Talanta 105 (2013) 360–365

Contents lists available at SciVerse ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Estimation of measurement uncertainty arising from manual sampling

of fuels
Dimitrios Theodorou n,1, Nikolaos Liapis, Fanourios Zannikos
National Technical University of Athens, School of Chemical Engineering, Laboratory of Fuels and Lubricants Technology, Iroon Polytechneiou 9, Athens 15780, Greece

a r t i c l e i n f o a b s t r a c t

Article history: Sampling is an important part of any measurement process and is therefore recognized as an important
Received 3 June 2012 contributor to the measurement uncertainty. A reliable estimation of the uncertainty arising from
Received in revised form sampling of fuels leads to a better control of risks associated with decisions concerning whether
14 October 2012
product specifications are met or not. The present work describes and compares the results of three
Accepted 18 October 2012
Available online 29 October 2012
empirical statistical methodologies (classical ANOVA, robust ANOVA and range statistics) using data
from a balanced experimental design, which includes duplicate samples analyzed in duplicate from 104
Keywords: sampling targets (petroleum retail stations). These methodologies are used for the estimation of the
Uncertainty uncertainty arising from the manual sampling of fuel (automotive diesel) and the subsequent sulfur
Sampling
mass content determination. The results of the three methodologies statistically differ, with the
Fuel
expanded uncertainty of sampling being in the range of 0.34–0.40 mg kg  1, while the relative
Classical ANOVA
Robust ANOVA expanded uncertainty lying in the range of 4.8–5.1%, depending on the methodology used. The
Range statistics estimation of robust ANOVA (sampling expanded uncertainty of 0.34 mg kg  1 or 4.8% in relative
terms) is considered more reliable, because of the presence of outliers within the 104 datasets used for
the calculations. Robust ANOVA, in contrast to classical ANOVA and range statistics, accommodates
outlying values, lessening their effects on the produced estimates. The results of this work also show
that, in the case of manual sampling of fuels, the main contributor to the whole measurement
uncertainty is the analytical measurement uncertainty, with the sampling uncertainty accounting only
for the 29% of the total measurement uncertainty.
& 2012 Elsevier B.V. All rights reserved.

1. Introduction four situations apparent for a case of compliance with an upper limit
and the conclusions drawn under the probabilistic and deterministic
The aim of sampling is to obtain a small portion of material approach (assuming that an upper limit is set with no allowance for
(sample) from a selected system (sampling target) within a uncertainty). EURACHEM/ CITAC Guide ‘‘Use of uncertainty informa-
container which is representative of the material in that system tion in compliance assessment’’ [5] covers the above matters
[1,2]. The sampling process should ensure that the sample is an extensively.
unbiased reflection of the composition of the sampling target [1]. Sampling becomes extremely important when considering the
Representative samples of petroleum and petroleum products are uncertainty of measurement. Until recently a ‘‘metrological gap’’
required for the determination of their chemical and physical existed between analysts and end-users concerning the interpre-
properties, which are often used to establish compliance with tation of measurement results and their associated uncertainties.
commercial and regulatory specifications [2]. Analysts concentrated on the analytical measurement process and
When a measurement result is compared with specified limits in estimated the uncertainty of the measurand of the sample
order to make a decision relating to conformance or compliance, it is received at the laboratory while the end user naturally inter-
very likely that measurement uncertainty will have implications for preted the measurement result together with its uncertainty in
the interpretation of the result. Not accounting for the uncertainty order to characterize the sampling target as a whole [7,8].
(deterministic approach) may lead to incorrect decisions i.e. false Therefore, the end user needs to know a precise estimate of an
positive or false negative classifications that may have financial, uncertainty that includes the uncertainty caused by sampling i.e.
health, environmental or other consequences [3,4]. Fig. 1 shows the the combined uncertainty from sampling and analysis [7,9,10].
Reliable estimations of the uncertainties of fuel sampling and
n
analysis are important as they are associated with the application
Corresponding author. Tel./fax: þ 30 21 0652 5976.
E-mail address: [email protected] (D. Theodorou).
of legal requirements and the identification of events of cross
1
Permanent Address: 57a Parnassou Street, 15234 Halandri, Athens, Greece. contamination of incompatible fuels and fuel adulteration.

