HomeMy WebLinkAboutStatistical Methods for Determining BTVs_01202017 REVISED DRAFT
Statistical Methods for
Developing Reference Background
Concentrations for
Groundwater and Soil
at Coal Ash Facilities
January 2017
Prepared By:
HDR Engineering, Inc.
440 S. Church St, Suite 1000
Charlotte, NC 28202
and
SynTerra Corporation
148 River Street, Suite 220
Greenville, South Carolina 29601
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
i | Page
CONTENTS
INTRODUCTION .......................................................................................................................... 1
PART I – DESCRIPTION OF BACKGROUND DATA SETS ................................................... 4
Groundwater .......................................................................................................................... 4
Soil ......................................................................................................................................... 4
PART II – PRELIMINARY DATA ANALYSIS .......................................................................... 5
1. Descriptive Statistics ............................................................................................... 5
2. Graphical Analysis ................................................................................................... 5
3. Identify Outliers ........................................................................................................ 6
4. Identifying Data Distributions ................................................................................... 6
5. Evaluating Background Groundwater Data ............................................................. 7
6. Autocorrelation ........................................................................................................ 7
7. Seasonality .............................................................................................................. 8
8. Trends ...................................................................................................................... 8
9. Additional Methods for Identifying Trends in Background Groundwater Data ........ 9
10. Determining Baseline Period for Background Wells .............................................. 10
PART III – TESTING FOR SUB-GROUPS IN BACKGROUND GROUNDWATER DATA ... 11
Graphical Analysis ............................................................................................................... 11
Analytical Tests for Comparing Sub-Groups ....................................................................... 12
Tests for Identifying Differences Among Sub-Groups ......................................................... 12
PART IV – DEVELOPMENT OF BTVs FOR CONSTITUENTS IN GROUNDWATER AND
SOIL ...................................................................................................................... 13
REFERENCES ........................................................................................................................... 15
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
ii | Page
FIGURES
Figure 1. Box-and-Whisker Plot
Figure 2. Quantile-Quantile (Q-Q) Plot
Figure 3. Scatter Plot of Time versus Concentration
Figure 4. Sample Autocorrelation Function
Figure 5. Scatter Plots of Time versus Concentration Illustrating Seasonality
Figure 6. Scatter Plots of Time versus Concentration Illustrating Trends
Figure 7. Piece-Wise Polynomial Regression Output Exhibiting Multiple Trends
Figure 8. Empirical Distribution Plot Comparing Constituent Concentrations between Two
Seasons
TABLES
Table 1. Chemical Parameters Analyzed in Groundwater and Soil
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
1 | Page
INTRODUCTION
In 2015, the North Carolina Coal Ash Management Act (CAMA) required the preparation of a
Comprehensive Site Assessment Report (CSA) for each regulated facility. The purpose of the
CSA was to identify the source and cause of exceedances of regulatory standards, potential
hazards to public health and safety, and identify receptors and exposure pathways. The CSA
was conducted in accordance with a conditionally approved Work Plan to meet the
requirements of 15A NCAC 02L .0106(g), which includes an assessment of the horizontal and
vertical extent of soil and groundwater contamination for all contaminants confirmed to be
present in groundwater in exceedance of groundwater quality standards.
Regulations regarding North Carolina groundwater quality standards provided in T15A NCAC
02L .0202. Section (b)(3) of the regulation state that:
Where naturally occurring substances exceed the established standard, the standard shall be
the naturally occurring concentration as determined by the Director.
For soil and groundwater assessments under the CAMA, naturally occurring concentrations of
constituents need to be determined in order to complete horizontal and vertical delineations
required as a basis for development of Corrective Action Plans. The horizontal and vertical
extent of constituent migration cannot be determined until naturally occurring background
concentrations are known.
This document serves as a framework for a consistent technical approach which will be utilized
for Duke Energy sites in North Carolina to determine proposed provisional background
threshold values (PPBTVs1) for naturally occurring constituents in groundwater and soil. For
the purpose of establishing background threshold values (BTVs1) at this time, the value which
represents the upper threshold value from the upper tail of the data distribution for a given
constituent will be considered the value representative of a naturally occurring concentration, or
the PPBTV. The process for evaluating background concentrations over time is iterative;
therefore, as additional background data is collected, the approach for developing BTVs may
be reviewed and potentially modified with consideration of expanded data sets, changes in data
set distribution, and input from the North Carolina Department of Environmental Quality
(NCDEQ).
