Professional Documents
Culture Documents
Abstract
This paper investigates the coverage probability of the uncertainty intervals determined in compliance with the GUM and EURACHEM Guide,
which are defined by expanded uncertainty U about the results uncorrected with the insignificant biases and corrected with the significant biases.
This coverage probability can significantly fall below the chosen level of confidence in some cases as Maroto et al. discovered by using the Monte
Carlo method. Their numerical results obtained provided that only the β errors have occurred in the test significance and findings that the coverage
reduction depends on the mutual proportions of the magnitudes of the systematic error, overall uncertainty and bias uncertainty are confirmed
in this paper by using probability calculus and numerical integration. This problem is also studied when all possible experimental biases, both
significant and insignificant, are considered. From this point of view, the reduction of the coverage probability turns out to be less severe than from
the previous one. The coverage probability is also investigated for some uncertainty intervals computed in different ways than the above mentioned
documents recommend. The intervals defined by U about the results corrected with both significant and insignificant bias give always the same
coverage probability equalling the chosen level of confidence. The intervals with some uncertainties modified or enlarged with the insignificant
biases remove or moderate the coverage reduction.
© 2006 Elsevier B.V. All rights reserved.
Keywords: Bias; β-Error; Correction for systematic error; Coverage probability; t-Test; Uncertainty
0039-9140/$ – see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.talanta.2006.02.018
V. Synek / Talanta 70 (2006) 1024–1034 1025
to identify such effects. The significant effect is characterised: obtained experimental biases were tested on significance
“significant in size relatively to the required accuracy of the
b
measurement”. However, this document does not establish how t= (1)
to prove this significance (I have not managed to find it there). u(b)
According to the EURACHEM Guide [2] and also many papers, When a found value of t was 1.96 or smaller, the experimental
e.g. [3–5], the experimental bias determined in a trueness study bias was assumed insignificant with a significance level of 0.05.
is significant provided that it differs significantly from zero with For each given δ, the frequency of such cases was recorded; it
respect to its standard uncertainty. This significance is tested indicated the probability of the β error. When the bias was found
by the t-test. When the experimental bias is found significant, it significant, a new value of b was again simulated. In the case
should be applied as a correction to compensate all future results of an insignificant bias, the authors simulated a result (denoted
of routine samples for the systematic effect. In the opposite case by c) obtained with the biased method for a routine sample. Its
(insignificant experimental bias), the correction is not necessary. value was drawn from a normal distribution. The mean of this
However, the bias uncertainty should be included in the overall distribution was (μ + δ), where μ denoted the true concentra-
combined uncertainty in both cases [1,2]. tion of the determined sample (the applied denotation is simpler
Due to the β error (type II error) in this significance test, than the origin one). The standard deviation of the distribu-
there is always a probability that the experimental bias is found tion (denoted by σ) equalled the standard uncertainty combined
insignificant, though in fact the evaluated systematic error is so from the uncertainty components reflecting all random errors of
large that it significantly affects the results of routine samples. In the applied hypothetical method. The obtained result was not
such cases, which can occur with a non-negligible probability, corrected because the experimental bias was insignificant. Its
the uncertainty interval about the uncorrected result can encom- overall standard uncertainty was
pass the true value of the measurand with a considerably lower
probability than the chosen level of confidence is (usually 95%). u = u2 (b) + σ 2 (2)
These consequences of the β error were disclosed and inves-
and its overall expanded uncertainty was
tigated by Maroto et al. [6,7]. These authors showed the condi-
tions under which the insignificant biases significantly decreased U =k·u (3)
the coverage probability of uncertainty intervals. They recom-
mended avoiding this underestimation of the uncertainty inter- where the applied coverage factor, k, was 1.96 so that the cov-
vals by including of such insignificant biases in the uncertainty erage probability was exactly 95%.
budget. This paper is concerned with the consequences of β error In the end the authors checked whether the true concentration
too, but it applies a different mathematical approach to this prob- μ was contained in the uncertainty interval defined by U about
lem and in addition a different evaluating view, which has been the result c and recorded the frequency of such cases. Since the
used in paper by O’Donnell and Hibbert [8]. numbers of these simulations were huge the relative frequency
gave the coverage probability of the uncertainty interval under
the given value of δ.
