Uncertainty in Cal EA-4-02
Uncertainty in Cal EA-4-02
Accreditation of Laboratories
Publication Reference
EAL-R2
PURPOSE
The purpose of this document is to harmonise evaluation of uncertainty of measurement
within EAL, to set up, in addition to the general requirements of EAL-R1, the specific de-
mands in reporting uncertainty of measurement on calibration certificates issued by accred-
ited laboratories and to assist accreditation bodies with a coherent assignment of best meas-
urement capability to calibration laboratories accredited by them. As the rules laid down in
this document are in compliance with the recommendations of the Guide to the Expression of
Uncertainty in Measurement, published by seven international organisations concerned with
standardisation and metrology, the implementation of EAL-R2 will also foster the global ac-
ceptance of European results of measurement.
Authorship
This document has been drafted by EAL Task Force for revision of WECC Doc. 19-1990
on behalf of the EAL Committee 2 (Calibration and Testing Activities). It comprises a
thorough revision of WECC Doc. 19-1990 which it replaces.
Official language
The text may be translated into other languages as required. The English language ver-
sion remains the definitive version.
Copyright
The copyright of this text is held by EAL. The text may not be copied for resale.
Further information
For further information about this publication, contact your National member of EAL:
Contents
Section Page
1 Introduction 4
2 Outline and definitions 5
3 Evaluation of uncertainty of measurement of input estimates 6
4 Calculation of the standard uncertainty of the output estimate 9
5 Expanded uncertainty of measurement 12
6 Statement of uncertainty of measurement in calibration certificates 13
7 Step-by-step procedure for calculating the uncertainty of measurement 14
8 References 15
Appendices 16
1 Introduction
1.1 This document sets down the principles of and the requirements on the evaluation of
the uncertainty of measurement in calibration and the statement of this uncertainty in
calibration certificates. The treatment is kept on a general level to suit all fields of cal-
ibration. The method outlined may have to be supplemented by more specific advice
for different fields, to make the information more readily applicable. In developing
such supplementary guidelines the general principles stated in this document should be
followed to ensure harmonisation between the different fields.
1.2 The treatment in this document is in accordance with the Guide to the Expression of
Uncertainty in Measurement, first published in 1993 in the name of BIPM, IEC, IFCC,
ISO, IUPAC, IUPAP and OIML [ref. 1]. But whereas [ref. 1] establishes general rules
for evaluating and expressing uncertainty in measurement that can be followed in most
fields of physical measurements, this document concentrates on the method most suit-
able for the measurements in calibration laboratories and describes an unambiguous
and harmonised way of evaluating and stating the uncertainty of measurement. It com-
prises the following subjects:
relationship between the uncertainty of measurement of the output quantity and the
uncertainty of measurement of the input quantities;
Worked out examples showing the application of the method outlined here to specific
measurement problems in different fields will be given in supplements. Evaluation of
uncertainty of measurement is also addressed in several of the EAL documents which
provide guidance on calibration methods, some of these documents containing specific
worked out examples.
1.3 Within EAL the best measurement capability (always referring to a particular quant-
ity, viz. the measurand) is defined as the smallest uncertainty of measurement that a
laboratory can achieve within its scope of accreditation, when performing more or less
routine calibrations of nearly ideal measurement standards intended to define, realize,
conserve or reproduce a unit of that quantity or one or more of its values, or when per-
forming more or less routine calibrations of nearly ideal measuring instruments de-
signed for the measurement of that quantity. The assessment of best measurement cap-
2.1 The statement of the result of a measurement is complete only if it contains both the
value attributed to the measurand and the uncertainty of measurement associated with
that value. In this document all quantities which are not exactly known are treated as
random variables, including the influence quantities which may affect the measured
value.
2.2 The uncertainty of measurement is a parameter, associated with the result of a meas-
urement, that characterises the dispersion of the values that could reasonably be attrib-
uted to the measurand [ref. 2]. In this document the shorthand term uncertainty is
used for uncertainty of measurement if there is no risk of misunderstanding. For typ-
ical sources of uncertainty in a measurement see the list given in Annex C.