For groundwater, non-filtered (total) results will be used to establish BTVs. In general,
groundwater data will not be included in the development of BTVs when turbidity of the
groundwater sample was reported to be greater than 10 nephelometric turbidity units (NTU) or
when pH is greater than 8.5. Professional judgment can be used to retain data that does not
meet these criteria. However, the decision to retain data that does not satisfy these criteria
must be documented; such as, concurrence with NCDEQ that naturally occurring pH is greater
than 8.5 in the unit being evaluated. Background locations for groundwater were identified for
each site in the CSA Reports and/or Corrective Action Plans (CAPs). Other wells unaffected by
1 The terms PPBTV and BTV are used interchangeably in this document. The term BTV is used in the EPA
ProUCL User guide.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
2 | Page
plant operations may also be used to augment the background data set with agreement from
NCDEQ.
For soil, only samples collected above the water table and at locations not influenced by Plant
operations will be included in the calculation of BTVs. Site-specific soil sampling locations and
intervals are described in the CSA Work Plans.
The methods for developing PPBTVs described in this document are based on the US
Environmental Protection Agency (USEPA) “Unified Guidance” (USEPA 2009), USEPA’S
Guidance for Comparing Background and Chemical Concentrations in Soil for CERCLA Sites
(USEPA 2002), and the ProUCL Technical Guide (USEPA 2015). In addition, the North
Carolina Division of Water Quality (NCDWQ) technical assistance document for Evaluating
Metals in Groundwater at DWQ Permitted Facilities (NCDWQ 2012) was also referenced.
The methods described in this document are intended to serve as guidelines to develop BTVs.
The use of the upper tolerance limit (UTL) to establish BTVs for constituents analyzed during
assessment monitoring is consistent with NCDEQ Guidance as well as the USEPA’s Unified
Guidance (2009). The UTL will be evaluated as the statistic for development of groundwater
and soil BTVs. BTVs will be developed for a select group of constituents derived from the list of
parameters investigated as part of CAMA (Table 1). The UTL will be used to represent an
upper limit for naturally occurring concentrations such that values exceeding this limit may be
indicative of groundwater and soil impacts.
Naturally occurring concentrations determined by the process presented in this document will
be submitted to the NCDEQ Division of Water Resources for determination of the PPBTVs.
Site-specific reports documenting the procedures, evaluations, and calculations will be
prepared and submitted to NCDEQ. Following NCDEQ’s approval of the PPBTVs, the PPBTVs
will be used as groundwater and soil standards when the values exceed concentrations
appearing in T15A NCAC 02L .0202(g) or the Interim Maximum Allowable Concentrations
(Appendix #1 to T 15A NCAC 02L) for groundwater or Preliminary Soil Remediation Goals as
described in Section 4 of the NCDEQ 2015 Inactive Hazardous Sites Program Guidelines for
Assessment and Cleanup (NCDEQ 2015) for soil.
This document consists of four parts describing the process for establishing BTVs for
constituents in groundwater and soil:
Part I – Description of Background Data Sets
Part I provides discussion of groundwater and soil sample collection, background data set
attributes, and preliminary treatment of background data.
Part II – Preliminary Data Analysis
Part II includes analyses used to assess and transform data (where necessary) for use in
producing appropriate UTLs. This analysis includes screening data sets for outliers, fitting
data sets to distribution models, assessing data for temporal variability, and
appropriateness of the period of record (sampling period).
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
3 | Page
Part III – Testing for Sub-Groups in Background Groundwater Data
Part III summarizes the approach for testing data sets for distinct sub-groups. If testing
indicates presence of subgroups, the same steps described in Part I can be applied to the
partitioned data to better understand the distribution of the samples within a sub-group for
each constituent.
Part IV – Development of BTVs for Constituents in Groundwater and Soil
Part IV documents the steps for producing UTLs for each constituent for groundwater and
soil.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
4 | Page
PART I – DESCRIPTION OF BACKGROUND DATA SETS
Background data sets vary by size at each of the sites. Background groundwater data has
been collected over a period of time at multiple locations per site. Background soil samples
were primarily collected as part of the CSA activities. The following sections describe the
groundwater and soil samples. Additional details regarding site-specific data sets have been
provided in the CSA, CAP 1 and CAP 2, supplemental reports and electronic data submittals
for each site. The data sets continue to be refined as additional data are available over time.
Sample results with a detection or reporting limit greater than the applicable standard will not
be included in the background data sets. Should the detection or reporting limit criteria impact
the data set such that sufficient data is not available for producing BTVs for particular
constituents, NCDEQ will be consulted to discuss alternative evaluation options for assessment
of background, such as groundwater fate and transport modeling.
