Step 3. Quantifying Uncertainty
7. Step 3. Quantifying Uncertainty
7.1. Introduction
7.1.1. Having identified the uncertainty sources as
• Plan to obtain the further data required
explained in Step 2 (Chapter 6), the next step is to For sources of uncertainty not adequately quantify the uncertainty arising from these
covered by existing data, either seek sources. This can be done by
additional information from the literature or • evaluating the uncertainty arising from each
standing data (certificates, equipment individual source and then combining them as
specifications etc.), or plan experiments to described in Chapter 8. Examples A1 to A3
obtain the required additional data. illustrate the use of this procedure.
Additional experiments may take the form of specific studies of a single contribution to
or uncertainty, or the usual method performance • by determining directly the combined
studies conducted to ensure representative contribution to the uncertainty on the result
variation of important factors. from some or all of these sources using
7.2.2. It is important to recognise that not all of method performance data. Examples A4 to A6
the components will make a significant represent applications of this procedure.
contribution to the combined uncertainty; indeed, In practice, a combination of these is usually
in practice it is likely that only a small number necessary and convenient.
will. Unless there is a large number of them, components that are less than one third of the
7.1.2. Whichever of these approaches is used, largest need not be evaluated in detail. A most of the information needed to evaluate the
preliminary estimate of the contribution of each uncertainty is likely to be already available from
component or combination of components to the the results of validation studies, from QA/QC
uncertainty should be made and those that are not data and from other experimental work that has
significant eliminated.
been carried out to check the performance of the method. However, data may not be available to
7.2.3. The following sections provide guidance on evaluate the uncertainty from all of the sources
the procedures to be adopted, depending on the and it may be necessary to carry out further work
data available and on the additional information as described in sections 7.10. to 7.14.
required. Section 7.3. presents requirements for the use of prior experimental study data,
7.2. Uncertainty evaluation procedure
including validation data. Section 7.4. briefly discusses evaluation of uncertainty solely from
7.2.1. The procedure used for estimating the individual sources of uncertainty. This may be overall uncertainty depends on the data available
necessary for all, or for very few of the sources about the method performance. The stages
identified, depending on the data available, and is involved in developing the procedure are
consequently also considered in later sections.
• Reconcile the information requirements
Sections 7.5. to 7.9. describe the evaluation of
with the available data
uncertainty in a range of circumstances. Section
7.5. applies when using closely matched First, the list of uncertainty sources should be
reference materials. Section 7.6. covers the use of examined to see which sources of uncertainty
collaborative study data and 7.7. the use of in- are accounted for by the available data,
house validation data. 7.8. describes special whether by explicit study of the particular
considerations for empirical methods and 7.9. contribution or by implicit variation within
covers ad-hoc methods. Methods for quantifying the course of whole-method experiments.
individual components of uncertainty, including These sources should be checked against the
experimental studies, documentary and other list prepared in Step 2 and any remaining
data, modelling, and professional judgement are sources should be listed to provide an
covered in more detail in sections 7.10. to 7.14. auditable record of which contributions to the
Section 7.15. covers the treatment of known bias uncertainty have been included.
in uncertainty estimation.
Page 16
Quantifying Uncertainty Step 3. Quantifying Uncertainty
7.3. Relevance of prior studies
7.5. Closely matched certified
7.3.1. When uncertainty estimates are based at
reference materials
least partly on prior studies of method
7.5.1. Measurements on certified reference performance, it is necessary to demonstrate the
materials are normally carried out as part of validity of applying prior study results. Typically,
method validation or re-validation, effectively this will consist of:
constituting a calibration of the whole • Demonstration that a comparable precision to
measurement procedure against a traceable that obtained previously can be achieved.
reference. Because this procedure provides information on the combined effect of many of
• Demonstration that the use of the bias data the potential sources of uncertainty, it provides obtained previously is justified, typically
very good data for the assessment of uncertainty. through determination of bias on relevant
Further details are given in section 7.7.4. reference materials (see, for example, ISO
N OTE : ISO Guide 33 [H.8] gives a useful account of
Guide 33 [H.8]), by appropriate spiking
the use of reference materials in checking
studies, or by satisfactory performance on
method performance.
relevant proficiency schemes or other laboratory intercomparisons.
