C L: A I R E

TB12
(November 2006)

technical bulletin

CL:AIRE technical bulletins describe specific techniques, practices and methodologies currently being employed on sites in
the UK within the scope of CL:AIRE technology demonstration and research projects. This bulletin provides a discussion on
the 'Mean Value Test' presented in Contaminated Land Report 7 and suggests alternative techniques to reduce the risk of
underestimating the true 95% Upper Confidence Limit of the sample mean.
Copyright CL:AIRE (Contaminated Land: Applications in Real Environments)

Statistical Assessment of Contaminated Land:

Some Implications of the 'Mean Value Test'
1.

BACKGROUND

Statistical methods are commonly used to guide decision-making in many

regulatory contexts. For the assessment of potentially contaminated land, it is
common practice to use the Mean Value Test approach defined in Appendix A
of Contaminated Land Report 7 (DEFRA and Environment Agency, 2002). This
defines the estimation of the 95% Upper Confidence Limit of the mean soil
concentration of a contaminant (95%UCL, also referred to as US95) and its use
as the appropriate value to be compared with the relevant soil guideline value
(SGV) or site-specific assessment criterion (SSAC). This 95%UCL is meant to
provide a reasonably conservative estimate of whether the measured
concentration is acceptable, considering the uncertainty and variability
associated with site investigations.
However, when applied to real environmental datasets, the Mean Value Test can
actually prove to be a non-conservative choice that may wrongly categorise a
potentially contaminated site as clean. This bulletin illustrates how such a
situation could arise, highlights the implications and introduces alternative
appropriate methods to derive the 95%UCL of the mean soil concentration.

2.

INTRODUCTION

An informed approach to decision-making takes this eventuality into account by

including in the estimation both the observed variability of determinand
concentrations and information on the distribution type. The decision can then
be based on a (conservative) surrogate of the true average concentration, the socalled Upper Confidence Limit (UCL) of the sample average.
The Mean Value Test presented in Appendix A of CLR7 makes use of the
95%UCL (referred to as US95 throughout CLR7) as a means to reduce the risk
of wrongly classifying a site as clean when it is truly contaminated. This
procedure, although theoretically sound, can prove inadequate in some
situations and provide insufficient information to the decision maker.
In the following sections, some of the available methods to calculate the
95%UCL of the mean for contaminated land datasets are presented. A brief
summary of the advantages and potential pitfalls of the different approaches is
illustrated by using examples and literature case studies.

3.

THE MEAN VALUE TEST AND THE METHODS TO DETERMINE

THE 95%UCL OF THE MEAN

The 95% Upper Confidence Limit is defined as the value such that the true
population mean is less than the 95%UCL with a confidence level of 95%.

In the UK contaminated land regime, the comparison of site data to an

appropriate threshold is adopted to support the decision of whether, taking due
account of the exposure pathways, the site poses a significant possibility of
causing significant harm to human health. This comparison is often influenced
by various sources of uncertainty, as discussed in CL:AIRE Technical Bulletin 7 Improving the Reliability of Contaminated Land Assessment using Statistical
Methods: Part 1 - Basic Principles and Concepts. As guidance for regulators,
consultants and all interested parties, Contaminated Land Report 7 (CLR7)
provides indications on the interpretation of site investigation data.
CLR7 suggests, for example, that a careful zoning of a site should be adopted
during investigation and that the average concentration in each site zone should
be compared to the relevant assessment criterion for a determinand to make a
judgement on the risk posed. In addition, CLR7 acknowledges that data
interpretation should be guided by statistical principles.
A statistical approach to data interpretation is based on the consideration that
it is not practicable to sample and analyse the entire soil mass within a site zone
and that the estimate of the mean soil concentration of a determinand therefore
has to rely on a (very) limited knowledge of the contaminant distribution.
Consequently, in making a decision, one needs to account for the probability
that the volume of the unsampled soil may contain concentrations higher than
the ones that have actually been measured. In such a case the sample average
may underestimate the true average soil concentration at the site (or zone) of
interest.