Groundwater
Groundwater samples are collected from monitoring wells screened in different flow layers
using low-flow sampling techniques in accordance with the USEPA Region 1 Purging and
Sampling Procedure for the Collection of Groundwater Samples from Monitoring Wells (revised
January 19, 2010) and the Groundwater Monitoring Program, Low Flow Sampling Plan, Duke
Energy Facilities, Ash Basin Groundwater Assessment Program, North Carolina, dated June
10, 2015. Groundwater samples have been analyzed for constituents listed in Table 1. Only
non-filtered sample results will be utilized for producing BTVs. Sample data associated with
turbidity reported to be greater than 10 NTUs, samples without a recorded turbidity, or samples
with a pH greater than 8.5 will be excluded from the background data set. Where site
conditions require, professional judgment can be used to retain data that does not meet these
criteria (such as where the naturally occurring groundwater pH is greater than 8.5). However,
the decision to retain data that does not satisfy these criteria must be documented. BTVs will
be calculated for each constituent within a flow layer using data pooled from all background
wells screened within that flow layer.
Soil
Discrete soil samples were collected from multiple depth intervals above and below the water
table during the CSA or other assessment events. The total number of samples and depths
collected vary by site. Soil samples have been analyzed for constituents listed in Table 1.
Only constituent concentrations in unsaturated soils above the water table will be utilized for
producing BTVs. Duke Energy and NCDEQ have agreed soil samples can be pooled from
multiple depth intervals. A statistical evaluation of the soil data sets will be performed to confirm
the approach is appropriate.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
5 | Page
PART II – PRELIMINARY DATA ANALYSIS
Preliminary data analysis includes eight steps and is summarized in the following sections.
1. Descriptive Statistics
Descriptive statistics are useful for characterizing data, increasing data set understanding, and
for assessing information quality. For each site, descriptive statistics will be calculated for
groundwater and soil data sets.
For groundwater, descriptive statistics will be calculated for each constituent within each
groundwater flow layer using pooled data from that groundwater flow layer.
Soil descriptive statistics will be calculated for each constituent using the pooled background
data set.
The following statistics will be calculated to describe each data set.
Sample Size Mean and median
Number of detects and non-detects Maximum and minimum
Percentage of non-detects Standard Deviation
Number of distinct observations Skewness
Number of distinct method detection
limits (MDL) Kurtosis
2. Graphical Analysis
Background groundwater data can be graphically portrayed using scatter plots, box-and-
whisker and quantile-quantile (Q-Q) plots (Figures 1 and 2), while background soil data can be
illustrated using box-and-whisker and Q-Q plots. The construction of scatter plots of
concentration versus time (Figure 3) for each constituent within each background monitoring
well or using the pooled data across all the background wells can assist in identifying potential
trends or seasonality within data. Box-and-whisker and Q-Q plots can be constructed for each
constituent within each groundwater flow layer using all data pooled from background wells
monitoring that flow layer to identify possible outliers and sub-groups in addition to assessing
data set distributions.
Instructions for interpreting box-and-whisker plots can be found on Figure 1. Q-Q plots (Figure
2) evaluate if a theoretical distribution can accurately model a sampled distribution. If the
sampled population is accurately modeled by the theoretical distribution, then quantiles from
the sampled distribution should plot along a straight line when plotted against the quantiles of
the theoretical distribution. Sampled values that plot markedly away from the straight line or
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
6 | Page
jumps or breaks in the plot may indicate the presence of multiple sample populations, potential
outliers, or non-normal sample distributions.
The graphical analysis provides information regarding a steady-state baseline period. Multiple
method detection limits over time will also be evaluated to determine if such variability affects
the quality of the data. This process will be used to determine if all data can be incorporated
into the analysis or if older historical data may need to be removed from the data set due to a
change in the data reporting protocols for samples over time.
3. Identify Outliers
Outliers are values that are not representative of the population from which they were sampled
and whose presence can significantly alter statistical results. Data sets will initially be screened
for potential outliers visually using box-and-whisker and Q-Q plots (Figures 1 and 2).
Following the visual assessment of data for potential outliers, data sets will be screened for
outliers quantitatively. While there are several tests available to test for possible outliers,
Dixon’s or Rosner’s Outlier tests are specifically identified in the Unified Guidance (USEPA
2009) for providing requisite statistical strength and power necessary to appropriately identify
potential outliers. Dixon’s Outlier Test is suitable for data sets containing less than 25 samples,
whereas Rosner’s test is applicable for data sets containing greater than 25 samples. Both
tests assume data are normally distributed.
Extreme outliers are of interest; therefore, outlier tests will be conducted using a significance
level of 0.01. Groundwater and soil constituent concentrations determined to be outliers will be
provided in the statistics report submitted to NCDEQ. If statistical outliers have been detected,
the project scientist will review the values to determine if they should be removed from the data
set or are representative of background and should be retained for statistical analysis.