7.6. Uncertainty estimation using prior
• Continued performance within statistical
collaborative method development
control as shown by regular QC sample
and validation study data
results and the implementation of effective
7.6.1. A collaborative study carried out to analytical quality assurance procedures. validate a published method, for example
7.3.2. Where the conditions above are met, and according to the AOAC/IUPAC protocol [H.9] or the method is operated within its scope and field
ISO 5725 standard [H.10], is a valuable source of of application, it is normally acceptable to apply
data to support an uncertainty estimate. The data the data from prior studies (including validation
typically include estimates of reproducibility studies) directly to uncertainty estimates in the
standard deviation, s R , for several levels of laboratory in question.
response, a linear estimate of the dependence of s R on level of response, and may include an
7.4. Evaluating uncertainty by
estimate of bias based on CRM studies. How this
quantification of individual
data can be utilised depends on the factors taken
components into account when the study was carried out.
During the ‘reconciliation’ stage indicated above
7.4.1. In some cases, particularly when little or no (section 7.2.), it is necessary to identify any method performance data is available, the most
sources of uncertainty that are not covered by the suitable procedure may be to evaluate each
collaborative study data. The sources which may uncertainty component separately.
need particular consideration are:
7.4.2. The general procedure used in combining • Sampling. Collaborative studies rarely include individual components is to prepare a detailed
a sampling step. If the method used in-house quantitative model of the experimental procedure
involves sub-sampling, or the measurand (see (cf. sections 5. and 6., especially 6.4.), assess the
Specification) is estimating a bulk property standard uncertainties associated with the
from a small sample, then the effects of individual input parameters, and combine them
sampling should be investigated and their using the law of propagation of uncertainties as
effects included.
described in Section 8. • Pre-treatment. In most studies, samples are
7.4.3. In the interests of clarity, detailed guidance homogenised, and may additionally be on the assessment of individual contributions by
stabilised, before distribution. It may be experimental and other means is deferred to
necessary to investigate and add the effects of sections 7.10. to 7.14. Examples A1 to A3 in
the particular pre-treatment procedures Appendix A provide detailed illustrations of the
applied in-house.
procedure. Extensive guidance on the application • Method bias. Method bias is often examined of this procedure is also given in the ISO Guide
prior to or during interlaboratory study, where [H.2].
possible by comparison with reference
Page 17
Quantifying Uncertainty Step 3. Quantifying Uncertainty
methods or materials. Where the bias itself, • Quantification of any uncertainties associated the uncertainty in the reference values used,
with effects incompletely accounted for in the and the precision associated with the bias
above overall performance studies. check, are all small compared to s R , no
Precision study
additional allowance need be made for bias uncertainty. Otherwise, it will be necessary to
7.7.2. The precision should be estimated as far as make additional allowances.
possible over an extended time period, and • Variation in conditions.
chosen to allow natural variation of all factors Laboratories participating in a study may tend
affecting the result. This can be obtained from towards the means of allowed ranges of
• The standard deviation of results for a typical experimental conditions, resulting in an
sample analysed several times over a period of underestimate of the range of results possible
time, using different analysts and equipment within the method definition. Where such
where possible (the results of measurements effects have been investigated and shown to
on QC check samples can provide this
be insignificant across their full permitted
information).
range, however, no further allowance is required.
• The standard deviation obtained from replicate • analyses performed on each of several
Changes in sample matrix. The uncertainty
samples.
arising from matrix compositions or levels of interferents outside the range covered by the
N OTE : Replicates should be performed at materially
study will need to be considered.
different times to obtain estimates of intermediate precision; within-batch
7.6.2. Each significant source of uncertainty not
replication provides estimates of repeatability
covered by the collaborative study data should be
only.
evaluated in the form of a standard uncertainty • From formal multi-factor experimental and combined with the reproducibility standard designs, analysed by ANOVA to provide deviation s R in the usual way (section 8.) separate variance estimates for each factor.
