1

A PROTOCOL FOR THE VALIDATION OF QUALITATIVE METHODS OF

ANALYSIS


Roy Macarthur
Food and Environment Research Agency, Risk and Numerical Sciences Team, Sand
Hutton, YO41 1LZ York, United Kingdom

Christoph von Holst
European Commission, Joint Research Centre, Institute for Reference Materials and
Measurements (EC-JRC-IRMM), Retieseweg 111, B-2440 Geel, Belgium


2

Abstract
This paper presents a draft protocol for analyzing the results of validation studies for
qualitative methods of analysis which is designed to meet three competing goals: 1)
to give correct answers; 2) have a broad scope of application; and 3) be accessible
to a wide range of users.
The draft protocol can be applied to the validation of methods by collaborative trial or
to single-laboratory studies. The protocol produces an estimate of the probability of a
positive response with a confidence interval within which 95% of laboratories (or
analytical runs) are expected to fall when the method is applied in practice. The
interval is calculated from the observed reproducibility (or within-laboratory
reproducibility) standard deviation. Then a simple plot of confidence intervals for the
probability of detection against the concentration of analyte is used to provide an
estimate of the range of limits of detection and false positive probability that we can
expect to see when the validated method is applied in practice. Use of the draft
protocol is demonstrated using results produced by three collaborative trials. A
simulation study showed that a conclusion that a method is fit for purpose that is
generated by the draft protocol is likely to be safe.

3

Introduction
A qualitative analytical method is one whose response can be interpreted as
detected or not detected, this may be because a response is inherently
qualitative, or where the response is quantitative it is uncorrelated (or very weakly
correlated with a short range of responses) with the quantity of analyte present in a
sample. Methods that deliver qualitative responses are commonly applied in many
analytical areas. For example rapid test methods, including qualitative tests are used
for the detection of microbiological contaminants, heavy metals, pesticides, foreign
bodies, mycotoxins, allergens and other analytes in food [1]. Their use, and in
particular their use by producers and consumers to test food and other products
against legislative of contractual limits, or by competent authorities for sanitary or
phytosanitray purposes, means that an internationally agreed guideline for validating
the performance of methods is needed. For quantitative methods the
IUPAC/AOAC/ISO Harmonized protocol [2] describes how to run validation studies,
and how to analyze results to describe important aspects of method performance,
although further work to clarify the relation between concentration and performance
may be required to validate methods against the specific requirements of the
decision-makers who use analytical results [3, 4, 5]. No internationally accepted
guidance on the validation of qualitative methods exists.
In 2006 IUPAC initiated a project to develop guidance for the validation of qualitative
methods (IUPAC project 2005-024-2-600 Establishment of guidelines for the
validation of qualitative and semi quantitative (screening) methods by collaborative
trial: a harmonized protocol) chaired by the European Commissions Institute for
Reference Materials and Measurements. The project became more active from 2008
onwards with the involvement of more experts and the use of real-world validation
studies via co-operation with European projects: (1) Safeedpap (SAFEED-PAP
(FOOD-CT-2006-036221, European Commission), on the Detection of presence of
species-specific processed animal proteins in animal feed
http://safeedpap.feedsafety.org/) and (2) MoniQA (Monitoring and Quality Assurance
in the Food Supply Chain, Food-CT-2006-036337, European Commission,
http://www.moniqa.org/). This paper describes a draft protocol for the design of
validation studies and the analysis of results to produce a description of method
performance that directly meet the needs of users and decision-makers.