Reasons as to why a particular statistical outlier should be included or excluded from either
groundwater or soil background data sets will be documented as part of the final reference
background concentration value documentation notes.
4. Identifying Data Distributions
Many statistical tests, such as UTLs, make an explicit assumption concerning the distribution of
sample data. Therefore, data must be fitted to a known distribution model (e.g., normal
distribution). Upon completion of screening data sets for outliers, groundwater and soil data will
be fitted to known distribution models using Goodness-of-Fit (GOF) tests. GOF tests assess
how closely a data set resembles a given distribution model. The distribution models under
consideration for the determination of groundwater and soil BTVs are normal, lognormal, and
gamma distributions.
In order to assess if data are normally or lognormally distributed, the Shapiro-Wilk or Lilliefors
GOF test will be used. The Shapiro-Wilk GOF test is applicable for data sets comprised 50 or
fewer samples, while the Lilliefors GOF test is appropriate for data sets containing more than
50 samples. To evaluate if data are gamma distributed, the Anderson-Darling or Kolmogorov-
Smirnov GOF test will be utilized. GOF tests will be performed using a significance level of
0.05.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
7 | Page
The software package developed by the US EPA, ProUCL, has incorporated these methods to
automatically test for either normal, lognormal, or gamma distribution types. If all GOF tests
fail, non-parametric estimation methods will be used.
The distribution of data will be evaluated as the data sets are established, with the
understanding that distributions may change over time.
5. Evaluating Background Groundwater Data
The following section applies to data sampled over time, such as groundwater data, and is not
applicable to soil data.
Constituent concentrations in groundwater sampled over time from multiple background well
locations may exhibit patterns which suggest concentrations are increasing or decreasing over
time. For background samples to be considered representative of areas unimpacted by human
activity and be meaningful in the production of the BTVs, constituent concentrations over time
should reflect a steady state, or ‘temporal stationarity’. In other words, a constituent’s
population characteristics (mean and variance) do not fluctuate over time (with consideration of
normal seasonal fluctuations). Another assumption regarding samples collected across multiple
background wells at a site is a constituent’s mean and variance are constant across
background wells, or ‘spatial stationarity’. If data collected from the background wells exhibit
temporal or spatial non-stationarity, pooling of background well data can result in an inflated
population variance and biased estimates of BTVs.
A comparison of multiple box-and-whisker plots (Figure 1) can be used to visually assess
whether background wells distributions have similar constituent concentration means and
variances. Based on visual inspection of box-and-whisker plots, further analysis (such testing
for differences in means or medians across background well locations) may be warranted to
determine if a background well should be considered representative of background.
Statistical tests for trends over time using the pooled data from the background wells should
show no statistical significance. However, before proceeding to test for trends in the
background samples, another assumption regarding constituent concentrations is the values
must be independent from one another. When values are related to each other over varying
time intervals, then values at any point in time can be expressed as a function of previous
value(s). This type of relationship is termed autocorrelation. When values express a seasonal
relationship, this type of autocorrelation is termed seasonality.
The presence of autocorrelation, seasonality, or trends indicates data are temporally non-
stationary. Assessment of background groundwater data should be performed to address
temporal stationarity prior to pooling background data for the production of BTVs.
Details for assessing data sets for temporal stationarity is summarized in the following sections.
6. Autocorrelation
Autocorrelation occurs when measurements collected at different points in time correlate with
one another. Sources of autocorrelation in groundwater data can be due to seasonality, trends,
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
8 | Page
or samples being collected too close to one another in time. Data that exhibit autocorrelation
can affect sample variance and can lead to biased estimates of BTVs.
Constituent concentrations in groundwater at a given background well will be checked for
autocorrelation using the sample autocorrelation function (USEPA 2009). The sample
autocorrelation function graphs correlation values between successive measurements against
the time lag between sampling events and assumes data can be fitted to a known distribution
model (Figure 4). Correlation values can be between zero and one, where one indicates a
perfect correlation (dependence) and zero represents no correlation (independence). The
sample autocorrelation function will be calculated using a significance level of 0.05.
Autocorrelated observations can be corrected by 1) reducing sampling frequency and
increasing the time between sample collection; 2) altering the statistical test used to analyze
the data; or, 3) removing temporal patterns using a technique such as deseasonalization.
7. Seasonality
Constituents in groundwater at background well locations may experience predictable recurring
increases and decreases in concentrations, termed seasonality (Figure 6). Seasonality within
a data set can introduce bias into the calculation of BTVs and result in falsely identifying a
seasonal effect as potential impacts.