7.6.3. For methods operating within their defined
7.7.3. Note that precision frequently varies scope, when the reconciliation stage shows that significantly with the level of response. For all the identified sources have been included in example, the observed standard deviation often the validation study or when the contributions increases significantly and systematically with from any remaining sources such as those analyte concentration. In such cases, the discussed in section 7.6.1. have been shown to be uncertainty estimate should be adjusted to allow negligible, then the reproducibility standard for the precision applicable to the particular deviation s R , adjusted for concentration if result. Appendix E.4 gives additional guidance on necessary, may be used as the combined standard handling level-dependent contributions to uncertainty.
uncertainty.
7.6.4. The use of this procedure is shown in
Bias study
example A6 (Appendix A)
7.7.4. Overall bias is best estimated by repeated
7.7. Uncertainty estimation using in-
analysis of a relevant CRM, using the complete
house development and validation
measurement procedure. Where this is done, and
studies the bias found to be insignificant, the uncertainty
associated with the bias is simply the combination
7.7.1. In-house development and validation of the standard uncertainty on the CRM value studies consist chiefly of the determination of the
with the standard deviation associated with the method performance parameters indicated in
bias.
section 3.1.3. Uncertainty estimation from these
N OTE :
Bias estimated in this way combines bias in
parameters utilises:
laboratory performance with any bias intrinsic
• to the method in use. Special considerations The best available estimate of overall
may apply where the method in use is
precision.
empirical; see section 7.8.1.
• The best available estimate(s) of overall bias • When the reference material is only and its uncertainty.
approximately representative of the test
Page 18
Quantifying Uncertainty Step 3. Quantifying Uncertainty
materials, additional factors should be
7.7.6. Overall bias can also be estimated by the considered, including (as appropriate)
addition of analyte to a previously studied differences in composition and homogeneity;
material. The same considerations apply as for reference materials are frequently more
the study of reference materials (above). In homogeneous that test samples. Estimates
addition, the differential behaviour of added based on professional judgement should be
material and material native to the sample should used, if necessary, to assign these
be considered and due allowance made. Such an uncertainties (see section 7.14.).
allowance can be made on the basis of: • Any effects following from different
• Studies of the distribution of the bias concentrations of analyte; for example, it is
observed for a range of matrices and levels of not uncommon to find that extraction losses
added analyte.
differ between high and low levels of analyte. • Comparison of result observed in a reference
7.7.5. Bias for a method under study can also be material with the recovery of added analyte in determined by comparison of the results with
the same reference material. those of a reference method. If the results show that the bias is not statistically significant, the
• Judgement on the basis of specific materials standard uncertainty is that for the reference
with known extreme behaviour. For example, method (if applicable; see section 7.8.1.),
oyster tissue, a common marine tissue combined with the standard uncertainty
reference, is well known for a tendency to co- associated with the measured difference between
precipitate some elements with calcium salts on digestion, and may provide an estimate of
methods. The latter contribution to uncertainty is ‘worst case’ recovery on which an uncertainty given by the standard deviation term used in the estimate can be based (e.g. By treating the significance test applied to decide whether the worst case as an extreme of a rectangular or difference is statistically significant, as explained
triangular distribution).
in the example below. • Judgement on the basis of prior experience.
EXAMPLE A method (method 1) for determining the
7.7.7. Bias may also be estimated by comparison
concentration of Selenium is compared with a
of the particular method with a value determined
reference method (method 2). The results (in
by the method of standard additions, in which
mg kg -1 ) from each method are as follows:
known quantities of the analyte are added to the test material, and the correct analyte
concentration inferred by extrapolation. The
Method 1
5.40 1.47 5 uncertainty associated with the bias is then normally dominated by the uncertainties
Method 2
4.76 2.75 5 associated with the extrapolation, combined
The standard deviations are pooled to give a
(where appropriate) with any significant pooled standard deviation s c contributions from the preparation and addition of
2 stock solution.