4

The aim of method validation for qualitative testing
When we use a qualitative method of analysis and we get a positive result we want
to be able to say with confidence that the analyte is present in the sample that we
tested. If we get a negative result then we want to be able to say with confidence that
if the analyte is present at all in the sample we tested then it must be at a
concentration below a specified low level. This is the purpose of qualitative testing.
The goal of validation of an analytical method is to provide assurance that a method
is likely to produce results that are fit for purpose. For qualitative methods two
quantities describe fitness for the purpose described above: 1) the probability that a
sample in which the analyte is absent will generate a negative response, and 2) the
lowest concentration for which the probability of a positive response is sufficiently
high (alternatively, the probability of a positive response for a given target
concentration of analyte). It is worth highlighting a few points about method
validation here:
1) A method validation study provides a prediction about the performance we
can expect to see when a method is applied in practice.
2) Analytical methods with the same probability of detection under the same
conditions are equivalent. For qualitative methods variation in performance
means a difference in the probability of detection.
3) To assess a qualitative methods fitness for purpose as defined above, the
most generally useful description of the performance is the relationship
between the probability of detection and quantity of analyte.
4) Hence, the goal of method validation can be achieved by producing an
estimate of the probability of detection across a range of quantities of analyte
with a interval that gives a prediction interval within which the probability is
likely to lie when a method is applied in a new laboratory (for multi-laboratory
studies), on a new day or with a new matrix etc. (for single-laboratory studies).
Traditional and standard statistical approaches to validation
A traditional approach to assessing the performance of a qualitative methods is
based on the analysis of samples which are known to be positive and samples
which are known to be negative and simply reporting the proportion of positive
samples that give positive responses, called the sensitivity of the method (or its

5

inverse the false negative rate), and the proportion of negative samples which give a
negative response, called the specificity of the method (or its inverse the false
positive rate). This approach may be most useful where the target of the analysis
itself is intrinsically qualitative such as disease status or pregnancy. However it does
not meet our needs for validation of a qualitative method of analysis because while
there is no such thing as a little bit pregnant a sample may contain a little bit of a
substance or it may contain a larger quantity of a substance. Hence, the
performance of qualitative methods used to detect the presence of a substance will
depend upon, and should be considered with, the quantity of analyte in samples.
A standard statistical approach to examining the relationship between a quantity and
a qualitative response (often a biological response) whose probability of occurrence
depends on the quantity is through a regression using a generalized linear model
(GLM), often using a logit or probit link function [6]. There are two significant
challenges for non-statisticians who try to use this approach for the validation of
qualitative methods of analysis. The first challenge is that in general we do not know
the shape of the relationship between the probability of detection and the quantity of
an analyte. This means that the standard statistical approach requires some
exploratory non-linear modeling to identify the right (or good enough) transformation
and link function [6] before it can be applied to a particular set of results. The second
challenge, which is also a challenge for statisticians, is that it is difficult to get to a
prediction interval using these methods. In practice this means that we can use the
standard statistical methods to get an estimate of the relation between the expected
average probability of detection and quantity of analyte, but it is difficult to use them
to get at the range of probabilities of detection that we might see when new
laboratories use a method that has been validated in a collaborative trial, or when we
use a qualitative method tomorrow that was validated in-house today.
Recent approaches to validation
Langton et al introduced the idea of using accordance (the proportion of pairs of
samples give the same response when analyzed under repeatability conditions) and
concordance (the proportion of pairs of samples that give the same response when
analyzed under reproducibility conditions) as measures of variation in qualitative
response [7]. A particular strength of this approach is that it gives an intuitively
graspable feel for how conclusive a single test result is, or whether undertaking more