2. The β error consequences determined by Maroto et Using tables and plots, the authors documented that when the
al. values of δ increased, the β error probability decreased admit-
tedly and the coverage probability decreased too. They showed
Maroto et al. investigated the β error consequences by using that this coverage depression depended on the relative size of
the Monte Carlo method; they assumed both constant errors [6] the bias uncertainty u(b) in comparison with the overall standard
and proportional errors [7]. In the former paper they inquired uncertainty u (ratio u(b)/u). When u(b) was higher than 30% of
three cases of models with a few uncertainty components aris- u, the coverage probability might fall too low (in extreme cases
ing from both random and systematic effects and with different lower than 80%). The authors made analogous conclusions when
values of those components. These components were considered they investigated proportional bias (more exactly recovery) and
constant in each investigated case. The authors studied how the relative uncertainty [7]. They meant that in such cases, the uncer-
magnitude of the true bias − true systematic error (denoted by tainty interval defined by the common expanded uncertainty U
δ) of a hypothetical analytical method would affect the coverage was too underestimated and consequently they recommended
probability of the uncertainty intervals of the results obtained applying different uncertainty intervals, which were modified
by this method. They took some selected values of the true with b. Particularly they proposed to calculate uncertainty inter-
bias δ and for each given δ they obtained possible values of vals obtained by so-called SUMU method or U(bias) method
its estimate − experimental bias (denoted by b). These values of [6,7], see below.
b originated from a normal distribution with a mean of δ and
a standard deviation equal to the uncertainty of the experimen- 3. The coverage probabilities of the intervals defined by
tal bias determination (denoted by u(b)); this distribution can U in case of the β error—probability calculus
be denoted by N[δ, u(b)]. The values of b were computed from
results generated by the Monte Carlo method (simulated results This chapter shows that the β error probability and the
obtained by the biased analytical method on a certified reference coverage probability can be obtained by using classic statis-
material (CRM) and simulated certified values of the CRM). The tic computations. (The standard uncertainties are assumed
1026 V. Synek / Talanta 70 (2006) 1024–1034
Table 1
Coverage probability of the uncertainty intervals defined by 1.96u about uncorrected results obtained by a biased analytical method; dependence on the method
characteristics (the computations consider only the insignificant values of experimental biases)
δ/u(b) u(b)/u β error probability
(δ) True systematic error, (u(b)) experimental bias uncertainty, (u) overall combined uncertainty, (σ) combined uncertainty of all random effects.
the results obtained with some selected values of the β error Maroto et al. [6] studied the coverage probability of these mod-
probability and the uncertainties applied in Case 1 of the paper ified intervals also by using the Monte Carlo method.
[6]; i.e. u(b) = 0.70, σ = 1.5, u = 1.655 (its value of 1.64 given in To calculate the coverage probabilities of these modified
Ref. [6] is a little incorrect). These results are the same as the uncertainty intervals in the cases of β error is a bit more difficult
ones obtained by the Monte Carlo method [6]. matter than the previous application of probability calculus. The
possible future results c have the distribution N[μ + δ, σ]. We
4. The coverage probabilities of the modified need the probability, with which an uncertainty interval rang-
uncertainty intervals—probability calculus ing from c − Ue− to c + Ue+ encompasses μ, or inversely the
probability that a future result c belongs between μ − Ue+ and
As stated above, Maroto et al. [6] found out that when μ + Ue− . Ue− and Ue+ are general denotations of the sides of the
the experimental biases were insignificant and their uncertain- modified uncertainty intervals; the widths of the left and right
ties were higher, the uncertainty intervals defined by U about sides can be different.
