Another top down MU method – ISO 11352 made simple

1.0 Introduction

There are several methods for estimating measurement uncertainty (MU) by

the holistic top down approaches, which study MU from the viewpoint of the
whole test method performance. We have discussed the use of ISO
21748:2010: Guidance for the use of repeatability, reproducibility and
trueness estimates in measurement uncertainty estimation which employs
the reproducibility of the test method from proficiency testing (PT)
program(s) as one of the main uncertainty components in the combined
standard uncertainty after confirming the method repeatability meets the
criteria set and the method bias (trueness) has been evaluated to be
satisfactory.

2.0 The ISO 11352:2012

An alternative method ISO 11352:2012: Water quality – Estimation of

measurement uncertainty based on validation and quality control data
recommends another avenue in estimating measurement uncertainty in
chemical analysis, by statistically taking a combination of precision estimate
(such as intermediate precision), and the method and laboratory bias into
one uncertainty measure. This approach has removed the difficulty in
conducting suitable inter-laboratory comparison (ILC) studies or finding a
proficiency testing (PT) program provider for a newly developed test method.
Instead, the data from own method validation and analytical quality control
protocols are utilized here.

Many terms have been used to describe the precision data through repeated
analysis on a stable laboratory control sample (LCS) over a period of time by
different analysts and/or different equipment and reagents within a
laboratory. They are known as intermediate precision, intermediate
reproducibility, within-laboratory precision, within-laboratory
reproducibility, etc.

This international standard advocates that if the measurement result

originates from a controlled analytical process, it is then not necessary to
estimate the MU of each individual measurement result. That means the
estimation of MU should apply to all of the measurement results under
controlled conditions with a quality assurance program, independently of,
say, sample matrix and analyst.

1
It further suggests that if the MU varies significantly, depending on sample
matrix and/or concentration range, the uncertainty estimation shall be made
separately for each matrix and/or concentration range.

Also, the analytical data obtained must be completely random and fall within
the normal probability distribution. The Anderson Darling statistic or other
suitable tests for data normality is used to confirm that this important
assumption is valid.

3.0 The uncertainty components to be considered

In short, for this international standard, we only need to consider two main
aspects of uncertainty contributors for a specific analytical method under
same controlled conditions as used when a routine analysis is carried out,
coupled with a robust quality assurance program, namely:

- intermediate precision estimate from validation process

- method and laboratory bias data estimate

3.1 Intermediate precision (within-lab reproducibility) standard

uncertainty, uR’

a) Use of quality control samples covering the whole analytical process

When the matrix and concentration range of a laboratory control sample

(LCS) are similar to those of the routine samples for analysis, the
intermediate precision standard deviation sR’ is equal to the intermediate
precision standard uncertainty uR’, i.e.

u R ' = sR ' [1]

b) Using synthetic standard solutions as quality control samples

If LCS with an identical matrix to test samples are not available, a laboratory
prepared synthetic standard solution can be used, which may have a matrix
which differs from that of routine samples.

Apart from subjecting this standard solution to the whole analytical process
to obtain the intermediate precision standard uncertainty, we need to
consider an additional uncertainty component due to possible increase
inhomogeneity of the analyte in the matrix.

This additional uncertainty as a within-laboratory repeatability component,

uRange can be estimated from a series of duplicate analyses of the actual test
samples over a range of concentrations.

2
For example, if, in a hypothetical experiment, we have a duplicate set of
analytical data 10.8mg/L, and 11.3mg/L, the difference between the
duplicate, D = 11.3 – 10.8 or 0.5mg/L and the mean result = 11.05mg/L.

The relative range standard uncertainty expressed as % relative standard

deviation, Rrel therefore equals to (0.5 x 100)/11.05 or 4.52.

Table 1 below shows 10 sets of duplicates on 10 test samples with different

concentrations, and their respective relative range standard deviations, Ri,rel.
__
The last column calculated the mean R rel of the 10 range values obtained.

Table 1: Evaluation of 10 duplicated analyses on different test samples

of varied concentrations in mg/L.

1 2 3 4 5 6 7 8 9 10

10.8 15.6 7.5 12.1 8.9 9.7 14.2 10.3 10.8 15.6

11.3 15.2 7.7 11.6 9.4 9.3 13.8 10.7 11.1 15.2

4.52 2.6 2.63 4.22 5.46 4.21 2.86 3.81 2.74 2.6 3.57

Then, the following equation [2] is used to estimate the intermediate precision
based on the repeatability uncertainty of the synthetic standard solution and
the uncertainty of testing different test samples with variable concentrations:

u R ' = u 2 R '( s tan d ) + u 2 r ( range ) [2]

where,

u R '( s tan d ) is the standard uncertainty of the synthetic standard solution

ur (range ) is the standard uncertainty of the range control which is:

__
__
R rel where R
u r ( range) = rel is the mean range of the duplicated analyses.
1.128

c) Unstable control sample

Where stable control samples are not available, the ISO standard suggests to
consider two uncertainty components:

3
 - the repeatability from the calculation of the mean of the ranges of
duplicate analyses, i.e. ur(range)
- the variation resulted from the means of different batches of analyses,
i.e. uR’(batch); in many cases, this component relies on scientific judgment
based on the analyst’s experience

The equation [3] is then used for the within-lab reproducibility, uR’ :

u R ' = u 2 r ( range ) + u 2 R '( batch ) [3]

3.2 Method and laboratory bias, ub

Generally speaking, sources for biased results should be investigated and

eliminated during the method validation or verification process. But an
observed bias does often exist for different matrices and different
concentrations.

