Professional Documents
Culture Documents
EURACHEM - UfS - 2007 PDF
EURACHEM - UfS - 2007 PDF
Measurement uncertainty
arising from sampling
A guide to methods and approaches
Produced jointly with
EUROLAB, Nordtest and the
UK RSC Analytical Methods Committee
Editors
Michael H Ramsey (University of Sussex, UK)
Stephen L R Ellison (LGC, UK)
EUROLAB members
Manfred Golze
Rdiger Kaus
Ulrich Kurfrst
Mikael Krysell
BAM, Germany
HS Niederhein, Denmark
FH Fulda, Denmark
Eurofins Denmark
CITAC member
Ilya Kuselman
Nordtest representatives
Acknowledgements
This document has been produced primarily by a
joint Eurachem/EUROLAB/CITAC/Nordtest/
AMC Working Group with the composition
shown (right) working in the period 20042006.
The editors are grateful to all these individuals
and organisations and to others who have
contributed comments, advice and assistance.
Production of this Guide was in part supported
under contract with the UK Department of Trade
and Industry as part of the National Measurement
System Analytical Measurement (VAM)
Programme, and also by the UK Food Standards
Agency and Environment Agency.
UfS:2007
Christian Grn
Bertil Magnusson
Astrid Nordbotten
AMC representatives
Michael Thompson
Jenny Lyn
University of London, UK
Food Standards Agency, UK
Observer
Marc Salit
NIST, USA
Page i
Foreword
Uncertainty of measurement is the most important single parameter that describes the
quality of measurements. This is because uncertainty fundamentally affects the
decisions that are based upon the measurement result. Substantial progress has been
made in devising procedures to estimate the uncertainty that originates in the
analytical portion of the measurement, and guidance on these procedures is available
[1]. However, a measurement almost invariably involves the process of taking a
sample. This is because it is usually impossible to analyse the entire bulk of the
material to be characterised (the sampling target). If the objective of the
measurement is to estimate the value of the analyte concentration in a sampling target,
then the uncertainty associated with the sampling process must inevitably contribute
to the uncertainty associated with the reported result. It has become increasingly
apparent that sampling is often the more important contribution to uncertainty and
requires equally careful management and control. The uncertainty arising from the
sampling process must therefore be evaluated. While existing guidance identifies
sampling as a possible contribution to the uncertainty in a result, procedures for
estimating the resulting uncertainty are not well developed and further, specific,
guidance is required.
Historically, measurement scientists have been primarily concerned with
measurements made within laboratories, and the process of sampling has been
conducted by, and the responsibility of, a different set of people who are often in
separate organisations. The measurement scientists knowledge of the sampling
process is then very limited. Conversely, the advent of in situ analytical techniques
sometimes enables the measurement scientist to make measurements at the sampling
site and in contact with the material to be sampled. Examples of this situation are
process analysis within industrial production, and in situ measurements on
contaminated land. The placing of the analytical sensor in these situations then
constitutes the taking of a sample, and the measurement scientist becomes not only
aware of, but responsible for, all stages of the measurement process, including the
sampling. Such an awareness of the whole process is important, irrespective of the
division of effort. Since analytical and sampling processes contribute to the
uncertainty in the result, the uncertainty can only be estimated if there is an
understanding of the complete process. Further, optimisation of the relative effort in
sampling and analysis is only possible where sampling and analytical processes are
both understood.
If the different stages are the responsibility of different people, there needs to be good
communication between all of the parties involved. Sampling planners and analytical
scientists need to optimise the whole measurement procedure, and to devise a strategy
to estimate the uncertainty. Both need to discuss the objectives of the measurements
with the customer. All three parties need guidance from the appropriate regulator on
how these estimates of uncertainty are to be acted upon, to ensure the reliability of the
decisions based upon the measurements. To underpin these decisions, all the parties
need reliable estimates of uncertainty, including that arising from sampling. Although
no general guidance can replace expert advice in complex or critical cases, this Guide
describes the methods needed to fulfil the need for reliable estimates of uncertainty
from sampling for most analytical measurement systems.
UfS:2007
Page ii
Summary
This Guide aims to describe various methods that can be used to estimate the
uncertainty of measurement, particularly that arising from the processes of sampling
and the physical preparation of samples. It takes a holistic view of the measurement
process to include all of these steps as well as the analytical process, in the case where
the measurand is defined in term of the value of the analyte concentration in the
sampling target, rather than in just the sample delivered to the laboratory. The Guide
begins by explaining the importance of knowing the total uncertainty in a
measurement for making reliable interpretation of measurements, and judging their
fitness for purpose. It covers the whole measurement process, defining each of the
component steps, and describing the effects and errors that cause uncertainty in the
final measurement.
Two main approaches to the estimation of uncertainty from sampling are described.
The empirical approach uses repeated sampling and analysis, under various
conditions, to quantify the effects caused by factors such as the heterogeneity of the
analyte in the sampling target and variations in the application of one or more
sampling protocols, to quantify uncertainty (and usually some of its component parts).
The modelling approach uses a predefined model that identifies each of the
component parts of the uncertainty, making estimates of each component, and sums
them in order to make an overall estimate. Models from sampling theory can
sometimes be used in this approach to estimate some of the components from a
knowledge of the characteristics of particulate constituents.
Worked examples are given of each of these approaches, across a range of different
application areas. These include investigations of the environment (of soil and water),
of food (at growing and processing) and of animal feed. The estimates of the total
uncertainty of measurement range from a few per cent up to 84% relative to the
measurand. The contribution of the sampling is occasionally small but is often
dominant (>90% of the total measurement variance). This suggests an increased
proportion of the expenditure needs to be aimed at the sampling, rather than the
chemical analysis, if the total uncertainty needs to be reduced in order to achieve
fitness for purpose.
Management issues addressed include the responsibility of the quality of the whole
measurement process, which needs to include the sampling procedure. Guidance is
given on the selection of the most appropriate approach for any application, and
whether one initial validation of the system is sufficient, or whether there is a need for
ongoing monitoring of the uncertainty from sampling using quality control of
sampling. The extra cost of estimating uncertainty is also considered in relation to the
cost savings that can be made by knowing the uncertainty of measurement more
reliably.
Such a Guide can never be fully comprehensive, and although there are appendices
with details of some of the statistical techniques employed and sources of more
detailed advice, there will often be a need for expert advice in more complex
situations. This Guide aims to be a useful introduction to this subject, but we hope it
will also stimulate further research into improved methods of uncertainty estimation.
UfS:2007
Page iii
Contents
Foreword
ii
Summary
iii
vii
Introduction
1.1
1.2
1.3
1.4
1.5
1.6
Terminology
Uncertainty of measurement
5.1
5.2
Specification of measurand
5.3
5.4
5.5
Sources of uncertainty
10
5.6
10
13
13
The measurand
14
14
UfS:2007
Page iv
Empirical approach
14
9.1
Overview
14
9.2
Uncertainty sources
14
9.3
15
9.4
16
9.5
18
9.6
19
20
10.1
Cause-and-effect modelling
20
10.2
20
24
24
24
25
13.1
25
13.2
26
27
14.1
Introduction
27
14.2
Expanded uncertainty, U
27
14.3
28
14.4
Contributions to uncertainty
28
14.5
Applicability of estimates
29
14.6
29
29
30
32
17.1
32
17.2
32
17.3
32
17.4
Fitness-for-purpose criteria
32
17.5
33
17.6
33
UfS:2007
Page v
Appendix A: Examples
34
35
40
46
Example A4: Vitamin A in baby porridge containing fruit and milled cereals
55
Example A6: Cadmium and phosphorous in agricultural top soil by modelling approach
74
Appendix B: Terminology
83
91
C1. Estimating bias between two sampling methods, by using paired samples 91
C2. Further description of sampling errors from sampling theory
93
96
97
97
98
UfS:2007
100
Page vi
UfS:2007
analysis of variance
analytical quality control
Bureau International des Poids et Mesures
European Committee for Standardization
constitution heterogeneity
certified reference material
collaborative trial in sampling
degrees of freedom
distribution heterogeneity
trade name of body that organises international proficiency tests
fundamental sampling error
global estimation error
graphite furnace atomic absorption spectrometry
grouping and segregation error
Guide to the Expression of Uncertainty in Measurement. ISO
high performance liquid chromatography
increment delimitation error
International Electrotechnical Commission
International Federation of Clinical Chemistry and Laboratory Medicine
increment and sample preparation error
International Organization for Standardization
International Union of Pure and Applied Chemistry
International Union of Pure and Applied Physics
increment extraction error
limit of detection
limit of quantification
measurement uncertainty
National Institute of Nutrition and Seafood Research
National Institute of Standards and Technology
International Organization of Legal Metrology
point selection error
point mineralisation error
proficiency test
quality assurance
quality control
robust analysis of variance
relative standard deviation
reference sampling target
standard deviation
sampling proficiency test
sum of squares
weighting error
total analytical error
total sampling error
Page vii
Introduction
Introduction
Rationale for the Guide
The main purpose of measurement is to enable decisions to be made. The reliability of these
decisions depends on knowing the uncertainty of the measurement results. If the uncertainty
of measurements is underestimated, for example because the sampling is not taken into
account, then erroneous decisions may be made that can have large financial consequences.
The fitness for purpose of measurement results can only be judged by having reliable
estimates of their uncertainty. For this reason it is essential that effective procedures are
available for estimating the uncertainties arising from all parts of the measurement process.
These must include uncertainties arising from any relevant sampling and physical preparation.
Judgements on whether the analytical contribution to the uncertainty is acceptable can only be
made with knowledge of the uncertainty originating in the rest of the measurement procedure.
1.2
1.2.1
The aim of this Guide is to explain the rationale, and practical application, of the
methods available for the estimation of uncertainty that includes the contribution from
sampling. The Guide does not aim to recommend individual sampling protocols, which are
often prescribed in other documents or regulations, but rather to consider the measurement
uncertainty generated by whatever protocol is employed.
1.2.2
The Guide also aims to explain the importance of sampling to the overall uncertainty
budget, and hence to the reliability of the consequent decisions made using the measurements.
As well as explaining how to estimate the uncertainty, the Guide will explain the justification
for including sampling in the overall management of the measurement process.
1.2.3
Unlike the assumption that is often made for estimates of uncertainty for an
analytical method, an estimate for one sampling protocol for one batch of material should not
be assumed as automatically applicable to any subsequent batch of material. For example,
depending on the sampling target, the degree of heterogeneity (i.e. inhomogeneity) may have
changed substantially. There will be a need, therefore, for routine monitoring of key
parameters of sampling quality to examine and update estimates of uncertainty for subsequent
batches.
1.3
UfS:2007
Page 1
Introduction
1.4
This Guide is intended primarily for specialists such as sampling planners and for analytical
chemists who need to estimate the uncertainty associated with their measurement results.
Other stakeholders should seek specialist advice for particular applications.
1.5
1.5.1
Current practice in the estimation of uncertainty for a broad range of measurements
follows the Guide to the Expression of Uncertainty in Measurement (the GUM) [2],
published in 1993 by ISO in collaboration with BIPM, IEC, IFCC, IUPAC, IUPAP and
OIML. The GUM sets out the concepts required, established the general principles, and
provided a procedure applicable to cases where an adequate model of the measurement
process is available. The application of this approach to chemical analysis was described in
1995 in a Eurachem Guide for Quantifying Uncertainty in Analytical Measurement [3], and
broadened to include the use of validation and method performance data in a second edition in
2000 [1]. Other useful contributions to the practical estimation of uncertainty of analytical
measurements using collaborative study data have been made by the Analytical Methods
Committee of the Royal Society of Chemistry in 1995 [4], and by ISO TC/69 in 2004 [5].
This Guide on sampling is consistent with the general principles established in the GUM.
1.5.2
Sampling theory has developed largely independently of analytical chemistry and
chemical metrology. Sampling quality has generally been addressed in sampling theory by the
selection of a correct sampling protocol, appropriate validation, and training of sampling
personnel (i.e. samplers) to ensure that this protocol is applied correctly [6]. It is then
assumed that the samples will be representative and unbiased, and the variance will be that
predicted by the model.
1.5.3
An alternative approach is to estimate the uncertainty of sampling for typical
materials, or for sampling targets, during validation of the sampling protocol, and to confirm
compliance in practice using ongoing quality control. This is more consistent with procedures
already in place for the rest of the measurement process. Interestingly, the quality of sampling
is only quantifiable through the measurements that are made upon the resultant samples.
1.5.4
Sampling protocols have been written to describe the recommended procedure for
the sampling of innumerable types of material and for many different chemical components.
These protocols are sometimes specified in regulation or in international agreementsa[7].
These procedures rarely identify the relative contributions of sampling and chemical analysis
to the combined uncertainty.b
1.5.5
There is accordingly a range of prior literature on the theory and practice of
sampling. As explained in Section 1.2.1, this Guide therefore does not seek to propose further
sampling protocols but rather to provide methodologies to quantify the uncertainty that arises
when a given protocol is used.
1.5.6 A handbook describing procedures for the estimation of uncertainty from sampling,
derived from this Guide but with further case studies, has been prepared by the Nordtest
group [8].
The acceptance sampling procedures are applied to the sampling of a wide range of materials [7].
Some concepts from sampling theory can be usefully adapted for estimation of uncertainty of measurement
(Section 10.2).
UfS:2007
Page 2
1.6
1.6.1
This document summarises the concepts necessary for understanding uncertainty in
sampling, and provides procedures that allow their practical implementation. The Guide
additionally covers issues related to management, quality assurance and reporting of results
with uncertainty. The scope and intended field of application are set out in Section 2, which
also summarises the approaches covered. Terminology is discussed in Section 3, and key
terms defined in Appendix B.
1.6.2
Fundamental concepts are covered in Sections 4 and 5. An overview of the
measurement process is provided in Section 4. This includes an explanation of the sampling
terminology used, and indicates which steps in the process are considered in detail in this
Guide. Measurement uncertainty and its sources are discussed further in Section 5.
1.6.3
Sections 6 to 10 describe methodologies for the estimation of uncertainty, with a
discussion of the merits of the various options. The two broad approaches available are
summarised in Section 6, and covered in detail in Sections 1 and 10 respectively. The intent is
to provide a range of options that may be applied, rather than to specify any particular
approach.
1.6.4
Management and quality issues are addressed in Sections 11 to 1. These include a
very brief discussion of responsibilities for quality in sampling (Section 11) before discussing
the selection of uncertainty estimation approach in Section 1. The use of sampling quality
control to monitor sampling performance is covered in Section 1. Reporting and use of
uncertainty, and its effect on the reliability of decisions, are discussed in Section 14. Cost is
an important factor and selection of the most cost-effective and appropriate method of
estimation is explained in Section 15. Knowing the value of the uncertainty helps to judge the
fitness for purpose of the measurement as a whole, and its component parts, and this is
discussed in Section 16.
1.6.5
A range of examples, a detailed glossary of terms and definitions used in this Guide,
some important statistical procedures and experimental designs, and a discussion of
improving sampling uncertainty using predictions from sampling theory are provided as
appendices.
2.1
The principles of this Guide are applicable to the estimation of uncertainty from the
full range of materials that are subject to analytical measurement (e.g. gaseous, liquid and
solid). These include environmental media (e.g. rock, soil, water, air, waste and biota), foods,
industrial materials (e.g. raw materials, process intermediaries and products), forensic
materials and pharmaceuticals. This approach is applicable to sampling by any protocol,
whether it uses single or composite samples, or single or multiple determinations.
2.2
The Guide describes the estimation of uncertainty using i) replicated measurement
and sampling (the empirical approach) and ii) modelling based on identified influence
quantities and theoretical considerations (the modelling approach).
2.3
The use of uncertainty estimates in the assessment of fitness for purpose and in the
optimisation of effort among individual parts of the measurement process is covered. Methods
of assessing fitness for purpose that are described include those based upon percentage of
total variance and others based on cost-benefit analysis.
UfS:2007
Page 3
Terminology
2.4
Although the general principles of this Guide apply, it does not specifically discuss
microbiological sampling. Nor does it discuss the estimation of uncertainty in spatial or
temporal information such as the location or size of areas of high analyte concentration.
Terminology
3.1
The precise definitions of many of the terms used in this Guide vary depending on
the area of application. A full listing of terms and their different definitions is given in
Appendix B. In this Guide, normative definitions of each term have been selected that are as
generally applicable as possible to all sectors of application. These terms are listed in
Appendix B and in bold on first use in the text.
UfS:2007
Page 4
Form of
material
Sampling
Sampling Target
Primary Sample
Sub-sample
Laboratory
sample
Test sample
Test portion
Test solution
Physical sample
preparation
Analysis
The figure shows a complete measurement process, starting with primary sampling and ending in the analytical
determination. There are many intermediary steps, such as transportation and preservation of samples, not all of
which are always present. Each step gives rise to a contribution towards the uncertainty of measurement. This
Guide concentrates on the process steps of sampling and physical sample preparation (shaded boxes), as the
last step is well covered in previous guidance [1]. Notice that two of the sampling steps occur within the
laboratory (light grey) and are frequently considered to be part of the analytical process. For definitions of terms
see Appendix B.