the uncorrected results c were underestimated, i.e. their cov-
erage probabilities were too low. The authors sampled four P(c − Ue− ≤ μ ≤ c + Ue+ ) = P(μ − Ue+ ≤ c ≤ μ + Ue− )
methods modifying the uncertainty intervals about the uncor-
μ + Ue− − (μ + δ) μ − Ue+ − (μ + δ)
rected results in order to get higher coverage probabilities. =Φ −Φ
These modified uncertainty intervals are defined (they are σ σ
termed by their specific denomination applied in literature −δ + Ue− −δ − Ue+
[6–10]): =Φ −Φ (11)
σ σ
U(bias) = U + |b| (7)
Since both Ue+ and Ue− depend on the found value of b, the
U(RSSu) = k (u2 + b2 ) (8) probability changes according to this value (conditional proba-
bility). Consequently, the coverage probability can be assumed
SUMU with two different sides invariable only for values of the experimental bias, which are
from an interval between b and b + db, where db is an infinitesi-
U+ = U − b and U− = U + b, provided |b| ≤ U (9)
mal difference of b. The values of b occur in conformity with the
probability density function, f(b), of the distribution N[δ, u(b)].
U(RSSU) = U 2 + b2 (10) The probability that a found value of b belongs to that infinitesi-
1028 V. Synek / Talanta 70 (2006) 1024–1034
mal interval equals f(b)·db. The coverage probabilities given by 5. Evaluation of coverage probability considering only
Eq. (11), which are individual for each infinitesimal interval of the β error cases
b, are used in the computation of the mean coverage probabil-
ity for the whole range of the insignificant experimental biases The plot in Fig. 2 depicts the relationships between the β error
(from −1.96u(b) to 1.96u(b)). Since each individual coverage probabilities and the coverage probabilities of the uncertainty
probability, ICP, occurs with its own probability, this mean must intervals defined by all the mentioned methods about uncor-
be computed as a weighted mean. The weights are those proba- rected results. This plot is identical with the one given in paper
bilities f(b)·db and their sum equals the total probability of those [6] though the depicted results in both plots were obtained by
insignificant b values, i.e. the β error probability (this probability two completely different methods. It shows that the Monte Carlo
can also be computed according to Eq. (5)). This mean cover- method and probability calculus give the same results and also
age probability, MCP, is computed over the whole range of the confirms the disturbing results published by Maroto et al. [6,7].
insignificant experimental biases for given values of δ and u(b) These results indicate that in case of the insignificant experimen-
and certain type of the uncertainty modified with insignificant tal biases, which are not applied in the corrections of results, the
bias coverage probabilities of the uncertainty intervals defined by
1.96u(b) expanded uncertainty U can be very low. Table 1 shows two
−1.96u(b) ICP · f (b) · db conditions for it: (1) the true bias δ must be high enough in
MCP = 1.96u(b) comparison with the uncertainty of the experimental bias u(b)
−1.96u(b) f (b) · db (parameter δ/u(b)) and (2) the u(b) must be high enough in com-
1.96u(b) −δ+Ue−
−δ−Ue+
parison with the overall uncertainties u (parameter u(b)/u). The
−1.96u(b) Φ σ − Φ σ · f (b) · db
= 1.96u(b) table also says that a higher ratio δ/u(b) brings about a lower
−1.96u(b) f (b) · db probability that an insignificant experimental bias (β error prob-
(12) ability) will occur and thereby also a lower probability of such
a high depression of the coverage probability. When the cov-
The general uncertainties Ue− and Ue+ in Eq. (12) can be erage probability of the computed uncertainty intervals is to be
replaced by those particular modified uncertainties defined by close to the chosen level of confidence in spite of the fact that
Eqs. (7)–(10). The definite integrals obtained in this way were the parameters δ/u(b) and u(b)/u of a given analytical method
solved numerically. The resulting mean coverage probabilities are high, those uncertainty intervals modified with the experi-
computed with the uncertainties of Case 1 from the article by mental biases should be computed (see Fig. 2). It means that it
Maroto et al. [6] were plotted in Fig. 2. is necessary to direct the uncertainty computations according to
As stated above, in paper [7], Maroto et al. also studied the the parameter u(b)/u, which depends on the ratio of both types
problem of coverage probability in case of proportional errors of uncertainty (u(b)/σ; see Table 1). This conclusion seems to
of the measured results. It means that they assumed the mag- cast doubt upon the thesis given in the GUM [1] that there is no
nitudes of random and systematic errors to increase with the inherent difference between the uncertainty components arising
concentration level of the analyte. Consequently, they replaced from random effects and from systematic effects.