0039-9140/$ - see front matter & 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.talanta.2012.10.058
 D. Theodorou et al. / Talanta 105 (2013) 360–365 361

2. Sampling protocol and experimental design

If true value >Tthen A balanced nested experimental design was used. Duplicate
false positive result
samples were taken from 104 petroleum retail stations, which
were selected at random and comprised the 10.9% of the 950
Analyte X Upper threshold
petroleum retail stations monitored by the laboratory. The
concentration value (T) scheme of sampling is shown in Fig. 2. The duplicated samples
(C) C+U X
were taken by repeating the same sampling protocol. The sam-
C pling protocol used was consistent with the standard method
If true value < T then ASTM D 4057 [2] concerning the manual sampling of petroleum
C-U false negative result
and petroleum products. Instructions were given to the samplers
to introduce variations to the sampling process provided that they
do not violate any requirement of the sampling protocol. These
Possibly Probably Non Probabilistic variations actually represent variations which may arise due to
Compliant
Compliant Compliant Compliant Classification the random nature of the sampling process. All automotive diesel
samples were maintained in special closed containers. During
Non Non Deterministic transport and storage samples were protected to prevent weath-
Compliant Compliant Compliant Compliant Classification ering or degradation from light, heat or other potential detri-
mental condition.
< 0.025 0.025-0.500 0.500-0.975 > 0.975 Probability of non
compliance (C>T)
3. Analytical method
Fig. 1. Deterministic and probabilistic classification for compliance assessment
against an upper limit. Adapted from [6].
The determination of sulfur mass content of diesel fuel
samples was carried out in the Laboratory of Fuels and Lubricants
Technology (National Technical University of Athens), which is
Sampling uncertainty is defined as the part of the total measure- operating under ISO 9001 [18] and ISO/IEC 17025 [19] manage-
ment uncertainty attributable to sampling [3,8]. Principles and ment systems and participates successfully in Proficiency Testing
procedures for estimating the uncertainty of measurement arising Schemes for a range of fuel quality parameters (sulfur mass
from sampling are described in the Guide published by Eurachem content included). The duplicated samples were analyzed in
and CITAC [11] as well as in the Nordtest handbook [8] which is duplicate under repeatability conditions for sulful mass content
intended for practical applications. determination. An ANTEK 9000S sulfur analyzer equipped with an
There are two broad approaches for the estimation of uncer- automatic sampler was employed in this work. This analyzer fully
tainty, the modeling method and the empirical method complies with ASTM D 5453 [16] and ISO 20846 [20]. Table 1
[12,13,10]. The modeling approach which is consistent with ISO presents the operating conditions of the instrument. The sample
GUM [14] and is described as a ‘‘bottom up’’ approach [15], sulfur content in nanograms per microlitre (ng mL  1) was
quantifies all sources of uncertainty individually, and then com-
bines (propagates) them through a mathematical model.
The implementation of the modeling approach reveals difficulties Analysis A1
in establishing reliable estimates for the input variables of the
model [12]. On the other hand the empirical approach, which is Sample A
described as ‘‘top down’’ approach [15], uses replicated measure-
ments in order to obtain a reliable estimate of the uncertainty, Analysis A2
without necessarily knowing any of the sources individually Sampling
[11–13]. One of the most commonly used empirical methods is target
the duplicate method with a balanced experimental design.
This method involves the formation of duplicate samples from Analysis B1
the sampling targets by applying the same sampling protocol and
duplicate analysis of samples under repeatability conditions. Sample B
Appropriate statistical analysis applied to the resulting data leads
to the estimation of the sampling uncertainty. Analysis B2
The aim of this work is to present and compare three
statistical approaches used for the estimation of the uncertainty Fig. 2. Balanced experimental design employed for the estimation of sampling
caused by manual sampling of fuels from petroleum retail uncertainty.
stations, utilizing the duplicate sampling method. Duplicate
samples of automotive diesel from 104 petroleum retail stations
(10.9% of the petroleum retail stations monitored for fuel quality Table 1
purposes) were analyzed in duplicate for the determination of Instrument parameters used for total sulfur determination in petroleum products.
sulfur content according to ASTM D 5453 [16]. The sulfur mass
Parameter Value
content is one of the most critical parameters associated with
automotive diesel specifications. The results of the measurements Volume injected (mL) 10
of the samples were analyzed using three statistical approaches, Syringe drive rate (mL s  1) 1
classical ANOVA, robust ANOVA and range statistics [8,17] and Furnace temperature (1C) 1080
the sampling uncertainty under each approach was calculated. Furnace oxygen flowmeter setting (mL min  1) 470
Inlet oxygen flowmeter setting (mL min  1) 15
Sampling (and analytical) bias has been assumed to be zero in Inlet carrier (Argon) flowmeter setting (mL min  1) 150
this study.
362 D. Theodorou et al. / Talanta 105 (2013) 360–365