The Mean Value Test, as presented in CLR7, calculates the 95%UCL of the
mean determinand concentration. This is then compared to the relevant SGV or
SSAC, as appropriate. The theory behind this method is based on the assumption
of normal distribution of the data but is inadequate to estimate the UCL for a
dataset with a distribution other than normal. Since many environmental
datasets are non-normal, alternative approaches were developed.
Land (1975) and Gilbert (1987), for example, describe methods to estimate the
upper confidence limit for a population that is lognormally distributed (the
method known as H-statistics). These and several other methods, referred to as
parametric and non-parametric, are available to properly estimate this key value.
So, the theory behind the Mean Value Test as presented in Appendix A of CLR7
is no longer the only possible way to estimate the 95%UCL of the population
mean nor is it appropriate when data are not normally distributed.
3.1

THEORETICAL BACKGROUND

All the formulae to estimate the [1-]%-UCL of a population mean can be

expressed in the same way. These formulae, in fact, can be easily linked to the
definition of pivotal quantity (see, for example, Casella and Berger, 1990) and be
adapted to a variety of cases provided that the appropriate meaning (and value)
is assigned to the symbols. Although formula [1] will appear familiar to those
who use Appendix A of CLR7, a family of methods is based on it, dependent
upon the meaning assigned to some of the key parameters:

If you have any questions about this Technical Bulletin or would like further information about other CL:AIRE publications please contact us at CL:AIRE
Email: enquiries@claire.co.uk Web site: www.claire.co.uk
This document is printed on recycled material containing 80% post-consumer waste and 20% totally chlorine free virgin pulp.

C L: A I R E

TB12 page 2

technical bulletin

s
(1 )%UCL = x + t

[1]
In [1],x and s represent an estimation of the population mean and standard
deviation, respectively, n represents the number of samples and t a statistical
quantity with a meaning depending on the distribution to which the dataset can
be fitted.
An intuitive review of formula [1] suggests that:
s: the estimated UCL depends on the variability of the dataset under
examination (i.e. the larger the standard deviation of the dataset, the more
the UCL is likely to differ from the mean of a given dataset);

n: the estimation of the UCL is more accurate for larger datasets (i.e. the
more data are available, the less the UCL is likely to differ from the sample
average for a given standard deviation);

regardless of the variability or the sample size, the UCL of the mean differs
from the sample average by a factor called t in [1].

Correct estimation of the values used in formula [1] is paramount for

appropriately estimating the UCL. Once this is done, however, the key factor that
will enable an appropriate UCL estimate is simply the term t. The rest of this
bulletin will focus on t to show the implications of assigning a correct or an
incorrect value to this term.
The factor t is a distribution specific term. In the case of a sample drawn from
a normal population, it stands for the upper th percentile of the Students t
distribution with n-1 degrees of freedom. In such a case, formula [1] assumes
the same meaning as that given in Appendix A of CLR7 to carry out the Mean
Value Test.
For a population which is non-normally distributed, which is the case for many
contaminated land datasets, t must have a different meaning as the Students
t distribution only applies to a normally distributed population. The approach
indicated in Appendix A of CLR7, consequently, is not appropriate for many
datasets.
For non-normally distributed datasets, however, formula [1] can still be used,
provided that one can appropriately estimate the multiplier of the quantity:
s

3.2

Estimating this multiplier for a non-normal population is not always

straightforward, however.
PARAMETRIC VERSUS NON-PARAMETRIC METHODS TO ESTIMATE
THE (1- )%UCL OF THE MEAN

In statistical terminology a parametric approach is a method based on

distributional assumptions. In order to apply such a method, therefore, the
practitioner is required to demonstrate with sufficient confidence that a dataset
follows a specific distribution.
Available parametric methods to estimate the UCL of a population mean exist
for normal, lognormal and gamma distributed data. Some discussion and
references on these methods are given in Masi and Morgan (2005).
When it comes to real environmental datasets, however, it is often hard to prove
that the data follow a particular distribution. Real datasets, in fact, include
values below the limit of detection or concentrations that are outliers. In the
case of geochemical datasets, outliers tend to assume very high values and make
the distribution of the data extremely skewed.
In many cases, the normal and many other distributions can prove inadequate
to fit the available data. When there is not sufficient evidence to eliminate
outliers from a dataset, or sufficient scope or budget to re-sample the suspected
locations, it is difficult to argue that the assumption of a parametric distribution
(normal, lognormal or gamma) still holds. If in these cases a parametric
distribution was forced to represent the pattern of the data, the analysis would
produce unreliable estimates increasing the risk of a wrong site classification.