A sampling target is the portion of material, at a particular time, that the sample (and therefore
the measurement result) is intended to represent. The sampling target needs to be defined
prior to the design of the sampling plan. It may be defined by regulation, such as the whole of
a batch, lot or consignment. If the properties and characteristics (e.g. analyte concentration) of
the material in a certain area, or time period, are of interest and must be known, then it can be
considered a sampling target. When the composition of a whole batch is required (e.g. of a
food material), then the whole batch constitutes the target. When the spatial (or temporal)
UfS:2007
Page 5
Uncertainty of measurement
variation of concentration is required (e.g. in finding hot spots within a contaminated
material), then each location where the concentration is required will be a separate sampling
target. Any one sampling target will ultimately generate one reported measurement result and
an uncertainty.
Primary samples are often made up of a number of increments, which are combined to form
a composite sample before a measurement is made. It is the uncertainty on this single
measurement value, made on this composite sample, and caused by all of the preparatory
steps, that is required. The value of this uncertainty will often be affected by the number of
increments that are taken. This contrasts with the situation when several distinct primary
samples (n) are taken from different parts of the sampling target, and measured separately. If
the composition of the sampling target is calculated by taking the mean value of these
separate measurements, then the uncertainty on the mean value is calculated using the
standard error of the mean (s/n). This is not the same as the uncertainty on the single
measurement, the estimation of which is the objective of this Guide.
4.1
The whole process of measurement (Figure 1) typically begins with the taking of the
primary sample from a sampling target. The resulting sample goes through one or more of a
series of steps prior to the analytical determination. All steps contribute to uncertainty in the
final result, when the analyte value required (i.e. measurand value or true value), is
expressed in terms of the analyte concentration in the sampling target. Guidance already
exists on the estimation of the analytical steps of the measurement process [1]. This will
certainly include the selection of the test portion, the chemical treatment preceding
measurement and the analytical determination, but may also include the physical preparation
of the laboratory sample by means such as drying, sieving, milling, splitting and
homogenisation.
4.2
In common practice, all the various portions of material in the second column of
Figure 1 are often referred to simply as a sample. It is clearly important to differentiate them
carefully in discussion, especially those considered particularly in this Guide (in shaded boxes
on Figure 1). This is discussed in more detail in Section 5.2.
4.3
Methods described in the Guide will help to identify the dominant source of the
uncertainty, such as the sampling rather than the chemical analysis, but will not necessarily
explain the cause. However, heterogeneity within the sampling target, either spatial or
temporal, is known to be a significant cause of uncertainty in many circumstances. Separate
studies would be needed to characterise the variability that contributes to the uncertainty. For
the purpose of this Guide, heterogeneity within the sampling target is treated as just one cause
of uncertainty in the final measurement. This is the case, whatever actions are taken to
minimise the effects of the heterogeneity by the application of any particular sampling
protocol.
5
5.1
Uncertainty of measurement
Definition of uncertainty of measurement
5.1.1
Uncertainty of measurement, or measurement uncertainty (MU), is defined in
metrological terminology [2] as:
Parameter, associated with the result of a measurement, that characterises the dispersion of the
values that could reasonably be attributed to the measurand.
UfS:2007
Page 6
Uncertainty of measurement
The definition includes several important features, which are discussed in the following
paragraphs.
5.1.2
The parameter may be, for example, a range, a standard deviation, an interval (like
a confidence interval) or half-interval (u is a statement of a half-interval) or other measure of
dispersion such as a relative standard deviation. Note that when MU is expressed as a
standard deviation, the parameter is known as standard uncertainty, usually given the
symbol u. Other forms of expression are considered in Section 14.
5.1.3
Uncertainty is associated with each measurement result. A complete measurement
result typically includes an indication of the uncertainty in the form xU, where x is the
measurement result and U an indication of the uncertainty (it will be seen that the symbol U
has a special meaning, in this Guide; it indicates an expanded uncertainty, which will be
discussed further in Section 14). This form of expressing a result is an indication to the enduser of the result that, with reasonable confidence, the result implies that the value of the
measurand is within this interval.
5.1.4
The measurand is simply a quantity, such as a length, mass, or concentration of a
material, which is being measured. The term value of the measurand is closely related to the
traditional concept of true value in classical statistical terminology. From this alternative
viewpoint uncertainty has also been defined [9] as:
An estimate attached to a test result which characterises the range of values within which the true
value is asserted to lie
This definition (which will be referred to as the statistical definition) has the advantage of
being easier to explain to decision makers, who often recognise the phrase true value as the
value of interest for their decision. It has the disadvantage that the true value itself can never
be known and this generally requires further explanation.
5.1.5
The metrological definition asserts that uncertainty expresses the dispersion of the
values that could reasonably be attributed to the measurand. This is a particularly important
phrase. It indicates that although the uncertainty is associated with a measurement result, the
range quoted must relate to the possible range of values for the measurand. For example, the
measurand could be the total mass of gold in a geological deposit. This is quite different from
a statement of precision, which would describe the range of results that might be observed if
the measurement were repeated. In requesting information about where the measurand value
might be, this definition of uncertainty implicitly requires the measurement scientist to
consider all the effects that might influence the measurement result. These effects obviously
include the causes of random variation from one measurement to the next over the
measurement timescale. But it is also essential to consider sources of bias during the
experiment, and very often, these generate larger effects than can be observed by repeated
measurement alone. That is, measurement uncertainty automatically asks for a range that
includes an allowance for both random and systematic effects.
5.1.6
To consider a simple analytical example, a simple measurement of concentration in a
solid will typically involve extraction of material, weighings, volumetric operations and
perhaps spectrometry or chromatography. Repeated measurement will show a spread of
values due to random variations in these operations. But all analysts know that extraction is
rarely complete, and, for a given material, that failure to extract material will lead to a
consistently low result. While good analytical practice always attempts to reduce such effects
to insignificance, some bias will remain. In expressing the uncertainty about the value of the
measurand, then, the analyst must take into account the reasonable possibility of bias from
UfS:2007
Page 7
Uncertainty of measurement
such causes. (Usually, this is done by considering such information as the range of analyte
recoveries observed on reference materials or from spiking experiments.)
5.1.7
The same considerations apply in the case of sampling. It is well known that different
samples taken from a bulk material will often show real variation in value, which is clear
from repeated measurement. It is also well known that sampling may be biased, for example
by differential removal of materials, inappropriate timing of sampling where temporal
fluctuations occur, or by access restrictions. These effects will influence the relationship
between the value of the measurand and the result that is observed. While good practice in
sampling is intended to reduce these effects to insignificance, a careful assessment of
uncertainty always considers the possibility of residual systematic effects.
5.1.8
Current guidance on measurement uncertainty [2] makes it clear that uncertainty of
measurement (Section 2.2 of reference [2]) is not intended to allow for gross error. This
would preclude, for example, mistakes caused by transcription errors or gross misuses of the
measurement protocol. Sampling can, however, produce high levels of uncertainty (e.g. 80%
of the concentration value), simply through the routine application of an accepted
measurement protocol to a highly heterogeneous material. Even when procedures are
nominally correct, there will also be slight variations in the actual procedures due to the
ambiguity in the measurement protocols, and the minor adaptations that are made to protocols
in real-world sampling situations. Whether these high levels of uncertainty lead to
unacceptable levels of reliability in the decisions that are based upon them, depends upon a
rigorous evaluation of fitness for purpose (see Section 16).
5.2
Specification of measurand
5.2.1
When an end-user is presented with a concentration result quoted for a bulk sample
in the form xU, they will very naturally interpret that interval as including the range of
values attributable to the concentration in the sampling target (e.g. a batch of material).
Implicit in this view is the idea that the measurand is the (true) concentration (of the analyte)
in the batch of material, and that the uncertainty includes any necessary allowance for
heterogeneity in the bulk. The analyst, by contrast, might refer to the concentration in the
laboratory sample analysed, implicitly ruling out the variation between laboratory samples.
Clearly, one viewpoint includes the effects of sampling, while the other does not. The effect
on the uncertainty can, of course, be very considerable. In metrological terms, this distinction
arises because the two views are considering different measurands. One is considering
concentration in the sampling target, the other concentration in the laboratory sample.
Another example might be contaminant concentration at a factory outlet at the time of
sampling, compared to the average contaminant concentration over a year.
5.2.2
These ambiguities in interpretation can be avoided only by careful specification of
the measurand. It is clearly necessary to state the quantity (mass, length, concentration etc.). It
is equally important to be clear on the scope of the measurement, by including information on
factors such as the time, location, or population to which the measurement result will be
assumed to apply. Some particular instances of measurand specification and their implications
for uncertainty estimation are discussed below.
It is never possible to avoid all ambiguity in implementing the wording of the sampling
protocol.
5.2.3
When a composite sample is taken by the combination of several increments from
across a sampling target, and analysed as a single primary sample, that single determination of
analyte concentration provides an estimate of the value of the measurand (i.e. the average
composition of the target), as discussed briefly in Section 4. The uncertainty on this single
UfS:2007
Page 8
Uncertainty of measurement
value reflects the uncertainty in the estimate of the measurand value. In contrast, if several
independent primary samples are taken from the target, each analysed once, and the mean
value calculated, this mean value will also be an estimate of the value of the measurand.
However, the uncertainty will not be that of the measurement (expressed as standard
deviation, s), but the standard error of the mean value (expressed as s/n). This later
uncertainty on the mean can be reduced by taking more primary samples,c whereas the
uncertainty on the measurement cannot.
5.3
5.3.1
Uncertainty is related to other concepts, such as accuracy, error, trueness, bias and
precision. Other guidance discusses the relationships in some detail [1, 2]. However, it is
worth repeating some of the important differences:
Uncertainty is a range of values attributable on the basis of the measurement result and
other known effects, whereas error is a single difference between a result and a true (or
reference) value
Uncertainty includes allowances for all effects that may influence a result (i.e. both
random and systematic errors); precision only includes the effects that vary during the
observations (i.e. only some random errors).
Uncertainty is valid for correct application of measurement and sampling procedures, but,
as noted in Section 5.1.8, it is not intended to make allowance for gross operator error
5.4
5.4.1
The act of taking a sample introduces uncertainty into the reported measurement
result wherever the objective of the measurement is defined in terms of the analyte
concentration in the sampling target and not simply in the laboratory sample.
5.4.2
Sampling protocols are never perfect in that they can never describe the action
required by the sampler for every possible eventuality that may arise in the real world in
which sampling occurs. The location in space (or time) for the taking of a sample is rarely
specified exactly (e.g. to the nearest millimetre or second). The sampler has to make such
decisions (ideally on objective criteria), but as heterogeneity is inevitable (in space or time)
such decisions will affect the estimated concentration. An appreciation of these sources of
uncertainty is important in the design and implementation of methods for the estimation of
uncertainty. When duplicate samples are taken, for example, taking them at exactly the same
place and time may not reflect the uncertainty of the measurement that really exists. This will
be discussed further in the description of methods of estimation (Sections 6 to 10), and in the
various worked examples (Appendix A).
5.4.3
Heterogeneity always gives rise to uncertainty. If the sampling target were perfectly
homogeneous then this contribution would be zero, but nearly all materials are heterogeneous
to some extent at some scale. If the test portion is a few micrograms, then nearly all material
will be heterogeneous and the sampling step will contribute to the uncertainty in the
measurement of an analyte concentration. Heterogeneity can be quantified in a separate
experiment, but if the aim is to estimate the analyte concentration in the larger sampling
target, then this heterogeneity is just one cause of measurement uncertainty (as discussed in
Section 4.2).
Assuming that the samples are random and independent, and assuming zero bias.
UfS:2007
Page 9
Uncertainty of measurement
5.4.4
Similar arguments can be made for the uncertainty that arises in the processes of
physical preparation (e.g. transportation, preservation, comminution, splitting, drying, sieving,
homogenisation) that happen after the act of sampling and before the chemical preparation of
the test sample (Figure 1). Each step can introduce errors from a range of mechanisms, such
as loss of analyte, loss of fine particles, or contamination from equipment or previous
samples. The methods employed, and training given, should aim to reduce these errors to a
minimum. In addition, however, procedures are required to estimate the uncertainty that all of
these steps, when applied in practice, generate in the final measurement value.
5.5
Sources of uncertainty
5.5.1
Uncertainty arises from a variety of sources, and these have been categorised in
different ways. For example, the Eurachem Uncertainty Guide identifies eight major
categories of effects that are important in estimating uncertainty have been identified [3], of
which the first two are sampling and sample preparation. Specific effects identifiable in these
two categories are shown in Table 1. A modelling approach might use these effects as the
basis for a mathematical model. Alternatively, sampling theory identifies eight distinct
sources of error in sampling (Table 2); each of these can also be reduced to a variety of causal
factors, which in turn can be used in various modelling approaches. A further alternative
approach is to consider all of the steps in the measurement process (Figure 1) as sources of
uncertainty that make some contribution to the uncertainty of the final measurement. In this
Guide, the simplest study designs treat uncertainty as arising from four classes of effect
(Table 3), and the classes are treated as sources of uncertainty in a simple statistical model;
this is consistent with the grouping of uncertainty sources explicitly suggested in reference
[3]. In its simplest form, this categorisation can be reduced to two categories: sampling
uncertainty and analytical uncertainty.
5.5.2
The important feature of each of these different classifications is that each is intended
to ensure that, however they are grouped and evaluated, all practically important effects are
taken into account in estimating the uncertainty. As long as this requirement is met, any
categorisation scheme may be applied to the estimation of uncertainty. The categorisation
schemes listed in Table 2 and Table 3 cover all practically important effects.
5.5.3
Each different categorisation of sources will generally lead to a different study
design, and very often to fundamentally different methods of evaluation of uncertainty
contributions. This results in substantially independent estimates of uncertainty via different
approaches. As noted elsewhere [5], grossly different estimates of uncertainty for the same
system suggest that at least one study methodology is in error. This forms the basis of a check
on the validity of an approach. Where practicable, therefore, comparison of uncertainty
estimates arising from independent evaluation approaches is recommended as a means of
validating particular estimates and of assessing the validity of different approaches.
5.6
5.6.1
IUPAC currently define both homogeneity and heterogeneity as The degree to
which a property or constituent is uniformly distributed throughout a quantity of material.
(see Appendix B for definitions). So defined, heterogeneity is among the most important
factors contributing to uncertainty associated with sampling. Increments from different
locations in the sampling target will have different concentrations of analyte in a
heterogeneous material and there will be a sample-to-sample variation in analyte
concentration usually visible as a contribution to the observed variation of results. In
general, the exact dependence of concentration on location is unknown, so no correction can
UfS:2007
Page 10
Uncertainty of measurement
be made. This results in uncertainty in any given result or, in general, any average of such
results.
5.6.2
IUPAC note, as an addendum to the above definition, that The degree of
heterogeneity (the opposite of homogeneity) is the determining factor of sampling error. The
note is a good indication of the importance of heterogeneity in sampling. There are other
sources of error and uncertainty in the general operation of sampling; for example, crosscontamination and imperfect stabilisation of samples, either of which can result in (unknown)
bias or additional variability. Yet heterogeneity and its effects such as random variability
and selection bias remain the largest problem in properly managed sampling and will
generally be the most significant source of uncertainty.
5.6.3
An alternative definition of homogeneity is sometimes used for particulate material,
which, if it consists of particles of different materials, cannot ever be homogeneous in the
sense defined by IUPAC. In this context, a mixture in which the probability of selection of
different types of particle is constant throughout the sampling target may be termed
homogeneous to denote that the expected concentration in would be the same in a sample
taken at any point in the material. Even here, however, it must be recognised that the
particulate nature of the material leads to sample-to-sample variation due to slightly different
composition of the increments actually taken; heterogeneity, as defined by IUPAC, still has
an effect under these circumstances, and consequently still contributes to the uncertainty.
Sample preparation
UfS:2007
- Dissolution
- Extraction
- Contamination
- Derivatisation (chemical effects)
- Dilution errors
- (Pre-)Concentration
- Control of speciation effects
Page 11
Uncertainty of measurement
Table 2: Sources of sampling uncertainty in sampling theory*
Source
Description
*This classification follows that of Gy [17] and others (discussed further in Section10).
** Excluded from uncertainty estimates as gross errors [2].
Effect class*
Random (precision)
Systematic (bias)
Analysis
Analytical variability
(combined contribution of
random effects)
Sampling
Sampling variability
(dominated by heterogeneity
and operator variations)
Sampling bias
(combined effect of selection bias,
operator bias etc.)
*The differentiation of random from systematic effects can depend on the context. A systematic effect in
measurements by one organisation (e.g. analytical bias) can also be considered a random effect when viewed in
the context of the consensus value from an inter-organisational proficiency test.
