edu/ improve/publications/graylit/028_xrf_mdl/asheville_xrf.pdf (accessed March 2012). (8) Hyslop, N. P.; White, W. H. An evaluation of Interagency Monitoring of Protected Visual Environments (IMPROVE) collocated precision and uncertainty estimates. Atmos. Environ. 2008, 42, 2691− 2705. (9) Hyslop, N. P.; White, W. H. Estimating precision using duplicate measurements. J. Air Waste Manage. Assoc. 2009, 59, 1032−1039. (10) Henrion, M.; Fischhoff, B. Assessing uncertainty in physical constants. Am. J. Phys. 1986, 54, 791−798. (11) White, W. H. Deteriorating air or improving measurements? On interpreting concatenate time series. J. Geophys. Res. 1997, 102, 6813− 6821. (12) Weatherhead, E. C.; Reinsel, G. C.; Tiao, G. C.; Meng, X. L.; Choi, D. S.; Cheang, W. K.; Keller, T.; DeLuisi, J.; Wuebbles, D. J.; Kerr, J. B.; Miller, A. J.; Oltmans, S. J.; Frederick, J. E. Factors affecting the detection of trends: statistical considerations and applications to environmental data. J. Geophys. Res. 1998, 103, 17149−17161. (13) Hyslop, N. P.; Trzepla, K.; White, W. H. Reanalysis of Archived IMPROVE PM2.5 Samples Previously Analyzed over a 15-Year Period. Environ. Sci. Technol. 2012, 46 (18), 10106−10113. (14) Helsel, R. Nondetects and Data Analysis: Statistics for Censored Environmental Data; Wiley-Interscience: New York, NY, 2005. (15) Chatfield, C. The Analysis of Time Series, 6th ed.; Chapman & Hall/CRC: Boca Raton, 2004. (16) Koenker, R. Quantile Regression; Cambridge University Press: Cambridge, 2005. (17) John, W.; Reischl, G. P. A cyclone for size-selective sampling of ambient air. J. Air Pollut. Control Assoc. 1980, 30, 872−876. (18) Eldred, R. A.; Cahill, T. A.; Feeney, P. J. Particulate monitoring at US national parks using PIXE. Nucl. Instrum. Methods Phys. Res., Sect. B 1987, B22, 289−295. (19) Trzepla-Nabaglo, K.; Wakabayashi, P. H.; White, W. H. Toward establishing the stability of multi-year monitoring of elements in airborne particles. J. Air Waste Manage. Assoc. 2009, 59, 1040−1044. (20) Hyslop, N. P.; White, W. H. An empirical approach to estimating detection limits using collocated data. Environ. Sci. Technol. 2008, 42, 5235−5240.

Figure 7. Trends in the ratios of the original to reanalysis concentrations for the elements with detection limits >60%. Different symbols are used to identify each site. A black circle around a point indicates the slope differs from zero at the 95% confidence level.

This 15 year reanalysis data set confirms that for welldetected (>80% detection rate) elements without previously identified measurement inconsistencies, the original data set provides reliable estimates of trends over the 15-year period from 1995 to 2010. All six elements meeting these criteria at MORA (S, K, Ca, Fe, Zn, Pb) and all eight elements meeting these criteria at PORE (S, K, Ca, Fe, Zn, Se, Sr, Pb) had statistically indistinguishable original and reanalysis trends, while eight of the nine elements meeting these criteria at GRSM (S, K, Ca, Fe, Cu, Zn, Se, Sr, but not Pb) had statistically indistinguishable original and reanalysis trends. We are particularly confident in the trend estimates from 1995 to 2010 for S, K, Ca, Fe, and Zn at sites where these elements are measured above the detection limit ≥80% of the time because the original and reanalysis trends were statistically indistinguishable at all three sites for these elements.

■

ASSOCIATED CONTENT

S Supporting Information *

Time-series graphs of the original, reanalysis, and ratios of original to reanalysis concentrations for all measured elements. The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.est.5b01572.

■

REFERENCES

AUTHOR INFORMATION

Corresponding Author

*Phone: (530)754-8979; e-mail: [email protected]. Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS This work was supported by National Park Service cooperative agreement P11AC91045. IMPROVE is a collaborative association of state, tribal, and federal agencies, and international partners. US Environmental Protection Agency is the primary funding source, with contracting and research support from the National Park Service. The Air Quality Group at the University of California, Davis is the central analytical laboratory, with ion 9254


Article


Environmental Science & Technology (21) Hand, J. L.; Schichtel, B. A.; Malm, W. C.; Frank, N. H. Spatial and Temporal Trends in PM 2.5 Organic and Elemental Carbon across the United States. Adv. Meteorol. 2013, 2013, Article ID 367674; DOI: 110.1155/2013/367674. (22) Hand, J. L.; Schichtel, B. A.; Malm, W. C.; Copeland, S.; Molenar, J. V.; Frank, N.; Pitchford, M. Widespread reductions in haze across the United States from the early 1990s through 2011. Atmos. Environ. 2014, 94, 671−679. (23) Hand, J. L.; Schichtel, B. A.; Malm, W. C.; Pitchford, M. Widespread Reductions in Sulfate Across the United States Since the Early 1990s. In Nucleation and Atmospheric Aerosols, AIP Conference Proceedings. 2013; 1527, 495−498. (24) White, W. H. S interference in XRF determination of Si, 2006. http://vista.cira.colostate.edu/improve/Data/QA_QC/Advisory/ da0011/da0011_S_Si.pdf. (25) Indresand, H.; Dillner, A. Experimental characterization of sulfur interference in IMPROVE aluminum and silicon XRF data. Atmos. Environ. 2012, 61, 140−147. (26) White, W. H. Data Advisory: Sporadic contamination by Zn, 2009. http://vista.cira.colostate.edu/improve/Data/QA_QC/Advisory/ da0005/da0005_blank_Zn_update1.pdf.

9255


Assessing the Suitability of Historical PM2.5 ... - ACS Publications

Recommend Documents