2.3 The measurands are the particular quantities subject to measurement. In calibration
one usually deals with only one measurand or output quantity Y that depends upon a
number of input quantities Xi (i = 1, 2 ,…, N) according to the functional relation-
ship
Y = f(X1, X2, …, X )
N (2.1)
The model function f represents the procedure of the measurement and the method of
evaluation. It describes how values of the output quantity Y are obtained from values
of the input quantities Xi. In most cases it will be an analytical expression, but it may
also be a group of such expressions which include corrections and correction factors
for systematic effects, thereby leading to a more complicated relationship that is not
written down as one function explicitly. Further, f may be determined experimentally,
or exist only as a computer algorithm that must be evaluated numerically, or it may be
a combination of all of these.
2.4 The set of input quantities Xi may be grouped into two categories according to the way
in which the value of the quantity and its associated uncertainty have been determined:
(a) quantities whose estimate and associated uncertainty are directly determined in the
current measurement. These values may be obtained, for example, from a single
observation, repeated observations, or judgement based on experience. They may
involve the determination of corrections to instrument readings as well as correc-
tions for influence quantities, such as ambient temperature, barometric pressure or
humidity;
(b) quantities whose estimate and associated uncertainty are brought into the measure-
ment from external sources, such as quantities associated with calibrated measure-
ment standards, certified reference materials or reference data obtained from hand-
books.
2.5 An estimate of the measurand Y, the output estimate denoted by y, is obtained from
equation (2.1) using input estimates xi for the values of the input quantities Xi
y f ( x1 , x 2 ,.., x N ) (2.2)
It is understood that the input values are best estimates that have been corrected for all
effects significant for the model. If not, the necessary corrections have been introduced
as separate input quantities.
2.6 For a random variable the variance of its distribution or the positive square root of the
variance, called standard deviation, is used as a measure of the dispersion of values.
The standard uncertainty of measurement associated with the output estimate or
measurement result y, denoted by u(y), is the standard deviation of the measurand Y. It
is to be determined from the estimates xi of the input quantities Xi and their associated
standard uncertainties u(xi). The standard uncertainty associated with an estimate has
the same dimension as the estimate. In some cases the relative standard uncertainty
of measurement may be appropriate which is the standard uncertainty of measure-
ment associated with an estimate divided by the modulus of that estimate and is there-
fore dimensionless. This concept cannot be used if the estimate equals zero.
Note: There are occasions, seldom met in calibration, when all possible values of a
quantity lie on one side of a single limit value. A well known case is the so-
called cosine error. For the treatment of such special cases, see ref. 1.
3.2.2 Assume that the repeatedly measured input quantity Xi is the quantity Q. With n statist-
ically independent observations (n > 1), the estimate of the quantity Q is q , the arith-
metic mean or the average of the individual observed values qj
(j = 1, 2, …, n)
1 n
q q
n j 1 j
(3.1)
(a) An estimate of the variance of the underlying probability distribution is the exper-
imental variance s²(q) of values qj that is given by
1 n
s2 (q) ( q q )2
n 1 j 1 j
(3.2)
Its (positive) square root is termed experimental standard deviation. The best estim-
ate of the variance of the arithmetic mean q is the experimental variance of the
mean given by
s2 (q)
2
s (q ) (3.3)
n
Its (positive) square root is termed experimental standard deviation of the mean.
The standard uncertainty u(q ) associated with the input estimate q is the experimental
standard deviation of the mean
u ( q ) s( q ) (3.4)
Warning: Generally, when the number n of repeated measurements is low (n < 10), the
reliability of a Type A evaluation of standard uncertainty, as expressed by equation
(3.4), has to be considered. If the number of observations cannot be increased, other
means of evaluating the standard uncertainty given in the text have to be considered.
