6/15/2006 Linsey DeBell Cooperative Institute for Research in the Atmospheres Interagency Monitoring of Protected Visual Environments Colorado State University Data Validation Historical Report Aerosol Sulfate Data Start: March 1988 Data End: December 2003 Retrieved from the VIEWS Database on Various Dates: 11/2004-3/2005 Background The standard IMPROVE sampler has four independent sampling modules: A, B, and C collect PM2.5 particles (0-2.5 um), and D collects PM10 particles (0-10 um). Module A utilizes a Teflon filter which is analyzed for gravimetric mass and elemental concentrations by XRF and PESA and prior to 12/2001 by PIXE. Module B utilizes a nylon filter and is analyzed primarily for anions by IC. Module C utilizes a quartz filter and is analyzed for organic and elemental carbon by TOR carbon analysis. IMPROVE analyzes the nylon filter for sulfate ([SO4] ug/m3) and the Teflon filter for sulfur ([S] ug/m3) concentrations by IC and PIXE/XRF respectively. In our data validation process it is assumed that all aerosol sulfur is in the form of sulfate and thus 3*[S] should equal [SO4] within measurement uncertainty. The most basic checks on these measurements include scatter plots of [SO4] versus [S] and time series plots of [SO4], [S] and [SO4]/[S]. General Observations Based on Visual Inspection of Time Series Site specific time series plots of [SO4], [S] and [SO4]/[S] indicate a substantial number of sample pairs with [SO4]/[S] ratios far away from the ideal value of 3 and persistent time periods where the central tendency of the ratio is either above or below 3. Most time periods exhibit a dominant bias direction that is visible at most operational sites. However, not all sites follow this network level pattern, possibly indicating that local conditions are overriding some network level phenomena. The time series can be broken into 9 time periods based on the dominant bias direction across the network: Table 1. Major trends in dominant bias direction observed in time series of [SO4]/[S] 1 Affected Time Period 1988-1989 Dominant Bias Direction [SO4]>3*[S] Data Indicator Central tendency of the [SO4]/[S]ratio 1 2 1990-1994 [SO4]>3*[S] 3 1995-early 1997 Early 1997-late 1997 1998-mid 2000 [SO4]<3*[S] Mid 2000-mid 2001 Mid 2001-2002 [SO4]>3*[S] Early 2003-mid 2003 Mid 2003-late 2003 [SO4]>3*[S] 4 5 6 7 8 9 No dominant direction [SO4]<3*[S] [SO4]<3*[S] [SO4]<3*[S] appeared to be greater than 3 at all operational sites Central tendency of the [SO4]/[S] ratio appeared to be greater than 3 at many sites Central tendency of the [SO4]/[S] ratio appeared to be less than 3 at many sites No clear network wide pattern to the typical central tendency of [SO4]/[S] Central tendency of the [SO4]/[S] ratio appeared to be less than 3 at many sites Central tendency of the [SO4]/[S] ratio appeared to be greater than 3 at most sites Central tendency of the S[SO4]/[S] ratio appeared to be less than 3 at many sites Central tendency of the [SO4]/[S] ratio appeared to be greater than 3 at many sites Central tendency of the [SO4]/[S] ratio appeared to be less than 3 at many sites Monthly averages of the [SO4]/[S] ratio for the whole network confirm that the generalized patterns observed in the site specific data validation charts and described in Table 1 are indeed dominant across the network (Figure 1). Site specific examples of the 8 time periods with a dominant bias direction are given below including maps depicting the spatial distribution of sites which followed the network level pattern. The associated PowerPoint file has slides documenting all bias problems at all sites, both those that conform to the network level pattern and those that appear to be local in nature. There are 8 sites along the west coast which have had persistently low [SO4]/[S] ratios from 1990-2003. The high [SO4]/[S] excursions during mid 2000-mid 2001 and early 2003-mid 2003 are visible at some of these sites. The sites involved are documented below under Anomaly 8. An additional observation made during this analysis was that of a cyclical pattern to the [SO4]/[S] ratio on roughly a 1 year cycle. This pattern is present at all sites to varying degrees, in its most extreme expression it is the dominant pattern visible at the site. An example is given below under Anomaly 9. 2 Figure 1. Monthly averages of the [SO4]/[S] ratio for the whole network confirm the generalized patterns observed in the site specific data validation charts. The nine time periods described in Table 1 are indicated by the blue lines. Monthly Network Average of the SO4/S Ratio 4 3.8 3.6 Average SO4/S 3.4 3.2 3 2.8 2.6 2.4 2.2 2 Jan-88 Jan-89 Jan-90 Jan-91 Jan-92 Jan-93 Jan-94 Jan-95 Jan-96 Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 3 Anomaly 1: Significant bias between [SO4] and [S] indicated by high [SO4]/[S] ratios from 1988-1994. Every site in operation in 1988-1989 showed very high [SO4]/[S] ratios for most sample pairs. During 1990- late 1994, the network norm was high [SO4]/[S] ratios, but not all sites had high [SO4]/[S] in all years and some sites had short to extended periods of the reverse situation. For details at each site the reader can refer to the associated PowerPoint file. Example: Figure 2. SAGO1 provides a good example of the high [SO4]/[S] ratios observed at many sites for the period 1988-1994. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 4 Figure 3. All sites identified with a blue circle visually had high [SO4]/[S] ratios for all or part of the period 1990-1994. All sites in operations from 1988-1989 had high [SO4]/[S] ratios. Some sites identified with black circles were not operational during this time period. 5 Anomaly 2: Significant bias between SO4 and S indicated by low SO4/S ratios from late 1995-early 1997 at the majority of the sites. For details at each site the reader can refer to the associated PowerPoint file. Example: Figure 4. UPBU1 provides a good example of the low [SO4]/[S] ratios observed at many sites for the period 1995-early 1997. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 6 Figure 5. All sites identified with a blue circle visually had low [SO4]/[S] ratios for all or part of the period 1995-early 1997. Some sites identified with black circles were not operational during this time period. 7 Anomaly 3: Significant bias between SO4 and S indicated by low SO4/S ratios from early 1998-mid 2000 at the majority of the sites. For details at each site the reader can refer to the associated PowerPoint file. Example: Figure 6. UPBU1 also provides a good example of the low [SO4]/[S] ratios observed at many sites for the period 1998-mid 2000. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 8 Figure 7. All sites identified with a blue circle visually had low [SO4]/[S] ratios for all or part of the period 1998-mid 2000. Some sites identified with black circles were not operational during this time period. 9 Anomaly 4: Significant bias between SO4 and S indicated by high SO4/S ratios during mid-2000 through mid-2001 at the majority of operating sites across most of the network. The sites identified as possibly affected had strong seasonal cycles that made this period look similar to surrounding years. For details at each site the reader can refer to the associated PowerPoint file. Example: Figure 8. REDW1 provides a good example of the high [SO4]/[S] ratios observed at most sites for the period mid 2000-mid 2001. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 10 Figure 9. All sites identified with a blue circle visually had high [SO4]/[S] ratios for the period mid 2000-mid 2001. Some sites identified with black circles were not operational during this time period. 11 Anomaly 5: Significant bias between SO4 and S indicated by low SO4/S ratios from mid 2001-2002 at the majority of the sites. For details at each site the reader can refer to the associated PowerPoint files. Example: Figure 10. KALM1 provides a good example of the low [SO4]/[S] ratios observed at many sites for the period mid 2001-2002. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 12 Figure 11. All sites identified with a blue circle visually had low [SO4]/[S] ratios for all or part of the period mid 2001-2002. Some sites identified with black circles were not operational during this time period. 13 Anomaly 6: Significant bias between SO4 and S indicated by high SO4/S ratios from early-mid 2003 at the majority of the sites. For details at each site the reader can refer to the associated PowerPoint file. Example: Figure 12. YELL2 provides a good example of the high [SO4]/[S] ratios observed at many sites for the period early 2003-mid 2003. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 14 Figure 13. All sites identified with a blue circle visually had high [SO4]/[S] ratios for the period early 2003-mid 2003. Some sites identified with black circles were not operational during this time period. 15 Anomaly 7: Significant bias between SO4 and S indicated by low SO4/S ratios from mid 2003 late 2003 at the majority of the sites. For details at each site the reader can refer to the associated PowerPoint files. Example: Figure 14. KALM1 provides a good example of the low [SO4]/[S] ratios observed at many sites for the period mid 2003-late 2003. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 16 Figure 15. All sites identified with a blue circle visually had low [SO4]/[S] ratios for the period mid 2003-late 2003. Some sites identified with black circles were not operational during this time period. 17 Anomaly 8: Significant persistent bias between SO4 and S indicated by low SO4/S ratios from 1990-2003. This anomaly was limited to west coast sites in Alaska, California, Oregon, and Washington. For details at each site the reader can refer to the associated PowerPoint files. Affected Sites Site Code Site Name State Denali NP Columbia River Gorge Mount Rainier NP Snoqualmie Pass Crater Lake NP AK WA LAVO1 Lassen Volcanic NP CA THSI1 Three Sisters Wilderness OR SOLA1 South Lake Tahoe CA DENA1 CORI1 MORA1 SNPA1 CRLA1 WA IMPROVE Region Alaska Columbia River Gorge Northwest WA Northwest OR Oregon and Northern California Oregon and Northern California Oregon and Northern California Sierra Nevadas 18 Figure 16. DENA1 provides a good example of the persistent low [SO4]/[S] ratios observed at a handful of sites along the west coast from 1990-2003. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 19 Figure 17. 20 Anomaly 9: Cyclical pattern to the SO4/S ratio on roughly a 1 year cycle. In its most extreme form the ratio switches from being consistently above 3 for ~1/2 the time to consistently below 3 the other ~1/2. This pattern is present to varying degrees at all sites; however the severity of the cyclicity is not constant through time at all sites. For details at each site the reader can refer to the associated PowerPoint files. Examples 21 Figure 18 a-b. VOYA2 and DENA1 provide good examples of the more extreme expressions of cyclicity in the [SO4]/[S] ratio observed at all sites. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 22 General Observations Based on Visual Inspection of Scatter Plots Additional analysis was done to better define at what sulfate concentrations the sample pairs with poor agreement between 3*[S] and [SO4] were occurring. The purpose of this analysis was to determine if the bulk of sample pairs with extreme [SO4]/[S] ratios were at very low concentrations. Very low concentrations near the mdl have much higher uncertainty and therefore a higher number of sample pairs with poor agreement might be expected. Symmetrical distribution of the [SO4]/[S] ratios above and below 3 was expected at all concentrations. Ideally a third independent measure of sulfate would be used to explore the relationship between the [SO4]/[S] ratios and aerosol sulfate concentration. Scatter plots of [SO4]/[S] versus [SO4] were examined to see at what sulfate concentrations the high and low ratios were occurring. Looking at network data for 1998-2003 aggregated by IMPROVE region (Figure 19), scatter plots of [SO4]/[S] versus [SO4] show a distinct bias of low ratios at low concentrations and high ratios at high concentrations for most regions. This relationship persists even if one restricts the dataset to those samples whose concentration are at least 10 times the reported minimum detection limit (mdl) and therefore should be well quantified. This relationship was unexpected and was initially suspected of being an artifact of using a dependent variable as the measure of aerosol sulfate concentration. For details for each region the reader is referred to the associated PowerPoint file. To test if the relationship was an artifact of dependent variables, scatter plots of [SO4]/[S] versus total PM10 Mass [PM10] were also examined. While this achieves independence between the “true” and the “predicted” variables in the analysis, it introduces a new set of possible complications since PM10 aerosol and sulfate aerosol may or may not be correlated for a given site or time period. Scatter plots of [SO4]/[S] versus [PM10] for the period 1998-2003 with the data grouped by IMPROVE region showed to a lesser degree the same pattern of low ratios at low concentrations and high ratios at high concentrations for most regions (Figure 20 ). For details for each region the reader is referred to the associated PowerPoint file. 23 Figure 19. Regional scatter plots of [SO4]/[S] versus [SO4] for 1998-2003 show a distinct bias of low ratios at low concentrations and high ratios at high concentrations. This pattern persists even if the dataset is restricted to those sample pairs where both [SO4] and [S] are greater than 10*mdl and therefore should be well quantified. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 24 Figure 20. Regional scatter plots of [SO4]/[S] versus [PM10] for 1998-2003 also show a distinct bias of low ratios at low concentrations and high ratios at high concentrations. The pattern is not as distinct as for [SO4]/[S] versus [SO4], but this could be because [PM10] is often not a direct analog for aerosol sulfate which is confined primarily to the fine aerosol mode. Legend: Sulfate concentration: SO4fVal = [SO4] Sulfur concentration: SfVal = [S] PM10 concentration: MTVAl = [PM10] Sulfate to Sulfur Ratio: SO4_S = [SO4]/[S] 25 Collocated data collection began in 2003 at a handful of sites by adding a fifth module to the standard IMPROVE sampler, referred to within IMPROVE as X modules. As of the writing of this report collocated A & B module data was available for 8 sites from the site specific installation dates through May 2004 (Table 2). The collocated data was analyzed to determine if the pattern of low ratios at low concentrations and high ratios at high concentrations was present for comparisons involving just [S] data or just [SO4] data. Scatter plots of [S]/[Sx] versus [S] and [SO4]/[SO4x] versus [S], where the x indicates the X module, were examined for this pattern. Table 2. IMPROVE Collocated QA Modules Site Name Site Module A Mesa Verde NP MEVE1 X Proctor Maple R. F PMRF1 X Olympic NP OLYM1 X Sac and Fox SAFO1 X Lassen Volcanic NP LAVO1 Mammoth Cave NP MACA1 Big Bend NP BIBE1 Gates of the Mountains GAMO1 Module B X X X X Start Date 8/13/2003 9/3/2003 11/8/2003 11/20/2003 4/21/2003 5/15/2003 9/3/2003 9/24/2003 The same general pattern is not readily apparent in scatter plots of [Sx]/[S] versus [S] (Figure 21). There is a slight indication of a bias towards high ratios at low concentrations but no indication of the reverse at high concentrations. So even if one reversed the ratio, the pattern of low ratios at low concentrations and high ratios at high concentrations is not present. In contrast, in scatter plots of [SO4x]/[SO4] versus [S] the pattern of low ratios at low concentrations and high ratios at high concentrations is clearly present at 2 of the 4 sites (Figure 22). The affected sites are Gate of the Mountains in MT (GAMO1) and Lassen Volcanic National Park (LAVO1) in CA, both of which typically have low sulfate concentrations that very rarely exceed 1.5 ug. Furthermore, scatter plots of Vol/Volx versus [S], where Vol is the regular network B module’s average air volume and Volx is the average air volume for the collocated B module, show the inverse pattern of high ratios at low concentrations and low ratios at low concentrations for the same 2 sites (Figure 23). The inverse relationship between [SO4x]/[SO4] and Vol/Volx, could indicate that there are data quality problems related to our sampling procedures or blank correction process which are affecting our [SO4] measurements and causing the observed bias between [SO4] and [S], particularly at low concentrations. 26 Collocated A Module S comparison 1.6 1.4 1.2 Sx/S 1 OLYM1 PMRF1 SAFO1 0.8 0.6 0.4 0.2 0 0 500 1000 1500 2000 2500 3000 S Figure 21. Scatter plots of [S]/[Sx] versus [S] for sites with collocated data for 20032004 do not show the same pattern as in [SO4]/[S] versus [SO4]. There is a slight indication of a bias towards high ratios at low concentrations but no indication of the reverse at high concentrations. Legend: Regular network sulfur and collocated sulfur concentrations: S = [S] and Sx=[Sx] Sulfur to collocated sulfur ratio: S/Sx = [S]/[ Sx] 27 Collocated B module SO4 comparison 1.5 1.4 SO4x/SO4 quantifiable 1.3 1.2 1.1 MACA1 BIBE1 GAMO1 LAVO1 1 0.9 0.8 0.7 0.6 0.5 0 1 2 3 4 5 6 7 S, ug/m3 Figure 22. A scatter plot of [SO4x]/[ SO4] versus [S] for sites with collocated data for 2003-2004 shows the same general pattern as in [SO4]/[S] versus [SO4]. Legend: Regular network sulfate and collocated