6

tests might yield more information. The approach was modified by van der Voet and
van Raamsdonk [8].
Wilrich and Wilrich have produced a method which is an example of the GLM-based
standard statistical approach [9]. It is designed to be specifically applicable to
microbiological qualitative tests in which the two sources of variation are the
presence or absence of colony forming units in a test aliquot and variation in the
probability of detection if organism is present which may be related to, for example,
the nature of the matrix. This method allows a relation between the average
probability of detection for each matrix and quantity of colony forming units to be
easily calculated. The size of uncertainty about the average probability of detection
for each matrix is also estimated.
Wilrich has also produced a method for estimating reproducibility and repeatability
standard deviations which treats qualitative results as quantitative results with value
0 or 1 [10]. An analysis of variance, similar to that used to produce estimates of
reproducibility and repeatability for quantitative analytical results (based on ISO
5725), is then applied to allow three quantities to be estimated: 1) the average
probability of detection, 2) the average repeatability standard deviation of responses
and 3) the between-laboratory standard deviation of responses. Methods for
detecting significant between-laboratory variation and producing an approximate
prediction interval (based on a normal approximation) for the performance of new
laboratories are given. Wilrich remarks that both the observed average probability of
detection and estimated average repeatability standard deviation are artefacts,
because no individual laboratory actually has average performance; the repeatability
standard deviation for each laboratory is a function of the probability of detection for
that laboratory and the number of replicate analyses used to estimate it and that, in
contrast with quantitative analyses, it is difficult for most users to understand what
values of repeatability and reproducibility standard deviation really mean.
Wilrich suggests that it may be fruitful to consider the response of qualitative
methods in a more explicitly quantitative manner as the proportion of a number (he
suggests a minimum of 10) of replicate analyses which give a positive result, i.e. that
we should understand the response of qualitative methods as a single observed
probability of detection rather than a set of replicate qualitative individual positive or
negative results. If we use this approach repeatability variation is expressed as a

7

quantity which describes uncertainty about our single estimate of the probability of
detection for an individual laboratory and reproducibility standard deviation is
estimated by the observed standard deviation across laboratories which includes
contributions from between-laboratory variation and the uncertainty associated with
the estimated probability of detection in each laboratory.
Wehling et. al. [11] produced a protocol for the validation of qualitative methods
which also includes the application of the ISO 5725 approach to qualitative results,
coded as the quantitative results 1 and 0, to get estimates of average repeatability
standard deviation and reproducibility standard deviation. Methods for producing
confidence intervals for the average probability of detection are based on a modified
Wilson interval (introduced by LaBudde [12]) for single-laboratory data and high and
low probabilities of detection for multi-laboratory data, and a normal approximation
for multi-laboratory data where the estimated probability of detection is closer to
50%. There is a strong focus on the use of the probability of detection to characterize
method performance in place of the traditional measures of qualitative performance,
and the use of graphs of the probability of detection against the quantity of analyte to
aid the interpretation of results.
Some points that we take from our reading of this recent work are:
1) Confirmation that the relation between probability of detection and quantity of
analyte is the most generally useful description of qualitative method
performance.
2) While the calculation of estimates for repeatability and reproducibility standard
deviations is as simple for individual qualitative results as it is for quantitative
results, the use of repeatability standard deviation as a measure of method
variation is problematic and the interpretation of what a particular value of
reproducibility standard deviation means is not as intuitive as it is for
quantitative results.
3) There is a lack of harmony about how the effect of between-laboratory
variation should be described. One approach estimates the uncertainty about
the global average probability of detection; another approach estimates an
interval within which we can expect the performance of a high proportion of
individual laboratories to lie.

8

4) Replicate analysis of a set of samples within a laboratory provides a single
estimate of the probability of detection for that laboratory. Hence, a simple
standard deviation, across laboratories, of these estimates will provide an
estimate of the reproducibility standard deviation of the probability of
detection.
5) Using the normal approximation to calculate confidence intervals works best
for probabilities close to 0.5, but will give intervals which are too short at the
lower end for probabilities close to 1 and too short at the upper end for
probabilities close to 0. Because these are precisely the intervals that are of
most relevance to the validation of qualitative methods a more plausible
distribution should be used such as the beta distribution [13].
6) Graphical representations of method performance are much more informative
than estimates of parameters describing method performance for most users.
A standard protocol for qualitative method validation
The protocol presented in this paper is based on the calculation of performance
characteristics as intervals within which we expect the great majority (95%) of
individual laboratorys performances to lie. This is important for method validation
because we want validated methods to perform well enough, with the claimed limit
of detection, in all laboratories that use them. This is in common with Wilrich[10] and
in contrast to Wehling et. al.[11] which uses estimates of the mean probability of
detection to describe performance.
The aim of this paper is to propose a protocol as a candidate to become a standard
protocol for the validation of qualitative methods of analysis. In common with
standard methods of analysis [5] the features we require of the protocol are that it
provides a way of meeting the goals of validation (gives estimates of 1: the
probability that a sample in which the analyte is absent will generate a negative
response, and 2: the lowest concentration for which the probability of a positive
response is sufficiently high) that is sufficiently robust so that it can be applied in a
wide range of situations, simple enough to be used by everyone who needs to, and
not requiring the use of particular equipment or software. Hence, like all standard
methods, this proposed protocol is a compromise between technical correctness,
wide scope of application and simplicity.