5 + 5 − 2 N OTE : To be directly relevant, the additions should be made to the original sample, rather than a
and a corresponding value of t:
prepared extract.
= 0 . 46 7.7.8. It is a general requirement of the ISO Guide
1 1 1 . 4 that corrections should be applied for all 2 . 205 +
recognised and significant systematic effects. Where a correction is applied to allow for a
t crit is 2.3 for 8 degrees of freedom, so there is
significant overall bias, the uncertainty associated
no significant difference between the means of
with the bias is estimated as paragraph 7.7.5.
the results given by the two methods. However,
described in the case of insignificant bias
the difference (0.64) is compared with a standard deviation term of 1.4 above. This
7.7.9. Where the bias is significant, but is
value of 1.4 is the standard deviation associated
nonetheless neglected for practical purposes,
with the difference, and accordingly represents
additional action is necessary (see section 7.15.).
the relevant contribution to uncertainty associated with the measured bias.
Page 19
Quantifying Uncertainty Step 3. Quantifying Uncertainty Additional factors
leachable metals in ceramics and dietary fibre in foodstuffs (see also section 5.2. and example A5)
7.7.10. The effects of any remaining factors should be estimated separately, either by
7.8.2. Where such a method is in use within its experimental variation or by prediction from
defined field of application, the bias associated established theory. The uncertainty associated
with the method is defined as zero. In such with such factors should be estimated, recorded
circumstances, bias estimation need relate only to and combined with other contributions in the
the laboratory performance and should not normal way.
additionally account for bias intrinsic to the method. This has the following implications.
7.7.11. Where the effect of these remaining factors is demonstrated to be negligible compared
7.8.3. Reference material investigations, whether to the precision of the study (i.e. statistically
to demonstrate negligible bias or to measure bias, insignificant), it is recommended that an
should be conducted using reference materials uncertainty contribution equal to the standard
certified using the particular method, or for which deviation associated with the relevant
a value obtained with the particular method is significance test be associated with that factor.
available for comparison.
EXAMPLE
7.8.4. Where reference materials so characterised are unavailable, overall control of bias is
The effect of a permitted 1-hour extraction time variation is investigated by a t-test on five
associated with the control of method parameters
determinations each on the same sample, for the
affecting the result; typically such factors as
normal extraction time and a time reduced by 1
times, temperatures, masses, volumes etc. The
hour. The means and standard deviations (in
uncertainty associated with these input factors
mg l -1 ) were: Standard time: mean 1.8, standard
must accordingly be assessed and either shown to
deviation 0.21; alternate time: mean 1.7,
be negligible or quantified (see example A6).
standard deviation 0.17. A t-test uses the pooled variance of
7.8.5. Empirical methods are normally subjected to collaborative studies and hence the uncertainty
can be evaluated as described in section 7.6.
to obtain
7.9. Evaluation of uncertainty for ad-
hoc methods
7.9.1. Ad-hoc methods are methods established to
carry out exploratory studies in the short term, or
This is not significant compared to t crit = 2.3.
for a short run of test materials. Such methods are
But note that the difference (0.1) is compared
typically based on standard or well-established
with a calculated standard deviation term of
methods within the laboratory, but are adapted
0 . 037 × ( 1 / 5 + 1 / 5 ) =0.12. This value is the
substantially (for example to study a different analyte) and will not generally justify formal
contribution to uncertainty associated with the effect of permitted variation in extraction time.
validation studies for the particular material in question.