(b) For a measurement that is well-characterised and under statistical control a com-
2
bined or pooled estimate of variance sp may be available that characterises the
dispersion better than the estimated standard deviation obtained from a limited
number of observations. If in such a case the value of the input quantity Q is de-
termined as the arithmetic mean q of a small number n of independent observa-
tions, the variance of the mean may be estimated by
sp2
2
s (q ) (3.5)
n
The standard uncertainty is deduced from this value by equation (3.4).
manufacturer’s specifications;
3.3.2 The proper use of the available information for a Type B evaluation of standard uncer-
tainty of measurement calls for insight based on experience and general knowledge. It
is a skill that can be learned with practice. A well-based Type B evaluation of standard
uncertainty can be as reliable as a Type A evaluation of standard uncertainty, espe-
cially in a measurement situation where a Type A evaluation is based only on a com-
paratively small number of statistically independent observations. The following cases
must be discerned:
(a) When only a single value is known for the quantity Xi, e.g. a single measured
value, a resultant value of a previous measurement, a reference value from the lit-
erature, or a correction value, this value will be used for xi. The standard uncer-
tainty u(xi) associated with xi is to be adopted where it is given. Otherwise it has to
be calculated from unequivocal uncertainty data. If data of this kind are not avail-
able, the uncertainty has to be evaluated on the basis of experience.
(b) When a probability distribution can be assumed for the quantity Xi, based on the-
ory or experience, then the appropriate expectation or expected value and the square
root of the variance of this distribution have to be taken as the estimate xi and the as-
sociated standard uncertainty u(xi), respectively.
(c) If only upper and lower limits a+ and a– can be estimated for the value of the quant-
ity Xi (e.g. manufacturer’s specifications of a measuring instrument, a temperature
range, a rounding or truncation error resulting from automated data reduction), a
probability distribution with constant probability density between these limits (rect-
angular probability distribution) has to be assumed for the possible variability of the
input quantity Xi. According to case (b) above this leads to
1
xi (a a ) (3.6)
2
for the estimated value and
1
u 2 ( xi ) (a a ) 2 (3.7)
12
for the square of the standard uncertainty. If the difference between the limiting values
is denoted by 2a, equation (3.7) yields
1
u 2 ( xi ) a 2 (3.8)
3
The rectangular distribution is a reasonable description in probability terms of one’s
inadequate knowledge about the input quantity Xi in the absence of any other informa-
tion than its limits of variability. But if it is known that values of the quantity in ques-
tion near the centre of the variability interval are more likely than values close to the
limits, a triangular or normal distribution may be a better model. On the other hand, if
values close to the limits are more likely than values near the centre, a U-shaped distri-
bution may be more appropriate.
N
u 2 ( y ) ui2 ( y ) (4.1)
i 1
Note: There are cases, seldom occurring in calibration, where the model function is
strongly non-linear or some of the sensitivity coefficients [see equation (4.2) and
(4.3)] vanish and higher order terms have to be included into equation (4.1). For a
treatment of such special cases see ref. 1.
f f
ci (4.3)
xi Xi X 1 x1 .. X N x N
4.2 The sensitivity coefficient ci describes the extent to which the output estimate y is in-
fluenced by variations of the input estimate xi. It can be evaluated from the model
function f by equation (4.3) or by using numerical methods, i.e. by calculating the
change in the output estimate y due to a change in the input estimate xi of +u(xi) and -
u(xi) and taking as the value of ci the resulting difference in y divided by 2u(xi). Some-
times it may be more appropriate to find the change in the output estimate y from an
experiment by repeating the measurement at e.g. xi u(xi).
4.3 Whereas u(xi) is always positive, the contribution ui(y) according to equation (4.2) is
either positive or negative, depending on the sign of the sensitivity coefficient ci. The
sign of ui(y) has to be taken into account in the case of correlated input quantities, see
equation (D4) of Annex D.
the output estimate according to equation (2.2) is given by the corresponding sum or
difference of the input estimates
N
y p i x i (4.5)
i 1
N
u 2 ( y ) pi2 u 2 ( xi ) (4.6)
i 1
the output estimate again is the corresponding product or quotient of the input estim-
ates
N
y c xipi (4.8)
i1
The sensitivity coefficients equal piy/xi in this case and an expression analogous to
equation (4.6) is obtained from equation (4.1), if relative standard uncertainties w(y) =
u(y)/y and w(xi) = u(xi)/xi are used,
N
w 2 ( y ) pi2 w 2 ( xi ) (4.9)
i 1
4.6 If two input quantities Xi and Xk are correlated to some degree, i.e. if they are mutu-
ally dependent in one way or another, their covariance also has to be considered as a
contribution to the uncertainty. See Annex D for how this has to be done. The ability
to take into account the effect of correlations depends on the knowledge of the meas-
urement process and on the judgement of mutual dependency of the input quantities. In
general, it should be kept in mind that neglecting correlations between input quantities
can lead to an incorrect evaluation of the standard uncertainty of the measurand.