9

The validation protocol is designed to be applied to results generated by
collaborative trial and single-laboratory studies. It provides an estimate of the
relationship between the quantity of analyte and the probability of detection with an
interval within which we can expect the probability to lie when the method is used in
practice. We can then use the resulting graphical probability profile to explore the
relationship between false positive rate and limit of detection, to demonstrate
whether a method is likely to produce results that are fit for purpose, and to consider
the relative contributions to uncertainty made by variation in method performance
and the uncertainty associated with the analysis of a finite number of samples. The
protocol is focused on providing information that is of immediate value to users of
analytical methods for meeting the goals of method validation. An absolute minimum
protocol (smallest number of samples) is also provided. This can be used to test
whether a method meets a fitness for purpose requirement for false positive
probability and limit of detection, but it will not allow users to determine a limit of
detection for a method that narrowly fails to meet the fitness for purpose requirement
nor will it give any indication that a limit of detection is much lower than the required
target.
The candidate protocol
The basic study design and the analysis of results are similar for validation by multi-
laboratory or single laboratory study. The validation study should be preceded by
some preliminary work to give an indication of what kind of design is best (minimum
or larger) and over what kind of concentration range is the study likely to produce
informative results. The validation protocol proceeds as follows:
1) For multi-laboratory studies a set of blind-replicate samples is sent to each
laboratory. Sets of n blind-replicate samples, each set at one of a range of
concentrations between zero to some level above the target limit of detection,
are sent to a number (L) laboratories.
2) Samples are analyzed in each laboratory under repeatability conditions and
the results reported. Where a method gives a quantitative response (e.g.
intensity of color change) this should also be reported.
3) Results are used estimate a prediction interval for the probability of detection
when the method is applied in a new laboratory at each concentration.

10

4) Prediction intervals are plotted to estimate upper and lower limits for the limit
of detection and false positive probability.
5) A method has demonstrated unequivocally that it is very likely to produce fit
for purpose results when upper limits for prediction intervals for false positive
probability and limit of detection are no greater than their target values.
6) Confidence intervals for the probability of detection per laboratory based on
zero between-laboratory variation are estimated. These are used to examine
the relative impact of between-laboratory variation and the limited number of
analyses undertaken per laboratory on the variation of results. This can be
used to consider whether further analyses are likely to improve estimates of
false positive probability and limit of detection.
7) The absolute minimum design requires n blind-replicate samples that do not
contain the analyte, and m blind-replicate samples that contain the analyte at
a concentration no higher than the target limit of detection sent to each of L
laboratories such that nL negative results confirm that the false negative
probability is sufficiently low and mL positive results confirm that the false
negative probability at the limit of detection is sufficiently low.
Each of L laboratories reports the number of positive results x
i
out of n
i
tests per
laboratory (i=1,2,L). Where a detection decision is made by applying a critical
value to a quantitative response, report the critical value and the response for each
test.
Hence at each concentration the estimated probability of detection for each
laboratory

is given by

and the estimated probability of detection across

laboratories is given by

Equation 1
Calculate the observed standard deviation s across the values of p
i
(i=1 to L). This
includes the contribution from between laboratory variation in the probability of
detection and uncertainty about the probability of detection in each laboratory which
is associated with the finite number of replicate analyses used to estimate each
probability.