The most interesting aspect of non-parametric approaches and the reason for
their being widely adopted in applied statistics is that they constitute a set of
distribution-free methods. Obviously this broader applicability comes at a cost,
which is a generally higher degree of conservatism. It could, however, be argued
that this is not excessive given the enormous difference between the total soil
mass and the sampled mass of a potentially contaminated site.
The most common non-parametric methods that can be applied for the
estimation of the 95%UCL are: the Central Limit Theorem (which can be
adjusted for skewed samples; Chen, 1995); the non-parametric version of the
(one-sided) Chebyshev Theorem; and the Bootstrap technique, in all its different
formulations (e.g. Davison and Hinkley, 1997).

4.

UNDERSTANDING DATA DISTRIBUTION: ADVANTAGES AND

POTENTIAL REPERCUSSIONS

The benefits of the application of parametric and non-parametric methods to

real datasets can be easily understood from simple case studies in which
different UCL formulae are applied to some typical datasets.
Let us consider a potentially contaminated site that has undergone site
investigation. Ninety-nine soil samples have been taken and analysed by a
certified laboratory for two contaminants of concern, one organic compound
(contaminant A) and one heavy metal (contaminant B). These two datasets will
be used to show some of the potential implications of inappropriate use of the
UCL based on the default assumption of normal distribution.
The summary statistics of the two complete datasets are reported in Table 1.
Table 1. Summary statistics for the two contaminants represented in the
hypothetical dataset
Contaminant A
(organic)

Contaminant B
(inorganic)

Units

Minimum

1.87

0.01

mg/kg

Maximum

1316.67

2873

mg/kg

Mean

134.26

181.99

mg/kg

Standard deviation

223.60

440.75

mg/kg

Although the data themselves span a range of three to five orders of magnitude,
initial review of the results did not show enough evidence for the presence of
multiple populations. In addition, scattering of high concentrations across the
site did not support any zoning of the site area.
The proprietary data analysis package CLR7_Stats.xls (ESI, 2006) has been used
to analyse the data, calculate the summary statistics, identify potential outliers
and the appropriate distributional assumption and correctly estimate the
95%UCL.
As suggested in Appendix A of CLR7, the Grubbs test (the so called Maximum
Value Test in DEFRA & Environment Agency, 2002; first reported in Grubb,
1969) has been applied to the two datasets of contaminant A and B
concentrations to screen for outliers. The test excluded the presence of any
outlier.
The two datasets have then been checked to identify possible underlying
distributions adopting a coupled visual and numerical assessment. The visual
assessment has been implemented through a q-q plot (e.g. Isaaks and
Srivastava, 1989) which is a simple and intuitive qualitative approach (see Masi
& Morgan, 2005, for an example application). The chosen numerical assessment
was performed by the Shapiro-Francia test (Shapiro and Francia, 1972) which
has been carried out at a level of significance of 5%. Both the methods have
been implemented through CLR7_Stats.xls.

C L: A I R E

TB12 page 3

technical bulletin

Figure 1 shows the q-q plot to assess normality of contaminant A distribution.

Normally distributed data would plot as a straight line; this is clearly not normally
distributed. Figure 2 shows an equivalent q-q plot to assess the lognormality of
the contaminant A dataset and shows that it is likely that this follows a
lognormal distribution.

q-q plot
6
5
Experimental standardised values

4
3

The same visual approach (Figures 3 and 4) seems to suggest that contaminant
B does not follow a normal distribution; however it might fit a lognormal
distribution.

2
1
0

The q-q plot results have then been confirmed by carrying out a Shapiro-Francia
test. The results are reported in Tables 2 and 3.

-1
-2

Table 2. Results of the Shapiro-Francia test for normality on contaminant A and

contaminant B

-3
-3

-1

Theoretical standardised values

ShapiroFrancia
Statistics

Critical Value

Level of
Significance

Normality
rejected

Contaminant A

0.5519

0.976

0.05

Yes

Contaminant B

0.4545

0.976

0.05

Yes

Figure 1. Normal distribution q-q plot of data for contaminant A.

q-q plot
3

Experimental standardised values

Table 3. Results of the Shapiro-Francia test for lognormality on contaminant A

and contaminant B

ShapiroFrancia
Statistics

Critical Value

Level of
Significance

Lognormality
rejected

Contaminant A

0.9846

0.976

0.05

No

Contaminant B

0.9443

0.976

0.05

Yes

-1

-2

-3
-3

-2

-1

Theoretical standardised values

Figure 2. Lognormal distribution q-q plot of data for contaminant A.