4.7 The covariance associated with the estimates of two input quantities Xi and Xk may be
taken to be zero or treated as insignificant if
(a) the input quantities Xi and Xk are independent, for example, because they have
been repeatedly but not simultaneously observed in different independent experi-
ments or because they represent resultant quantities of different evaluations that
have been made independently, or if
4.8 The uncertainty analysis for a measurement — sometimes called the uncertainty
budget of the measurement — should include a list of all sources of uncertainty to-
gether with the associated standard uncertainties of measurement and the methods of
evaluating them. For repeated measurements the number n of observations also has to
be stated. For the sake of clarity, it is recommended to present the data relevant to this
analysis in the form of a table. In this table all quantities should be referenced by a
physical symbol Xi or a short identifier. For each of them at least the estimate xi, the
associated standard uncertainty of measurement u(xi), the sensitivity coefficient ci and
the different uncertainty contributions ui(y) should be specified. The dimension of each
of the quantities should also be stated with the numerical values given in the table.
4.9 A formal example of such an arrangement is given as Table 4.1 applicable for the case
of uncorrelated input quantities. The standard uncertainty associated with the measure-
ment result u(y) given in the bottom right corner of the table is the root sum square of
all the uncertainty contributions in the outer right column. The grey part of the table is
not filled in.
Y y u(y)
U = ku(y) (5.1)
In cases where a normal (Gaussian) distribution can be attributed to the measurand and
the standard uncertainty associated with the output estimate has sufficient reliability,
the standard coverage factor k = 2 shall be used. The assigned expanded uncertainty
corresponds to a coverage probability of approximately 95%. These conditions are
fulfilled in the majority of cases encountered in calibration work.
5.2 The assumption of a normal distribution cannot always be easily confirmed experi-
mentally. However, in the cases where several (i.e. N 3) uncertainty components,
derived from well-behaved probability distributions of independent quantities, e.g.
normal distributions or rectangular distributions, contribute to the standard uncertainty
associated with the output estimate by comparable amounts, the conditions of the Cent-
ral Limit Theorem are met and it can be assumed to a high degree of approximation
that the distribution of the output quantity is normal.
5.3 The reliability of the standard uncertainty assigned to the output estimate is determined
by its effective degrees of freedom (see Annex E). However, the reliability criterion is
always met if none of the uncertainty contributions is obtained from a Type A evalu-
ation based on less than ten repeated observations.
5.4 If one of these conditions (normality or sufficient reliability) is not fulfilled, the stand-
ard coverage factor k = 2 can yield an expanded uncertainty corresponding to a cover-
age probability of less than 95%. In these cases, in order to ensure that a value of the
expanded uncertainty is quoted corresponding to the same coverage probability as in
the normal case, other procedures have to be followed. The use of approximately the
same coverage probability is essential whenever two results of measurement of the
same quantity have to be compared, e.g. when evaluating the results of an inter-labor-
atory comparison or assessing compliance with a specification.
5.5 Even if a normal distribution can be assumed, it may still occur that the standard un-
certainty associated with the output estimate is of insufficient reliability. If, in this
case, it is not expedient to increase the number n of repeated measurements or to use a
Type B evaluation instead of the Type A evaluation of poor reliability, the method
given in Annex E should be used.
5.5 For the remaining cases, i.e. all cases where the assumption of a normal distribution
cannot be justified, information on the actual probability distribution of the output es-
timate must be used to obtain a value of the coverage factor k that corresponds to a
coverage probability of approximately 95%.
ded uncertainty assigned to the measurement result. For the process of rounding, the
usual rules for rounding of numbers have to be used (for further details on rounding
see ISO 31-0:1992, Annex B). However, if the rounding brings the numerical value of
the uncertainty of measurement down by more than 5%, the rounded up value should
be used.
(a) Express in mathematical terms the dependence of the measurand (output quantity)
Y on the input quantities Xi according to equation (2.1). In the case of a direct
comparison of two standards the equation may be very simple, e.g.
Y = X1+X2.
