Professional Documents
Culture Documents
ebook download (eBook PDF) Advanced Mathematical And Computational Tools In Metrology And Testing X all chapter
ebook download (eBook PDF) Advanced Mathematical And Computational Tools In Metrology And Testing X all chapter
http://ebooksecure.com/product/ebook-pdf-the-tools-of-
mathematical-reasoning/
https://ebooksecure.com/download/advanced-methods-and-
mathematical-modeling-of-biofilm-ebook-pdf/
https://ebooksecure.com/download/network-performance-and-
security-testing-and-analyzing-using-open-source-and-low-cost-
tools-ebook-pdf/
http://ebooksecure.com/product/ebook-pdf-load-testing-of-bridges-
proof-load-testing-and-the-future-of-load-testing-structures-and-
infrastructures-book-13/
(eBook PDF) Translational Medicine in CNS Drug
Development, Volume 29
http://ebooksecure.com/product/ebook-pdf-translational-medicine-
in-cns-drug-development-volume-29/
https://ebooksecure.com/download/progress-in-heterocyclic-
chemistry-ebook-pdf/
https://ebooksecure.com/download/concepts-of-advanced-zero-waste-
tools-present-and-emerging-waste-management-practices-ebook-pdf/
http://ebooksecure.com/product/original-pdf-mathematical-proofs-
a-transition-to-advanced-mathematics-4th-edition/
https://ebooksecure.com/download/build-your-own-cybersecurity-
testing-lab-low-cost-solutions-for-testing-in-virtual-and-cloud-
based-environments-ebook-pdf/
Contents
Foreword v
Measurement models
A. Possolo 70
vii
viii
Measurement
A. Possolo 273
O. Bodnar
Physikalisch-Technische Bundesanstalt, Berlin, Germany
and Mälardalen University, Västerås, Sweden
E-mail: olha.bodnar@mdh.se, olha.bodnar@ptb.de
C. Elster
Physikalisch-Technische Bundesanstalt, Berlin, Germany
We propose a statistical method for analyzing key comparisons with two trans-
fer standards measured in two petals. The new approach is based on an exten-
sion of the established random effects model. A full Bayesian analysis based
on the reference prior is developed and analytic expressions for the results are
derived. One benefit of the suggested approach is that it provides a comprehen-
sive assessment of the laboratory biases in terms of their posterior distributions.
Another advantage is that it can easily be applied in practice. The approach
is illustrated for the CCM.M-K7 key comparison data.
1. Introduction
Drawing inferences from data, which themselves are the results of analyses,
is known as meta-analysis. Meta-analysis has become an important statis-
tical tool in many applications. Examples comprise the combination of re-
sults from clinical trials 1,2 , the determination of fundamental constants 3–5 ,
or the analysis of interlaboratory comparisons 6 .
1
September 3, 2018 8:43 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-01 page 2
laboratory measures at the beginning and at the end of each petal. Some
of the laboratories may participate in both petals, and some only in one of
them. The goal of this paper is to propose an analysis for data arising from
such a scenario, and to provide the degrees of equivalence.
λX |σ ∼ Nn (0, σ 2 In ) , (3)
2
λY |σ ∼ Nm−1 (0, σ Im−1 ) , (4)
The derivation of the reference prior for the extended random effects
model is similarly to the derivation of the reference prior for the random
September 3, 2018 8:43 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-01 page 4
effects model given in 9 . The resulting prior depends on σ only and is given
by
r 2
−1
π(σ) ∝ σ tr (In+m−1 + σ 2 LT V−1 L) LT V−1 L . (7)
with
−1
T 1 T X
µλ|σ = L RL + 2 In+m−1 L R ,
σ Y
−1
T 1
Vλ|σ = L RL + 2 In+m−1 ,
σ
and
π(σ)
π(σ|X, Y) ∝ p p
det (σ L V L + In+m−1 ) det(KT (V + σ 2 LLT )−1 K)
2 T −1
T !!
1 X X T −1
× exp − R − µλ|σ V µλ|σ (9)
2 Y Y λ|σ
where
and
1n 0n
K= .
0m 1m
where ei is the i-th basis vector in IRn+m−1 and the marginal posterior of
σ is given in (9).
September 3, 2018 8:43 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-01 page 5
a) 5kg b) 100g
c) 10g d) 5g
Fig. 1. Measurement data from CCM.M-K7 with two transfer standards measured in
two petals. The dashed line in the figure separates the two petals. The pilot laboratory,
KRISS, participated in two petals.
Figure 1 shows the data. The pilot laboratory KRISS performed mea-
surements in both petals, while all other national laboratories participated
in one petal only.