UfS:2007
Page 12
6.1
There are two broad approaches to the estimation of uncertainty. One of them,
described as empirical, experimental, retrospective, or top-down, uses some level of
replication of the whole measurement procedure to give a direct estimate of the uncertainty
for the final result of the measurement. This approach is called the empirical approach in
this Guide. The second, variously described as modelling, theoretical, predictive or
bottom-up, aims to quantify all of the sources of uncertainty individually, and then uses a
model to combine them. It will accordingly be referred to as the modelling approach. These
approaches are not mutually exclusive. The empirical method can be adapted to estimate
contributions to uncertainty from one or more effects or classes of effect. Both approaches can
usefully be used together to study the same measurement system, if required. The
applicability of the two approaches varies between the different materials to be sampled.
6.2
The approach taken in this Guide is to describe in detail the empirical approach,
which has the widest applicability to the broadest range of measurement systems and
applications (e.g. gaseous, liquid and solid). Modelling approaches are described for
particular situations to which they apply (e.g. particulate solids). Advice will also be given on
how a combination of these different approaches can be used to give more reliable and costeffective estimates of uncertainty in a range of measurement systems. This dual approach is
intended to enable a user of the Guide to select the most appropriate method of uncertainty
estimation for their particular situation. (Section 1 provides guidance on selection of the
approach.)
6.3
Reference [5] notes that the modelling approaches and the type of empirical study
used in collaborative trial are extremes of a continuum:
Note, however, that observed repeatability or some other precision estimate is very
often taken as a separate contribution to uncertainty even in the [modelling] approach.
Similarly, individual effects are usually at least checked for significance or quantified
prior to assessing reproducibility. Practical uncertainty estimates accordingly often
use some elements of both extremes.
In referring to either extreme, therefore, it is important to be aware that these are extremes and
that many practical estimates involve elements of both approaches.
6.4
The overall objective of any approach is to obtain a sufficiently reliable estimate of
the overall uncertainty of measurement. This need not necessarily require all of the individual
sources of uncertainty to be quantified, only that the combined effect be assessed. If, however,
the overall level of uncertainty is found to be unacceptable (i.e. the measurements are not fit
for purpose) then action must be taken to reduce the uncertainty. Alternatively, the
uncertainty may be unnecessarily small, in which case there may be justification for
increasing the analytical uncertainty, and thereby decreasing the cost of analysis. Methods for
modifying uncertainty are discussed in Appendix E. At this stage, however, it is essential to
have information on which general part of the measurement procedure is causing the
dominant contribution to the overall uncertainty, and it may then be necessary to evaluate
individual effects. The advantage of detailed early study is that this information is already
available; the disadvantage is that it is costly to obtain and may prove unnecessary if
uncertainty is acceptable. Planners should accordingly consider the level of detail required in
UfS:2007
Page 13
The measurand
7.1
In the following discussion, it is assumed that the measurand is an average value
representing the composition of the whole sampling target, and that the measurand is to be
estimated through a process of sampling and analysis. This relates to the specification of the
measurand (Section 5.2) and the definition of the sampling target (Section 4.1).
8.1
Analytical work should be undertaken under an appropriate quality system, including
validated analytical methods, proficiency testing, internal quality control and external
assessment where appropriate. Validation procedures should include all the steps normally
undertaken within the laboratory (including any sub-sampling of test samples), and should
include checks on bias using certified reference materials, or other methods, for the estimation
of analytical bias. Note that the uncertainty estimation methods described in this Guide can
also be applied to the estimation of uncertainties associated with sub-sampling.
8.2
Laboratories undertaking the chemical analysis should report the concentration
estimates exactly as found; in particular, values must not be censored, truncated or reported as
less than a reporting limit, whether below the limit of detection (LOD) or below zero.
Failing to report negative or sub-LOD observations will result in an underestimate of the
uncertainty.
9
9.1
Empirical approach
Overview
9.1.1
The empirical (top-down) approach is intended to obtain a reliable estimate of the
uncertainty, without necessarily knowing any of the sources individually. It relies on overall
reproducibility estimates from either in-house or inter-organisational measurement trials. It is
possible to describe the general type of source, such as random or systematic effects, and to
subdivide these as those arising from the sampling process or the analytical process. Estimates
of the magnitude of each of these effects can be made separately from the properties of the
measurement methods, such as sampling precision (for random effects arising from
sampling) or analytical bias (for systematic effects arising from chemical analysis). These
estimates can be combined to produce an estimate of the uncertainty in the measurement
result. This approach is illustrated in detail in Examples A1, A2, A3 and A4.
9.2
Uncertainty sources
9.2.1
It is possible to consider uncertainty of measurements to arise from four broad
sources of error. These four sources are the random errors arising from the methods of both
the sampling and analysis, and also the systematic errors arising from these methods. These
errors have traditionally been quantified as the sampling precision, analytical precision,
sampling bias and the analytical bias respectively (Table 4). If errors from these four sources
are quantified, separately or in combinations, it is possible to estimate the uncertainty of the
measurements that these methods produce. Methods for the estimation of three of the four
errors are well established. Sampling and analytical precision can be estimated by duplication
UfS:2007
Page 14
Empirical approach
of a proportion (e.g. 10%) of the samples and analyses respectively. Analytical bias can be
estimated by measuring the bias on well-matched certified reference materials, and assuming
that this bias represents that present for the test material, or by taking it directly from the
validation of the analytical method.
9.2.2
Procedures for estimating sampling bias include the use of a reference sampling
target [10, 24] (the sampling equivalent of a reference material). Alternatively they utilise
measurements from inter-organisational sampling trials, in which the sampling bias
potentially introduced by each participant is included in the estimate of uncertainty based on
the overall variability (Section 9.5). Although some of the components of uncertainty
associated with systematic effects may be difficult to estimate, it may be unnecessary to do so
if there is good evidence that systematic effects are small and under good control. Such
evidence may be qualitative, as in prior knowledge of the chemical or physical nature of the
sampling target, or quantitative, such as information, for example from prior measurements
on complete batches. (See Examples A3 and A4, Appendix A.)
Table 4: Estimation of uncertainty contributions in the empirical approach
Process
Effect class
Random (precision)
Systematic (bias)
Analysis
Sampling
Duplicate samples
Four classes of effects that contribute to the uncertainty of measurements, and methods for their estimation.
9.3
In order to design experimental methods to estimate uncertainty using the empirical approach
it is necessary to have a statistical model describing the relationship between the measured
and true values of analyte concentration. This random effects model considers a single
measurement of analyte concentration (x), on one sample (composite or single), from one
particular sampling target:
x = X true + sampling + analysis
where X true is the true value of the analyte concentration in the sampling target (i.e. equivalent
to the value of the measurand). For example, this could be the total mass of the analyte in the
target divided by the total mass of the target. The total error due to sampling is sampling and the
total analytical error is analysis
In an investigation of a single sampling target, if the sources of variation are independent, the
measurement variance 2meas is given by,
UfS:2007
Page 15
Empirical approach
s 2 meas = s 2 sampling + s 2 analytical
The standard uncertainty (u) can be estimated using smeas , which is therefore given by
Equation 1
Variance caused by physical sample preparation can be included into the sampling variance,
or expressed as a separate term if required.
In a survey across several sampling targets, which is recommended for the estimation of
sampling uncertainty (Section 9.4.2), the model needs to be extended to
is also a useful parameter in assessing fitness for purpose; this is discussed further in section
16.2. For practical purposes the population variances are replaced by their estimates s2 to
give
s 2 total = s 2 betweentarget + s 2 sampling + s 2 analytical
9.4
Equation 2
9.4.1
Four types of method are applicable to the estimation of uncertainty using the
empirical approach (Table 5). A fifth variographic method is described briefly in Section 9.6.
The main method described further in this Guide is the duplicate method (#1). If one
sampler uses several sampling protocols in Method #2, any bias between the protocols can be
detected. If multiple samplers all apply one protocol (Method #3, which is equivalent to a
collaborative trial in sampling CTS, or method performance test), then bias between
different samplers can be detected and included in the uncertainty estimate. If multiple
samplers apply different protocols that are selected to be the most appropriate for the stated
objective, in their professional opinion (Method #4, which is equivalent to a sampling
proficiency test SPT), then any sampling bias introduced by either the sampling protocol, or
the sampler, can be detected and included in the estimate of uncertainty.
UfS:2007
Page 16
Empirical approach
Method
description
Samplers
(persons)
Protocols
Component estimated
Sampling
Analytical
Precision
Bias
Precision
Bias
No
Yes
No1
Duplicates
Single
Single
Yes
Protocols
Single
Multiple
Between protocols
Yes
No 1
CTS
Multiple
Single
Between samplers
Yes
Yes 2
SPT
Multiple
Multiple
Between protocols
+between samplers
Yes
Yes 2
Analytical bias information may be obtained by including certified reference materials in the analytical run (see
Example A2, Appendix A).
2
Analytical bias is partially or completely included in collaborative exercises where multiple laboratories are
involved.
9.4.2
The duplicate method is the simplest and probably most cost-effective of the four
methods described in Table 5. It is based upon a single sampler duplicating a small proportion
(i.e. 10%, but no less than eight targets) of the primary samplesd [11, 12]. Ideally the
duplicates are taken from at least eight sampling targets, selected at random to represent the
typical composition of such targets. If only one target exists, then all eight duplicates can be
taken from it, but the uncertainty estimate will only be applicable to that one target. The
duplicated samples are taken by repeating the same nominal sampling protocol, with
permitted variations that reflect the ambiguity in the sampling protocol and the effect of
small-scale heterogeneity of the analyte of interest on the implementation of that protocol. For
example, in a W design for collecting a composite sample over a bay of lettuces, the initial
starting point and orientation of the W is altered for the duplicate sample; for a grid design,
again, starting point and orientation are altered (ExampleA1, Appendix A). The duplicate
samples are obtained using a single sampling protocol and by a single person (sampler). Both
of the duplicated samples are subject to physical preparation resulting in two separate test
samples. Duplicate test portions are drawn from both of the test samples and analysed in
duplicate (i.e. duplicate chemical analysis). This system of duplicated sampling and chemical
analysis is known as a balanced design (Figure 2). Note that the duplicate method does not
include any contribution from sampling bias, which must be either assumed to be negligible,
or estimated separately using, for example, multiple samplers, multiple protocols and/or interorganisational sampling trials as in the other three methods.
Note: Although the duplicate method is generally described in terms of a single sampler and
protocol, the same design can be used with different samplers to incorporate the betweenoperator contribution to uncertainty (equivalent to Method #3).
A higher level of replication can be used, but duplication is usually the most effective form of replication in
sampling studies. It is better to take duplicates from 12 sample targets, than take triplicates from eight targets, for
example, as although each estimate of the uncertainty of sampling (ssampling) has a lower standard error, the
estimate is based on a smaller proportion of the entire population of sample targets, and is therefore less
representative. The minimum number of eight duplicates is required to provide sufficiently reliable estimates of
the uncertainty [12].
UfS:2007
Page 17
Empirical approach
Figure 2: A balanced design
10% of targets in whole
survey n 8
between-target variance
Sampling
target
Sample 1
Sample 2
between-sample variance
sampling precision s samp
-between-analysis variance
Analysis 1
Analysis 2
Analysis 1
Analysis 2
analytical precision s
anal
Balanced experimental design for empirical estimation of uncertainty (i.e. two-stage nested design), using the
duplicate method.
9.4.3
The test portions are then chemically analysed anonymously by an appropriate
analytical method under repeatability conditions (e.g. distributed randomly within an
analytical batch). If estimates of the analytical portion of the measurement uncertainty have
been made independently by the laboratory, this will be useful for comparison with estimates
made by this method. Variance caused by physical sample preparation can be included into
the sampling variance by having independent preparation on each of the sample duplicates.
Alternatively this variance can be estimated separately by inserting an extra level of
replication in the experimental design (Appendix D).
9.4.4
The balanced design proposed here will only give the repeatability standard deviation
of the analytical measurements. In order to estimate the other part of the analytical
uncertainty, an allowance has to be made for potential analytical bias. The limitations of this
approach, and a worked example, are given in Section 6 of Example A2. One alternative is to
ask the measuring laboratory for the repeatability and measurement uncertainty, and then to
check that the repeatability obtained in this study is similar to that claimed by the laboratory.
If this is the case, we can use the measurement uncertainty given by the lab as u(analytical)
(normally U/2). A second alternative is to use the estimation of analytical bias made from the
well-matched certified reference materials contained in the analytical batch. This bias
estimate can then be combined with the repeatability to obtain the measurement uncertainty
[1, 30].
9.5
9.5.1
The random component of the uncertainty can be estimated by applying analysis of
variance (ANOVA)e or range calculationf to the measurements of concentration on the
duplicated samples. The estimation is based upon the model described in Section 9.3, applied
to whatever measurement protocol that is being employed (with its specified number of
sample increments and analytical replicates).
There is often a small proportion (i.e. <10%) of outlying values in the frequency distributions of the analytical,
within-sample and between-sample variability. This requires the use of some method of down-weighting the
effect of the outlying values on classical ANOVA, such as the use of robust statistical methods. This gives a
more reliable estimate of the variances of the underlying populations. A fuller explanation of these methods is
given in the worked example in Appendix A1 (p40), A2 (p45).
UfS:2007
Page 18
Empirical approach
9.5.2
The values of ssamp and sanal from the ANOVA are estimates of sampling precision
and analytical precision respectively. The random component of the measurement uncertainty
is calculated by the combination of these two estimates (Equation 1). The expanded
uncertainty, for approximately 95% confidence for example, requires this value to be
multiplied by a coverage factor of 2. The expanded uncertainty (U) is then calculated using
U = 2 smeas
Equation 3
U can also be expressed relative to the reported value x and expressed in terms of a
percentage, as a relative expanded uncertainty U':
U ' = 100
2 s meas
%
x
Equation 4
The relative uncertainty is more widely applicable than the standard uncertainty, as it does not
change appreciably as a function of concentration at values well above the analytical detection
limit (>10 times). Other coverage factors can be selected as appropriate. The improvement of
this estimate of uncertainty to include the systematic error from the chemical analysis is
discussed in Example A2, Appendix A.
The relative expanded uncertainty for the sampling or analysis alone can similarly be
expressed as
U samp '= 100
2 ssamp
x
and
2sanal
9.5.3
Because the uncertainty of many measurement systems is dominated by
heterogeneity within the sampling target, the use of the simplest duplicate method often
gives a reasonably reliable estimate of uncertainty. Studies of environmental systems have
shown that between-operator and between-protocol effects are often much smaller than those
caused by heterogeneity [43]. Further information on selection of the most effective method
for uncertainty estimation is provided in Section 1. Examples of applications of the duplicate
method are given in Examples A1 and A2, Appendix A.
9.5.4
In addition to an initial single estimate of uncertainty for a particular sampling
protocol applied to a particular sampling target, routine application of the duplicate method
is also useful as a way of monitoring the ongoing sampling quality (Section 13). This can
allow for the effect on uncertainty of changes in the heterogeneity of the sampling target
between different applications of the same sampling protocol. Quantitative evidence of the
quality of sampling can then be gained, rather than relying solely on the assumption that
samples are representative, if taken by a correct protocol.
9.6
9.6.1
Variography has also been suggested as a further empirical means of estimating the
uncertainty of measurement from the combined sources of sampling and analysis [13]. It is
particularly useful in situations where there is large-scale spatial and/or temporal variation in
contaminant concentration that can be quantified and modelled. This is the case for some
instances of rock and soil geochemistry, and in emission control of (e.g. waste water), when
UfS:2007
Page 19
UfS:2007
Page 20
GEE
GEE
TSE
TAE
Point
PointMaterialization
MaterializationError
Error
Weighting Error
PME
PME
Increment
Increment DelimiDelimitation
tationError
Error
IDE
IDE
Increment
IncrementExtraction
Extraction
Error
Error
IXE
IXE
SWE
Increment
Incrementand
andSample
Sample
Preparation
PreparationError
Error
IPE
IPE
PSE
Long Range
Point Selection Error
Periodic
Point Selection Error
PSE1
PSE2
Fundamental
Sampling Error
Grouping and
Segregation Error
FSE
GSE
10.2.2 The classification of errors of sampling forms a logical and useful framework for
designing and auditing sampling procedures. Those that are central to the estimation of
uncertainty (e.g. FSE in Figure 3) are discussed below, and others (SWE, PSE and GSE) in
Appendix C.
10.2.3 The total determination error, which Gy calls the global estimation error (GEE), is
the sum of the total sampling error (TSE) and total analytical error (TAE). The components of
TSE can be divided into two major groups: 1) errors of incorrect sampling, 2) errors of correct
sampling. Some incorrect sampling errors arise from what the GUM [2] refers to as gross
errors, and as such would be excluded from estimates of uncertainty. The errors of correct
sampling occur within good practice and can be considered for inclusion within estimates of
uncertainty following the GUM approach [2].