(c) List all sources of uncertainty in the form of an uncertainty analysis in accordance
with Section 4.
(d) Calculate the standard uncertainty u(q ) for repeatedly measured quantities in ac-
cordance with sub-section 3.2.
(e) For single values, e.g. resultant values of previous measurements, correction val-
ues or values from the literature, adopt the standard uncertainty where it is given
or can be calculated according to paragraph 3.3.2 (a). Pay attention to the uncer-
tainty representation used. If no data are available from which the standard uncer-
tainty can be derived, state a value of u(xi) on the basis of scientific experience.
(f) For input quantities for which the probability distribution is known or can be as-
sumed, calculate the expectation and the standard uncertainty u(xi) according to
paragraph 3.3.2 (b). If only upper and lower limits are given or can be estimated,
calculate the standard uncertainty u(xi) in accordance with paragraph 3.3.2 (c).
(g) Calculate for each input quantity Xi the contribution ui(y) to the uncertainty associ-
ated with the output estimate resulting from the input estimate xi according to
equations (4.2) and (4.3) and sum their squares as described in equation (4.1) to
obtain the square of the standard uncertainty u(y) of the measurand. If input quant-
ities are known to be correlated, apply the procedure given in Annex D.
(h) Calculate the expanded uncertainty U by multiplying the standard uncertainty u(y)
associated with the output estimate by a coverage factor k chosen in accordance
with Section 5.
(i) Report the result of the measurement comprising the estimate y of the measurand,
the associated expanded uncertainty U and the coverage factor k in the calibration
certificate in accordance with Section 6.
8 References
[1] Guide to the Expression of Uncertainty in Measurement, first edition, 1993, corrected
and reprinted 1995, International Organization for Standardization (Geneva, Switzer-
land).
[2] International Vocabulary of Basic and General Terms in Metrology, second edition,
1993, International Organization for Standardization (Geneva, Switzerland).
[3] International Standard ISO 3534-1, Statistics - Vocabulary and symbols - Part I: Prob-
ability and General Statistical Terms, first edition, 1993, International Organization
for Standardization (Geneva, Switzerland).
Appendix A
A3 With 'more or less routine calibrations' is meant that the laboratory shall be able to
achieve the stated capability in the normal work that it performs under its accredit-
ation. Obviously there are instances where the laboratory would be able to do better
as a result of extensive investigations and additional precautions but these cases are
not covered by the definition of best measurement capability, unless it is the out-
spoken policy of the laboratory to perform such scientific investigations (in which
case these become the 'more or less routine' type calibrations of the laboratory).
A4 Inclusion of the qualifier 'nearly ideal' in the definition means that best measure-
ment capability should not be dependent on the characteristics of the device to be
calibrated. Inherent in the concept of being nearly ideal is thus that there should be
no significant contribution to the uncertainty of measurement attributable to phys-
ical effects that can be ascribed to imperfections of the device to be calibrated.
However, it should be understood that such a device should be available. If it is es-
tablished that, in a particular case, even the most 'ideal' available device contributes
to the uncertainty of measurement, this contribution shall be included in the de-
termination of the best measurement capability and a statement should be made that
the best measurement capability refers to calibration of that type of device.
A5 The definition of best measurement capability implies that within its accreditation
a laboratory is not entitled to claim a smaller uncertainty of measurement than the
best measurement capability. This means that the laboratory shall be required to
state a larger uncertainty than that corresponding to the best measurement capabil-
ity whenever it is established that the actual calibration process adds significantly
to the uncertainty of measurement. Typically the equipment under calibration may
give a contribution. Obviously the actual uncertainty of measurement can never be
smaller than the best measurement capability. When stating the actual uncertainty,
the laboratory shall be asked to apply the principles of the present document.
A6 It should be pointed out that according to the definition of best measurement capab-
ility the concept is applicable only to results for which the laboratory claims its
status as accredited laboratory. Thus, strictly speaking the term is of an administrat-
ive character and does not necessarily need to reflect the real technical capability of
the laboratory. It should be possible for a laboratory to apply for accreditation with
a larger uncertainty of measurement than its technical capability if the laboratory
has internal reasons for doing so. Such internal reasons usually involve cases where
the real capability has to be held in confidence to external customers, e.g. when do-
ing research and development work or when providing service to special custom-
ers. The policy of the accreditation body should be to grant accreditation on any
applied level if the laboratory is capable of carrying out calibrations on that level.