Data should be assessed for seasonality once an adequate number of background
groundwater samples have been collected. Useful diagnostic tools for evaluating data sets for
seasonality are autocorrelation and scatter plots (Figures 4 and 5). When sufficient
observations are available, then a side-by-side comparison of multiple box-and-whisker plots
constructed by season are informative. If constituent concentrations within a given background
well appears to experience seasonal fluctuations, the seasonal component within the data can
be removed for the purpose of testing for trends. If seasonality is not addressed prior to testing
for trends, then the statistical tests for trends may be misleading (i.e., fail to detect a trend when
one is actually present or may indicate a significant trend when in fact, no trend exits). Strong
evidence for the cause of seasonality within a data set should exist prior to removing seasonal
components from background data sets.
8. Trends
Wells installed at background locations monitor natural groundwater quality unaffected by
anthropogenic activities. Therefore, a key assumption regarding background is constituent
concentrations in groundwater should demonstrate stationary conditions through time, free of
any trends (Figure 6). Background data exhibiting trends (upward or downward) violate the
assumption of temporal stationarity. Trending constituent concentrations in background wells
may identify potential anthropogenic impacts (resulting in the well no longer being considered
background), seasonality, or altering groundwater conditions. Furthermore, presence of trends
in background data can lead to overestimation of variances which result in inflated BTVs.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
9 | Page
Prior to the calculation of BTVs, background well data will be evaluated for the presence of
trends. Depending on the presence of non-detects (NDs) and seasonality, background data
sets can be assessed for trends using one of three tests:
Mann-Kendall trend test
Seasonal Kendall Regression
Maximum likelihood estimation (MLE) regression
The Mann-Kendall (MK) trend test is a nonparametric test method that can be used to identify
linear trends within data sets that do not adhere to specific distribution models, do not exhibit
seasonality, and contain NDs. The MK trend test can only be utilized to evaluate data sets
containing only one MDL. Seasonal Kendall regression is similar to the MK test (data sets do
not have to adhere to specific distribution models and can contain NDs as long as they are
represented by a single MDL), except it accounts for seasonality. MLE Regression is a
parametric method that estimates parameters of a statistical model and for fitting a statistical
model to data. MLE Regression can be performed on data sets that can be fitted to a specific
distribution model, do not demonstrate seasonality, and contain NDs.
In cases where trending background constituent concentrations are identified, further analysis
is recommended to rule out if the trend is more of an artefact related to the length of time
available for the analysis and/or the small sample sizes. For example, if less than 10 samples
have collected from a background well over a short duration (less than two years), then an
observed trend in the well may not necessarily indicate changes in the natural variability of
groundwater quality and may be representative of natural variation. If sufficient data is available
for a constituent (> 20 observations), a statistical method called the piece-wise polynomial
model can be used to inform the overall trend results. A description of this approach is as
follows.
9. Additional Methods for Identifying Trends in Background
Groundwater Data
The piece-wise polynomial model is a useful tool for assessing constituent concentrations that
have experienced multiple trends throughout monitoring. Piece-wise polynomial models
attempt to find an appropriate mathematical function that expresses the relationship between
the constituent concentrations and the sampling dates by using piece-wise regressions. Two
types of piece-wise models can be used to evaluate trends, the linear-linear and linear-linear-
linear regression models.
The linear-linear regression model assumes and identifies one structural break in a time-series,
in which the two portions of the data separated by the break point exhibit two different trends
modeled by two different linear equations. Similarly, the linear-linear-linear regression model
attempts to identify two structural breaks to assess three different linear trends.
Piece-wise polynomial models can be informative, but it have the disadvantage of not being
able to account for NDs in within data sets. Therefore, it is recommended implementing piece-
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
10 | Page
wise polynomial models in conjunction with MLE regression. Piece-wise models can also serve
as a visual guide when selecting the baseline sampling periods for statistical analysis.
For example, in Figure 7 the MLE regression suggested that the overall trend in constituent
concentrations over time are steadily increasing, whereas the polynomial piece-wise regression
with two structural breaks indicates concentrations have experienced upward and downward
trends.
10. Determining Baseline Period for Background Wells
This step provides information to make a determination of whether the entire period of record
from which the background samples were collected is representative of natural background
conditions and represents a baseline against which downgradient constituent concentrations in
groundwater can be tested. If trend analysis indicates that over time the observations are
steadily increasing or decreasing, then review of the data will be performed to determine if a
sub-segment of the data set better represents the background period. For values to be
considered representative of background, they should demonstrate temporal stationarity.
A minimum of eight samples will be collected from background wells screened in each
groundwater flow layer prior to statistical analysis. Analytical results from eight or more
sampling events will be used for the statistical determination of BTVs for constituents in
groundwater.