10.2.4 Incorrect sampling errors arise from sampling equipment and procedures that do not
follow the rules of sampling correctness defined in the sampling theory. In Figure 3 these
errors are shown in shaded boxes. Increment delimitation error (IDE) is error that is generated
if the shape of sample is not correct. For example, from a process stream the correct sample is
a complete slice of equal thickness cut through the process stream. The sampling device
should be designed so that it can extract the intended sample profile (i.e. all constituents have
an equal chance to end up in the sample). Otherwise sample or increment extraction error
UfS:2007
Page 21
MS ML
r2 = Cd 3
Equation 5
where
r =
a
aL
UfS:2007
Page 22
C is a sampling constant that depends on the properties of the material sampled; C is the
product of four parameters:
C = f gc
Equation 6
= liberation factor (see Figure 4). For materials where the particles are completely
x
L
liberated, = 1. For unliberated material an empirical equation, = , is used, where
d
values of x ranging from 0.5 to 1.5 have been suggested.
c = constitution factor and can be estimated if the necessary material properties are available
by using:
2
aL
1
a
c + 1 L m
c=
aL
Equation 7
Here aL is the average concentration of the lot, the concentration of the analyte in the
critical particles, c the density of the critical particles and m the density of the matrix or
diluent particles.
10.2.9 If the material properties are not available and they are difficult to estimate, the
sampling constant C can always be estimated experimentally. Certified reference materials,
for example, are a special group of materials for which the sampling constant can be
estimated from existing data.
10.2.10 An example of how the fundamental sampling error model can be used in practice is
given in Example A5, Appendix A.
Figure 4 Estimation of factors for the estimation of fundamental sampling error.
f= 1
f= 0,524
f= 0,5
f= 0,1
L=d
L
L
d
=1
The particle shape factor f (upper part), and liberation factor for unliberated material (lower
left) and liberated material (lower right). L is the liberation size of the critical particles.
UfS:2007
Page 23
The empirical approach includes all sources of uncertainty, without the scientist having to
know their identity in advance. For example, it is automatically applicable to the
particular contaminants and mineralogy present at a geological site. The calculations do
not require any prior knowledge of the nature of the material (e.g. grain size, analyte
speciation, degree of heterogeneity). It is relatively quick and easy to apply practically
(especially for the duplicate method). There are at least four options available to allow
progressively more accurate (and more expensive) estimates of uncertainty, as appropriate
(Table 5). Some of these methods can allow for systematic error (such as sampling bias)
within the estimate of uncertainty. Sampling proficiency tests and reference sampling
targets are still in the early stages of development, but already show considerable promise
for this application.
Among the disadvantages of the empirical approach is that it does not necessarily quantify
any of the individual components of uncertainty (although this knowledge can be added
with limited resolution). It is not based on a theoretical model of particulate sampling, but
this may be an advantage in applications to materials that are not particulate in form (e.g.
gaseous, liquids, biota). The empirical approach only gives an approximate value of
uncertainty, which is assumed to be constant over the target, but this is also true of the
modelling approach. Extreme values in the replicate measurements may lead to an
overestimate of the uncertainty value, which is not representative of most measurements.
This effect can be minimised, however, by the use of robust statistics.
The principal advantage of the modelling approach is that it allows the largest source of
uncertainty to be readily identified, if it was in the model. It gives a transparent method
showing which components of uncertainty have been considered in the summation of
uncertainty. Finally, where prior information is available, modelling approaches can be
less costly than extensive experimental studies.
The disadvantages of the modelling approach include that the theoretical predictions of
uncertainty may require detailed prior measurements of the mineralogy, grain size and
analyte speciation of the material to be sampled (e.g. soil), and how these vary across the
UfS:2007
Page 24
UfS:2007
Page 25
Validation
Quality control
13.1.4 The need for internal quality control of sampling is not widely recognised at present,
and methods for executing it are not well established, except in some specialised areas such as
geochemical prospecting [21]. Specific suggestions for sampling quality control are given for
some environmental sampling matrices in [22]. However, no new principles are involved;
with minor qualification, the principles of internal quality control of analysis are applicable to
sampling [23, 24, 25]. Moreover, the methods used in validation are, with some
simplification, applicable to internal quality control. The reason for the simplification is that
validation needs to provide a good estimate of uncertainty, while quality control merely needs
to demonstrate consistency, over space and time, with the uncertainty established at
validation.
(u
2
s
+ u a2 . Consequently, a one-
sided range control chart can be constructed with a control limit (at the 95% confidence
interval) of 2.83u meas and an action limit (at the 99% confidence interval) of 3.69u meas [25]
(Figure 5). An out-of-control value of d shows that the result should be scrutinised as possibly
unfit for purpose. Such a result is not diagnostic and may stem from a disturbance in either
sampling or analysis; the latter should be detected by standard methods of analytical quality
control.
UfS:2007
Page 26
1.4
1.2
1.0
0.8
0.6
0.4
0.2
0.0
14-Oct
20-Oct
26-Oct
29-Oct
5-Nov
17-Nov
24-Nov
30-Nov
10-Dec
Date of analysis
For the construction of this R-chart see the Nordtest Guide [25].
13.2.3 The data from quality control can also be used to update sampling method precision
as obtained in method validation using the same methods, ANOVA or relative difference
calculations.
13.2.4 In some instances, the extra cost of duplicate sampling can be eliminated by use of
the SAD (Split Absolute Difference) method in which the normal number of increments to be
combined as the sample is segregated at random into two equal sized sub-sets, each of which
is processed and analysed separately [27, 28]. The difference between such results has an
uncertainty of
Page 27
In the modelling approach, the number of degrees of freedom is formally derived from the
degrees of freedom for contributing uncertainties according to a published formula [1, 2],
or approximated from the number of degrees of freedom for the dominant contribution
[1]. More commonly, the number of degrees of freedom is assumed to be sufficiently
large to justify a choice of k=2 for approximately 95% confidence.
For most practical purposes, k=2 is considered acceptable, and is sometimes mandatory [29].
However, it is important to state the value of k used, and the approximate level of confidence
that k implies, when reporting expanded uncertainty.
14.3 Relative uncertainty statements
14.3.1 It is often found that the standard uncertainty from sampling increases approximately
proportionally with the value of the result. Under these circumstances, it is often most
practical to quote the uncertainty in a relative form, such as a relative standard deviation (u/x)
or percentage interval using Equation 4 (e.g. 10%). The relative value quoted is usually
based on an estimate of uncertainty for one or more representative results, but is applicable
over a greater range of concentration values.
14.3.2 It is important not to extrapolate a simple relative standard deviation to zero
concentration, as uncertainty does not normally disappear entirely at very low levels and the
proportionality assumption is no longer valid. More general approaches to these situations can
either specify a range of concentration over which the relative uncertainty value applies [25],
or else express the uncertainty as a function of concentration [1, 42].
14.4 Contributions to uncertainty
14.4.1 The exact steps that are included in each contribution to the measurement uncertainty
need to be stated. Depending on the method of estimation employed, the details of the
experimental design, and the person to whom the information is intended, it is possible to
quantify some specific components of the measurement uncertainty. For example, the
experimental design in Figure 2 will give separate estimates of two components called
sampling and analysis. When the details of this particular implementation of the design are
examined it becomes evident that uncertainty from physical sample preparation is included
within the general title of sampling, whereas that from chemical preparation is included
within analysis. If required, it is possible to insert a further level of duplication of physical
preparation within the experimental design to estimate the separate contribution which that
particular step introduces [30]. The exact steps that are included in each contribution to the
measurement uncertainty need to be documented. For less experienced users of analytical
UfS:2007
Page 28
For proof of compliance, the result and its uncertainty interval must be entirely within the
permitted range.
For proof of non-compliance, the result and its uncertainty interval must be entirely
outside the permitted range.
For shared risk approaches, set a range for acceptable measurement results based on the
permitted interval, adjusted to provide a specified probability of false acceptance and false
rejection rates. Recent guidance gives useful details of the procedure [32].
UfS:2007
Page 29
This concept is equally applicable to situations where materials have regulated minimum analyte
concentrations, in which case the terms false compliance and false non-compliance are applicable.
UfS:2007
Page 30
Line A shows the costs of measurement. Line B shows costs of incorrect decisions. The sum
of these two lines (the total cost shown by the highest line) shows a minimum cost at point C,
which is the uncertainty that can be regarded as fit for purpose.
16.5
The optimal apportionment of resources between sampling and analysis is also a
matter of costs. Even an elementary consideration (excluding costs) shows that the
uncertainties of sampling and analysis should be roughly balanced. For example, if the
uncertainties of sampling and analysis are 10 and 3 units respectively, the overall uncertainty
of measurement is 10 2 + 32 = 10.4 . The overall uncertainty is hardly affected by a reduction
of the uncertainty of analysis: if it is reduced to (say) 1 unit, the overall uncertainty is reduced
to 10 2 + 12 = 10.05 , an inconsequential change. A more sophisticated approach takes into
account the different costs of analysis and sampling. If the unit costs of sampling and analysis
are A and B for the same specific level of uncertainty, the optimum ratio of sampling
uncertainty u samp to analytical uncertainty u anal is given by
usamp
uanal
14
A
= .
B
This ratio provides the minimum expenditure for a given overall uncertainty of
2
2
usamp
+ uanal
Page 31
where the observed or expected frequency distributions are not normal, for example where
the results contain more than 10% outliers, or where the results show markedly
asymmetric distributions;
where confidence intervals are needed on the estimates of uncertainty or, for more
complex sampling plans, on the measurement results;
where the sampling strategy is more complex than simple random sampling with
replicated measurements, for example in implementing stratified sampling.
UfS:2007
Page 32
UfS:2007
Page 33
Appendix A: Examples
Appendix A: Examples
Introduction
The most effective way to explain the methodologies described in the main text of this Guide
is to show worked examples. These examples are not intended to cover all circumstances, but
to show how the general principles can be applied to a variety of situations across a range of
different sectors. These include food (production and retail), animal feed, and environment
(soil and water). The examples are all structured using the same basic format, so as to aid
comprehension and comparability.
UfS:2007
Page 34
Example A1
Unit1
Sector/
Matrix
Nitrate/Hot
mg kg-1
Food/
water
as received Lettuce
extraction and
determination
by HPLC
Uncertainty estimation
Sampling
target(s)
Purpose
1 bay of
Iceberg
lettuce
grown under
glass
Uncertainty
total
measurement,
sampling and
analytical
Design
Statistics
Empirical Robust
- duplicate ANOVA
method
1 Scope
Estimate the measurement uncertainty, and contributions from sampling and analysis, for
routine monitoring of glasshouse grown lettuce, using a standard sampling protocol.
3 Sampling protocol
The accepted protocol for this purpose specifies that one composite sample is prepared from
10 heads of lettuce harvested from each bay of lettuce [36]. The lettuces are selected by
walking a W shape or five-point die shape through the bay under investigation. This protocol
is applied to all bays regardless of the size. Samples were taken in the morning and
transported to the contracted analytical laboratory in ice-packed cool boxes to arrive within 24
hours of sampling.
UfS:2007
Page 35
Example A1
Figure A1.1: Example of the 'duplicate method'
Example of how the duplicate method can be applied. Using the W design as an example, the protocol
stipulates the design but not the position or orientation. The W is equally likely to start on the left or
the right. Ten heads are taken along the line of the W to create a composite sample for one target.
6 Results
The best estimates of the nitrate concentration at each of the eight target locations are shown
in Table A1.1.
Table A1.1: Measurements of the concentration (mg kg-1) of nitrate in eight duplicated samples. The duplicate
samples are labelled S1 and S2. Likewise, duplicate analyses are labelled A1 and A2. Hence, DS1A2 (value
4754 mg kg-1) is analysis 2, from sample 1 from sampling target D
Sample
S1A1
S1A2
S2A1
S2A2
3898
4139
4466
4693
3910
3993
4201
4126
5708
5903
4061
3782
5028
4754
5450
5416
4640
4401
4248
4191
5182
5023
4662
4839
3028
3224
3023
2901
3966
4283
4131
3788
target
UfS:2007
Page 36
Example A1
Before applying statistical methods it is useful to inspect the data to ascertain the general
levels of variability. The analytical duplicates (e.g. BS1A1 and BS1A2) are generally within
300 mg kg-1 of each other, suggesting an analytical precision of less than 10%. The sample
duplicates (e.g. DS1 and DS2) agree less well, but generally differ by less than 20%.
However, one target (C) displays a greater difference, suggesting an outlying value.
Quantification of the random component of the measurement uncertainty and two of its main
components (sampling and analysis) was made using robust analysis of variance (RANOVA,
Appendix C3, with output in Figure.A1.2. Robust ANOVA was used here as outlying targets
are, in this relatively well-controlled environment, considered likely to be anomalies, rather
than reflecting the underlying population statistics, and as a precaution against analytical
outliers.
Note: Robust methods should not be used where apparent outliers arise as part of the typical
population of sampled increments or targets, unless the specific implementation allows for
non-normal distributions for part of the assumed error structure.
Figure A1.2
CLASSICAL ANOVA RESULTS
Mean = 4345.5625
Standard Deviation (Total) = 774.5296
Sums of Squares = 12577113 4471511
Between-target
351320
Sampling
Analysis
Standard Deviation
556.2804
518.16089
148.18063
Percentage Variance
51.583582
44.756204
3.6602174
Sampling
Analysis
Measurement
Standard Deviation
565.39868
319.04834
167.94308
360.5506
Percentage Variance
71.090791
22.636889
6.2723172
28.909209
Relative Uncertainty
14.474814
7.6193626
16.357719
(% at 95% confidence)
The output of ANOVA for data produced form a balanced experimental design (n = 8, Table A1.1). Both robust
and classical estimates are given for comparison. Standard deviation estimates are computed for betweentarget (sbetween-target), within-target (ssamp) and within-chemical analysis (sanal). Results are in the same units of
concentration as the input data (i.e. mg kg-1 in this case).
Page 37
Example A1
This can be used as an estimate of the random component of the standard uncertainty (u).
The expanded relative uncertainty is given by Equation 3 as:
7 Comments
This uncertainty estimate does not include any estimate of the possible sampling bias.
UfS:2007
Page 38
Example A1
10 Summary
Measurement uncertainty
Sampling
Analytical
Total
14.5%
7.6%
16.4%
Table A1.2
Sample
S1A1
Expanded
Uncertainty
3898
639.3
3910
641.2
5708
936.1
5028
824.6
4640
761.0
5182
849.8
3028
496.6
3966
650.4
target
The nitrate concentrations associated with S1A1 (routine sample) are shown with the associated measurement
uncertainty (calculated from U = 16.4%). As an example, Target F has a value of the measurand (or true value)
between 4332 mg kg-1 and 6032 mg kg-1.
UfS:2007
Page 39
Example A2
Unit
Sector/Matrix
mg kg-1 Environmental/
dry
Top soil
basis
Uncertainty estimation
Sampling
target(s)
Purpose
100 targets
each of area
30 m x 30 m,
with depth of
0150 mm
Uncertainty
total
measurement,
sampling and
analytical
Design
Statistics
Empirical Robust
- duplicate ANOVA
method
1 Scope
Estimate the measurement uncertainty, and contributions from sampling and analysis, at each
of 100 different sampling targets within one site, using a common sampling protocol.
3 Sampling protocol
One hundred samples of top soil (nominal depth 0150 mm) were taken with a hand auger
(diameter 25 mm) at 100 locations. These locations were distributed on a regular grid with
sample spacing of 30 m (Table A2.1), and therefore each is intended to represent an area 30 m
by 30 m. The surveying was conducted with a measuring tape and compass.
Page 40
Example A2
taken for dissolution with nitric and perchloric acids, prior to determination of lead by ICPAES [39]. The measurements were subject to full analytical quality control (AQC), and
corrected for reagent blank concentrations where these values were statistically different from
zero. The raw measurement values use for the estimation of uncertainty had no rounding or
suppression of values less than either zero, or the detection limit.
6 Results
The best estimates of the lead concentration at each of the 100 target locations are shown in
the format of a map (Table A2.1).
Table A2.1: Measured lead concentrations at each target on the sampling grid (mg kg-1), shown by the actual
coordinates used in the regular sampling grid (spacing 30 m). They show a high degree of variability betweenlocations of roughly a factor of 10. The variability within 10 of these locations selected at random (i.e. A4, B7,
C1, D9, E8, F7, G7, H5, I9 and J5) was used for the estimation of uncertainty from sampling (Table A2.2). This
within-target variation is substantial (e.g. a factor of 2) but substantially less than the between-target variability.
Row
474
287
250
338
212
458
713
125
77
168
378
3590
260
152
197
711
165
69
206
126
327
197
240
159
327
264
105
137
131
102
787
207
197
87
254
1840
78
102
71
107
395
165
188
344
314
302
284
89
87
83
453
371
155
462
258
245
237
173
152
83
72
470
194
82.5
162
441
199
326
290
164
71
101
108
521
218
327
540
132
258
246
72
188
104
463
482
228
135
285
181
146
10
89
366
495
779
60
206
56
135
137
149
Four measurements from the balanced design for each of the 10 sample targets selected for
duplication were used for the estimation of uncertainty (Table A2.2). Visual inspection of the
data allows an initial qualitative assessment of the relative importance of the two sources of
measurement uncertainty. The low level of agreement between the concentration values from
some of the sample duplicates is indicative of a high level of sampling uncertainty (e.g. S1
compared to S2 for target D9). The agreement between the analytical duplicates (A1 and
A2) is however generally much better for most samples (< 10% difference) than that between
the sample duplicates.