Closer examination of the Shapiro-Francia test statistics for contaminant B (Table

3) shows that the value is close to the critical value of the statistical test. This
condition suggests that a second look is taken at the q-q plot (Figure 4) where
it can be seen that the data points actually follow an s-shaped pattern around
the straight line. The closer visual assessment therefore supports the results of
the statistical test in rejecting the hypothesis of lognormality of the contaminant
B dataset.

q-q plot
7

Experimental standardised values

6
5
4
3
2

The conclusion of the analysis of the data distribution is therefore that the
lognormal theory could be applied to calculate the 95%UCL for contaminant A
but both normal and lognormal calculation methods would be inappropriate for
contaminant B. For the latter dataset the large skewness suggests that the nonparametric Chebyshev method is appropriate to supply a reasonably conservative
estimate of the 95%UCL (Singh et al., 1999).

1
0

-1
-2
-3
-3

-1

Theoretical standardised values

The correct estimates of the 95%UCL for the two datasets are reported in Table
4 together with the values that would have resulted from an inappropriate
application of the normal theory. The H-statistics estimate and the Chebyshev
method are reported as well.

Figure 3. Normal distribution q-q plot of data for contaminant B.

q-q plot
3

Table 4. Comparison of 95%UCL for contaminants A and B calculated using

appropriate statistical approaches and an incorrect assumption of normal
distribution

Experimental standardised values

95%UCL 95%UCL
95%UCL
(normal
(HDifference*
(Chebyshev)
assumption) statistics)

-1

Units

Contaminant A

171.58

192.14

232.02

60.44

mg/kg

Contaminant B

255.56

N.A.

374.69

119.13

mg/kg

-2

-3
-3

-2

-1

Theoretical standardised values

Figure 4. Lognormal distribution q-q plot of data for contaminant B.

N.A. = Not Applicable

* 95%UCL (Chebyshev) 95%UCL (normal assumption)

5.

C L: A I R E

TB12 page 4

DISCUSSION

technical bulletin
6.

A close look at Table 4 can be used to exemplify what, in statistical terms, is

called a false negative response. The 95%UCL estimated by the non-normal
assumption, be it the lognormal theory (or H-statistics) or the one-sided
Chebyshev theorem, are always greater than the ones estimated by means of
the normal theory. This is in agreement with theoretical studies (e.g. Singh et al.,
1997) showing that non-normal assumptions usually offer a better coverage of
the 95%UCL in the case of skewed populations. Since environmental datasets
are often characterised by high skewness, it follows that an estimation of the
95%UCL based on the normal assumption is often likely to be inappropriate for
sound decision-making.
In the case that contaminant B in the example above had a SSAC of 300 mg/kg,
the difference between the results of the normal and the non-normal theory
would be large enough to allow an assessor unaware of the limitation of the
normal theory to wrongly conclude that the site does not pose a significant risk.
A second and deeper level of observations can be made specifically from the Hstatistics. It has been found that contaminant A follows a lognormal distribution,
and therefore it should be appropriate to adopt the H-statistics estimation rather
than the Chebyshev theorem to estimate the 95%UCL for this contaminant.
However, the H-statistics estimation is characterised by some instability (i.e. a
little change in the data might imply a large change in the estimated UCL) and
can therefore be an unreliable basis for risk management related decisions.
To show this feature of H-statistics, the minimum of the dataset of contaminant
A concentrations (1.87 mg/kg) has been replaced with the lower value of
0.01 mg/kg. Since this substitution affects only the lower end of the dataset, it
should be harmless and not significantly alter any statistics. However, the Hstatistics estimate of the UCL increases from 192.14 mg/kg to 275.72 mg/kg.
The effect of a small change to the low end of the dataset has therefore been a
significant change in the estimated 95%UCL. This effect, which has been
previously observed (e.g., Singh et al., 1997; 1999), suggests that the Hstatistics, although based on sound theory, is not always an appropriate basis for
informed risk management. It follows that more general methods, such as the
Chebyshev theorem, are to be preferred.
Some analyses have been carried out on the potential implications of an
incorrect evaluation of the 95%UCL in different contexts. Masi & Morgan (2005)
give a specific example of the possible impact of a false negative response with
specific reference to CLR7. In that paper, a number of hypothetical site
investigations were simulated on a hypothetical contaminated site. The gram by
gram soil concentration of a contaminant of concern, described by a distribution
with an average greater than the relevant SSAC, was sampled thousands of
times to simulate thousands of hypothetical site investigations. The 95%UCL has
then been estimated on each one of the datasets using the normal theory, the
Chebyshev theorem and the modified Central Limit Theorem, another nonparametric method (Chen, 1995). The number of cases in which each estimate
failed in identifying the potential risk posed at the site was recorded.
Table 5 shows a summary of the result of this study in terms of percentage of
failure of each method adopted to estimate the 95%UCL. This table clearly
shows that the Chebyshev theorem outperformed the other two methods.
Table 5. Percentage failure of each statistical method in drawing a true
conclusion from the sampled dataset
Method