Table 2
Calculation of sampling and analysis uncertainty components by one-way classical ANOVA using data from measurements from a balanced experimental design with n
targets (i¼1,2,y,n), 2 samples (j¼ A, B) from each target and 2 analyses (k¼ 1, 2) of each sample [8,12].

Source of variation

Sampling Analytical

a,b,c    
Differences from the mean values (D) DiðxÞ ¼ X i x iA  ¼ X i x iB  DiAðxÞ ¼ jxiA1 x iA j ¼ jxiA2 x iA j DiBðxÞ ¼ jxiB1 x iB j ¼ jxiB2 x iB j
P 2 P
Sum of squares (SS) SSsampling ¼ 4 DiðxÞ SSanalysis ¼ 2 ðD2iAðxÞ D  2 
iB xÞ
i i
Degrees of freedom (df) 2n–n ¼n 2  2 n  2n ¼2n
Mean square (MS) MSsampling ¼
SSsampling
MSanalysis ¼
SSanalysis
df sampling df analysis
Variances (V) V sampling ¼
MSsampling MSanalysis V analysis ¼ MSanalysis
2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Uncertainty parameter (s) ssampling ¼ V sampling sanalysis ¼ V analysis

a
X i : mean value of target i (two samples–four analyses).
b
x ij : mean value of the 2 analyses of sample j (A or B) of target i.
c
xijk : measured value from target i, sample j and split (analysis) k (1 or 2).

calculated using a calibration curve. Six calibration standards calculated using the following equations:
(VHG Labs, Petrochemical Test Standard, Range set 1) with U measurement ¼ 2smeasurement ð4Þ
concentrations of 0, 1.00, 2.50, 5.01, 7.50 and 10.0 ng mL  1 were
analyzed in triplicate. The instrument responses were recorded U sampling ¼ 2ssampling ð5Þ
and a calibration curve with 18 points was constructed. The
samples and the calibration standards were injected using a 10 mL U analysis ¼ 2sanalysis ð6Þ
syringe. Using density measurements the results were converted
into milligram per kilogam (mg kg  1; ppm m/m). The majority of The expanded uncertainties can also be expressed relative to
the automotive diesel sample measurement results were found to the reported value x (as a percentage), as relative expanded
be under or close to 10 mg kg  1 which is the EU regulatory limit uncertainties U(%):
for sulfur mass content. 2smeasurement
U measurement ð%Þ ¼ 100 % ð7Þ
x

2ssampling
U sampling ð%Þ ¼ 100 % ð8Þ
4. Data analysis methods x

4.1. Statistical model of the empirical approach and estimation 2sanalysis

U analysis ð%Þ ¼ 100 % ð9Þ
of uncertainties x

The statistical model describing the relationship between the 4.2. Estimation of uncertainty using classical analysis of variance
measured and the true values of analyte concentration based on a
single measurement, x, has the following form [11]: Analysis of variance (ANOVA) is a statistic technique by which
x ¼ X true þ esampling þ eanalysis ð1Þ variations associated with different sources can be isolated and
estimated [22]. The simplest type of ANOVA is the one-way
where Xtrue, is the true average concentration of the analyte at the classical ANOVA, which deals with one independent variable
sampling target, esampling, is the total error due to sampling, and one dependent variable. Classical ANOVA may be applied to
associated with a variance of s2sampling and eanalysis, is the total the data produced by the implementation of the balanced
analytical error, associated with a variance of s2analysis. If the experimental design in order to estimate the sampling uncer-
sources of variation are independent [21], then the total variance tainty. Classical ANOVA estimations are based on the differences
of the measurement, s2measurement, for a single sampling target is from the mean values, not on the range as in the approach of the
given by: range statistics. Table 2 presents the one way classical ANOVA
calculations used for the estimation of sampling and analytical
s2measurement ¼ s2sampling þ s2analysis ð2Þ uncertainty.