UfS:2007
Page 41
Example A2
Table A2.2: Measurements of the concentration (mg kg-1) of a lead on 10 duplicated samples from the total of
100 targets in a survey of contaminated land (Table A2.1). The duplicate samples are labelled S1 and S2.
Likewise, duplicate analyses are labelled A1 and A2. Hence, D9S1A2 (value 702 mg kg-1) is analysis 2, from
sample 1 from sampling target D9. Values shown are rounded for clarity, and used for subsequent calculations,
but generally un-rounded values are preferable for these calculations.
Sample
S1A1
S1A2
S2A1
S2A2
A4
787
769
811
780
B7
338
327
651
563
C1
289
297
211
204
D9
662
702
238
246
E8
229
215
208
218
F7
346
374
525
520
G7
324
321
77
73
H5
56
61
116
120
I9
189
189
176
168
J5
61
61
91
119
target
Quantification of the random component of the measurement uncertainty and two of its main
components (sampling and analysis) was made using robust analysis of variance (RANOVA),
with typical output shown in Figure A2.1. Robust statistics were selected to allow for the
outlying values that are evident in this data (e.g. target A4, sample duplicate D9S1/S2,
analytical duplicate B7S2A1/A2), and in most similar data sets (but see the Note in Example
A1, section 6). The estimates of uncertainty are averaged over the 10 targets, assuming that
the uncertainty is not varying significantly over this range of concentration. The uncertainty is
expressed in relative terms so that it is applicable over this range of concentration (Section
14.3).
Extracting the robust estimates from this output gives:
UfS:2007
Page 42
Example A2
Figure A2.1: The output of ANOVA for data produced from a balanced experimental
design (n = 10, Table A2.2)
CLASSICAL ANOVA RESULTS
Mean = 317.79999
Standard Deviation (Total) = 240.19238
Sums of Squares = 1738031.9 370075.5 6473
Between-target
Standard Deviation
Sampling
Analysis
197.55196
135.43246
17.990274
31.792678
0.5609926
Sampling
Analysis
Measurement
Standard Deviation
179.67409
123.81386
11.144044
124.31436
Percentage Variance
67.62655
32.113293
0.26015487
32.373447
Relative Uncertainty
83.289726
7.4966113
83.626415
(% at 95% confidence)
Both robust and classical estimates are given for comparison. Standard deviation estimates are computed for
between-target (sbetween-target), within-target (ssamp) and within-chemical analysis (sanal). Results are in the same
units of concentration as the input data (i.e. mg kg-1 in this case).
For comparison the expanded uncertainty for the analytical contribution (random component)
is given by
Page 43
Example A2
random component within one organisation. In this case, the standard relative analytical
uncertainty is increased to 5.24% [ = (3.752 + 3.412 + 1.342)0.5]. The expanded analytical
uncertainty (10.48%) is then greater than the analytical target value of 10%, but it can also
usefully be compared with an independent estimate of the analytical measurement uncertainty
made within the laboratory. The expanded uncertainty for the whole measurement is thereby
increased to 83.95% [ = (83.292 + 10.482)0.5], which is practically identical to the purely
random component of 83.63%.
Table A2.3: Measured and certified lead concentration values for CRMs for the
estimation of the bias of the analytical method
CRM name
(n=4)
Mean
(mg kg-1)
Standard
Deviation
(mg kg-1)
Certified
value
(mg kg-1)
U on
certified
value
(95% conf.)
NIST2709
19.7
3.2
18.9
0.5
NIST2710
5352.0
138.0
5532.0
80.0
NIST2711
1121.4
14.7
1162.0
31.0
BCR141
34.4
3.9
29.4
2.6
BCR142
36.2
4.6
37.8
1.9
BCR143
1297.5
33.0
1333.0
39.0
7 Comments
This estimate of uncertainty does not make allowance for any undetected sampling bias
(Section 9.4.2). However, because the uncertainty is often dominated by the heterogeneity of
the sampling target, the extra uncertainty introduced by bias in sampling can often be
assumed to be insignificant by comparison (as shown for the analytical bias). Where the
highest quality of uncertainty estimate is required, due perhaps to potentially large financial
consequences from underestimating the uncertainty, it may be preferable to use one of the
more elaborate methods using multiple samplers and/or protocols (Table 5).
If the measurand (or true value) had been defined as the mean concentration of lead across the
whole site, the uncertainty would have had to include the contribution from the standard error
on the calculated mean value, expressed as stotal / n . For this example stotal is 403 mg kg-1, n
= 100 and the uncertainty on the mean (291.9 mg kg-1) is therefore 27.6% of the value, at 95%
confidence. This value can be calculated without knowing the individual contribution of the
uncertainty from either the sampling or the analysis, and is often dominated by sbetween-sample.
Page 44
Example A2
10 Summary
Measurement uncertainty*
Sampling
Analytical
Total
83.3%
10.5%
83.9%
UfS:2007
Page 45
Example A3
Unit
Sector/Matrix
mg l-1 Environment/
groundwater
Uncertainty estimation
Sampling target
Purpose
Design
Empirical
duplicates
used in
validation
and quality
control
Statistics
Range
1 Scope
The scope is determination of the total uncertainty of the measurement of dissolved iron in a
sampling validation study and subsequent control of sampling uncertainty during monitoring.
UfS:2007
Page 46
Example A3
Dissolved iron
mV
mg l-1
Mean
-123
1.11
27%
56%
Filtering
The chemical data suggest that the groundwater composition is quite uniform over time and
space with respect to the main components (data not shown, relative standard deviation 1.9
16%), whereas the variability is high for the redox parameters (oxygen, redox potential and
dissolved iron). The expected main causes of uncertainty are indicated in the table for the two
key parameters and the causes were controlled during sampling.
3 Sampling protocol
Sampling was done according to the rhus County groundwater monitoring protocol, with
permanent, dedicated pumps (Grundfos MP1) set in the middle of the screened interval of
each well. Pump rates were 12 m3 h-1 (well purging) with a 10% reduction just before
sampling. Two of the six wells were large-diameter abstraction wells equipped with high
yield pumps. These were pumped with 4060 m3 h-1 for well purging followed by pump rate
reduction just before sampling. During well purging, the development in water quality was
followed with on-line measurements of oxygen, pH, electrical conductivity and redox
potential until stable readings and then samples were taken. A field report was filled in during
the sampling, including pump yields and pumping times as well as water table measurements.
UfS:2007
Page 47
Example A3
4.1 Validation
The objective of the validation programme was to ensure that a measurement uncertainty
meeting the set quality objective could be obtained and to describe the components of
uncertainty in order to identify points of improvement, if required. The validation programme
was set up with sampling of six wells, two independent samplings per well and two subsamples per sample analysed, see Figure A3.1.
Groundwater body
Well 1
Well 2
Well 3
Sample 1
Analysis 1
Well 4
Well 5
Well 6
Sample 2
Analysis 2
A total of 12 samples were taken and 24 sub-samples were sent for analysis in one sampling
round as a validation study.
4.2 Quality control
The objective of the quality control programme for the operational monitoring was to ensure
that measurement uncertainty did not increase over time during the monitoring. The quality
control programme was set up after careful evaluation of the results from the validation study.
Quality control was designed to include duplicate sampling, each with duplicate analysis, on
one of the two annual sampling occasions of the monitoring programme, see Figure A3.2. In
total, six sampling occasions with 12 samples and 24 sub-samples analysed were included in
the first phase of the quality control programme.
UfS:2007
Page 48
Example A3
Figure A3.2 Design outline for quality control, shown for one sampling occasion
Groundwater body
Monitoring well
Sample 1
Analysis 1
Sample 2
Analysis 2
Pre-treatment
Redox potential
Dissolved iron
On-line analysed
The sample stream was pumped through an on-line measuring array of a flow-through cell
with sensors set up in series. The sensor used for redox potential is described in Table A3.3.
UfS:2007
Page 49
Example A3
Table A3.3: On-line sensor used for redox potential measurements
Parameter
Instrument
Cell
Instrument
accuracy
Calibration and
control
Redox potential
pH 340
Sensolyt Pt
2 mV
Daily service
Table A3.4: Methods and performance data from quality control for laboratory analyses
Method
Iron
ICP-AES
0.95%
4.2%
Total
reproducibi
-lity
Total
expanded
uncertainty
4.3%
8.6%
Detection
limit
0.01 mg l-1
The certified reference material (CRM) VKI Metal LL2, nominal 0.200 mg Fe l-1 was used
for quality control with 101.9% recovery (mean for 92 control results).
5.4 Calculation methods
The replicate data were treated using the range method (ISO 3085). For comparison,
uncertainty estimates were calculated by analysis of variances (ANOVA) and robust ANOVA
(RANOVA) using ROBAN version 1.0.1 (Appendix C3).
The applied calculation methods are demonstrated in Section 7 (below). The range
calculations are easily done using standard spreadsheets, and an example can be downloaded
from http://team.sp.se/analyskvalitet/sampling/default.aspx.
The occurrence of systematic sampling errors was not assessed quantitatively, but the
consistency of the obtained results was used as a qualitative control of systematic errors. As
an example, if dissolved iron was found above 0.1 mg l-1 in the same sample as oxygen was
determined to be above 0.1 mg l-1, this would indicate a systematic sampling and/or pretreatment error. Similarly, redox potential and oxygen contents were checked to correspond in
order to control systematic errors.
6 Results
The data set from the validation study is shown in Table A3.8 for dissolved iron with the
range calculations. The calculations for redox potential in the validation study and for both
dissolved iron and redox potential during quality control were done similarly.
The data from the validation study (six different wells) using range calculations are shown in
Table A3.5.
UfS:2007
Page 50
Example A3
Table A3.5: Relative expanded uncertainty (%, coverage factor 2) for analysis, sampling
and between-target (between wells) as obtained during validation using range
calculations
Range calculations
Analyses
Sampling
Between-target
Redox potential
5.2%
15%
14%
Dissolved iron
2.1%
10%
70%
For comparison, the statistical estimates are shown in Table A3.6 as obtained using ANOVA
and RANOVA.
Table A3.6: Relative expanded uncertainty (%, coverage factor 2) for analysis, sampling
and between-target (between wells) as obtained for dissolved iron during validation
using ANOVA and RANOVA calculations
Dissolved iron
Analyses
Sampling
Between-target
ANOVA
1.6%
9.6%
70%
RANOVA
1.8%
9.9%
72%
The statistical estimates obtained with the range statistics during quality control (six sampling
occasions) are shown in Table A3.7.
Table A3.7: Relative expanded uncertainty (%, coverage factor 2) for analysis, sampling
and between-target (between occasions) as obtained during quality control using range
calculations
Analyses
Sampling
Between-target
Redox potential
18%
3.8%
23%
Dissolved iron
2.5%
3.6%
9.9%
UfS:2007
Page 51
Example A3
Table A3.8: Results and range calculations for the validation study, dissolved iron, basic data in bold, symbols used to describe
calculations only (T: target, S: sample, A: analysis, R: absolute differences, r: relative differences, n: numbers)
Well
number
S1A1
S1A2
S2A1
R1 = S1A1 S1A2
S2A2
mg l-1
0.815
1.80
1.69
2.62
1.66
1.52
99.474
99.468
99.469
99.916
99.327
99.371
mg l-1
0.834
1.83
1.68
2.61
1.63
1.53
mg l-1
0.912
1.94
1.79
2.83
1.58
1.47
mg l-1
0.019
0.030
0.010
0.010
0.030
0.010
mg l-1
0.893
1.93
1.77
2.84
1.59
1.50
S1 =
mg l-1
0.825
1.82
1.69
2.62
1.65
1.53
%
2.30
1.65
0.593
0.382
1.82
0.656
R2 = S 2 A1 S 2 A2
S2 =
mg l-1
0.019
0.010
0.020
0.010
0.010
0.030
S 2 A1 + S 2 A2
2
mg l-1
0.903
1.94
1.78
2.84
1.59
1.49
r = 7.413
1
Analysis
rA =
r + r
1
n r1 + n r 2
Sampling
rS + A
r
=
Between-target
ST + S + A
nr
S
=
nr
7.413 + 6.750
= 1.18
6+6
rS + A =
35.36
= 5.89
6
ST +S + A =
10.32
= 1.72
6
CV A =
CV A =
rS + A
1.128
CVS + A =
sT + S + A = s S k
*100 S = S1 + S 2
2
%
2.11
0.517
1.12
0.353
0.631
2.02
r=
S1 S 2
S
mg l-1
0.864
1.88
1.73
2.73
1.62
1.51
%
9.03
6.40
5.48
8.07
3.72
2.66
nr = 6
nr = 6
*100
S = 10.32 r = 35.36
= 6.750
nr 2 = 6
rA i
1.128
CVS + A =
S2
nr1 = 6
rA =
R2
r2 =
1.18
= 1.05
1.128
5.89
= 5.22
1.128
sT + S + A = 0.604
CV S = CV S + A
CVT + S + A =
CV A 2
sT + S + A
*100
ST +S + A
CVT = CVT + S + A
CV S + A 2
The standard deviation can be obtained from the mean of relative differences between duplicate measurements by division with the statistical factor 1.128.
CVS = 5.22 2
CVT + S + A =
1.05 2
= 5.17
2
0.604
*100 = 35.1
1.72
CVS = 35.12
5.17 2
= 34.9
2
CV A 2
with the factor on CV A 2 due to the mean of duplicate analyses being used.
2
s: standard deviation with n-1 degrees of freedom as obtained from most standard calculators and spreadsheets.
UfS:2007
Page 52
Example A3
7 Comments
Overall, the validation data show that the variability in the aquifer (between-target)
was dominating the total uncertainty for dissolved iron, whereas sampling and
between-target uncertainties were of the same size for dissolved iron. Analytical
uncertainties were small (25%), and for dissolved iron were comparable to the
repeatability obtained in laboratory quality control (expanded uncertainty 2.1% as
compared to 1.9% respectively). If different wells were sampled, the sampling
uncertainty was 1015%.
For dissolved iron measured during validation, the use of ANOVA and RANOVA
calculations did not provide statistical estimates more than slightly different from
those obtained with the simple range calculations.
In the quality control scheme of monitoring, the variability between sampling
occasions (between-target, 9.9%) was dominating the total uncertainty for parameters
analysed as laboratory analysis (dissolved iron, 2.5% uncertainty), whereas the
analytical uncertainty (18%) was almost as important as the between-target
uncertainty (23%) for on-line measurements (redox potential). The reason for the
large contribution from on-line measurements is that during quality control, duplicate
on-line measurements were done with two different instruments in contrast to the
validation study done with one single instrument for both duplicate measurements.
Accordingly, the analytical uncertainty (instrument to instrument variation) for redox
potential was considerably larger in the quality control (18%) than in the validation
study (5.2%). For dissolved iron, the analytical uncertainty was comparable in
validation and in the subsequent quality control (2.1% and 2.5% respectively). The
sampling uncertainty was lower when sampling just one well on different occasions
during quality control (3.63.8%) than when sampling different wells at the same time
during validation (1015%). The uncertainty between-target (variation from one
sampling occasion to the next) during quality control was small for dissolved iron
(9.9%), but larger for redox potential (23%).
If a continuous control of sampling uncertainty had been required, the control data
could have been plotted in control charts in order to obtain an early warning of
excessive uncertainty (random errors) for each sampling occasion.
The number of replicates (six) in this study was less than used in most cases and the
risk of a decreased confidence in the uncertainty estimates should be considered in
evaluation of the results.
The uncertainty contribution from sampling bias was only addressed through
evaluation of the consistency of the measurements obtained from different,
interrelated chemical parameters (oxygen, dissolved iron, redox) and the evaluation
supported the conclusion that sampling and sample pre-treatment had succeeded in
avoiding bias due to oxygen impact and filter clogging.
8 Summary
The measurement uncertainty (% uncertainty with coverage factor 2) is summarised
below for dissolved iron.
UfS:2007
Page 53
Example A3
The data show that the requirement for less than 20% expanded measurement
uncertainty could be fulfilled for dissolved iron (sampling validation), and that the
required measurement uncertainty was in reality achieved during the routine
monitoring (sampling quality control). Furthermore, the data show that if an
improvement of the certainty of monitoring was required, the obvious point of
improvement would be increased monitoring density for dissolved iron (betweentarget uncertainty dominating), whereas improvement of the on-line measurement
uncertainty could help for redox potential (large contribution of analysis uncertainty).
Dissolved iron
in groundwater
Between-target
variability
Sampling
Analysis
Measurement
Validation
10%
2.1%
10%
35%l
Quality control
3.6%
2.5%
4.4%
9.9%m
9 Acknowledgements
The work presented here has been supported by Nordic Innovation Centre, the Soil
and Ground Water Contamination Committee of the Danish Council of Technical
Sciences and rhus County, Denmark. Fieldwork has skilfully been done by Mogens
Wium, GEO.
UfS:2007
Page 54
Example A4
Unit
Sector/
Matrix
Food/
Baby
porridgepowder
containing
fruit
Uncertainty estimation
Sampling
target
Purpose
Design
Statistics
Empirical One-way
Produced Total
batch
measurement duplicate ANOVA
method
uncertainty
1 Scope
The scope is to estimate the measurement uncertainty and contributions from
sampling and analyses. The estimates are based on samples from one type of baby
porridge taken from 10 different batches using a sampling protocol collecting
duplicate samples from each batch.