A minimum of eight additional sampling events should be completed before evaluating if new
background data should be combined with previous background data to produce revised BTVs.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
11 | Page
PART III – TESTING FOR SUB-GROUPS IN BACKGROUND
GROUNDWATER DATA
The following sections summarize the methodology for identifying sub-groups in groundwater
data resulting from spatial or temporal variability and will not be applicable for the assessment
of soil data.
Part III summarizes steps to validate if statistically significant differences in background
concentrations exist across potential sub-groups. Sub-groups represent distinct populations
with statistically significant differences in mean or median concentrations among potential
groups within a data set. An example of possible sub-groups is a difference in concentrations
among constituents detected in background wells monitoring bedrock groundwater that were
installed in different rock types. Each rock type has its own chemical characteristics that can
influence groundwater chemistry and result in differing concentrations for constituents across
background wells.
In order to test for differences across potential sub-groups, sufficient sample size of at least
eight to 10 samples is recommended for each potential sub-group (USEPA 2009, 2015).
Testing for potential sub-groups within background data will be completed in three steps:
Graphical analysis
Analytical test for comparing sub-group differences
Tests for distinguishing which sub-groups are different
Statistical tests utilized to test for potential sub-groups will be performed using a significance
level of 0.05.
Graphical Analysis
Graphical representation of data is an effective tool for depicting patterns and relationships
within data.
Background groundwater data can be assessed for sub-groups using box-and-whisker and Q-
Q plots (Figures 1 and 2). Multiple box-and-whisker and Q-Q plots can be constructed for
comparing constituent concentrations and variability across perceived sub-groups.
Another useful visual test assessing potential sub-group differences is the Empirical Distribution
Function (EDF). EDFs compute summary statistics, generate EDF plots (Figure 9), and
compute hypothesis tests appropriate for comparing two or more groups for data containing
NDs (provided less than 50 percent of the results are NDs).
Figure 8 of an EDF plot demonstrates that the two sub-groups representing samples taken
during two different seasons show similar distributions or no differences in constituent
concentrations between the two seasons.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
12 | Page
Analytical Tests for Comparing Sub-Groups
The following methods can be used to test for differences across sub-groups:
T-test and One-way Analysis of Variance (ANOVA)
Wilcoxon rank-sum and Kruskal-Wallis (KW) tests
Kaplan-Meier (KM) (log-rank) test
All three types of tests can be used to test data sets containing NDs.
The t-test and One-way ANOVA are parametric statistical analyses that test for differences in
means among groups. T-tests are used to test for differences in means among two groups,
whereas ANOVA is used to test for differences in means across three or more groups. Both
tests assume data are normally distributed with normally (or lognormally) distributed residual
values and the variances among groups being compared are roughly the same.
The Wilcoxon rank-sum test and KW Test are nonparametic equivalents of the parametric t-test
and One-way ANOVA. Both the Wilcoxon rank-sum and KW Test analyze the ranks of the data
rather than the actual concentrations and test for difference among average ranks between
groups. The Wilcoxon rank-sum test compares the average rank between two groups, while
the KW test compares the average rank across three or more groups.
The KM (log-rank) test is a nonparametric test that compares the survival distribution between
two or more groups. The KM (log-rank) test is useful for data sets that cannot be fitted to a
discernible distribution model and contain a large percentage of NDs concentrations.
Testing for potential sub-groups within background groundwater data sets will be performed
using a significance level of 0.05.
Tests for Identifying Differences Among Sub-Groups
If results from One-way ANOVA, KW, and KM (log-rank) tests indicate a statistically significant
difference during comparison of three or more groups, additional tests need to be performed to
compare all possible pairs of sub-group means or average ranks to determine which ones are
different from one another. These tests are referred to as ‘post-hoc’ tests because they are
performed after the fact. The Tukey-Kramer and Dunn’s test are post-hoc tests that should be
used to compare possible pairs of sub-groups means or average ranks. The Tukey-Kramer
test is parametric and should be used to evaluate One-Way ANOVA results, whereas Dunn’s
test is nonparametric and should be utilized to assess KW and KM (log-rank) test results. Post-
hoc analysis will be performed using a significance level of 0.05.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
13 | Page
PART IV – DEVELOPMENT OF BTVs FOR CONSTITUENTS IN
GROUNDWATER AND SOIL
The USEPA Unified Guidance (2009) recommends using UTLs to estimate BTVs for
constituents evaluated during assessment monitoring as opposed to the use of other statistical
intervals such as confidence limits (UCL) and upper prediction limits (UPL). UTLs represent
fixed values that do not rely on future observations, unlike UPLs, and are constructed using
background data as opposed to a fixed health-based standard, unlike UCLs (USEPA 2009).