7.7.12. Where an effect is detected and is statistically significant, but remains sufficiently
7.9.2. Since limited effort will be available to small to neglect in practice, the provisions of
establish the relevant uncertainty contributions, it section 7.15. apply.
is necessary to rely largely on the known performance of related systems or blocks within
7.8. Evaluation of uncertainty for these systems. Uncertainty estimation should
accordingly be based on known performance on a
empirical methods
related system or systems. This performance
7.8.1. An ‘empirical method’ is a method agreed information should be supported by any study upon for the purposes of comparative
necessary to establish the relevance of the measurement within a particular field of
information. The following recommendations application where the measurand
assume that such a related system is available and characteristically depends upon the method in
has been examined sufficiently to obtain a use. The method accordingly defines the
reliable uncertainty estimate, or that the method measurand. Examples include methods for
consists of blocks from other methods and that
Page 20
Quantifying Uncertainty Step 3. Quantifying Uncertainty
the uncertainty in these blocks has been § By modelling from theoretical principles established previously.
§ Using judgement based on experience or
7.9.3. As a minimum, it is essential that an informed by modelling of assumptions estimate of overall bias and an indication of
These different methods are discussed briefly precision be available for the method in question.
below.
Bias will ideally be measured against a reference material, but will in practice more commonly be
7.11. Experimental estimation of
assessed from spike recovery. The considerations of section 7.7.4. then apply, except that spike
individual uncertainty
recoveries should be compared with those
contributions
observed on the related system to establish the
7.11.1. It is often possible and practical to obtain relevance of the prior studies to the ad-hoc
estimates of uncertainty contributions from method in question. The overall bias observed for
experimental studies specific to individual the ad-hoc method, on the materials under test,
parameters.
should be comparable to that observed for the related system, within the requirements of the
7.11.2. The standard uncertainty arising from study.
random effects is often measured from repeatability experiments and is quantified in
7.9.4. A minimum precision experiment consists terms of the standard deviation of the measured of a duplicate analysis. It is, however,
values. In practice, no more than about fifteen recommended that as many replicates as practical
replicates need normally be considered, unless a are performed. The precision should be compared
high precision is required.
with that for the related system; the standard deviation for the ad-hoc method should be
7.11.3. Other typical experiments include: comparable.
• Study of the effect of a variation of a single
N OTE : It recommended that the comparison be based
parameter on the result. This is particularly
on inspection. Statistical significance tests
appropriate in the case of continuous,
(e.g. an F-test) will generally be unreliable
controllable parameters, independent of other
with small numbers of replicates and will tend
effects, such as time or temperature. The rate
to lead to the conclusion that there is ‘no
of change of the result with the change in the
significant difference’ simply because of the
parameter can be obtained from the
low power of the test.
experimental data. This is then combined
7.9.5. Where the above conditions are met directly with the uncertainty in the parameter unequivocally, the uncertainty estimate for the
to obtain the relevant uncertainty contribution. related system may be applied directly to results
N OTE : The change in parameter should be sufficient
obtained by the ad-hoc method, making any
to change the result substantially compared to
adjustments appropriate for concentration
the precision available in the study (e.g. by
dependence and other known factors.
five times the standard deviation of replicate measurements)
7.10. Quantification of individual
• Robustness studies, systematically examining
components
the significance of moderate changes in
7.10.1. It is nearly always necessary to consider parameters. This is particularly appropriate for some sources of uncertainty individually. In some
rapid identification of significant effects, and cases, this is only necessary for a small number of
commonly used for method optimisation. The sources; in others, particularly when little or no
method can be applied in the case of discrete method performance data is available, every
effects, such as change of matrix, or small source may need separate study (see examples 1,2
equipment configuration changes, which have and 3 in Appendix A for illustrations). There are
unpredictable effects on the result. Where a several general methods for establishing
factor is found to be significant, it is normally individual uncertainty components:
necessary to investigate further. Where insignificant, the associated uncertainty is (at
§ Experimental variation of input variables least for initial estimation) that obtained from § From standing data such as measurement and
the robustness study.
calibration certificates
Page 21
Quantifying Uncertainty Step 3. Quantifying Uncertainty
• Systematic multifactor experimental designs would give the standard uncertainty. Of course, intended to estimate factor effects and
systematic deviation from traceable assigned interactions. Such studies are particularly
values and any other sources of uncertainty (such useful where a categorical variable is
as those noted in section 7.6.1.) must also be involved. A categorical variable is one in
taken into account.