If statistical estimates, s2, are used to approximate variances,

4.3. Estimation of uncertainty using robust analysis of variance
s2, then Eq. (2) becomes
s2measurement ¼ s2sampling þ s2analysis ð3Þ The term ‘‘robustness’’ in the statistics is used to describe
methods designed to be insensitive to distributional assumptions
The components of measurement variation, the sampling and (such as normality) and tolerate a certain amount of unusual
the analytical variance, which represent the corresponding uncer- observations (outliers). Robust statistics are characterized by the
tainties, can be separated and estimated using appropriate accommodation rather than the rejection of outlying values
statistical methods, such as the statistical approaches presented within a certain dataset, avoiding the risk—often associated with
in this work. classical statistics—of skewing the statistics [23,13]. The valid
In order to obtain a coverage interval (expanded uncertainty) application of classical ANOVA is based on three assumptions
corresponding to an approximately 95% coverage probability, summarized by Ramsey [24] as: (i) the variances should be
standard deviations, s, are multiplied by a coverage factor of independent, (ii) each level of variance should be homogeneous,
two. The expanded uncertainties for the measurement, sampling not varying systematically within one level and (iii) the distribution
and analysis, Umeasurement, Usampling and Uanalysis, respectively, are of errors within each level of variance should be approximately
 D. Theodorou et al. / Talanta 105 (2013) 360–365 363

Gaussian. The application of robust ANOVA which utilizes robust 5. Results and discussion
statistics has been shown to be particularly appropriate for provid-
ing estimated of variances, in cases where the validity of classical The main objective of this work was to estimate the uncertainty
ANOVA is doubtful [11]. Robust ANOVA uses robust estimates of the components resulting from sampling of fuels from petroleum retail
mean and standard deviation which are calculated by an iterative stations. The results of the evaluation of the measurement results
process [23,24]. In this process, also known as Huber’s method, (104 duplicate sample analyzed twice) using three statistical meth-
extreme values that exceed a certain distance (product of a constant odologies are presented in Table 4 and Fig. 3.
c and the standard deviation) from the sample mean are down- The expanded uncertainty of sampling is in the range of 0.34–
weighted or brought in. Actually these data are assigned a new value 0.40 mg kg  1, while relative expanded uncertainty lies in the
equal to that distance. A value of c¼1.5 is widely accepted as range of 4.8–5.1%, depending on the statistical methodology used.
optimal for datasets containing a small proportion of outliers (upto The differences between the results of the 3 methodologies
10% of outliers). Initial values for sample mean and standard concerning sampling and analysis variances were evaluated using
deviation estimates can be obtained by classical or robust statistics F-test (with 208 and 104 degrees of freedom for the variances of
(e.g. median, mean absolute deviation). After the population analysis and sampling, respectively). Table 5 presents the results
has been modified, new sample mean and standard deviation are of the F-tests performed. F-test results should be treated with
estimated. This process is repeated until the estimated values caution as the test is used ‘‘out-of-its-scope’’ and possibly some of
converge to the so called robust estimates. its assumptions (e.g. normality, independence) are violated.
In the present work, robust ANOVA was implemented using a Nevertheless, it may provide a quick and gross evaluation of the
specifically written computer program called, Roban.exe, devel- significance of the differences of the estimated variances. The
oped from the Analytical Methods Committee (AMC) in Great differences are statistically significant for all cases compared,
Britain [25]. with the exception of the sampling variances estimated by
range statistics and classical ANOVA. Classical ANOVA and range
statistics are typical tools of classical statistics that are strongly

4.4. Estimation of uncertainty using range statistics Table 4

Results calculated using range statistics, classical ANOVA and Robust ANOVA. All
Range statistics may also be used for calculating standard uncertainties U are estimated using a coverage of 2 which corresponds to an
deviations by treating data produced by a balanced experimental approximately 95% coverage probability.