UTLs allow for a “suitably high enough level above current background to allow for reversal of
the test hypothesis and are the preferred statistical interval (USEPA 2009). In nearly all cases,
the UTL is computed because the concern is generally for exceedances greater than the value.
The only parameter that may require both upper and lower tolerance limits is pH. Site-specific
BTVs for select constituents from Table 1 in groundwater and soil will be produced using UTLs.
Tolerance intervals test the null hypothesis that concentrations in downgradient wells or at
impacted soil sampling locations are similar to that of background and are constructed using
the mean, standard deviation, and tolerance factor. For the estimation of BTVs for constituents
in groundwater and soil, a coverage of 95 percent (p) and a confidence interval of 95 percent
(1-α) will be used. This means, there is a 95 percent probability that 95 percent of background
sample concentrations will fall within this limit. The formulation of the UTL may vary slightly
with the details of the test to be made and the characteristics of the data involved (see chapters
3 and 5 of ProUCL’s Version 5.1.02 Technical Guide for the full specifications of the UTL
formula under differing parametric and non-parametric assumptions), but the basic form for the
(1-α)*100 percent UTL with coverage coefficient, p, under normal distribution assumptions in
general is:
ܷܶܮ ൌ ݔ̅ܭ∗ݏ
Where
ݔ̅ = baseline (historical data) sample mean; and,
s = baseline (historical data) standard deviation.
K represents a special function called the tolerance factor. It depends on the sample size (n),
the confidence coefficient (1 – α), and the coverage proportion (p). For selected values of n, p,
and (1-α), values of the tolerance factor (K) have been tabulated extensively in the statistical
literature. ProUCL will be utilized to produce UTLs for each constituent. The type of UTL
produced is a factor of distribution type, the desired confidence interval, coverage, and the
percentage of NDs.
Following completion of the preliminary data analysis described in Part II and applicable steps
in Part III, the steps below will be completed for selection of appropriate UTLs.
1. UTLs will be produced for constituents in groundwater and soil using the statistical
software program ProUCL. The first step in constructing UTLs using ProUCL is to
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
14 | Page
categorize constituents based on the presence or absence of NDs. ProUCL calculates
UTLs differently depending on whether NDs are present within a data set. The
algorithms in ProUCL use imputation and modeling techniques to address NDs.
ProUCL does not substitute values (e.g., multiplying the MDL by a constant) for NDs, as
this method introduces bias into the estimation of UTLs. Some constituent data sets
may be represented 50 percent or more NDs. A large percentage of NDs make it
difficult to fit data to distribution models. For data sets containing 50 percent or more
NDs, UTLs will be constructed utilizing nonparametric techniques.
2. Produce UTLs using a coverage (p) and confidence level (1- α) of 95 percent.
3. Record all UTLs under all parametric and non-parametric distribution models. If data
sets used for producing UTLs can be fitted to multiple distribution models, preference is
given to one distribution type over another. In cases where data fit normal, log normal,
and gamma distribution models, preference should be given to UTLs calculated
assuming data were gamma distributed since this distribution has the advantage of not
having extremely long tails to the right. This mitigates the chance that UTLs will be
extremely high relative to the median of the distribution as could be the case if one used
either the lognormal or normal distribution assumptions for producing UTLs. If data are
normally and lognormally distributed, but not gamma distributed, preference should be
given to UTLs produced assuming data are lognormally distributed provided the logged
data have a standard deviation less than or equal to one. If the standard deviation of
the logged data is greater than one, then the gamma distribution is preferred.
Otherwise, preference should be given to UTLs produced assuming normally distributed
data. If data cannot be fitted to a discernible distribution, UTLs are produced using
nonparametric techniques. Data set distributions will continue to be evaluated as
additional samples are collected, with the understanding that distributions may change
over time.
4. It has been demonstrated that if there are insufficient samples sizes, the non-parametric
UTL cannot achieve the desired confidence coefficient of 95 percent. Depending on
background sample size, a different order statistic is selected to produce UTLs. For
constituent data sets containing eight to 11samples, UTLs will be produced using a
coverage of 85 percent (i.e., the 85th percentile) and a confidence coefficient of 95
percent. For data sets containing 29 or more samples but less than 59 samples, UTLs
will be calculated using coverage of 90 percent and confidence coefficient of 95 percent,
and data sets containing more than 59 samples will be calculated using a coverage of
95 percent and a confidence coefficient of 95 percent.