sulfate concentrations: SO4 = [SO4] and SO4x=[ SO4x] Collocated sulfate to sulfate ratio: SO4x/ SO4 = [SO4x]/[ SO4] 28 Collocate B Module Flow Rate Comparison 1.5 1.4 1.3 VOLx/VOL 1.2 1.1 MACA BIBE GAMO LAVO 1 0.9 0.8 0.7 0.6 0.5 0 1 2 3 4 5 6 7 S, ug/m3 Figure 23. A scatter plot of [Volx]/[ Vol] versus [S] for sites with collocated data for 2003-2004 shows the reverse pattern as in [SO4]/[S] versus [SO4]. Legend: Regular network sulfur and collocated sulfur concentration: S = [S] Regular network B module flow rate and collocated B module flow rate: network flow = Vol and collocated flow=Volx collocated flow ratio to network flow: VOLx/ VOL Summary of Time Series and Scatter Plots The combined analyses of time series and scatter charts above indicates that the degree of agreement and the direction of bias in the [SO4] to [S] relationship for a sample pair is dependent on when the samples were collected and/or analyzed, where the samples were collected and the sample sulfate concentrations. To quantify these general observations the spatial, temporal and concentration distribution of sample pairs where [SO4] and [S] do not agree within 3σ uncertainty were investigated and are described below. These 29 results were also compared to statistical expectations for a population of samples with accurate measurements and well quantified measurement uncertainties. Statistical Background for Quantitative Checks on Measurement Comparability The Z test and the T test both follow the same general formula and can be used to test if two numbers are equal within their uncertainty. The T test differs from the Z test in that it allows for the populations to have unknown variances. If the underlying populations from which the two numbers are drawn are normal than the test scores will follow a standard normal in the case of known variability and a t distribution with properly calculated degrees of freedom in the case of estimated variability. The formulas for calculating the Z score and T score are: Z score= ([SO4]-3*[S])/√(SO42+(3*S)2) Where represents a known measurement uncertainty T score= ([SO4]-3*[S])/√( SO42+(3* S)2) Where represents a statistically estimated measurement uncertainty In our case we are comparing two independent measurements, [SO4] and 3*[S], with the assumption that their difference should be 0 within measurement uncertainty. We are treating each measurement as an estimate of the true atmospheric sulfate value and the uncertainty that is uniquely reported for each sample as an estimate of the true variance for that measurement reported in terms of the standard deviation. Since our measurement uncertainties are based on a theoretical understanding we are treating the samples as having known variability and therefore are loosely referring to them as Z scores. The Z scores indicate how many standard deviations of the difference apart [SO4] is from 3*[S]. This interpretation of the Z score is independent of any assumptions about distributions of the underlying populations or the test scores. Z scores can be positive or negative. A positive Z score indicates that the [SO4] value is greater than 3*[S]. A negative Z score indicates that [SO4] is less than 3*[S]. If the measurement errors are symmetrical, than the Z scores will also be symmetrical. Furthermore, if the measurement errors are distributed normally than the Z scores will follow a standard normal distribution. If neither is the case, than the Z scores will still represent a standardized score which measures the distance, in standard deviations of the difference, between the paired samples. However no assumptions about the distribution of the Z scores can be made independent of assumptions about the sample populations. In this case the population of interest is not our time series of [SO4] and [S], which follow an approximately log normal distribution, but the theoretical population of all potential [SO4] and [S] samples that could have been collected at the same point in time and space as our sample date of interest. On a theoretical level, assuming all S is in the form of sulfate and a well mixed air mass, we would expect these potential measurements to both pull from a single population. Additionally, we would expect our measurements to only have unbiased random errors associated with sampling and analysis. So under ideal 30 sampling and analytical conditions, we would expect the calculated Z scores to minimally follow a symmetrical distribution and possibly a standard normal distribution. Additionally, according to Chebychev's rule, in any distribution the proportion of scores between the mean and k standard deviations is at least 1-1/k2 scores. So even if our test scores do not follow a normal distribution, if all or our parameters are reasonably accurate than we can minimally expect that at least 89% of the scores would reside symmetrically between the mean, 0, and ±3. If the test scores do follow a normal distribution than 99% of the scores would reside between ±3. For the purposes of this report, sample pairs with calculated Z scores outside of the range [-3, 3], which is comparable to pairs that are not equivalent within 3 uncertainty, are being referred to as “outlier” pairs. Expectations The dataset was explored using calculated Z scores for all samples with reported [S], S, [SO4] and SO4 to see if certain assumptions were met. It was assumed that given accurate measurements and well estimated measurement uncertainty that the following would be true: At most 10% of the sample pairs should be outliers The Z scores should be symmetrically distributed above and below 0 This symmetry should persist through time, space and all quantifiable (concentration>10*mdl) aerosol concentrations Methods To test if the outlier sample pairs were distributed evenly in bias direction, time, space, and aerosol sulfate concentration the