Page 55
Example A4
3 Sampling protocol
Normally a spot sampling approach is employed in which one sample (one package)
of a batch is used as a screening sample by comparing its content against the declared
values and legal limits.
Validation In this study two samples are collected from each of 10 different batches
of one type of baby porridge powder (i.e. 10 sampling targets). Each sample is equal
to one package of approximately 400 g powder.
Quality control Quality control (QC) of sampling from different types of baby
porridge is done by collecting two samples from each of eight batches of different
types of baby porridges (i.e. eight sampling targets). All the types of porridges contain
fruit in addition to milled cereals.
To ensure the quality in each package of the product at the time of the best before
date of the porridge powder, the producer wraps the product in an air-tight and lightprotecting bag. It is therefore assumed the degradation of the vitamin A is negligible
during normal shelf life. The sampling for the validation was performed at the place
of production. For QC, the samples were purchased partly at the place of production,
and partly at the retailers. When the samples were collected from retailers, care was
taken to collect the two samples (of each product) at different retailers but in addition
to assure the samples had the same batch marking. This is important to avoid adding
between-batch variations to the apparent sampling uncertainty, as the sampling
protocol in this case specifies sampling from a particular batch.
Samples are collected on-line (just after the filling operation of packages) at random
times. Two samples (two packages, each of approximately 400 g) are collected from
each of 10 production units (batches) of one type of baby porridge powder.
UfS:2007
Page 56
Example A4
Figure A4.1: Sampling for validation. Two samples are taken from each of 10
production units/batches of the same type of baby porridge
Batch 1
S2
.. Etc.
Batch 10
S1
S2
S1
Sample B10S2
Sample B1S2
Sample B10S1
Sample B1S1
Test sample 1
Test sample 1
Test sample 1
Test sample 2
Test sample 1
Test sample 2
Test sample 2
Test sample 2
For quality control (QC) two samples are collected from one batch of each of eight
different types of baby porridges, containing fruit and milled cereals. The porridges
are products from three different producers. The samples (except for two types of
porridges) were provided by two of the producers. The rest was bought at the retailer.
Figure A4. 2: Sampling for QC. Two samples are taken from one batch of each
of eight different types of baby porridge
Product 1
S2
S1
Product 8
.. Etc.
S2
S1
Sample P1S1
(1 package)
Sample P1S2
(1 package)
Test sample 1
Test sample 1
Test sample 2
Test sample 2
UfS:2007
Page 57
Example A4
The laboratory participates in Laboratory Proficiency Tests (FAPAS and Bipea) with
good results (in the period 20002005, |Z-score|<1). The method was validated using
a certified reference material (CRM). Data concerning the laboratory performance is
given in Table A4.1.
Table A4.1: Methods and performance data from quality control laboratory
analyses
Parameters
Method
Repeatability
2RSD (%) = 6
Within-reproducibility
2RSD(%) = 8
Measurement uncertainty
Recovery
0.14 mg kg-1
CRM used
A mechanical sample divider (Retsch) is used to split the samples. From each of the
primary samples, four test samples are collected: two portions of approximately 35 g
and two portions of approximately 4050 g.
UfS:2007
Page 58
Example A4
Figure A4.3: Splitting of the primary sample to make four test samples
Batch 1
Batch 10
S2
.. Etc.
S1
Sample B1S2
(1 package)
S2
S1
Sample B1S1
5.2 Analyses
UfS:2007
Page 59
Example A4
Table A4.2: Product data provided by the producer
Product
Weight of batch, including premix
1092 kg
1.228 kg
Vitamin A declared as
7 Results
Test sample 40 g baby porridge
Table A4.3: Validation data from the same product, results given in g 100 g-1
powder
Batch
S1A1
S1A2
S2A1
S2A2
B1
402
325
361
351
B2
382
319
349
362
B3
332
291
397
348
B4
280
278
358
321
B5
370
409
378
460
B6
344
318
381
392
B7
297
333
341
315
B8
336
320
292
306
B9
372
353
332
337
B10
407
361
322
382
S1 and S2: Primary samples from sampling location 1 and 2 of one production batch
A1 and A2: Analyses of duplicate test samples of a primary sample S
Analysed mean value (test sample 40 g): 348 g 100 g-1
UfS:2007
Page 60
Example A4
Test sample 4 g baby porridge
Table A4.4: Validation data same product, results given in g 100 g-1 powder
Batch
S1B1
S1B2
S2B1
S2B2
B1
400
491
323
355
B2
413
159
392
434
B3
315
391
252
454
B4
223
220
357
469
B5
462
343
262
293
B6
353
265
305
456
B7
298
234
152
323
B8
425
263
417
353
B9
622
189
291
272
B10
292
397
142
568
S1 and S2: Primary samples from sampling location 1 and 2 of one production batch
B1 and B2: Analyses of duplicate test samples of a primary sample S
Analysed mean value (test sample 4 g): 341 g 100 g-1
7.1 Calculations
The ANOVA calculation can be done by using available tools in Excel, Minitab,
SPSS etc. In this study the calculations are done in an excel spreadsheet and the
details are shown in Section 10 ANOVA calculations.
Calculation of uncertainty of analyses, one-way ANOVA, test sample 40 g
Table A4.5: Results from ANOVA calculations uncertainty of analyses sum
of squares of differences, within-groups (SS-Error). For details see Section A4.11
SSE-Anal
(g 100 g-1)2
Degrees of
freedom (df)
Variance
= SSE-Anal /df
(g 100 g-1)2
16595
20
829.75
Standard deviation,
SDanal
= (SSEAnal) /df
(g 100 g-1)
28.805
X a)*100%
8.28
(g 100 g )
14231
UfS:2007
Degrees of
freedom (df)
10
Variance
VSamp=
Standard deviation,
SDsamp
(SSS/dfS SSEAnal/dfA)/2
(g 100 g-1)2
296.7
17.22
Relative standard
deviation RSDsamp(%)
= (SD /
X s)*100%
4.95
Page 61
Example A4
Calculation of measurement uncertainty 40 g test sample
The RSD (%) value from the ANOVA calculation can be used as an estimate of the
standard uncertainty u (%). The analytical laboratory has estimated the analytical
standard uncertainty to be 7%, which is lower than the random analytical component
for this sample type of 8.28%. The higher value of these two is used in the
calculations. Combining the RSD values from tables A4.5 and A4.6 with Equation 1,
the results can be written as in Table A4.7.
umeas =
(u samp ) 2 + (u anal ) 2
(Equation A1)
Sampling
Analytical
Total
4.95
8.28
9.7
9.9
17
20
The same calculations are used as for the test sample size of 40 g (see Table A4.14, in
Section 11 of this example).
Table A4.8: Results from ANOVA calculations uncertainty of analyses, 4 g test
sample sum of squares of differences, within groups (SS-Error)
SSE
(g 100 g-1)2
312206.5
Degrees of
freedom (df)
Variance
(N*2-N)=20
(g 100 g-1)2
20
= SSE /df
15610.325
Standard deviation,
SDanal
=
SSE /df
(g 100 g-1)
124.9413
X a)*100%
36.6800
(g 100 g )
Degrees of
freedom (df)
VSamp=
(SSS/dfS
SSEAnal/dfA)/2
(g 100 g-1)2
102860.25
UfS:2007
10
-2662.15
Standard deviation,
SDsamp
= V Samp
(g 100 g-1)
- 2662.15
Set to zero
X s)*100%
Page 62
Example A4
The same calculations are used as for the test sample size of 40 g (Table A4.15,
Section 11 of this example).
The negative value of VSamp in Table A4.9 indicates that SDSamp is small compared to
the calculated value of SDanal. In this case, the estimates of SDanal and SDsamp using
robust ANOVA confirmed the smaller sampling standard deviation; the robust
ANOVA estimates were uSamp(%) = 6.9% and uAnal(%) = 30%.
As the sampling is identical for the experiments with 40 g and 4 g test samples (and
the uncertainty of sampling therefore should be the same), an RSDsamp (%) = 5%
( 4.95 see table A4.7) is used as an estimate.
Calculation of measurement uncertainty 4 g test sample
Using the calculated RSD (%) value in Tables A4.5 and A4.6 as an estimate of the
measurement uncertainty and combining with Equation A1 the results can be written
as follows (Table A4.10).
Table A4.10: Measurement uncertainty 4 g test sample
Analytical
Measurement
36.7
37
10
73.4
74
The baby porridge powder looks homogeneous, and therefore a low measurement
uncertainty (u) is expected. However, analyses of the powder indicated a surprisingly
large u when using a test sample size of 4 g (the CEN-standard EN-12823-1 indicates
a test sample size of approximately 210 g). The producers recommended using a test
sample size of 4050 g.
The validation tests gave the following results
Table A4.11: Comparing measurement uncertainty test samples of 40 g and 4 g
Measurement
uncertainty (umeas)
Expanded measurements
uncertainty Umeas
40 g test sample
9.7%
20%
4 g test sample
37%
74%
It can be concluded that u40g << u4g. A Umeas of approximately 20% is acceptable,
using the manufacturers criterion of 2030%, while a Umeas of 74% is considered to
be too high, taking into account the matrix and production conditions of this type of
product.
UfS:2007
Page 63
Example A4
It can therefore be concluded that a test sample weight of 4 g is not fit for purpose
when analysing vitamin A (retinol) in baby porridge powder containing milled cereals
and fruit. A test sample size of 4050 g is recommended. This also supports the theory
that the vitamin is unevenly distributed in the product, possible as local hot spots
due to electrostatic interactions.
7.3 Quality control
According to Section 13.2.2 of this Guide, the principal tool in quality control is
replication. This is minimally executed by taking two samples from each target by a
complete (and suitably randomised) duplication of the sampling protocol. There is
only a need to analyse the sample once and the difference between the results
D = x1 x2 is calculated. In this study each sample was analysed twice, but the
comparisons were made between one analyses of each sample (double set).
In the quality control study, test portions of 40 g were used. According to
declarations, the products contain different amounts of vitamin A.
Table A4.12: Quality control data for test portion 40 g of different products
Product Producer
Porridge powder ingredients
S1A1 S1A2 S2A1 S2A2
P1
1
Oat, rice and pear
322 319 350 375
P2
1
Oat, rye, rice and pear
332 317 358 393
P3
1
Wheat, banana and apple
443 430 461 388
P4
1
Wheat and apple
318 383 390 334
P5
2
Oat, rice and banana
252 219 265 227
P6
2
Wheat and apple
274 239 233 217
P7
2
Oat, rice and apple
206 225 198 195
P8
3
Wheat, spelt, oat and apple (organic product) 392 335 375 416
S1 and S2: Primary samples (laboratory samples) from sampling locations 1 and 2 of
one batch from each product.
A1 and A2: Analyses on two test samples from each laboratory sample.
The validated uncertainties of sampling and analysis are usamp and uanal respectively.
The construction of a control chart is described in Section 13.2. In the case of baby
porridge (40 g test sample) the following calculations can be made:
Warning limit: WL = 2.38 * u 2 anal + u 2 samp = 2.83 *
Action limit:
AL = 3.69 *
UfS:2007
(4.95
(4.95
(4.95
+ 8.282 )% = 27%
+ 8.28 2 )% = 36%
2
+ 8.28 2 )% = 11%
Page 64
Example A4
Table A4.13: Quality control: calculation of differences D and D (%) between
samples from a batch
Sample S1
Sample S2
D(%) = ( D / x) * 100%
Product
Analyses
XS1
XS2
D = xS1 xS 2
P1
A1
322
350
28
336
P2
332
358
26
345
P3
443
461
18
452
P4
318
390
72
354
20
P5
252
265
13
259
P6
274
233
41
254
16
P7
206
198
202
P8
392
375
17
384
319
375
56
347
16
P2
317
393
76
355
21
P3
430
388
42
409
10
P4
383
334
49
359
14
P5
219
227
223
P6
239
217
22
228
10
P7
225
195
30
210
14
P8
335
416
81
376
22
P1
A2
The calculated D (%) in Table A4.13 can be compared directly with the action limit,
or the results can be presented in a control chart, see Figure A4.4.
UfS:2007
Page 65
Example A4
Figure A4.4: Control chart: quality control analyses of vitamin A in baby
porridge containing cereals and fruits
40
30
Calculated: D (%)
Action limit (36 %)
20
10
P8-A2
P7-A2
P6-A2
P5-A2
P4-A2
P3-A2
P2-A2
P1-A2
P8-A1
P7-A1
P6-A1
P5-A1
P4-A1
P3-A1
P2-A1
P1-A1
Product/analyses
The control chart in Figure A4.4 shows that when collecting duplicated samples from
the same batch, the difference between analytical results D (%) is smaller than the
action limit. All the calculated difference are in fact smaller than the warning limit of
27%.
The measurement uncertainty determined in the validation step is therefore considered
suitable for the quality control of the sampling of baby porridge containing milled
cereals and fruit.
If the normal procedure is to analyse one sample from each batch, it is recommended
that duplicate samples are collected from the same batch at least in one of ten of the
sampled batches.
Measurement uncertainty
Sampling uncertainty
Calculations from the validation study gave an expanded sampling uncertainty Usamp
(%) = 10% (40 g test sample see Table A4.7). The calculated uncertainty does not
include contributions to the uncertainty due to between protocol and between
samplers differences.
Analytical uncertainty
Calculation from the validation study gave an expanded measurement uncertainty of
analyses (Uanal) of 17% 40 g test sample. The laboratory reports their own
estimation of the analytical uncertainty (see Table A4.1): 2*RSDinlab(%) = 14%.
2*RSDinlab(%) is used as an estimate of Uanal in the laboratory. The Uanal found in the
validation study was at the same level but still a little bigger than the Uanal reported by
the laboratory.
The certified reference material (CRM) used is 2383 (NIST) baby food composite.
The CRM is a mix of different foods of plant and animal origins and the uncertainty
found when analysing the CRM might not be identical with that found when analysing
baby porridge powder. Laboratory data for the CRM 2383 is included in the table
below.
UfS:2007
Page 66
Example A4
Mean value mg
CRM 2383 kg-1
Laboratory
U (%)95% bias (%)
Certified
0.80 0.15
18.8
Analysed
0.77 0.14
18.2
- 3.75
The measurement uncertainty and the bias determined for the CRM could be allowed
for in the analytical measurement uncertainty (as in the NordTest UFS Guide,
Example 2), but as the matrix in the validation study is different from that for the
CRM used, we chose not to include it in this study.
Total measurement uncertainty
Calculations from the validation study gave an expanded measurement uncertainty
Umeas(%) = 20% (40 g test sample see Table A4.7).
Systematic bias
The laboratory reports a recovery of normally 90110%. Recovery based on
laboratory proficiency tests 19992005: 88113%. The results for the PT indicate no
(or very small) systematic bias. Analyses of CRM 2383 in the laboratory gives a mean
analysed value of 96.3% of the certified value which indicates a small bias (-3.7%).
As the matrix of the CRM baby food composite is different to the baby porridge,
and the analytical method includes an extraction, the bias determined when analysing
the CRM might not be representative for the analyses of baby porridge.
In the validation study, the mean value of retinol was determined to be 348 g 100 g-1
(when using a test sample of 40 g). According to data provided by the producer (see
Table A4.2), the true value for retinol was calculated to be 349 g 100 g-1 porridge
powder. This gives a recovery of 99.7% of the true value. This indicates that the
systematic error due to sampling and analyses is small and might be negligible when
analysing baby porridge-powder containing milled cereals and fruits on the
condition that a test sample of 4050 g is used.
8 Comments
When a test sample of approximately 40 g is used, the retinol concentration C in baby
porridge-powder containing milled cereals and fruit should be reported with the
expanded measurement uncertainty, i.e. C20% of the measured value C (95%
confidence).
When baby porridge-powder containing milled cereals and fruit is to be analysed, it is
recommended to use a relatively large test sample of approximately 4050 g and not
210 g as indicated in the official CEN method (EN-12823-1). As the analytical
uncertainty (40 g test sample) was bigger than the normal analytical uncertainty of the
laboratory, even larger samples than 40 g might be considered.
UfS:2007
Page 67
Example A4
9 Summary
Measurement uncertainty for 40 g test samples
Sampling Analytical
Sample
Total
Uncertainty u (%) =
RSD (%)
#
Expanded uncertainty
U (%) = 2*u
#
4.95
8.3
9.7
8.2
9.9
16.6
19.4
16.4
10 Acknowledgements
Nestl (Norway) is thanked for its enthusiastic cooperation and in addition for
providing samples to the project (validation and quality control study). Also Smaafolk
Tine Norske Meierier is thanked for kindly offering us samples for the quality
control study. The National Institute of Nutrition and Seafood Research (NIFES) is
thanked for the analytical contribution (analyses and information on the laboratory
QA-system). The study was done with financial support from the Nordic Innovation
Centre and the Norwegian Food Safety Authority.