5. A minimum of eight valid background groundwater samples should be obtained prior to
producing BTVs for each constituent in each flow layer. If it is deemed necessary to
produce BTVs prior to obtaining eight valid samples, NCDEQ will be consulted and the
maximum observation may need to be used as a BTV. In addition, a minimum of eight
additional samples should be obtained prior to evaluating if new background data should
be combined with previous data to produce revised BTVs.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
15 | Page
REFERENCES
NCDEQ DWR, 2012. Evaluating Metals in Groundwater at DWQ Permitted Facilities: A
Technical Assistance Document for DWQ Staff.
http://digital.ncdcr.gov/cdm/ref/collection/p16062coll9/id/251181.
USEPA, 1992. Supplemental Guidance to RAGS: Calculating the Concentration Term.
Publication 9285.7-081.
USEPA, 2002. Guidance for Comparing Background and Chemical Concentrations in Soil for
CERCLA Sites. EPA 540-R-01-003.
USEPA, 2009. Statistical Analysis of Groundwater Monitoring Data at RCRA Facilities – Unified
Guidance, March 2009. EPA 530-R-09-007.
USEPA, 2015. Statistical Software ProUCL 5.1.002 for Environmental Applications for Data
Sets with and without Nondetect Observations. EPA/600/R07/041.
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
FIGURES
FIGURE 1
BOX-AND-WHISKER PLOT
75th Percentile
50th Percentile
25th Percentile
90th Percentile
10th Percentile
Possible Outliers
FIGURE 2
QUANTILE-QUANTILE (Q-Q) PLOT
FIGURE 3
SCATTER PLOTS COMPARING TIME VERSUS
CONCENTRATION BETWEEN TWO WELLS
0
50
100
150
200
250
300
350
400
450
1/1/2010 8/24/2011 4/15/2013 12/6/2014
CO
N
C
E
N
T
R
A
T
I
O
N
DATE
MW-1
MW-2
FIGURE 4
SAMPLE AUTOCORRELATION OUTPUT
FIGURE 5
SCATTER PLOT OF TIME VERSUS
CONCENTRATION ILLUSTRATING
SEASONALITY
0
10
20
30
40
50
60
3/3/2010 5/11/2012 7/20/2014 9/27/2016
PA
R
A
M
E
T
E
R
(
C
O
N
C
E
N
T
R
A
T
I
O
N
)
DATE
FIGURE 6
SCATTER PLOTS OF TIME VERSUS
CONCENTRATION ILLUSTRATING TRENDS
0
20
40
60
80
100
120
3/3/2010 5/11/2012 7/20/2014 9/27/2016
PA
R
A
M
E
T
E
R
(C
O
N
C
E
N
T
R
A
T
I
O
N
)
DATE
0
20
40
60
80
100
120
3/3/2010 5/11/2012 7/20/2014 9/27/2016
PA
R
A
M
E
T
E
R
(C
O
N
C
E
N
T
R
A
T
I
O
N
)
DATE
FIGURE 7
PIECE-WISE POLYNOMIAL REGRESSION
OUTPUT EXHIBITING MULTIPLE TRENDS
Upward Trends Downward Trend
FIGURE 8
EMPERICAL DISTRIBUTION PLOT COMPARING
CONSTITUENT CONCENTRATIONS BETWEEN
TWO SEASONS
REVISED DRAFT
Duke Energy Carolinas, LLC
Statistical Methods For Developing Reference Background Concentrations For
Groundwater and Soil At Coal Ash Facilities
January 2017
TABLE
TABLE 1
CHEMICAL PARAMETERS ANALYZED IN GROUNDWATER AND SOIL
FIELD PARAMETERS RADIONUCLIDES
pH*†Radium 226*
Specific Conductance*Radium 228*
Temperature*Uranium (233, 234, 236, 238)*
Dissolved Oxygen*ANIONS/CATIONS/OTHER
Oxidation Reduction Potential*Alkalinity (as CaCO3)*
Eh*Bicarbonate*
Turbidity*Calcium
INORGANICS Carbonate*
Aluminum Chloride
Antimony Magnesium
Arsenic Nitrate (as N)†
Barium Nitrate + Nitrite*
Beryllium Potassium
Boron Percent Moisture†
Cadmium Methane*
Chromium Sodium
Cobalt Sulfate
Copper Sulfide*
Iron Total Dissolved Solids*
Lead Total Organic Carbon
Manganese Total Suspended Solids*
Mercury
Molybdenum
Nickel
Selenium
Strontium
Thallium (low level)
Vanadium (low level)
Zinc
NOTES:
* = Indicates parameter analyzed in groundwater only.
† = Indicates parameters analyzed in soil only.
Metals in groundwater were analyzed for total and dissolved concentrations.
Soil pH measured at 25 degrees C.
Analysis of groundwater and soil samples for Chromium (VI) began after initial samples were collected as part
of CSA.
Page 1 of 1