11

Prediction interval across laboratories case 1: different values for p
i

The expected range of probabilities of detection across laboratories can be
calculated as follows. First calculate the standard deviation s of the estimates of

across laboratories (if all estimates of

are the same then see below). Next

calculate the values v
s
, w
s
, v
h
, and w
h
using

1
Equation 2

Equation 3

0.5
Equation 4

0.5
Equation 5

N is the total number of results and X is the total number of positve results, both
parameters being pooled from all laboratories. v
s
and w
s
are the parameters of a
beta distribution with mean equal to and standard deviation equal to s. v
H
and w
H

are the parameters of a beta distribution that describes the uncertainty associated
with the average probability of detection [13] assuming no between-laboratory
variation, reflecting exclusively the sampling error.
A prediction interval for the expected range of probabilities of detection across
laboratories can be calculated using the inverse beta distribution with shape
parameters v and w. The inverse beta distribution is available in Excel, in the free
open source software packages R and OpenOffice.org, and in many other packages
(see Appendix 1).
Calculate lower and upper limits for the prediction from the observed between-
laboratory variation using

12

Equation 6

,
Equation 7
and from the pure sampling error using

Equation 8

Equation 9
Where LC is the percentile value of the lower end of a confidence interval (usually
5%) and UC is the percentile value of the upper end of a confidence interval (usually
95%).
and then

Equation 10

Equation 11
Hence upper limits are taken as the maximum of the estimated upper end of the
confidence intervals (equation 7 or 9) for the average probability of detection
calculated by equation 1 and the probability of detection in a single laboratory, lower
limits are taken as the minimum of the estimated lower end of the confidence
intervals (equation 6 or 8) for the average probability of detection and the probability
of detection in a single laboratory.
Range across laboratories case 2: all p
i
=0 or all p
i
=1.
All laboratories participating in a trial may achieve 100% positive results for samples
containing higher concentrations of analyte and, for a good method, 100% negative
results for samples that do not contain the analyte. For these sets of results use [13]

13

for 1



Equation 10b
1
Equation 11b
for 0
0
Equation 10c
1 1


Equation 11c
Plotting the prediction intervals
The performance of the analytical method is described graphically by plotting results,
the limits for probabilities of detection and using these to estimate the values of the
limit of detection and false positive probability, as shown in Figure 1 to 3.
1) For each concentration plot the estimates of the probability of detection in
each laboratory p
1
, p
2
, p
L
, (Figure 1)


2) plot , upperlimit and lowerlimit and join the averages and limits across
concentrations by linear interpolation (Figure 2)


3) The estimated average false positive probability is given by the intercept of
the line on the y-axis. The upper 95% confidence for the false positive
probability in an individual line is given by the intercept of the upperlimit line
on the y-intercept (0.032 in Figure 3). The estimated average limit of detection
is given by the concentration at which the line crosses the required
probability of detection. For instance, at 95 % confidence level, the estimated
average limit of detection is 0.089 % MBM (Figure 3). The estimated upper

14

95% confidence interval for the limit of detection in an individual laboratory is
given by the concentration at which the lowerlimit line crosses the required
probability of detection. (0.122% MBM in Figure 3)

Assessment of between laboratory variation
4) The variation displayed by estimates of the probability of detection across
laboratories (s) has two sources: between laboratory variation s
L
and the
binomial sampling variation associated with the finite number of analyses
used to estimate p
i
(s
B
). The use of an analysis of variance to test whether
there is significant between-laboratory variation is demonstrated by Wilrich
[10] and Wehling at al [11].
An alternative graphical approach is used here where an interval is plotted
which gives the variation we can expect to see in results if there is no
between-laboratory variation; the probability of variation in all laboratories is
equal to the average probability of detection across laboratories; and the
observed variation in probability of detection is just what we would expect to
see when using a small number of tests to estimate it in each laboratory. The
size of the interval is estimated using the beta-binomial distribution (Appendix
2) which gives estimates of confidence intervals for the number of positive
results x
L
(lower limit) and x
U
(upper limit) per laboratory where there is no
between-laboratory variation. Plot

and

at each concentration level,

where n is the number of analyses per laboratory (Figure 4).
Comparison of the interval given by

and

and upperlimit and

lowerlimt enables a direct assessment of the contribution to variation in results
made by between-laboratory variation and the study design, and whether
increasing the number of replicates might change the conclusions drawn from
the validation study. For example, where all or much of the variation comes
from the study design,

and

and we may
conclude that increasing the number of analyses will lead to better estimates
of the probability of detection.