samples were sorted and aggregated in the following ways: Bias Direction--The dataset was aggregated into two groups, those with Z<-3 indicating that 3*[S]>> [SO4] and those with Z>3 indicating the reverse. The sample pairs with Z in [-3,3] were excluded from this analysis. Time--The data was aggregated by year and then year and month to look at how the total number of outliers pairs, outlier pairs with Z<-3 and outlier pairs with Z>3 relative to the number of valid sample pairs changed through time. Space--The data was aggregated by year and site and then month and site for 2003 to look at how the total number of outliers pairs, outlier pairs with Z<-3 and outlier pairs with Z>3 relative to the number of valid sample pairs changed through space. Aerosol Concentration--Based on visual inspection of the [SO4]/[S] versus [SO4] and [PM10] scatter plots, arbitrary concentration cut points of 1 ug and 10 ug were selected for [SO4] and [PM10] respectively. These values seemed to roughly coincide with the inflection points in the curves where the bias shifted from [SO4]/[S]<3 to [SO4]/[S]>3. In each case for [SO4] and [PM10], the data was aggregated into two groups based on these 31 cut points and the total number of outliers pairs, outlier pairs with Z<-3 and outlier pairs with Z>3 relative to the number of valid sample pairs were compared between the groups. Results When the dataset was taken as a whole, that is all sites for the period 1988-2003, the expectations of less than 10% of the valid samples being outliers and even distribution of the outliers in bias direction and across [SO4] and [PM10] concentrations were met. However, when the dataset was further broken down by site, year and/or month or concentration then the expectations for even distribution and percent outliers were no longer met for the data subsets. Consistent with our expectations of good measurements: Less than 10% of all valid samples outliers for whole dataset Across the network for the period 1988-2003, 8.8% of the valid sample pairs (those with all 4 necessary parameters reported) were outliers, indicating that on the whole we do not have an unexpected number of outlier sample pairs (Tables 3 and 4). Symmetrical distribution of low and high Z scores for whole dataset Across the network for the period 1988-2003, the outlier sample pairs were fairly evenly split between those with Z scores above and those below 0 (Tables 3 and 4). Of the outlier sample pairs, 53% of the pairs had Z<-3 and 47% had Z>3. Symmetrical distribution of outlier samples between low and high [SO4] concentrations Total counts of outliers were fairly evenly split when the outlier population is split into two populations based on [SO4] being below or above 1 μg/m3 with some indication of bias towards higher numbers of outlier pairs at lower [SO4] concentrations (Tables 3 and 4). This distribution could easily be shifted by picking a lower [SO4] concentration as the split point. Of the outlier sample pairs, 59% and SO4<1 μg/m3 and 41% had [SO4]>1 μg/m3. The percentages of samples classified as high or low concentration were similar in the non-outlier population (Tables 3 and 4). Symmetrical distribution of outlier samples between low and high [MT] concentrations Total counts of outliers were also fairly evenly split when the outlier population is split into two populations based on [PM10] being below or above 10 ug (Tables 3 and 4). Of the outlier sample pairs, 52% of the pairs had [PM10]<10, 43% had [PM10]>=10 and the remainder had null [PM10] values. The percentages of samples classified as high or low concentration were similar in the non-outlier population (Tables 3 and 4). Summary Statistics for the period 1988-2003 Table 3. Counts of samples meeting certain criteria including tests on existence, concentration, and Z score values. Dataset Subset Conditions Network Count for 1988-2003 Low Outliers Samples with Z<-3 5090 High Outliers Samples with Z>3 4499 32 Total Outliers Low Outliers with Low PM10 mass High Outliers with Low PM10 mass Total Outliers with Low PM10 mass Low Outliers with High PM10 mass High Outliers with High PM10 mass Total Outliers with High PM10 mass Low Outliers with Low sulfate mass High Outliers with Low sulfate mass Total Outliers with Low sulfate mass Low Outliers with High sulfate mass High Outliers with High sulfate mass Total Outliers with High sulfate mass Total Valid Samples Total Potential Samples Total Non-Outlier Samples Total Non-Outlier Samples with Low PM10 mass Total Non-Outlier Samples with High PM10 mass Total Non-Outlier Samples with Low sulfate mass Total Non-Outlier Samples with High Sulfate mass Valid Samples with Low PM10 mass Valid Samples with High PM10 mass Valid Samples with Low sulfate mass Valid Samples with High Sulfate mass Total Outliers: Samples with Z not in [-3,3] Samples with Z<-3 and PM10<10 ug Samples with Z>3 and PM10<10 ug Samples with Z not in [-3,3] and PM10<10 ug Samples with Z<-3 and PM10>10 ug Samples with Z>3 and PM10>10 ug Samples with Z not in [-3,3] and PM10>10 ug Samples with Z<-3 and SO4<1 ug 9589 Samples with Z>3 and SO4<1 ug 1634 Samples with Z not in [-3,3] and SO4<1 ug Samples with Z<-3 and SO4>1 ug 5631 Samples with Z>3 and SO4>1 ug 2864 Samples with Z not in [-3,3] and SO4>1 ug Samples with SO4, SO4UNC, S, and SUNC all not null Total Records (all records) Samples with Z in [-3,3] Samples with Z in [-3,3] and PM10<10 ug Samples with Z in [-3,3] and PM10>10 ug Samples with Z in [-3,3] and SO4<1 ug Samples with Z in [-3,3] and SO4>1 ug Samples with PM10<10 ug 3957 Samples with PM10>10 ug 44511 Samples with SO4<1 ug 62263 Samples with SO4>1 ug 46989 3348 1641 4989 1514 2649 4163 3997 1093 109257 144336 99668 55222 40348 56632 43032 60211 33 Table 4. Relative counts of samples, expressed as percentages, meeting certain criteria including tests on existence, concentration, and