Percentage failure

Normal theory (Mean Value Test)

28.2%

Chebyshev Theorem (non-parametric) 7.7%

Central Limit Theorem (adjusted)

23.2%

Table 5 suggests that inappropriate data treatment is more likely to predict that
the 95%UCL of the mean is below the screening criterion, even when the actual
population mean is higher than the screening criterion.

CONCLUSIONS

A rigorous and scientifically sound approach to decision-making is expected

from site assessment to ensure that decisions are correctly made and to avoid
unnecessary expenditure. Together with a sound conceptual model and site
zoning, appropriate statistical treatment of site data is essential. By means of
some theory and practical examples, this bulletin has illustrated some possible
consequences of incorrect statistical assessment. This can include both underand overestimating the true mean concentration.
CLR7 outlines a way to carry out the Mean Value Test which is designed for a
specific data distribution (normal) and therefore is often inappropriate in real
world cases.
Commercial statistical packages make a wide range of statistical tests readily
available to the interested practitioner. However, spreadsheets are more than
adequate to allow professionals to develop their own tool with little
programming effort. Appropriate techniques, including the q-q plot, can easily
be implemented in a spreadsheet. As the above examples show, these methods
can have major benefits in data interpretation and should be used on a routine
basis in the interpretation of site data, as they will significantly increase the
confidence in decision-making.
ACKNOWLEDGEMENTS
This bulletin was prepared by CL:AIRE staff from an original text provided by
Paolo Masi and Philip Morgan of ESI.
REFERENCES
Casella, G., and Berger, R. L. 1990. Statistical Inference, Belmont, CA,
Duxbury.

Chen, L. 1995. Testing the Mean of Skewed Distributions, J Am Stat Assoc,
90, 767-772.

Davison, A.C. and Hinkley, D.V. 1997. Bootstrap Methods and Their
Applications. Cambridge University Press, Cambridge.

DEFRA and Environment Agency. 2002. Assessment of Risk to Human
Health from Land Contamination: An Overview of the Development of Soil
Guideline Values and Related Research. R&D Publication CLR7.

ESI. 2006. CLR7_Stats.xls User Manual

Gilbert, R.O. 1987. Statistical Methods for Environmental Pollution
Monitoring. NY, Van Nostrand Reinhold.

Grubb, F.E. 1969. Procedures for detecting outlying observations in
samples. Technometrics, 11, 1 - 21.

Isaaks E. H., and Srivastava, R.M. 1989. An introduction to applied
geostatistics. Oxford University Press, Inc. NY

Land, C.E. 1975. Tables of Confidence Limits for Linear Functions of the
Normal Mean and Variance. In Sel Tab Mathem Stat, 3, American
Mathematical Society, Providence, R.I., pp.385-419

Masi, P. and Morgan, P. 2005. Appropriate methods for the Mean Value Test
for soil contamination datasets. Land Contamination & Reclamation, 13 (4).

Shapiro, S.S. and Francia, R.S. 1972. An approximate analysis of variance
test for normality. J. Amer. Statist. Assoc. 63. 1343 - 1372

Singh, A.K., Singh, A., Engelhardt, M. 1997. The lognormal distribution in
Environmental Applications. Technology Support Center EPA/600/R-97/006

Singh, A.K., Singh, A., Engelhardt, M. 1999. Some practical aspects of
sample size and power computations for estimating the mean of positively
skewed distributions in environmental applications. Technology Support
Center EPA/600/s-99/006.

For further information please contact Paolo Masi and Philip Morgan at ESI:
New Zealand House,
160-162 Abbey Foregate,
Shrewsbury,
SY2 6BZ, UK
Email: PaoloMasi@esinternational.com; PhilMorgan@esinternational.com