15

5) Report a) the scope of the validation, b) estimates of likely range of false
positive rates and limits of detection when the method is used in practice c) an
opinion on whether undertaking further analyses is likely to change the
conclusions of the study.
6) For results from an absolute minimum study design calculate the mean upper
(5%) and lower (95%) limits of the prediction interval for the false positive
probability and probability of detection at the target limit of detection using
Equations 1 to 11
Examples
Example 1: Detection of meat and bone meal in animal feed
A collaborative trial (18 laboratories, 20 replicate samples at 7 levels including zero)
of a PCR-based method to detect the presence of meat and bone meal (MBM) in
animal feed yielded the following results (Table 1).
Note that we do not recommend that validation results are reported using the format of Table 1
because information about the performance of individual laboratories is lost. This format is
used here because it is a compact way of giving data which can be reconstituted to allow
readers of this paper to repeat the data analysis methods used in the draft validation protocol.
For example Table 2 gives reconstituted results for results produced by the analysis of
samples containing 0.01% MBM.
Table 2 shows an example of individual test results, reconstituted from Table 1,
which can be used as inputs for the candidate validation protocol.
Table 3 and 4 show the results of applying Equations 1 to 11 from the draft protocol
to results reconstituted from Table 1.
Figures 1 3, which we used to illustrate the draft protocol, are based on this
example. Figure 1 shows the observed probabilities of detection for laboratories
across the seven concentrations of MBM used in the validation study. Figure 3
shows that the observed probabilities of detection tell us that we can expect limits of
detection for individual laboratories using this method to be no higher (95%
confidence) than 0.12 % MBM extracted from animal feed and may be as low as
0.036% MBM extracted from animal feed for some laboratories. We can expect that
the false positive probability will be less than 0.032 for 95% of laboratories.

16

Figure 4 shows that there is more variation in results than can be explained by the
uncertainty about the mean probability of detection and binomial sampling variation
associated with 20 replicates per laboratory for samples containing up to 0.1% MBM.
Hence, the size of the interval within which the limit of detection may lie is likely to be
driven by variation in method performance (probability of detection changing
between laboratories) rather than uncertainty about performance that could be
reduced by further measurements.
Example 2: Detection of peanut protein
A collaborative trial (18 laboratories, 5 replicate samples at 7 levels including zero) of
a dipstick to test to an allergen (peanut) in cookies [14] yielded the following results
(Table 5).
Estimates of statistical parameters and prediction intervals for the probability of a
positive response are shown in Tables 6 and 7
Figure 5 shows that the observed probabilities of detection tell us that we can expect
limits of detection for individual laboratories using this method to be no higher (95%
confidence) than 29 mg/kg peanut in cookies, and may be as low as 13 mg/kg in
some laboratories. We can expect that the false positive probability will be less than
0.14 for 95% of laboratories.
Figure 6 shows that a large part of the observed variation can be explained by the
uncertainty about the mean probability of detection and the binomial sampling
variation associated with 5 replicates per laboratory for most concentrations. Hence,
if the uncertainty associated with the estimated limit of detection is too large then
estimates of method performance may be improved by undertaking further studies
using more replicate samples. This might be the case if the target limit of detection
were around 20 mg/kg.
The effect of the low number of replicates is particularly clear for results produced by
the analysis of cookies that did not contain peanut. Here the 2 out of 90 samples
gave a positive response, giving an estimated average false positive probability of
0.022 with a 95% confidence interval of 0.005 to 0.069. The problem is that the
estimated probability of a positive response for those two laboratories that each
produced a single positive response is 0.2, and these high estimates have a big
impact on the upper limit for the estimated probability of detection. In short the

17

results are consistent with a low-between-laboratory-variation false positive
probability of less than 1% (probably good enough for a screening method), or a
much higher (estimated upper 95% interval of 0.14) false positive rate for some
laboratories, and we cant tell the difference because the number of replicates per
laboratory is low. In order to defend against these kinds of outcomes where
preliminary work leads to the expectation of a few false positive, or false negative
results a larger number of replicate analyses should be undertaken in each
laboratory (see Assessment of draft protocol performance and study design).