Z score values. Network percentages for Formula Values 1988-2003 Percentage of Total Outliers (# Samples where Z<-3)/( # Samples where Z 53.1% that are Low Outliers not in [3,3])*100% Percentage of Total Outliers (# Samples where Z>3)/( # Samples where Z 46.9% that are High Outliers not in [-3,3])*100% Percentage of Valid Samples ( # Samples where Z not in [-3,3])/(# Samples 8.8% that are Low or High Outliers where Z can be calculated) *100% Percentage of Total Outliers (# Samples where Z<-3 and MT<10)/ (# 67.1% with Low PM10 Mass that are Samples where Z not in [-3,3] and MT<10) Low Outliers *100% Percentage of Total Outliers (# Samples where Z>3 and MT<10)/ (# 32.9% with Low PM10 Mass that are Samples where Z not in [-3,3] and MT<10) High Outliers *100% Percentage of Total Outliers (# Samples where Z not in [-3,3] and 52.0% that have Low PM10 Mass MT<10)/(# Samples where Z not in [3,3])*100 Percentage of Total Outliers (# Samples where Z<-3 and MT>10)/ (# 36.4% with High PM10 Mass that Samples where Z not in [-3,3] and MT>10) are Low Outliers *100% Percentage of Total Outliers (# Samples where Z>3 and MT>10)/ (# 63.6% with High PM10 Mass that Samples where Z not in [-3,3] and MT>10) are High Outliers *100% Percentage of Total Outliers (# Samples where Z not in [-3,3] and 43.4% that have High PM10 Mass MT>10)/(# Samples where Z not in [3,3])*100 Percentage of Total Outliers (# Samples where Z<-3 and SO4<1)/ (# 71.0% with Low Sulfate Mass that Samples where Z not in [-3,3] and SO4<1) are Low Outliers *100% Percentage of Total Outliers (# Samples where Z>3 and SO4<1)/ (# 29.0% with Low Sulfate Mass that Samples where Z not in [-3,3] and SO4<1) are High Outliers *100% Percentage of Total Outliers (# Samples where Z not in [-3,3] and 58.7% that have Low Sulfate Mass SO4<1)/(# Samples where Z not in [3,3])*100 Percentage of Total Outliers (# Samples where Z<-3 and SO4>1)/ (# 27.6% with High Sulfate Mass that Samples where Z not in [-3,3] and SO4>1) are Low Outliers *100% Percentage of Total Outliers (# Samples where Z>3 and SO4>1)/ (# 72.4% with High Sulfate Mass that Samples where Z not in [-3,3] and SO4>1) are High Outliers *100% Percentage of Total Outliers (# Samples where Z not in [-3,3] and 41.3% 34 that have High Sulfate Mass Percentage of Potential Samples that are Valid Samples Percentage of Non-Outlier Samples that have low PM10 mass Percentage of Non-Outlier Samples that have High PM10 mass Percentage of Non-Outlier Samples that have low Sulfate mass Percentage of Non-Outlier Samples that have High Sulfate mass Percentage of Valid Samples that have Low PM10 mass Percentage of Valid Samples that have High PM10 mass Percentage of Valid Samples that have Low Sulfate mass Percentage of Valid Samples that have High Sulfate mass SO4>1)/(# Samples where Z not in [3,3])*100 (# Samples where Z not null)/(# of potential samples)*100% 75.7% (# samples where Z in [-3,3] and MT<10)/(# samples where Z in [-3,3]) *100% 55.4% (# samples where Z in [-3,3] and MT>10)/(# samples where Z in [-3,3]) *100% 40.5% (# samples where Z in [-3,3] and SO4<1)/(# samples where Z in [-3,3]) *100% 56.8% (# samples where Z in [-3,3] and SO4>1)/(# samples where Z in [-3,3]) *100% 43.2% (# samples where MT<10)/(# Samples where Z not null) (# samples where MT<10)/(# Samples where Z not null) (# samples where MT<10)/(# Samples where Z not null) (# samples where MT<10)/(# Samples where Z not null) 55.1% 40.7% 57.0% 43.0% Inconsistent with our expectations of good measurements: Non-symmetrical distribution of low and high Z scores within low and high concentration groups While the total number of outliers was fairly evenly distributed between the low and high concentration groups, the bias direction of the outliers within the groups is unevenly distributed regardless of whether [SO4] or [PM10] is used to define the concentration groups. At low concentrations, ~70% of the sample pairs had Z<-3 whereas at high concentrations ~30% of the sample pairs had Z<-3 (Tables 3 and 4). In other words, outlier pairs at low concentrations typically had 3*[S] >>[SO4] and the reverse is true at higher concentrations. This is in line with the general observations made looking at scatter plots of [SO4]/[S] versus [SO4] or [PM10]. Relative number of outliers more than 10% for a quarter of the months from 19882003 The total number of outliers, expressed as a percentage of the valid samples, has not been consistent over time (Figure 24). Aggregating the data by year and month, the total number of outliers has ranged from 1.4% to 100% of the valid samples for a given month with a median value of 6.1%. Of the 190 months of data collection, 25% of them have had outlier sample pairs makeup over 10% of the sample pairs. That is 25% of the time 35 more than 10% of the valid sample pairs consist of [SO4] and 3*[S] measurements which are not equivalent within 3σ uncertainty. Additionally, the clustering of outliers in particular months does not appear to be random. There is a roughly seasonal cycle to the percentage of total outliers, with the peak percentages typically occurring in summer or fall. Non-symmetrical distribution of low and high Z scores for most months Aggregating the data by year, month and by bias direction results in even more drastic temporal patterns (Figure 24). During a given month the outlier sample pairs are rarely evenly distributed in terms of bias direction. The period 1988-2003 can be broken into about 10 distinct periods lasting from months to years based on which direction of bias was dominant as indicated by the percentage of outlier sample pairs with Z<-3 or those with Z>3. These are the same time periods give or take a couple months as those identified through visual inspection of time series of 1) [SO4]/[S]for each site and (Table 1) 2) monthly network averages of [SO4]/[S] (Figure 1). The only exception