which the value of the variable is unrelated to
7.12.3. Quality Assurance (QA) data. As noted the size of the effect; laboratory numbers in a
previously it is necessary to ensure that the study, analyst names, or sample types are quality criteria set out in standard operating examples of categorical variables. For procedures are achieved, and that measurements example, the effect of changes in matrix type on QA samples show that the criteria continue to (within a stated method scope) could be
be met. Where reference materials are used in QA estimated from recovery studies carried out in checks, section 7.5. shows how the data can be
a replicated multiple-matrix study. An analysis used to evaluate uncertainty. Where any other of variance would then provide within- and stable material is used, the QA data provides an between-matrix components of variance for estimate of intermediate precision (Section observed analytical recovery. The between- 7.7.2.). QA data also forms a continuing check on
matrix component of variance would provide a the value quoted for the uncertainty. Clearly, the standard uncertainty associated with matrix
combined uncertainty arising from random effects variation.
cannot be less than the standard deviation of the QA measurements.
7.12. Estimation based on other results or data
7.12.4. Suppliers' information. For many sources of uncertainty, calibration certificates or suppliers
7.12.1. It is often possible to estimate some of the catalogues provide information. For example, the standard uncertainties using whatever relevant
tolerance of volumetric glassware may be information is available about the uncertainty on
obtained from the manufacturer’s catalogue or a the quantity concerned. The following paragraphs
calibration certificate relating to a particular item suggest some sources of information.
in advance of its use.
7.12.2. Proficiency testing (PT) schemes. A laboratory’s results from participation in PT
7.13. Modelling from theoretical
schemes can be used as a check on the evaluated
principles
uncertainty, since the uncertainty should be
7.13.1. In many cases, well-established physical compatible with the spread of results obtained by theory provides good models for effects on the that laboratory over a number of proficiency test result. For example, temperature effects on rounds. Further, in the special case where volumes and densities are well understood. In
• the compositions of samples used in the such cases, uncertainties can be calculated or scheme cover the full range analysed
estimated from the form of the relationship using routinely
the uncertainty propagation methods described in • section 8.
the assigned values in each round are traceable to appropriate reference values, and
7.13.2. In other circumstances, it may be • necessary to use approximate theoretical models
the uncertainty on the assigned value is small combined with experimental data. For example, compared to the observed spread of results
where an analytical measurement depends on a timed derivatisation reaction, it may be necessary
then the dispersion of the differences between the to assess uncertainties associated with timing. reported values and the assigned values obtained This might be done by simple variation of elapsed in repeated rounds provides a basis for a good time. However, it may be better to establish an estimate of the uncertainty arising from those approximate rate model from brief experimental parts of the measurement procedure within the studies of the derivatisation kinetics near the scope of the scheme. For example, for a scheme concentrations of interest, and assess the
operating with similar materials and analyte uncertainty from the predicted rate of change at a levels, the standard deviation of differences
given time.
Page 22
Quantifying Uncertainty Step 3. Quantifying Uncertainty
7.14. Estimation based on judgement
the extent to which a proclaimed standard substance reasonably resembles the nature of
7.14.1. The evaluation of uncertainty is neither a the samples in a particular situation. routine task nor a purely mathematical one; it
depends on detailed knowledge of the nature of • Another source of uncertainty arises when the the measurand and of the measurement method
measurand is insufficiently defined by the and procedure used. The quality and utility of the
procedure. Consider the determination of uncertainty quoted for the result of a
"permanganate oxidizable substances" that measurement therefore ultimately depends on the
are undoubtedly different whether one understanding, critical analysis, and integrity of
analyses ground water or municipal waste those who contribute to the assignment of its
water. Not only factors such as oxidation value.
temperature, but also chemical effects such as matrix composition or interference, may
7.14.2. Most distributions of data can be have an influence on this specification. interpreted in the sense that it is less likely to
observe data in the margins of the distribution
A common practice in analytical chemistry than in the centre. The quantification of these
calls for spiking with a single substance, such distributions and their associated standard
as a close structural analogue or isotopomer, deviations is done through repeated
from which either the recovery of the measurements.
respective native substance or even that of a whole class of compounds is judged. Clearly,
7.14.3. However, other assessments of intervals the associated uncertainty is experimentally may be required in cases when repeated
assessable provided the analyst is prepared to measurements cannot be performed or do not
study the recovery at all concentration levels provide a meaningful measure of a particular
and ratios of measurands to the spike, and all uncertainty component.