(This consideration refers not only to the best measurement capability but to all
parameters that define the scope of a calibration laboratory.)
A10 The best measurement capability should normally be stated numerically. Where the
best measurement capability is a function of the quantity to which it refers (or any
other parameter) it should be given in analytical form but in this case it may be il-
lustrative to support the statement by a diagram. It should always be unequivocally
clear whether the best measurement capability is given in absolute or relative
terms. (Usually the inclusion of the relevant unit gives the necessary explanation
but in case of dimensionless quantities a separate statement is needed.)
A11 Although the assessment should be based on the procedures of this document, in
the main text there is the requirement that the assessment normally shall be 'suppor-
ted or confirmed by experimental evidence'. The meaning of this requirement is
that the accreditation body should not rely on an evaluation of the uncertainty of
measurement only. Interlaboratory comparisons that substantiate the evaluation
have to be carried out under the supervision of the accreditation body or on its be-
half.
Appendix B
B20 sensitivity coefficient associated with an input estimate (from [ref. 1] Sec-
tion 5.1.3)
The differential change in the output estimate generated by a differential change in
an input estimate divided by the change in that input estimate.
Appendix C
(c) non-representative sampling — the sample measured may not represent the
defined measurand;
(h) inexact values of constants and other parameters obtained from external
sources and used in the data-reduction algorithm;
C2 These sources are not necessarily independent. Some of the sources (a) to (i) may
contribute to (j).
Appendix D
Correlated input quantities
D1 If two input quantities Xi and Xk are known to be correlated to some extent — i.e. if
they are dependent on each other in one way or another — the covariance associ-
ated with the two estimates xi and xk
u( xi , x k ) u( xi ) u( x k ) r ( xi , x k ) (i k ) (D.1)
has to be considered as an additional contribution to the uncertainty. The degree of
correlation is characterised by the correlation coefficient r(xi, xk) (where i k and
r 1).
N N1 N
u ( y ) c u ( x i ) 2
2 2
i
2
c c u( x , x
i k i k ) (D.3)
i 1 i 1 k i 1
N N1 N
u 2 ( y ) ui2 ( y ) 2 u ( y )u
i k ( y ) r ( xi , x k ) (D.4)
i 1 i 1 k i 1
with the contributions ui(y) to the standard uncertainty of the output estimate y res-
ulting from the standard uncertainty of the input estimate xi given by equation (4.2).
It should be noted that the second summation of terms in equation (D.3) or (D.4)
may become negative in sign.
D4 In practice, input quantities are often correlated because the same physical refer-
ence standard, measuring instrument, reference datum, or even measurement
method having a significant uncertainty is used in the evaluation of their values.
Without loss of generality, suppose that two input quantities X1 and X2 estimated by
x1 and x2 depend on the set of independent variables Ql (l = 1,2,,L)
X 1 g1 ( Q1 , Q2 ,.., QL )
(D.5)
X 2 g 2 ( Q1 , Q2 ,.., QL )
although some of these variables may not necessarily appear in both functions. The
estimates x1 and x2 of the input quantities will be correlated to some extent, even if
the estimates ql (l = 1,2,…,L) are uncorrelated. In that case the covariance u(x1,x2)
associated with the estimates x1 and x2 is given by
L
u( x1 , x 2 ) c1l c2 l u 2 ( ql ) (D.6)
l 1
where c1l and c2 l are the sensitivity coefficients derived from the functions g1 and
g2 in analogy to equation (4.3). Because only those terms contribute to the sum for
which the sensitivity coefficients do not vanish, the covariance is zero if no vari-
able is common to functions g1 and g2. The correlation coefficient r(x1,x2) associ-
ated with the estimates x1 and x2 is determined from equation (D.6) together with
equation (D.1).
D5 The following example demonstrates correlations which exist between values at-
tributed to two artefact standards that are calibrated against the same reference
standard.
Measurement Problem
The two standards X1 and X2 are compared with the reference standard QS by means
of a measuring system capable of determining a difference z in their values with an
associated standard uncertainty u(z). The value qS of the reference standard is
known with standard uncertainty u(qS).