One of the main outputs of the developed Bayesian approach are the
joint posterior and the marginal posteriors of the laboratory biases as
September 3, 2018 8:43 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-01 page 6
a) 5kg b) 100g
c) 10g d) 5g
Fig. 2. Estimated laboratory biases together with 95% probability symmetric credible
intervals for the CCM.M-K7 data with two transfer standards measured in both petals.
presented in Section 2.2. Using the analytical expressions for the marginal
posteriors, we computed the estimates for laboratory biases as the corre-
sponding posterior means together with the corresponding probabilistically
symmetric 95% credible intervals (see, Figure 2). All of the constructed
credible intervals cover zero, which indicates that all participating labora-
tories measure without biases.
a) 5kg b) 100g
c) 10g d) 5g
Fig. 3. Marginal posterior for heterogeneity parameter σ for the CCM.M-K7 data from
Figure 1. The red lines show the posterior mean and the limits of the probabilistically
symmetric 95% credible interval.
4. Conclusion
The determination of the laboratory biases is one of the most important
tasks in key comparisons. Recently, a random effect model has been applied
successfully for this purpose 8,9 . We extend that approach to the case of two
transfer standards which are measured in two petals. The new approach
is based on the Bayesian reference analysis and infers laboratory biases by
their joint and marginal posteriors. The proposed method utilizes an ana-
lytical solution and requires only a one-dimensional numerical integration.
Therefore, the proposed approach can easily be used in practice.
References
1. A. J. Sutton and J. Higgins, Recent developments in meta-analysis,
Statistics in Medicine 27, 625 (2008).
September 3, 2018 8:43 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-01 page 8
1. Introduction
A lot of key comparisons of national measurement standards have been
arranged since Mutual recognition of national measurement standards and
of calibration and measurement certificates issued by national metrology
institutes (CIPM MRA) was signed in 1999 1 . Key comparison (KC) is a
special type of interlaboratory comparison, which, on the one hand, results
in establishing degree of equivalence (DoE) of measurement standards and,
on the other hand, serves as a tool for assessment of measurement and
calibrations capabilities provided by national metrology institutes (NMIs).
Such assessment is arranged as a confirmation of measurement uncertain-
ties declared by KC participants where conventional statistical criteria of
measurement data consistency are applied. A lot of publications address
the issue of studying KC data consistency as well as items concerning eval-
uation of a key comparison reference value (KCRV) and evaluation of a
DoE of measurement standards 2–10 .
In this paper conventional procedures for the KC data consistency based
on a chi-square test and En scores are considered. The En scores are treated
as a normalized deviation from KCRV as well as normalized pair-wise de-
viation of measurement results reported by participants. The conventional
procedure for KC data evaluation consists of three steps and every step im-
plies checking the consistency of the measurement data. At the first step the
data consistency is checked using a chi-square test with the aim to validate
9
September 3, 2018 8:45 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-02 page 10
10
11
If the test is passed then the KCRV xref is established equal to x̄w
with the associated uncertainty uw :
xref = x̄w , uref = uw .
• Calculating DoEs di = xi − xref , corresponding extended uncer-
tainties U0.95 (di ), and confirming the declared uncertainties using
En scores:
|xi − xref |
En(i) = q ≤ 2.
u2i − u2ref
• Calculating the pair-wise DoEs which are also used for confirmation
of declared uncertainties:
|xi − xj |
En(i,j) = q ≤ 2.
u2i + u2j
12
n K0.95 cond
K0.95
5 2.55 2.33
10 2.80 2.64
15 2.92 2.80
20 3.01 2.91
13
n K0.95 cond
K0.95
5 2.73 2.35
10 3.16 2.92
15 3.39 3.19
20 3.54 3.37
(i,j)
The analysis of the above conditional probability P cond (En ) shows
that the consistency with the reference value for all the participants does
not guarantee pairwise consistency (and the probabilities of pair-wise con-
sistency are lower as compared with the consistency probabilities with the
KCRV). So the indication of pair-wise DoEs for the KC seems reasonable.
cond
It should be noted also that the values K0.95 , K0.95 for given n vary widely
enough (the values for the conditional probabilities are lower, as expected).
September 3, 2018 8:45 ws-procs9x6-9x6 WSPC Proceedings - 9in x 6in 11100-02 page 14
14
The following notes can be made about the distributions of the follow-
ing statistics in presence of the outlier (assuming the normality of the Xi
distributions and the equality of all the weights with the exception of wi0 ):
|x −x |
• Statistic Yi,j = √ i 2 j 2 has a normal distribution,
ui +uj
− 12
w
λ 1 + wii0 j = i0 ,
E(Yi,j ) = Var(Yi,j ) = 1.
0, j 6= i ,
0
the minimum point for the above probability P {Yi ≤ 2}. Values of wi∗0 for
some n are presented in the Table 3 as well as the corresponding weight
w
and uncertainty ratios wii0 , uuii .
0
Another random document with
no related content on Scribd:
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back
back