is time period # 3 which was not identified through visual inspection and is only a couple months long. Table 5. Major trends in the dominant bias direction observed in the time series of the percentage of outlier samples with Z<-3 or with Z>3 calculated each month for the whole network. These time periods are marked with blue lines in figure 24. Time Affected Time Dominant Bias Data Indicator Period # Period Direction 1 1988- late 1994 [SO4]>>3*[S] Majority of outliers have Z>3 2 Late 1994-late [SO4]<<3*[S] Majority of outliers have Z<-3 1995 3 Late 1995 [SO4]>>3*[S] Majority of outliers have Z>3 4 1996-early [SO4]<<3*[S] Majority of outliers have Z<-3 1997 5 Early 1997No dominant bias outliers evenly split between those that early 1998 direction have Z>3 and those that have Z<-3 6 Early 1998-mid [SO4]<<3*[S] Majority of outliers have Z<-3 2000 7 Mid 2000-mid [SO4]>>3*[S] Majority of outliers have Z>3 2001 8 Mid 2001-2002 [SO4]<<3*[S] Majority of outliers have Z<-3 9 Early 2003-mid [SO4]>>3*[S] Majority of outliers have Z>3 2003 10 Mid 2003-late [SO4]<<3*[S] Majority of outliers have Z<-3 2003 This network wide look at the dominant direction of bias in the outlier sample pairs captures all the major trends identifiable in the site specific time series plots of [SO4]/[S]. Additionally, relative percentages of outlier pairs may provide an objective metric for testing overall performance in terms of [SO4] and [S] comparability. 36 The relative number of outliers for a year was more than 10% for a changing subset of sites even when the relative number of outliers for the network as a whole was less than 10%. Additionally, there was non-symmetrical distribution of low and high Z scores for most sites with more than 10% outliers. Aggregating the data by year, site, % total outliers, and by dominant bias direction indicates that within a given time period the outlier pairs are clustered at particular sites (see appendix A for details). In addition, at the sites where outliers make up more than 10% of the valid samples, a dominant bias direction is usually apparent with between 60100% of the outlier pairs being on one side of the distribution. The dominant bias direction of particular sites is not always in-line with the network wide dominant bias direction. For example in 1991, the network wide dominant bias direction was SO4>>3*S but Washington DC, which had over 10% relative outliers, had over 60% of the outliers have SO4<<3*S (see Appendix A). Local deviations from the network wide pattern like this one suggest that local conditions can override whatever is causing the network wide pattern in terms of dominant bias direction. There are some sites that have significant numbers of outlier pairs most years and there are also sites which have never exceeded 10% (see appendix A and B for details). The regional patterns shift from year-to-year in terms of which regions have sites with high outlier counts and which sites have a particular bias direction. There is no obvious pattern between the type of [SO4] to 3*[S] disagreement and site location. Maps displaying which sites had over 10% outlier sample pairs and the dominant bias direction for that site are shown for each year in Appendix A at the end of this report. A table listing all sites with no years with % total outliers≥10% is included in Appendix B at the end of this report. 37 % of valid samples that are outliers % of valid samples that are outliers with Z<-3 % of valid samples that are outliers with Z>3 SO4 artifact 30 5 25 20 % Outliers -5 15 -10 10 -15 5 0 Jan-88 Jan-89 Jan-90 Jan-91 Jan-92 Jan-93 Jan-94 Jan-95 Jan-96 Jan-97 Jan-98 Jan-99 Jan-00 Jan-01 Jan-02 Jan-03 Date Figure 24. Looking network wide, the percentage of valid sample pairs with Z<-3 or Z>3 (Navy), with Z<-3 (red) and Z>3 (pink) for a given month has varied significantly over time. Rarely are the outlier pairs evenly distributed between those with SO4>>3*S (pink) and those with SO4<<3*S (red). This network wide look at the outliers captures all the major bias trends observable in site specific time series. From 6/2002 forward when the blank corrections began being applied to monthly rather than quarterly data batches, there is a rough correlation (R2=0.5) between the SO4 artifact concentration and the % valid samples with Z<-3. Other Interesting Observations A major shift in dominant bias direction from [SO4]>>3*[S] (Z>3) to [SO4]<<3*[S] (Z<-3) occurred in November 1994. It has been proposed that many of the sample pairs with [SO4]>>3*[S] from east coast sites with high aerosol sulfate concentrations were due to S loss from the Teflon filters due to the mask which reduced the exposed surface area of the filter. However, this shift precedes the removal of masks from the east coast sites, which did not begin until spring 1995. 38 -20 SO4 artifact ug/filter 0 Starting in 6/2002, when the switch was made from quarterly to monthly blank subtraction, there is a rough (R2=0.5) correlation between the % valid samples with Z<-3 and the blank correction concentration (ug/filter) (Figure 24). Looking at the collocated [SO4] data for 2003-2004, there appears to be a negative correlation between the sulfate offsets ([SO4x]/[SO4]) and the flow offsets (Volx/Vol), particularly at low values. This suggests there may be a relationship between flow offsets, blank corrections and how accurately our reported [SO4] concentrations represent the true atmospheric aerosol SO4 concentrations at low concentrations. Using [SO4] and [S] comparability as a metric of data quality, our performance is degrading rather than improving. Looking at the percentage of samples which qualify as outliers, 2003 at 13% is the worst year since 1989 at 24% when there were known [SO4] problems. Looking only at the post-validation data, it cannot be determined if this is the result of changes in data reporting