design [8,26]. Range statistics, like classical ANOVA assume

Range Classical Robust
normal distribution and the calculations are done from the Statistics ANOVA ANOVA
differences between duplicate measurements. The relation
between standard deviation and differences requires the applica- Mean (mg kg  1) 7.988 7.988 7.079
sanalysis (mg kg  1) 0.205 0.378 0.265
tion of a certain factor depending on the replication chosen, e.g.
ssampling (mg kg  1) 0.202 0.200 0.169
1.128 for duplicate, 1.693 for triplicate etc. [27]. Actually the smeasurement (mg kg  1) 0.288 0.427 0.314
variance of sampling is calculated indirectly as the difference of Uanalysis (mg kg  1) 0.411 0.755 0.529
the variances of measurement and analysis. Table 3 presents the Uanalysis (%) 5.1 9.5 7.5
range statistics calculations used for the estimation of sampling Usampling (mg kg  1) 0.404 0.401 0.337
Usampling (%) 5.1 5.0 4.8
and analytical uncertainty.
Umeasurement (mg kg  1) 0.576 0.855 0.628
Umeasurement (%) 7.2 10.7 8.9
Table 3 Analysis uncertainty 51 78 71
Calculation of sampling and analysis uncertainty components by range statistics contribution (%)
using data from measurements from measurements from a balanced experimental Sampling uncertainty 49 22 29
design with n targets (i¼ 1,2,y,n), 2 samples (j ¼A, B) from each target and contribution (%)
2 analyses (k¼1, 2) of each sample [8].

Parameter Equation of calculation

 
Differences of duplicates of DiA ¼ xiA1 xiA2 
sample A (DiA)a
 
Differences of duplicates of DiB ¼ xiB1 xiB2 
sample B (DiB)a
 
Differences of the means of the Di ¼ x iA x iB 
two measurements (Di)b
P
Mean range of measurement Di

(Dmeasurement) Dmeasurement ¼ i n
P P !
Mean range of analysis (Danalysis) DiA DiB
1
Danalysis ¼ 2
i
n þ i
n

Standard deviation of analysis Danalysis

sanalysis ¼ 1:128
(sanalysis)
Standard deviation of Dmeasurement
smeasurement ¼ 1:128
measurement (smeasurement)
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2ffi
Standard deviation of sampling s
ssampling ¼ s2measurement  analysis pffiffi The
(ssampling) 2
sanalysis is divided by a square root of
2 because the result of the analysis of each
sample is the mean of two
measurements—standard error of the
mean. Fig. 3. Graphical comparison of the relative uncertainty estimates obtained by
range statistics, classical ANOVA and robust ANOVA. The relative uncertainties are
a
xijk : measured value from target i, sample j and split (analysis) k (1 or 2). estimated using a coverage of 2 which correspond to an approximately 95%
b
x ij : mean value of the 2 analyses of sample j (A or B) of target i coverage probability.
364 D. Theodorou et al. / Talanta 105 (2013) 360–365

Table 5 samples leads to uncertainty estimates that are often fit for

Results of the F-test used for the comparison of the results of the three statistical purpose [12].
methodologies.
For the purposes of this work sampling bias has not been
Methodologies F-statistic for analysis F-statistic for sampling considered. It is widely accepted that it is difficult to establish the
compared variances (probability) variances (probability) presence of bias in a sampling protocol because it is difficult to
define a reference sampling method or a reference sampling
Range statistics– 3.38n (o 0.001) 1.01 (0.470) target. This often leads to the characterization of the sampling
ANOVA
Classical ANOVA– n
2.04 (0.001) 1.41n (0.039)
methods as ‘‘empirical’’ (by analogy to empirical analytical
robust ANOVA method). This means that the results are dependent on the
Range statistics– 1.66n (0.001) 1.43n (0.034) method (and are usually reported with reference to the method),
robust ANOVA which by definition is associated with a zero bias [7]. It has to be
n noted though that there are methods available that may be used
Indicating significant difference (probabilityo 0.05).
for the estimation and the inclusion of analytical or sampling bias
in the uncertainty estimates [11,30].