To evaluate the uncertainty associated with method and laboratory bias, two
components are to be estimated:

i. the test result bias itself, expressed as the difference between the test
result and the nominal, certified or accepted reference value;
ii. the uncertainty of the nominal or certified reference value

Note 1: the bias uncertainty component can be neglected if ub < uR’/3.

a) Analysis of suitable reference materials

The following equations [4] and [5] are used for uncertainty component
associated with method and laboratory bias:

i. For nr number of certified reference materials (CRM’s) studied:

2

ub =
∑b i
__ 2
+u Cref [4]
nr
ii. If only one reference material is available, the repeated results of
analyses of this reference material are treated as the best available
estimate for the measurement uncertainty component associated with
method and laboratory bias, ub.

4
 So, when only one CRM is used, the bias uncertainty component is:
__ 2 2
s2
ub = b + b + u Cref [5]
nM
where ：
bi is the ith bias which is the difference between the mean measured
value and the accepted reference value (ARV) of the ith reference
material
b is the difference between the mean measured value and the
accepted reference value of the single CRM
nr is the number of CRM’s
nM is the number of bias measurements on the reference material
__
u Cref is the mean value of all uCref, which are the estimated standard
uncertainties of the CRM’s given
sb is the standard deviation of the measured values of the reference
material

b. The use of PT or inter-laboratory comparison (ILC) results

Results from PT programs or inter-laboratory comparisons may be used

in the same way from analysis of reference materials, if the assigned value
in the inter-laboratory comparison is a sufficiently good estimate of the true
value. An important condition to note is that the laboratory’s participation
in such studies must be satisfactory (i.e. not being labelled as outlier with z-
score amongst the participants being less than 2 for 95% confidence).
Alternatively, the Mandel’s h & k statistic test can be applied to confirm this
assumption.

The differences, Di between the test results and the assigned values of the
different samples are calculated and squared before making an estimate of
root mean square of the differences, Drms:
2

Drms =
∑D i
[6]
nilc
where,
nilc is the number of inter-laboratory comparison samples analyzed

Note 2: If the individual differences and the uncertainties of the assigned

values vary significantly, it may be necessary to separately estimate
uncertainties for the different cases.

5
The mean uncertainty of the assigned values of the inter-laboratory
comparison samples can be calculated as (a) median or robust mean, or (b)
arithmetic mean from the results of the participating laboratories (i.e.
__
consensus value), u Cref as follows:

__
u Cref =
∑u Cref ,i
[7]
nilc
where,
if the median or robust mean is used as consensus value:
sR ,i
uCref ,i = 1.25 × [8]
Li

or,
if the arithmetic mean is used as consensus value:
sR ,i
uCref ,i = [9]
Li
where,
uCref,i is the uncertainty of the assigned value of the inter-
laboratory sample i;
sR,i is reproducibility standard deviation from the inter-
laboratory comparison for sample i;
Li is the number of participating laboratories for sample i

Finally, the standard uncertainty component associated with method and

laboratory bias, ub, is calculated as in equation [10]:
__
ub = D 2 rms + u Cref [10]
Where,
Drms is the root mean square of the differences
__
u Cref is the mean uncertainty of the assigned values of the inter-
laboratory comparison or PT samples

c. Making use of the discovery experiments

A recovery experiment checks for the recovery of a known amount of analyte

added to a previously analyzed sample. Such experiment can also be used to
evaluate bias. To be statistical relevance, this recovery experiment should be
performed with at least six different samples of the relevant matrix.

6
In this case, the uncertainty components associated with method and method
bias, ub are:

- the difference between observed and known concentration of the

analyte
- the uncertainty in the concentration of the analyte added to the test
sample

The standard uncertainty associated with method and laboratory bias, ub,
estimated from recovery experiments is:

ub = b 2 rms + u 2 add [11]

where
uadd is the uncertainty in the concentration of the analyte added,
which should cover two uncertainty components, i.e. the volume
made up and the amount of analyte added;

brms is the root mean square of the deviations from the recovery
experiments, which is calculated from the equation below:

brms =
∑b i
[12]
nRe c
where
bi is the deviation from the 100% recovery of the ith
recovery experiment or from the mean recovery, if
the results are corrected with this mean recovery;
nRec is the number of recovery experiments

4.0 Calculation of the combined standard uncertainty and expanded

uncertainty

Under 95% confidence limit, the combined standard uncertainty, uc, for uR’ and
ub, and expanded uncertainty, U, can be calculated by the following equations:

U = 2uc = 2 u 2 R ' + u 2 b [13]

U rel = 2uc ,rel = 2 u 2 R ', rel + u 2b, rel [14]

where,
uR’,rel and ub,rel are the relative standard deviation of uR’ and ub,
respectively

7
5.0 Conclusion

It is obvious that the holistic top down approach in MU evaluation is relatively

simpler than using the GUM (so called bottom up) method which considers
all uncertainty components in each of the analysis steps of the test method.
This top down approach makes use of the readily available method validation
data and outcomes from PT or ILC studies participated. Another advantage is
that the MU estimate by this manner is dynamic and current as the quality
control data are being updated regularly.