or to changes in measurement quality. Table 6. Percentage of valid samples which qualify as outliers for each calendar year Year # Samples with Z<-3 # Samples # Samples with Z>3 with Z<-3 or Z>3 # Valid % Outlier Samples Sample Pairs 1988 11 543 554 2286 24.2% 1989 6 213 219 2928 7.48% 1990 18 129 147 3515 4.2% 1991 40 126 166 3949 4.2% 1992 26 161 187 4442 4.2% 1993 46 246 292 4583 6.4% 1994 69 227 296 4969 6.0% 1995 140 125 265 5174 5.1% 1996 225 32 257 5199 4.9% 1997 115 109 224 5434 4.1% 1998 330 65 395 4992 7.9% 1999 211 74 285 4905 5.8% 2000 434 415 849 8104 10.5% 2001 646 590 1236 13607 9.1% 2002 1229 649 1878 16873 11.1% 2003 1544 795 2339 18297 12.8% 39 Qualitatively checking the relative number of outliers with Z<-3, the relative number of outliers with Z>3 and the relative number of outliers with Z<-3 or Z>3 seems to be a valid way of quickly detecting all significant bias problems. Various data aggregates can be used to look at problems on a site specific, regional or network wide basis. They also provide a reasonably defensible basis for setting standard expectations or acceptance criteria for valid data. Conclusions Taking the IMPROVE dataset as a whole, the sulfur measurements do not have an unexpected number of pairs that are more than 3σ apart. However, once the temporal, spatial, and concentration dimensions of the population are taken into account the outlier sample pairs are clustered in certain spatial or temporal subsets of the data. The clusters are typically not symmetrical in terms of bias direction. The bottom line is that the expectation of random distribution of the outliers is violated and the expectation of at most 10% of the population being in extreme disagreement is also violated for all dataset fractions that take the key dimensions of the population into account. Therefore, I would suggest that the poor agreement in many of these sample pairs is not due to random chance but the reflection of real analytical and/or sampling problems specific to certain conditions. More generally, these results indicate that either the SO4 and/or the S measurements are not accurate under some conditions and/or one or both of the estimated measurement uncertainties are an under prediction of the true uncertainty under some conditions. However, they also indicate that the measurements are consistently accurate under some conditions and that the estimated uncertainties are accurate or even an overestimate of the actual uncertainty under some conditions. The network wide patterns in terms of dominant bias direction are likely due to problems in our analytical process or to sampling media. It appears that local conditions can override the network wide patterns suggesting that sampling conditions are a key component to producing comparable [SO4] and [S] measurements. Sampling conditions could alternately play the role of enhancing the network level signal at particular sites or damping it—either reducing or increasing the comparability of those particular [SO4] and [S] measurements. Additional investigation is required to understand what factors in terms of analytical and sampling equipment and procedure might be negatively or positively impacting the comparability of our measurements. The fact that low Z scores dominate at low concentrations and high Z scores dominate at high concentrations suggests there may be different underlying problems causing sample pairs with [SO4]<<3*[S] (low Z scores) and those with [SO4]>>3*[S] (high Z scores). Furthermore, the correlation between sulfate offsets and flow offsets in the collocated data and the correlation between the percentage of valid samples with Z<-3 and blank concentration hint at a connection between flow problems, blank corrections and the poor agreement between [SO4] and [S] at low concentrations, even those well over 10*mdl, for at least the recent past. 40 Appendix A. Spatial Distribution of Sites with High Outlier Counts All sites identified with a pink, red or blue circle had at least 10% of their valid sample pairs for the calendar year have Z Scores outside of the range [-3,3]. Legend: ● 0-40% of the outlier pairs had Z<-3 (SO4<<3*S) ● 41-60% of the outlier pairs had Z<-3 (SO4<<3*S) ● 61-100% of the outlier pair had Z<-3 (SO4<<3*S) ● IMPROVE site, not necessarily operational during specified time period 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 Appendix B. Sites with no site-year combinations with over 10% outliers (excluding 1988-1989) Site Code Site Name State IMPROVE Region ARCH1 Arches NP UT Colorado Plateau BADL1 Badlands NP SD Northern Great Plains BALD1 Mount Baldy AZ Mongollon Plateau BAND1 Bandelier NM NM Colorado Plateau BOAP1 Bosque del NM Mongollon Apache Plateau CHER1 Cherokee OK Mid South Nation CRES1 Crescent Lake NE Central Great Plains CRMO1 Craters of the ID Hells Canyon Moon NM ELDO1 El Dorado MO Central Great Springs Plains GICL1 Gila NM Mongollon Wilderness Plateau GRCA2 Hance Camp AZ Colorado at Grand Plateau Canyon NP HOOV1 Hoover CA Sierra Nevadas IKBA1 Ike's AZ Mongollon Backbone Plateau ISRO1 Isle Royale MI Boundary NP Waters JEFF1 Jefferson NF VA Appalachia MAVI1 Martha's MA Northeast Vineyard PEFO1 Petrified AZ Mongollon Forest NP Plateau PINN1 Pinnacles NM CA California Coast QUCI1 Quaker City OH Ohio River Valley RAFA1 San Rafael CA California Coast RMHQ1 Rocky CO Central 56 STAR1 WHPE1 WIMO1 YELL1 ZICA1 Mountain NP HQ Starkey Wheeler Peak Wichita Mountains Yellowstone NP 1 Zion Canyon Rockies OR NM OK WY UT Hells Canyon Central Rockies Mid South Northern Rockies Colorado Plateau 57 Appendix D Associated files for additional detail: Year-Site_Zscore.xls Network_Zscore.xls 2003_Monthly_BiasMaps.ppt Yearly_BiasMaps.ppt ZScore_TimeSeries_88-03.ppt SO4_SvsSO4_Scatter_Regional_98-03.ppt SO4_SvsMT_Scatter_Regional_98-03.ppt SO4_S_TimeSeries_88-03.ppt SO4_S_TimeSeries_98-03.ppt 58