the absolute uncertainties changing in this instance with the
invariable relative standard uncertainties, i.e. with coefficients 6. The coverage probabilities of uncertainty intervals
of variation (relative standard deviations) and the changing for the whole distribution of b
true biases with the invariable true recoveries. (This problem
was not studied by the probability calculus in this paper. I The investigation of coverage probability in the previous
think this problem to be mathematically very similar to the chapters considered only the β error cases. It is true that (1)
previous one.) When the authors applied the methods defying when δ differs from zero, each experimental bias insignificant
those special uncertainty intervals modified or enlarged with by the t-test is a case of the β error and also that (2) practically
the insignificant uncorrected experimental biases they distin- each analytical method has its own systematic error δ, mwhich
guished between the mathematical method of uncertainty cal- of course can be negligible and undiscovered. From this point of
culation and the model of errors [6,7]. It means that the same view such a narrow base of the coverage probability evaluation
mathematical method applied in combination with the two dif- seems to be justifiable. Nevertheless, I think that this evalua-
ferent error models (either constant or proportional) was always tion should take into account all possible results obtained for
taken as the only method, though both combinations would give the whole distribution of experimental biases b conditioned by
different uncertainty intervals being applied on the same mea- a given value of the parameter δ/u(b). This distribution includes
sured results (of course, only one error model could fit a set both significant biases applied in the correction and insignificant
of observed results). Also paper [10] grasps a “method defin- biases not applied in it. Consequently besides those underes-
ing uncertainty intervals” in this way. However, O’Donnell and timated uncertainty intervals defined by U about uncorrected
Hibbert [8] take the same mathematical method combined with results there always must be some uncertainty intervals defined
the two error models as two different methods of uncertainty by U about corrected results, which can be properly estimated
computations (e.g. they distinguish the RSSu method and the or maybe even overestimated. In addition the uncertainty inter-
method denoted by names of its authors Barwick and Ellison vals with those problematically low coverage probabilities are
[11]). less probable because they appear only under high values of δ
V. Synek / Talanta 70 (2006) 1024–1034 1029
Table 2
Coverage probability of the uncertainty intervals defined by 1.96u about uncorrected results obtained by a biased analytical method; dependence on the method
characteristics (the computations consider only the insignificant values of experimental biases)
δ/u(b) u(b)/u β error probability
(δ) True systematic error, (u(b)) experimental bias uncertainty, (u) overall combined uncertainty, (σ) combined uncertainty of all random effects.