6. Conclusions

Three alternative statistical approaches for data analysis con-

cerning the estimation of measurement uncertainty of manual
sampling of fuel were described and compared in this work. A
balanced experimental design was used, which included duplicate
samples of automotive diesel from 104 petroleum retail stations
(10.9% of the petroleum retail stations monitored for fuel quality
purposes) and duplicate analyses of these samples for sulfur mass
content determination. The sulfur mass content is a critical fuel
quality parameter associated with automotive diesel specifica-
tions and is often used for identifying fuel cross contamination or
fuel adulteration incidents. The results were treated using classi-
cal ANOVA, robust ANOVA and range statistics. The three meth-
odologies gave statistically different estimates with the expanded
uncertainty of sampling being in the range of 0.34–0.40 mg kg  1,
Fig. 4. Box Plot of the average values of the 104 datasets (produced by PASW 18 while the relative expanded uncertainty lying in the range
[28]). The box length represents the interquantile range. Values which are more of 4.8–5.1%. The fact that the robust ANOVA leads to different
than three box lengths from either end of the box are extreme values are denoted results compared to the other two methodologies is an indication
by an asterisk. Values which are between one and a half and three box lengths
that the assumptions of classical ANOVA and range statistics are
from either end of the box are extreme values and are denoted by a circle.
not justified. This fact is also confirmed by the presence of a small
but unignorable number of outliers (8.7%) within the data used
affected by the presence of outlying values. Indeed, as shown in for the calculations. Therefore, robust ANOVA, which is not
the Box Plot diagram (produced by PASW 18 [28]) of Fig. 4 created influenced by less than 10% outliers, is considered as the method
using the averages of the 104 datasets used for the calculations, providing the most reliable estimates for the sampling expanded
there exist seven outliers and extreme values (between target uncertainty (0.34 mg kg  1 or 4.8% in relative terms). The results
outliers). These seven values, were also confirmed as outliers by of robust ANOVA show that the analytical measurement uncer-
applying the Grubbs test. Moreover, three datasets were identi- tainty accounts for the 71%, while the sampling measurement
fied as analytical or sampling outlier by applying the Cochran test. uncertainty accounts only for the 29% of the total measurement
In total 9 out of 104 datasets (8.7%) were flagged by the outlier tests uncertainty. Thus, minimizing or reducing repeatability errors
(one dataset was flagged by both Cochran and Grubbs test). There- may lead to substantial reduction of the total measurement
fore, the results of robust ANOVA, which is insensitive to small uncertainty. Finally, it has to be noted that besides providing
number of outliers (less than 10%) can be considered as more realistic estimates of uncertainty which allow end users make
reliable than the results of range statistics and classical ANOVA. more informed decisions, the estimates of sampling and analy-
It is obvious from the results of robust ANOVA that the tical variance may be also used for establishing an internal quality
measurement uncertainty is dominated by the analytical variance. control method that will monitor the whole measurement pro-
In fact the analysis uncertainty accounts for the 71% of the cedure, including sampling, ensuring that it remains in statistical
measurement uncertainty. This leaves ‘‘room’’ for an effective control.
reduction of the measurement uncertainty. Uncertainty reduction
may be accomplished by making more measurements and calculat- References
ing their average, instead of making a single measurement. Then the
standard deviation of the mean gets smaller as the number of data [1] K. Heydorn, K. Esbensen, Accred. Qual. Assur. 9 (2004) 391–396.
increases leading to smaller random error uncertainty contributions. [2] ASTM Standard D 4057-06, Standard Practice for Manual Sampling
It has to be noted that multiple samplings and analyses of Petroleum and Petroleum Products, ASTM International, West Consho-
hocken, PA, 2011.
required for collecting data used in uncertainty calculations are
[3] J.B. Hansen, C. Grøn, U. Lund, B. Magnusson, A. Nordbotten, A. Oberender,
often associated with unreasonably high costs. These costs cost M.H. Ramsey, Accred. Qual. Assur. 12 (2007) 377–381.
may be reduced, if the classical ANOVA and robust ANOVA are [4] K.A. Boon, P. Rostron, M.H. Ramsey, Geostand. Geoanal. Res. 35 (2011)
implemented using results from an unbalanced design [29,11], 353–367.
[5] S.L.R. Ellison, A. Willims (Eds.), EURACHEM/CITAC Guide, Use of Uncertainty
which is more economical as it requires 33% less analyses. Information in Compliance Assessment, 2007 (Available from the Eurachem
Moreover, it has been shown that a minimum of eight duplicate secretariat at /http://www.eurachem.orgS).
 D. Theodorou et al. / Talanta 105 (2013) 360–365 365