UfS:2007
Page 68
Example A4
-1
Analyses (g 100 g )
(g 100 g-1)
Squares of differences
within groups (g 100 g-1)2
2
Sample
A1 =xi j =xi1
A2 = xi j = xi2
xi = ( xi1+ xi2)/2
(xi- x i)
B1-S1
402
325
363.5
1482.25
B2-S1
382
319
350.5
992.25
B3-S1
332
291
311.5
420.25
B4-S1
280
278
279
B5-S1
370
409
389.5
380.25
B6-S1
344
318
331
169
B7-S1
297
333
315
324
B8-S1
336
320
328
64
B9-S1
372
353
362.5
90.25
B10-S1
407
361
384
529
B1-S2
361
351
356
25
B2-S2
349
362
355.5
42.25
B3-S2
397
348
372.5
600.25
B4-S2
358
321
339.5
342.25
B5-S2
378
460
419
1681
B6-S2
381
392
386.5
30.25
B7-S2
341
315
328
169
B8-S2
292
306
299
49
B9-S2
332
337
334.5
6.25
B10-S2
322
382
352
900
X a =1/20 *
SSE
(g 100 g-1)2
SS-Error (SSE):
20
20
i =1
i =1
Degrees of
freedom (df)
Variance
(N*2-N)=20
(g 100 g-1)2
20
829.75
16595
= SSE /df
20
2* (xi- x i)2
i =1
Standard deviation,
SDanal
= SSE /df
(g 100 g-1)
28.80538
8.280978
UfS:2007
20
20
i =1
i =1
Page 69
Example A4
If the number of test samples analysed is greater than two, the squares of differences will be not be equal and the
20
(x
i =1
j=1
ij
xi )
2. df = (N*n-N)=(20*2-20)= 20 where N is the number of samples and n is the number of test samples analysed of
each batch.
i
S1A1=x
i1
S1A2=xi2
S2A1=xi3
S2A2=xi4
xi
xi1 + xi 2
xi
402
325
361
351
359.75
14.0625
14.0625
382
319
349
362
353
6.25
6.25
332
291
397
348
342
930.25
930.25
280
278
358
321
309.25
915.0625
915.0625
370
409
378
460
404.25
217.5625
217.5625
344
318
381
392
358.75
770.0625
770.0625
297
333
341
315
321.5
42.25
42.25
336
320
292
306
313.5
210.25
210.25
372
353
332
337
348.5
196
196
407
361
322
382
368
256
256
xi 3 + xi 4
xi
2
2
2
2
10
x + xi 2
x + xi 2
x + xi 4
x + xi 4
SS Samp = i1
x i + i1
xi + i 3
xi + i 3
xi
2
2
2
2
i =1
2
2
10
x + xi 2
x + xi 4
= 2 * i1
xi + 2 * i 3
x i = 14231
2
2
i =1
Vsamp = 17.224
x + xi 2
x of the two values i1
and
x i 3 + xi 4
to each of
SS Samp = 4 * d i
i =1
2
xi1 + xi 2
= 4 *
xi
2
i =1
10
2. dfs= (NB*n-NB)=(10*2-10)= 10 where NB is the number of batches and n is the number of primary samples (=
laboratory samples) analysed for each batch.
UfS:2007
Page 70
Example A5
Unit1
% m/m
(i.e. mass
fraction)
Sector/
Matrix
Uncertainty estimation
Sampling
target
Food &
25 kg bag
Feed/
Chicken feed
Purpose
Design
Statistics
Total
uncertainty
(weak links
in
measurement
chain)
Modelling
with
sampling
theory (of
Gy)
Summation
of
component
variances
1 Scope
The scope is to estimate the sampling uncertainty with the given sampling protocol by
applying Gys sampling theory (Section 10.2). The analyte is an added enzyme
ingredient in the feed. Sampling theory provides realistic estimates only if all
sampling and sample splitting operations are carried out obeying the rules of sampling
correctness; it is assumed in this example that no gross errors are present and that
incorrect sampling errors are negligible.
Page 71
Example A5
5 Results
To estimate the fundamental sampling error (FSE, Section 10.2.7, Figure 4) of the two
sampling steps, we have the to evaluate the material properties (Table A5.1).
Table A5.1: Input values for estimation of sampling uncertainty by the modelling
approach, using sampling theory
Primary
sample
M1 = 500 g
ML1= 25,000
g
d1 = 0.1 cm
g1 = 0.5
Secondary sample
Comment
M2 = 2.0 g
Sample sizes
ML2 = 500 g
d2 = 0.05 cm
g2 = 0.25
Both samples
aL = 0.05% m/m
= 100% m/m
Particle sizes
Estimated size distribution factors
Mean concentration of enzyme in the lot
Enzyme concentration in enzyme particles
c = 1.08 g cm-3
m = 0.67 g cm-3
f = 0.5
=1
These material properties give for the constitution factor (Equation 7) the value c =
2160 g cm-3 and for the sampling constants (Equation 6) C values
C1 = 540 g cm-3 and C2 = 270 g cm-3
Equation 5 can be used to give estimates of the standard deviation for each sampling
step (as estimates of the standard uncertainty).
sr1 = 0.033 = 3.3%
. Primary sample
. Secondary sample
sr3 = 0.05 = 5%
. Analytical determination
The total relative standard deviation (st, combined uncertainty) can now be estimated
by applying the rule of propagation of errors; for i errors we have:
st =
2
ri
= 0.143 = 14.3 %
6 Comments
The largest source of uncertainty in the whole measurement process is identified as
that generated in preparing the test portion (2 g) for the extraction of the enzyme.
No additional allowance has been made for uncertainties associated with systematic
effects during analysis, and incorrect sampling errors (and sampling bias) have been
assumed to be negligible.
UfS:2007
Page 72
Example A5
9 Summary
Measurement uncertainty*
Sampling
Analytical
Total
26.8% (rel)
10.0% (rel)
28.6% (rel)
UfS:2007
Page 73
Example A6
Unit
mg kg-1
to air dried
basis
Sector/
Matrix
Environmental/
agricultural
top soil
Uncertainty estimation
Sampling
target
Arable soil
143 x 22 m,
depth 30 cm
P: Ca-Acetate
Lactate (CAL)
method
Purpose
Total
uncertainty
(with
contributions
from each
sampling
effect)
Design
Statistics
Modelling
approach
(using
exploratory
measurements for
single
effects)
Summation
of
component
variances
1 Scope
Estimation of the overall uncertainty of measurement by summation of individual
uncertainty contributions from sampling, sample preparation and analysis using the
modelling approach.
3 Sampling protocol
The target area was sampled using a stratified protocol, with a sampling density of
approximately 20 increments per hectare, to a depth of 30 cm, using a soil auger.
The sampling variance of the analyte content between composite samples from
different locations characterises the statistical distribution of the analyte over the
target area. This value often depends on the distance between sampling
points/sampling locations.
UfS:2007
Page 74
Example A6
If the spatial pattern of the analyte on the area is not represented by the sampling
pattern (sampling strategy), sampling bias may occur.
With the use of a sampling tool, different effects may appear, such as point
materialisation error (Figure 3). This may occur due to an ill-defined reference level
of the soil (e.g. due to undulation of the soil surface or difficulties in the definition of
the horizons), or variation in the actual sample depth or in soil density (e.g. by
moisture content), or by selective loss of soil material from the sampling device.
These effects only lead to an uncertainty contribution, if there is a depth gradient in
the analyte content (a third dimension to the target body). For this reason, these
effects, which are difficult to determine one by one, are summarised collectively as
the depth-effect.
4.1.2 Sample preparation
The physical sample preparation comprises the step from the field sample to the
laboratory sample. Mechanical treatment, such as disaggregation, sieving, grinding
and splitting steps, reduce the amount of soil material. With these steps errors may
arise due to variation in duration and forces of mechanical treatment, heterogeneity,
segregation of different soil (particle) fractions and particles size distribution. A
periodic point selection error (Figure 3) may occur due to variation in moisture
content of the dried soil sample by sorption/desorption of water from air to an
equilibrium state (depending on the humidity and properties of the sample material,
e.g. particle size).
4.1.3 Analysis
The analysis is the third step of the measurement process, which is connected with
different kinds of effects that give rise to uncertainty contributions. The analytical
uncertainty of the laboratory samples can be estimated by previously published
procedures [1, 35]. The separation of the laboratory sample into analytical test
samples will add to the sampling uncertainty; specifically, another fundamental error
may occur. However, the random component of this sampling effect is included in the
analytical repeatability precision between test samples. A significant systematic
component should be avoided by proper mixing of the sampling powder.
4.2 Cause-and-effect diagram
Figure A6.1 shows the cause-and-effect diagram for the measurement process. In the
sampling and sample preparation steps the sources of uncertainty contributions are
given; for the analysis, only the analytical quality parameters are indicated.
4.3 Model equation
The input quantities of the sampling effects discussed above are not constituent parts
of the equation from which the measurement result is calculated. An appropriate
model equation for the overall measurement process can be established, however, by
introducing respective nominal correction factors on the analytical result:
x site = x anly f b-loc f strat f depth f prep f dry
where
xsite
= measurement result
x analy
UfS:2007
Page 75
Example A6
fb-loc
fstrat
fdepth
fprep
fdry
Figure A6.1: Cause-and-effect diagram for soil sampling on arable land (Rw is
within-laboratory reproducibility)
Long range
point selection
Point
materialisation
0 - level
depth
sampling pattern
loss of material
moisture content
number of increments
sampling strategy
"depth effect"
sample preparation
xsite
RW
heterogeneity
selective loss
bias
mech. force
drying
humidity
C ref
temperature
material properties
Mechanical sample
preparation
Analysi s
If no significant bias is detected, all correction factors can be set to unity so that the
best estimate for the measurand is given by:
x site = x anly
Because of the simplicity of the model equation (only factors), and assuming
independence between the factors, the combined uncertainty can be achieved by
variance addition of the relative standard uncertainties from the various effects:
u site =
2
2
2
2
u anly
+ u 2b-loc + u strat
+ u depth
+ u 2prep + u dry
UfS:2007
Page 76
Example A6
The area was divided into nine squares (A, B, C 1, 2, 3), and three increments are
taken from each of five squares (crosswise over the area). The increments from each
square are combined, resulting in five separate composite samples. These samples are
treated and analysed separately. The mean of the single results constitutes the
measurement result in agreement with the specification of the measurand.
The analytical results for both analytes under investigation are shown in Table A6.1.
The standard deviation between these values (ssqr) reflects the variation between the
composite samples for each nominate square.
The standard uncertainty in the overall mean value (i.e. the measurement result) due
to this effect can be estimated by considering the number of samples between
locations using the standard error on the mean:
u b-loc =
s sqr
n b-loc
Cd
Square
mg kg-1
mg kg-1
A1
0.270
124
A3
0.285
112
B2
0.343
120
C1
0.355
118
C3
0.343
105
x sqr
0.319
116
s sqr
0.039
8.0
(12%)
(6.5%)
5.4%
2.9%
ub-loc
The table shows the mean value across the five squares (the measurement result), the standard
deviation calculated from these values (ssqr), and the estimated uncertainty contribution from
the standard error on the mean (ub-loc).
Inspection of the analyte contents between the squares (Table A6.1) shows no notable
differences for phosphorus in any direction (neither vertical, nor horizontal, nor
diagonal). So, no significant bias (e.g. 0.5% ) in the measurement result can be
expected for this analyte from this source.
UfS:2007
Page 77
Example A6
For cadmium both A squares show a considerably lower analyte content than the B
and C squares. Such a gradient was not unexpected for this particular area because the
C squares lay on a forest boundary, while the A squares border on grassland and the 1
and 3 squares lay between other arable land areas. It is well known that in the upper
horizon of forest soils accumulation of heavy metal occurs, which can influence
adjacent areas.
A hypothesis-based sampling pattern was applied to look for such an effect.
However, the values measured with this sampling strategy only detected a minor
systematic effect. A standard uncertainty of 1% is therefore inserted into the
uncertainty budget for cadmium for sampling strategy.
6.3 Depth effect
For revealing the collection of effects referred to as the depth effect, the following
exploratory experiment was performed.
Increment cores are taken for a depth of 35 cm within the five test squares. From
these cores segments of 2530 cm and of 3035 cm are separated and combined.
Table A6.2 shows the analytical results for these samples.
Table A6.2: Depth experiments
Cd
mg kg-1
mg kg-1
c- (2530 cm)
0.14
47
c+ (3035 cm)
0.10
35
x-
0.34
124
x+
0.30
109
0.04
15
udepth
3.5%
3.7%
The table shows the average content of the depth horizons from five cores from different
locations, the calculated content limits and the estimated uncertainty contribution
Both analytes show a statistically significant negative gradient with respect to depth.
The uncertainty due to the depth effect was estimated by considering the analyte
content of the soil layers below and above the reference depth (c-, c+) by the following
model.
The maximum variation in the sampling depth is assumed to be not more than 10%
(i.e. 2733 cm). From these data the lower and upper content limits (x -, x+), related to
the mean content of an auger core of nominal depth, are estimated according to:
x_=
x - 0.1 c _
0.9
x+ =
x + 0.1 c +
1.1
x depth / 2
3
Page 78
Example A6
6.4 Splitting
The primary field samples were split in half, seven times by a coning and quartering
procedure resulting in a laboratory sample that was 1/64 of the original mass.
To reveal the splitting effect the following exploratory experiment was performed.
In the first splitting step the second half of the material was not discarded, but
considered as a duplicate sample, which was treated like the original sample and
analysed separately. Table A6.3 shows the relative standard deviations between the
duplicates of each of the five squares for both analytes.
As a simple approximation, the mean of the relative standard deviations is taken as the
standard uncertainty of the splitting step
u split = ssplit
Note: The observed large spread of standard deviations between the duplicates must
be expected. The 2-distribution for df = 1 shows high probability for very low values
and a moderate probability for large values.
Table A6.3: Relative standard deviations between duplicate split samples and the
mean of these standard deviations for both analytes
Cd
Square
(%)
(%)
A1
0.44
1.49
A3
9.17
2.80
B2
5.32
0.84
C1
3.24
8.88
C3
0.44
1.81
3.7
3.3
ssplit
6.5 Drying
For the drying effect no experiment was performed, but information from the
literature was used to estimate the effect. A moisture content between 1 and 3% has
been found for a large number of air-dried soil samples [44]. According to the
sampling protocol, the measurand refers to air-dried soil material. Consequently, no
correction for moisture content is required for the concentration measurements.
However, a range of xdry = 2% difference in moisture content must be considered.
Assuming a rectangular distribution across this range, the standard uncertainty for
both analytes can be estimated as:
u dry =
x dry / 2
UfS:2007
= 0 .6 %
Page 79
Example A6
6.6 Analysis
The uncertainty from the analytical process for cadmium and phosphorus (Tables
A6.4 and A6.5) were estimated from quality control data, using the Nordtest-approach
[25].
Table A6.4: Standard uncertainty components and combined uncertainty in the
analysis of the soil sample for cadmium
RW
uRw = 3.6%
cref
uref = 2.7%
bias
sbias
ubias = 2.7%
uanly = 5.2%
RW
uRw= 1.7%
cref
bias
ubias = 9.5%
uanly = 9.7%
sbias
Table A6.6 lists the evaluated standard uncertainty from the effects under
consideration. The combined uncertainty is calculated from these contributions.
UfS:2007
Page 80
Example A6
Table A6.6: Relative standard uncertainties from the considered effects and the
combined uncertainty for both analytes
Relative standard
uncertainty (%)
Effect
Cd
5.4
2.9
Sampling strategy
1.0
0.5
Depth
3.5
3.7
Splitting
3.7
3.3
Drying
0.6
0.6
Analysis
5.2
9.7
Combined uncertainty
9.1
11.3
Measurement result:
Cd:
0.32
0.06 mg kg-1
P:
116
26 mg kg-1
7 Comments
7.1 Contribution of effects
Page 81
Example A6
10 Summary
Measurement uncertainty*
Analyte
Sampling
Analytical
Total
Cd
15.0%
10.4%
18.2%
11.6%
19.4%
22.6%
UfS:2007
Page 82
Appendix B: Terminology
Appendix B: Terminology
Accuracy
Duplicate
(replicate) sample
Adapted from ISO 11074-2: 2.14 (1998) [45], ISO 1998 was
formally adapted from ISO 3534-1 (1993) [9], AMC (2005) [50]
Error of result
The test result minus the accepted reference value (of the
characteristic).
Note: Error is the sum of random errors and systematic errors.
UfS:2007
Page 83
Appendix B: Terminology
Fitness for purpose The degree to which data produced by a measurement process
enables a user to make technically and administratively correct
decisions for a stated purpose.
Laboratory sample Sample as prepared for sending to the laboratory and intended for
inspection or testing.
Measurement
Uncertainty
UfS:2007
Page 84
Appendix B: Terminology
Precision
UfS:2007
Page 85
Appendix B: Terminology
Random sampling; The taking of n items from a lot of N items in such a way that all
possible combinations of n items have the same probability of being
simple random
chosen.
sampling
Notes:
1. Random selection can never be replaced by ordinary haphazard or
seemingly purposeless choice; such procedures are generally insufficient
to guarantee randomness.