Mathematical Model
The estimates x1 and x2 depend on the value qS of the reference standard and the ob-
served differences z1 and z2 according to the relations
x1 qS z1
(D.7)
x2 qS z2
Standard uncertainties and covariances
The estimates z1, z2 and qS are supposed to be uncorrelated because they have been
determined in different measurements. The standard uncertainties are calculated
from equation (4.4) and the covariance associated with the estimates x1 and x2 is
calculated from equation (D.6), assuming that u(z1) = u(z2) = u(z),
u2 ( x1 ) u2 ( qS ) u2 ( z )
u2 ( x2 ) u2 ( qS ) u2 ( z ) (D.8)
2
u( x1 , x2 ) u ( qS )
u2 ( qS )
r ( x1, x2 ) 2 (D.9)
u ( qS ) u2 ( z )
Its value ranges from 0 to +1 depending on the ratio of the standard uncertainties
u(qS) and u(z).
D6 The case described by equation (D.5) is an occasion where the inclusion of correla-
tion in the evaluation of the standard uncertainty of the measurand can be avoided
by a proper choice of the model function. Introducing directly the independent
variables Ql by replacing the original variables X1 and X2 in the model function f in
accordance with the transformation equations (D.5) gives a new model function
that does not contain the correlated variables X1 and X2 any longer.
D7 There are cases however, where correlation between two input quantities X1 and X2
cannot be avoided, e.g. using the same measuring instrument or the same reference
standard when determining the input estimates x1 and x2 but where transformation
equations to new independent variables are not available. If furthermore the degree
of correlation is not exactly known it may be useful to assess the maximum influ-
ence this correlation can have by an upper bound estimate of the standard uncer-
tainty of the measurand which in the case that other correlations have not to be
taken into account takes the form
2
u 2 ( y ) u1 ( y ) u2 ( y ) ur2 ( y ) (D.10)
with ur(y) being the contribution to the standard uncertainty of all the remaining in-
put quantities assumed to be uncorrelated.
Note: Equation (D.10) is easily generalised to cases of one or several groups with
two or more correlated input quantities. In this case a respective worst case
sum has to be introduced into equation (D.10) for each group of correlated
quantities.
Annex E
E2 The procedure for calculating an appropriate coverage factor k when the conditions
of the Central Limit Theorem are met comprises the following three steps:
(a) Obtain the standard uncertainty associated with the output estimate accord-
ing to the step by step procedure given in Section 7.
(b) Estimate the effective degrees of freedom eff of the standard uncertainty
u(y) associated with the output estimate y from the Welch-Satterthwaite for-
mula
u4 ( y )
eff N
ui4 ( y ) , (E.1)
i
i1
where the ui(y) (i=1,2,,N), defined in equation (4.2), are the contributions
to the standard uncertainty associated with the output estimate y resulting
from the standard uncertainty associated with the input estimate xi which are
assumed to be mutually statistically independent, and i is the effective de-
grees of freedom of the standard uncertainty contribution ui(y).
For a standard uncertainty u(q) obtained from a Type A evaluation as dis-
cussed in sub-section 3.1, the degrees of freedom are given by i = n-1. It is
more problematic to associate degrees of freedom with a standard uncer-
tainty u(xi) obtained from a Type B evaluation. However, it is common
practice to carry out such evaluations in a manner that ensures that any un-
derestimation is avoided. If, for example, lower and upper limits a– and a+
are set, they are usually chosen in such a way that the probability of the
quantity in question lying outside these limits is in fact extremely small.
Under the assumption that this practice is followed, the degrees of freedom
of the standard uncertainty u(xi) obtained from a Type B may be taken to be
i .
(c) Obtain the coverage factor k from the table of values given as Table E.1 of
this annex. This table is based on a t-distribution evaluated for a coverage
probability of 95,45%. If eff is not an integer, which will usually be the
case, truncate eff to the next lower integer.
Table E.1: Coverage factors k for different effective degrees of freedom eff.
eff 1 2 3 4 5 6 7 8 10 20 50
k 13,97 4,53 3,31 2,87 2,65 2,52 2,43 2,37 2,28 2,13 2,05 2,00