"relevant" matrices. But frequently this
7.14.4. There are numerous instances in experimentation is avoided and substituted by analytical chemistry when the latter prevails, and
judgements on
judgement is required. For example: • the concentration dependence of • An assessment of recovery and its associated
recoveries of measurand, uncertainty cannot be made for every single
• the concentration dependence of sample. Instead, an assessment is made for
recoveries of spike,
classes of samples (e.g. grouped by type of matrix), and the results applied to all samples
• the dependence of recoveries on (sub)type of similar type. The degree of similarity is
of matrix,
itself an unknown, thus this inference (from • the identity of binding modes of native type of matrix to a specific sample) is
and spiked substances. associated with an extra element of
uncertainty that has no frequentistic
7.14.5. Judgement of this type is not based on interpretation.
immediate experimental results, but rather on a subjective (personal) probability, an expression
• The model of the measurement as defined by which here can be used synonymously with the specification of the analytical procedure
"degree of belief", "intuitive probability" and is used for converting the measured quantity
"credibility" [H.11]. It is also assumed that a to the value of the measurand (analytical
degree of belief is not based on a snap judgement, result). This model is - like all models in
but on a well considered mature judgement of science - subject to uncertainty. It is only
probability.
assumed that nature behaves according to the specific model, but this can never be known
7.14.6. Although it is recognised that subjective with ultimate certainty.
probabilities vary from one person to another, and even from time to time for a single person, they
• The use of reference materials is highly are not arbitrary as they are influenced by
encouraged, but there remains uncertainty common sense, expert knowledge, and by earlier regarding not only the true value, but also
experiments and observations. regarding the relevance of a particular
reference material for the analysis of a
7.14.7. This may appear to be a disadvantage, but specific sample. A judgement is required of
need not lead in practice to worse estimates than
Page 23
Quantifying Uncertainty Step 3. Quantifying Uncertainty
those from repeated measurements. This applies
7.15. Significance of bias
particularly if the true, real - life, variability in experimental conditions cannot be simulated and
7.15.1. It is a general requirement of the ISO the resulting variability in data thus does not give
Guide that corrections should be applied for all
a realistic picture. recognised and significant systematic effects.
7.15.2. In deciding whether a known bias can long-term variability needs to be assessed when
7.14.8. A typical problem of this nature arises if
reasonably be neglected, the following approach no collaborative study data are available. A
is recommended:
scientist who dismisses the option of substituting
i) Estimate the combined uncertainty without subjective probability for an actually measured
considering the relevant bias. one (when the latter is not available) is likely to
ii) Compare the bias with the combined ignore important contributions to combined
uncertainty.
uncertainty, thus being ultimately less objective than one who relies on subjective probabilities.
iii) Where the bias is not significant compared to
the combined uncertainty, the bias may be For the purpose of estimation of combined
neglected.
uncertainties two features of degree of belief estimations are essential:
iv) Where the bias is significant compared to the • combined uncertainty, additional action is
degree of belief is regarded as interval valued required. Appropriate actions might: which is to say that a lower and an upper
bound similar to a classical probability • Eliminate or correct for the bias, making distribution is provided,
due allowance for the uncertainty of the • correction.
the same computational rules apply in combining 'degree of belief' contributions of
• Report the observed bias and its uncertainty to a combined uncertainty as for
uncertainty in addition to the result. standard deviations derived by other
N OTE : Where a known bias is uncorrected by methods. convention, the method should be considered
empirical (see section 7.8).
Page 24
Quantifying Uncertainty Step 4. Calculating the Combined Uncertainty