[6] M.H. Ramsey, A. Argyraki, Sci.Total Environ. 198 (1997) 243–257. [19] ISO/IEC 17025, General Requirements for the Competence of Testing and
[7] M. Thompson, Accred. Qual. Assur. 3 (1998) 117–121. Calibration Laboratories, International Organization for Standardization,
[8] C. Grøn, J.B. Hansen, B. Magnusson, A. Nordbotten, M. Krysell, K.J. Andersen, Geneva, Switzerland, 2005.
U. Lund, N.T. Technical Report 604, Uncertainty from Sampling. A Nordtest [20] ISO 20846, Petroleum products—Determination of Sulfur Content of Auto-
Handbook for Sampling Planners on Sampling Quality Assurance and Uncer- motive Fuels—Ultraviolet Fluorescence Method International Organization
tainty Estimation, Nordisk Innovation Centre, Oslo, Norway, 2007. for Standardization, Geneva, Switzerland, 2011.
[9] M.H. Ramsey, M. Thompson, Accred. Qual. Assur. 12 (2007) 503–513. [21] C.E. Efstathiou, Talanta 52 (2000) 711–715.
[10] M.H. Ramsey, Accred. Qual. Assur. 9 (2004) 727–728. [22] S.L.R. Ellison, V.J. Barwick, T.J.D. Farrant, Practical Statistics for the Analytical
[11] M.H. Ramsey, S.L.R. Ellison (Eds.), Eurachem/EUROLAB/CITAC/Nordtest/AMC Scientist: A Bench Guide, second ed., The Royal Society of Chemistry,
Guide, Measurement Uncertainty Arising From Sampling: A Guide to Meth- Cambridge, UK, 2009.
ods and Approaches, 2007 (Available from the Eurachem secretariat at [23] Royal Society of Chemistry—Analytical Methods Committee, Analyst 114
/http://www.eurachem.orgS).
(1989) 1693–1697.
[12] J.A. Lyn, M.H. Ramsey, A.P. Damant, R. Wood, Analyst 132 (2007) 1231–1237.
[24] M.H. Ramsey, J. Geochem. Explor. 44 (1992) 23–36.
[13] M.H. Ramsey, J. Anal. Atom. Spectrom. 13 (1998) 97–104.
[25] Roban Royal Society of Chemistry—Analytical Methods Committee, /http://
[14] ISO/IEC Guide 98-3, Uncertainty of Measurement-Part 3: Guide to the
www.rsc.org/Membership/Networking/InterestGroups/Analytical/AMC/Soft
Expression of Uncertainty in Measurement (GUM:1995), International Orga-
ware/ROBAN.aspS (accessed 02.05.12).
nization for Standardization, Geneva, Switzerland, 2008.
[26] ISO 3085, Iron ores—Experimental Methods for Checking the Precision of
[15] J. Galbán, C. Ubide, Talanta 71 (2007) 1339–1344.
[16] ASTM Standard D 5453-09, Standard Test Method for Determination of Total Sampling, Sample Preparation and Measurement, International Organization
Sulfur in Light Hydrocarbons, Spark Ignition Engine Fuel, Diesel Engine Fuel, for Standardization, Geneva, Switzerland, 2002.
and Engine Oil by Ultraviolet Fluorescence, ASTM International, West Con- [27] H. Hovind, B. Magnusson, M. Krysell, U. Lund, I. Mäkinen, NT Technical Report
shohocken, PA, 2009. 569, third ed., Internal Quality Controll—Handbook for Chemical Labora-
[17] E.V. Reiter, M.F. Dutton, A. Agus, E. Nordkvist, M.F. Mwanza, P.B. Njobeh, tories, Nordisk Innovation Centre, Oslo, Norway, 2007.
D. Prawano, P. Häggblom, E. Razzazi-Fazeli, J. Zentek, M.G. Andersson, [28] PASW Statistics 18, Release Version 18.0.0, SPSS, Inc., Chicago, IL, 2009,
Analyst 136 (2011) 4059–4069. /www.spss.comS.
[18] ISO 9001, Quality Management Systems—Requirements, International Orga- [29] P.D. Rostron, M.H. Ramsey, Accred. Qual. Assur. 17 (2012) 7–14.
nization for Standardization, Geneva, Switzerland, 2008. [30] M.H. Ramsey, B. Geelhoed, A.P. Damant, Analyst 136 (2011) 1313–1321.