but it is calculated as the β error probability (see Table 2). The intervals defined by SUMU method about the uncorrected results
significant biases, lying outside this interval, are the cases of α are computed. The integration is computed continuously along
error (type I error) in the significance test. both lines 2 and 2 in Figs. 3 and 4 over the whole range of
It is evident that the equations to calculate the mean coverage possible biases
probability from the ICPs must differ only in the parts solving ∞
b−δ+U b−δ−U
the region of the insignificant experimental biases. The simplest MCP = Φ −Φ
case arises when the intervals defined by expanded uncertainty −∞ σ σ
U about always corrected results are computed. It is identical ·f (b) · db (14)
with the case when for the range of the insignificant biases the
method when the insignificant biases occur (in case of signifi- icance and that this coverage depression depends on these
cant biases the intervals defined by U about the corrected results three quantities (their relative magnitudes): the true bias δ and
are applied). This line expressing the relationship between the the uncertainty of the experimental bias u(b) and the overall
mean coverage probability and δ deviates both positively and uncertainty u (or the combined uncertainty of the components
negatively from the expected value (0.95) like line 1, but the arising from the random errors σ). The performed computa-
negative deviations are less. tions apply only two parameters expressing the relative mag-
nitudes of these quantities: δ/u(b) and u(b)/u. The paper tries
8. Evaluation of coverage probability considering the to show that the approach by Maroto et al. [6], which took
whole distribution of b into account only the possible insignificant experimental biases
under a given value of the true bias, does not give a com-
The obtained results confirms when the uncertainty intervals plete view. It prefers the approach by O’Donnell and Hibbert
are calculated according to the GUM [1] and the EURACHEM [8] evaluating the coverage probabilities of uncertainty inter-
Guide [2], the β error can causes that the mean coverage proba- vals from the viewpoint of all possible experimental biases,
bility evaluated over the whole distribution of the experimental both significant and insignificant. This viewpoint does not show
biases for a given value of δ is lower than 0.95. Table 2 shows that the coverage probability depression to be as severe as Maroto
if the bias uncertainty is extremely high (values u(b)/u higher et al. indicated. The lowest coverage (below 0.9) appears for
than 0.5) the mean coverage probability can fall bellow 0.90. the parameter δ/u(b) of a middle size (about 2) and with an
This coverage depression is not such dramatic as the previous extremely high value of u(b)/u (higher than 0.5). When δ/u(b)
approach indicates, nevertheless it can be considered too large. is low then the coverage probability is higher than 0.95 and
The mean coverage probability can be increased by application when δ/u(b) is extremely high the coverage reaches a value
of those uncertainty intervals modified with the experimental of 0.95. When the results are corrected with both significant
bias in case of the insignificant biases. The SUMU uncertainty and insignificant biases the coverage probability equals 0.95
is shown as the best because it keeps the mean coverage probabil- at all values of δ. It indicates that from this point of view
ity at a level of 0.95, but the obtained intervals are asymmetric. besides the t-test an additional criterion of the experimen-
The other studied methods are also acceptable, either U(bias) tal bias significance should be taken into account. This crite-
method, which makes only positive deviations of the mean rion would discover other experimental biases, which should
coverage probability, which might be regarded as too high, or be applied in the result correction. Then the coverage prob-
Ue (95%) method, which moderately deviates the mean cover- ability would not need to be increased by calculating some
age probability from 0.95 to both sides. Very important findings, uncertainty intervals modified with the insignificant biases,
revealed by O’Donnell and Hibbert [8] and confirmed in this e.g. by using U(bias) or SUMU methods, as Maroto et al.
paper, are that the uncertainty intervals, which are defined by recommended.
U about the results always corrected with both significant and At the end of Part 1 of this work it is necessary to note that
insignificant biases, keep the mean coverage probability at a there is a shortcoming of the used computations evaluating cov-
level of 0.95. This finding states that a measured result should erage probability. The coverage probability here is quantified
be corrected not only when the experimental bias is statistically for given values of true bias, which are in fact unknown. The
significant with respect to its u(b) (on the base of the t-test; see analyst who knows only an experimental bias of the applied
Eq. (1)) but also when the experimental bias is too high—really analytical method (an estimate of that true value) would need
significant. The question is what means ‘really significant’ in the coverage probability expressed as a function of this known
this case. Provided that significance is grasped in this way, the quantity. Then, he could compute the coverage probabilities of
idea saying that there is no inherent difference between both various types of uncertainty intervals for his given case and he
types of uncertainty components (from random and systematic could make a decision on it. Part 2 of this work tries to solve this
effects) is saved. However, only the GUM allows this interpreta- problem.
tion of significance since it asks to correct all systematic effects
significant in size relatively to the required accuracy of a mea-
Acknowledgements
surement [1]; the EURACHEM Guide defines significance only
by the t-test [2].