2. The phrase random sampling applies also to sampling from bulk or
continuous materials but its meaning requires specific definition for each
application.
Sample
UfS:2007
Page 86
Appendix B: Terminology
Sample preparation The set of material operations (such as reduction of sizes, mixing,
dividing etc.) that may be necessary to transform an aggregated or
bulk sample into a laboratory or test sample.
Note: The sample preparation should not, as far as possible, modify the
ability of the sample to represent the population from which it was taken.
Sample size
Sampler
Sampling location
Sampling plan
UfS:2007
Page 87
Appendix B: Terminology
Sampling point
Sampling precision The part of the total measurement precision attributable to the
sampling.
Sampling target
Sub-sample
Page 88
Appendix B: Terminology
Sub-sampling
(sample division)
Systematic error of A component of the error which, in the course of a number of test
results for the same characteristic, remains constant or varies in a
result
predictable way.
Test portion
Test sample
Sample, prepared from the laboratory sample, from which the test
portions are removed for testing or analysis.
IUPAC (1990) [46], ISO 11074-2: 3.16 (1998) [45], AMC (2005)
[50]
Trueness
UfS:2007
Page 89
Appendix B: Terminology
Uncertainty
(of measurement)
UfS:2007
Page 90
Design of experiment to estimate the bias between two sampling methods A and B, by collecting paired
samples at each target.
The first stage of the examination of the results is to check whether the paired
differences are dependent on the concentration of the analyte. This is particularly
likely to happen if the concentration range encountered in successive targets is wide.
A scatterplot provides a useful visual check. Where there is no dependence, the bias
estimate is the mean of the signed paired differences and this mean can be tested for
significant difference from zero in the usual fashion. In the example shown in Figure
C1.2, there is no apparently significant dependence between the signed difference and
the concentration, and the bias between the methods is not significantly different from
zero at the 95% level of confidence by the two-sample t-test. Where there is a clear
bias that is dependent on concentration, as in Figure C1.3, the bias should be
expressed as a function of concentration. In the instance illustrated, there is evidence
(established by the functional relationship method [40]) of a significant rotational bias
with a trend expressed by the equation Result (B) = Result (A) 1.2.
.
UfS:2007
Page 91
Differences between results of two sampling protocols applied to 25 targets, as a function of the
concentration. There is no significant bias and no suggestion of a dependence of bias on concentration.
Differences between results of two sampling protocols applied to 33 targets, plotted as a function of the
concentration. There is a significant bias (because 27/33 results are negative) and the absolute bias
increases with increasing concentration.
UfS:2007
Page 92
Random sampling: s (a L ) =
Stratified sampling: s (a L ) =
sp
N
s strat
Systematic sampling: s (a L ) =
N
s sys
N
sstrat and ssys are standard deviation estimates, where the autocorrelation has been
taken into account.
UfS:2007
Page 93
Normally the order is sp > sstrat > ssys, except when in systematic sampling the
sampling frequency is a multiple of process frequency. In this case the systematic
sampling is the worst choice and the mean may be biased.
CONCENTRATION
Random selection
50
0
10
15
20
25
30
CONCENTRATION
TIME
100
50
0
Stratified selection
0
10
15
20
25
30
CONCENTRATION
TIME
100
50
0
Systematic selection
0
10
15
20
25
30
TIME
Ten samples selected from the target by using random, stratified random and
systematic stratified sample selection.
ai a L M i
( i =1, 2,K, N )
aL M
M a
M
i i
i
1
N
Mi
( M )a
Page 94
N j
1
Vj =
hi hi
2( N j ) i = 1 + j
j = 1,2, K ,
N
2
The variogram has to be integrated to estimate the PSE for different sampling
strategies. Gy uses a robust numerical integration.
UfS:2007
Page 95
UfS:2007
Page 96
The general balanced design for the empirical estimation of uncertainty (Figure 2)
includes the uncertainty from the physical sample preparation in with the sample
step. An alternative experimental design (Figure D.1) can be used to make a separate
estimate of the uncertainty from this source (sprep). Two sub-samples from both of the
two primary samples are prepared separately (grey boxes in Figure D.1). Duplicate
test portions are taken from these sub-samples so that the analytical contribution can
also be estimated. The standard robust ANOVA can be used to separate all of these
sources of variance (Figure A1.2, and Appendix C3), by selecting two different subsets of four measurements, shown in Figure D.1. Full details of the application of this
design to food sampling are given elsewhere [30].
Figure D.1: Experimental design utilised for the estimation of uncertainty from
sample preparation, as well as that from sampling and analysis
Sampling
Sample 1
(S1)
2kg
Prep. 1
(P1)
1kg
Analysis 1
(A1)
S1P1A1
Analysis 2
(A2)
S1P1A2
S1P1A1
Sample 2
(S2)
2kg
Prep. 2
(P2)
1kg
Analysis 1
(A1)
Analysis 2
(A2)
Prep. 1
(P1)
1kg
Prep. 2
(P2)
1kg
Analysis 1
(A1)
Analysis 1
(A1)
S2P1A1
S2P2A1
S1P2A1 S1P2A2
S1P2A1
* The upper section depicts the three-layered and unbalanced experimental design. The additional layer in this experimental
design, required for the evaluation of sprep, is shown by the grey boxes. The lower section (shaded) shows the data groupings
required for the application of ANOVA so as to provide estimates of ssamp, sprep and sanal, i.e. the statistical design. Figure taken
from [30] with permission of Royal Society of Chemistry.
UfS:2007
Page 97
(a)
Sample 1
Sample 2
Analysis 1
Analysis 1
Measurement Uncertainty
Sampling Target
(b)
Sampling Uncertainty
Analytical Uncertainty
Sample 1
Analysis 1
Analysis 2
Sample 2
Analysis 1
The simplified design (Figure D.2a) has the same duplicated samples as in the full
balanced design (Figure 2), but does not include and duplicated chemical analyses.
The uncertainty estimated using this design gives the total measurement uncertainty,
without any values for the components of the uncertainty from the sampling or the
analysis. If these components are required, the analytical uncertainty can be estimated
externally by the laboratory, and removed from the total uncertainty, to give a
separate estimate of the sampling uncertainty, using Equation 1. The main advantage
of this design is that the analytical cost of implementation is only half of that for the
full balanced design, for the same number of duplicated samples. Alternatively, twice
the number of duplicated samples can be taken, from twice the number of targets to
increase their representativeness for the same expenditure on chemical analysis.
The unbalanced design (Figure D.2b) is intermediate between these two designs, with
only one analytical duplicate carried out on one of the duplicated samples. This has
the advantage of giving estimates of the sampling and analytical components of the
uncertainty, as well as the total measurement uncertainty (with the same caveats
expressed as for the full balanced design in Section 9.4.2). The analytical costs are
reduced by 25% compared with those for the fully balanced case. The degrees of
freedom in this case are similar for both the analytical and sampling estimates of
variance, which is more cost-effective than the extra degrees of freedom for the
analytical uncertainty in the fully balanced case.
Classical ANOVA can be applied to the output of both of these designs using many
different spreadsheet software packages (Appendix C3), but robust ANOVA has not
yet been developed for this case.
UfS:2007
Page 98
..(Equation E1)
This approach can usefully be illustrated using the case study of nitrate in lettuce in
Example A1. The sampling uncertainty was shown not to be fit for purpose (by the
method in Section 16.3), and the optimal uncertainty required was calculated to be
lower by a factor of approximately 2. Equation E1 predicts that this should be
achieved by increasing the sample mass by a factor of 4 (i.e. 22). The implementation
of this prediction by increasing the number of increments from 10 heads to 40 heads
of lettuce per batch, did achieve the predicted reduction in the sampling uncertainty in
this case (i.e. by a factor of 1.80, which is not statistically significantly different from
the predicted improvement of 2.0) [38]. Such successful predictions are not always
achieved in practice. In a different example for the determination of moisture in
butter, a predicted reduction of 3.7 in the usamp, was calculated to require an increase
in ms by a factor of 14. In practice this increase in sample mass only produced an
experimental improvement of 1.3. The inability of this model to predict the change in
sampling uncertainty was probably due to the nature of the heterogeneity of the
analytes in this particular material [51].
UfS:2007
Page 99
Appendix F: References
Appendix F: References
1 Ellison S L R, Roesslein M, Williams A (eds) (2000) Eurachem/CITAC Guide: Quantifying
Uncertainty in Analytical Measurement, Eurachem, 2nd edition, ISBN 0 948926 15 5. Available from
the Eurachem secretariat, or from LGC Limited (London).
2 ISO (1993) Guide to the Expression of Uncertainty in Measurement (GUM). Geneva (2nd printing
1995).
3 Ellison S L R, Roesslein M, Williams A (eds) (1995) Eurachem Guide: Quantifying Uncertainty in
Analytical Measurement, Eurachem, 1st edition.
4 Analytical Methods Committee (1995) Uncertainty of measurement: implications of its use in
analytical science. Analyst, 120, 23032308.
5 ISO/TS 21748 (2004) Guidance for the use of repeatability, reproducibility and trueness estimates in
measurement uncertainty estimation. ISO, Geneva.
6 Gy P M (1979) Sampling of Particulate Materials Theory and Practice. Elsevier, Amsterdam,
431pp.
7 Codex (2004) General Guidelines on Sampling. CAC/GL-2004
(http://www.codexalimentarius.net/download/standards/10141/CXG_050e.pdf)
8 Nordtest (2007) Uncertainty from sampling. A Nordtest handbook for sampling planners and
sampling quality assurance and uncertainty estimation. NT tec 604/TR604 (www.nordicinnovation.net)
9 ISO 3534-1: 1993 Statistics Vocabulary and Symbols, International Organization for
Standardization, Geneva
10 Ramsey M H, Squire S, Gardner M J (1999) Synthetic reference sampling target for the estimation
of measurement uncertainty. Analyst, 124 (11), 17011706.
11 Ramsey M H (1998) Sampling as a source of measurement uncertainty: techniques for
quantification and comparison with analytical sources. Journal of Analytical Atomic Spectrometry, 13,
97104.
12 Lyn J A, Ramsey M H, Coad D S, Damannt A P, Wood R, Boon K A The duplicate method of
uncertainty estimation: are 8 targets enough? (Submitted to Analyst)
13 De Zorzi P, Belli M, Barbizzi S, Menegon S, Deluisa A (2002) A practical approach to assessment
of sampling uncertainty. Accreditation and Quality Assurance, 7, 182188.
14 Kurfurst U, Desaules A, Rehnert A, Muntau H (2004) Estimation of measurement uncertainty by the
budget approach for heavy metal content in soils under different land use. Accreditation and Quality
Assurance, 9, 6475.
15 Minkkinen P (2004) Practical applications of sampling theory. Chemometrics and Intelligent Lab.
Systems, 74, 8594.
16 Gy P M (1998) Sampling for Analytical Purposes. John Wiley & Sons Ltd, Chichester.
17 Gy P M (1992) Sampling of Heterogeneous and Dynamic Material Systems. Elsevier, Amsterdam.
18 Gy P M (2004) Proceedings of First World Conference on Sampling and Blending. Special Issue of
Chemometrics and Intelligent Lab. Systems, 74, 770.
19 Pitard F F (1993) Pierre Gy's Sampling Theory and Sampling Practice. CRC Press, Boca Raton,
2nd edition.
20 Smith P L (2001) A Primer for Sampling Solids, Liquids and Gases Based on the Seven Sampling
Errors of Pierre Gy. ASA SIAM, USA.
21 Ramsey M H (1993) Sampling and analytical quality control (SAX) for improved error estimation
in the measurement of heavy metals in the environment, using robust analysis of variance. Applied
Geochemistry, 2, 149153.
22 Nordtest Sampler Certification, Version 1-0, Nordic Innovation Centre (2005).
23 Thompson M, Wood R (1995) Harmonised guidelines for internal quality control in analytical
chemistry laboratories. Pure and Applied Chemistry, 67, 649666.
UfS:2007
Page 100
Appendix F: References
References (Continued)
24 Thompson M, Ramsey M H (1995) Quality concepts and practices applied to sampling an
exploratory study. Analyst, 120, 261270.
25 Magnusson B, Naykki T, Hovind H, Krysell M (2003) Handbook for Calculation of Measurement
Uncertainty. NORDTEST report TR 537 (project 1589-02). Can be obtained from
www.nordicinnovation.net/nordtest.cfm under link Rapporter.
26 Ramsey M H, Squire S, Gardner M J (1999) Synthetic reference sampling target for the estimation
of measurement uncertainty. Analyst, 124 (11), 17011706.
27 Thompson M, Coles B J, Douglas J K (2002) Quality control of sampling: proof of concept.
Analyst, 127, 174177.
28 Farrington D, Jervis A, Shelley S, Damant A, Wood R, Thompson M (2004) A pilot study of routine
quality control of sampling by the SAD method, applied to packaged and bulk foods. Analyst, 129,
359363.
29 Codex (2004) Guidelines on measurement uncertainty CAC/GL 54-2004.
30 Lyn J A, Ramsey M H, Fussel RJ, Wood R (2003) Measurement uncertainty from physical sample
preparation: estimation including systematic error. Analyst, 128 (11), 1391 1398
31 Uncertainty and Assessment of Compliance Eurachem (in preparation).
32 Measurement Uncertainty and Conformance Testing: Risk Analysis. Joint Committee on General
Metrology, Working group 1/Subcommittee 3 (in preparation).
33 Ramsey M H, Thompson M, Hale M (1992) Objective evaluation of precision requirements for
geochemical analysis using robust analysis of variance. Journal of Geochemical Exploration, 44, 23
36.
34 Thompson M, Fearn T (1996) What exactly is fitness for purpose in analytical measurement?
Analyst, 121, 275278.
35 Commission Regulation (EC) No 563/2002 of 2 April 2002 amending Regulation (EC) No
466/2001 setting maximum levels for certain contaminants in foodstuffs, Official Journal of the
European Communities, L 86/5 to L 86/6.
36 European Directive 79/700/EEC. OJ L 207, 15.8.1979, p26.
37 Ramsey M H, Lyn J A, Wood R (2001) Optimised uncertainty at minimum overall cost to achieve
fitness-for-purpose in food analysis. Analyst, 126, 17771783.
38 Lyn, J A, Palestra I M, Ramsey M H, Damant A P, Wood R (2007) Modifying uncertainty from
sampling to achieve fitness for purpose: a case study on nitrate in lettuce. Accreditation and Quality
Assurance: Journal for Quality, Comparability and Reliability in Chemical Measurement, 12, 67-74.
39 Thompson M, Walsh J N A (1989) Handbook of Inductively Coupled Plasma Spectrometry.
Blackie, Glasgow.
40 Analytical Methods Committee (2002) Technical Brief No.10, Royal Society of Chemistry, London.
41 Ramsey M H, Taylor P D, Lee J C (2002) Optimized contaminated land investigation at minimum
overall cost to achieve fitness-for-purpose. Journal of Environmental Monitoring, 4 (5), 809814.
42 Lee J C, Ramsey M H (2001) Modelling measurement uncertainty as a function of concentration: an
example from a contaminated land investigation. Analyst, 126 (10), 17841791.
43 Ramsey M H, Argyraki A (1997) Estimation of measurement uncertainty from field sampling:
implications for the classification of contaminated land. Science of the Total Environment, 198, 243
257.
44 Dahinden R, Desaules A (1994) Die Vergleichbarkeit von Schwermetallanalysen in Bodenproben
von Dauerbeobachtungsflchen, Eidgenssische Forschungsanstalt fr Agrikulturchemie und
Umwelthygiene. Liebefeld, Switzerland, p27.
45 ISO 11074-2: 1998 Soil Quality Vocabulary. Part 2: Terms and definitions related to sampling,
International Organization for Standardization, Geneva.
46 IUPAC (1990) Nomenclature for sampling in analytical chemistry (Recommendations 1990),
prepared for publication by Horwitz W, Pure and Applied Chemistry, 62, 11931208.
UfS:2007
Page 101
Appendix F: References
References (Continued)
47 ISO Standard 78-2: Chemistry Layouts for Standards Part 2: Methods of Chemical Analysis
(Second Edition, 1999).
48 ISO 7002: 1986 Agricultural food products Layout for a standard method of sampling from a lot,
First Edition, International Organization for Standardization, Geneva.
49 IUPAC (2005) Terminology in Soil Sampling (IUPAC Recommendations 2005), prepared for
publication by De Zorzi P, Barbizzi S, Belli M, Ciceri G, Fajgelj A, Moore D, Sansone U, and Van der
Perk M. Pure and Applied Chemistry, 77 (5), 827841.
50 AMC (2005) Analytical Methods Committee Technical Brief No 19. Terminology the key to
understanding analytical science. Part 2: Sampling and sample preparation
(http://www.rsc.org/images/brief19_tcm18-25963.pdf).
51 Lyn J A, Ramsey M H, Damant A, Wood R (2005), Two stage application of the OU method: a
practical assessment, Analyst, 130, 1271-1279
UfS:2007
Page 102
Copyright 2007
ISBN 978 0 948926 26 6