I am grateful to Dr. Pavel Janoš (UJEP Ústı́ nad Labem)
for his suggestions that improved this article and to Miss Soňa
9. Conclusions
Hýbnerová for her assistance in preparing the text of this
article.
This paper confirms the results of the coverage probabil-
ity evaluations obtained by Maroto et al. [6] and also some
results published by O’Donnell and Hibbert [8]. It proves that Appendix A
these results made by the Monte Carlo method can also be
obtained by probability calculus. It confirms that the coverage The mean coverage probabilities in Eqs. (14) and (15) are in
probabilities of the uncertainty intervals defined in agreement fact computed on the base of the total probability law [12,13]. Let
with GUM [1] and EURACHEM Guide [2] can fall bellow B1 , B2 , . . ., Bn be mutually exclusive events whose union is the
0.95 because of the β error in the t-test of the bias signif- whole sample space of an experiment. Then for the probability
V. Synek / Talanta 70 (2006) 1024–1034 1033
of an event A, P(A), this formula is valid appendix specifies only the relationships, which can be derived
n from the most extensive formula expressed by Eq. (15). They
P(A) = P(A|Bi )P(Bi ) compute the MCPs when the whole distribution of b is con-
i=1
sidered. The obtained equations consist of three integrals; two
integrals for the uncertainty intervals defined by U about the
where P(Bi ) denotes the probability of Bi and P(A|Bi ) denotes results corrected with the significant biases are common for all
the probability of A given Bi (conditional probability). (It means cases, the third one alternates and it is expressed separately for
each studied case. The general equation is
−1.96u(b)
b − δ + 1.96u b − δ − 1.96u 1.96u(b) −δ + U− −δ − U+
MCP = Φ −Φ · f (b) · db + Φ −Φ
−∞ σ σ −1.96u(b) σ σ
+∞
b − δ + 1.96u b − δ − 1.96u
·f (b) · db + Φ −Φ · f (b) · db
+1.96u(b) σ σ
Appendix B References
The complete relationships to compute the mean coverage [1] ISO, Guide to the Expression of Uncertainty in Measurement, International
probabilities can be derived from Eqs. (12), (14) and (15). This Organisation for Standardisation, Geneva, 1993.
1034 V. Synek / Talanta 70 (2006) 1024–1034
[2] L.R. Ellison, M. Rősslein, A. Williams (Eds.), EURACHEM/CITAC [8] G.E. O’Donnell, D.B. Hibbert, Analyst 130 (2005) 721.
Guide, Quantifying Uncertainty in Analytical Measurement, second ed., [9] S.D. Phillips, K.R. Eberhardt, J. Res. Natl. Inst. Technol. 102 (1997) 577.
2000. [10] V. Synek, Talanta 65 (2005) 829.
[3] S.L.R. Ellison, Accred. Qual. Assur. 3 (1998) 95. [11] V.J. Barwick, S.L.R. Ellison, Analyst 124 (1999) 981.
[4] U. Kurfűst, Accred. Qual. Assur. 3 (1998) 406. [12] C. Clapham, The Concise Oxford Dictionary of Mathematics, second ed.,
[5] W. Hässelbarth, Accred. Qual. Assur. 3 (1998) 418. Oxford University Press, Oxford, 1996, p. 277.
[6] A. Maroto, R. Boqué, J. Riu, F.X. Rius, Accred. Qual. Assur. 7 (2002) 90. [13] Wikipedia, the free encyclopedia, http://en.wikipedia.org/wiki/Law of
[7] A. Maroto, R. Boqué, J. Riu, F.X. Rius, Analyst 128 (2003) 373. total probability, modified May 6, 2005.