Example 3: A collaborative trial for the detection of salmonella in ground beef
A collaborative trial (11 laboratories, 6 replicate samples at 3 levels including zero) of
a method for the detection of salmonella in ground beef [11] yielded the following
results (Table 8), following the removal of results produced by one laboratory whose
performance was judged to be inconsistent with the other laboratories.
Calculation of the upper and lower limits (5

and 95%) for estimated probability of
detection at each concentration is shown in Tables 9 and 10. We were unable to use
the results to provide an estimate of the limit of detection where the mean probability
of detection was 95% or an upper limit for the expected limit of detection when the
method is applied in a new laboratory (Figure 7). We were able to estimate that the
limit of detection, when the method is applied in a new laboratory, can be expected
(95% confidence) to be higher than 8 cfu/25g and that the false positive probability
was likely (95% confidence) to be less than 0.05.

The observed variation between laboratories can be explained by the uncertainty
about the mean probability of detection and the binomial sampling variation
associated with 6 replicates per laboratory for all concentrations (Figure 8). Hence
estimates of method performance may be improved by undertaking further studies
using more replicate samples, possibly including some higher concentration
samples. However, we cant be confident that this will be fruitful unless a limit of
detection larger than 8 cfu/25g is considered fit for purpose.



18


Example 4: The minimum study
Minimum studies may be designed using the relation
N
log 1
log1 F

Equation 12
where N is the number of tests undertaken, F is the maximum acceptable false
detection probability for a fit for purpose method, and C is the confidence required
that the false response probability is below F. The target false positive rate is
confirmed if all negative samples give a negative result and the target limit of
detection is confirmed if all positive samples give a positive response.
For example if we require 95% confidence that a false positive probability is less
than 0.05 then:
N
log 1 0.95
log1 0.05
58.4
Which we round up to the nearest whole number of replicates: 59, and finally
distribute replicates evenly between laboratories which may make some further
rounding-up necessary.
Results of a fictional study designed to test whether false positive probability given
by two methods was less than 0.05, and that the probability of detecting 1 mg/kg of
substance was at least 0.95. (10 laboratories, 6 replicates per laboratory at two
levels: zero and target limit of detection), are shown in Table 11. Despite the tests
being fictional, one laboratory failed to provide results for negative samples analyzed
using method B and two laboratories observed one false negative result each using
method B.
Probabilities of detection were calculated using Equations 1 to 11 in the usual way
(Tables 11-15). All method A analyses produced expected results giving an
estimated false positive probability of less than (95% confidence) 0.0487 and an
estimated probability of detection for 1 mg/kg of analyte of at least (95% confidence)
0.9513. Hence a limit of detection of no more than 1 mg/kg was estimated.

19

The missing results from one laboratory using method B mean that the results do not
confirm that the false positive probability is less than 0.05. Similarly the two negative
responses produced by the analysis of the positive samples have a large effect on
the estimated range of values for the probability of detection that we may expect to
see when the method is used in practice (95% confidence interval 0.82 1.00).
Hence the absolute minimum design is most useful where both the false positive
probability and false negative probability at the target limit of detection are thought to
be low enough such that one or more unexpected result is unlikely. Low enough
means very-close-to-zero, e.g. less than 0.0005 where 60 samples are analyzed at
the limit of detection and zero concentrations.
Assessment of draft protocol performance and validation study size
We assessed the performance of the draft protocol, and how it is influenced by
validation study size, by examining how well the upper 95% confidence limit of the
estimated probability of a positive response represented the expected 95
th
percentile
of simulated laboratories. This gave us an estimate of how reliable the upper limit of
the estimated false positive probability would be when assessing the fitness for
purpose of the method. By symmetry, this also gave us an estimate of how reliable
the lower limit of the estimated probability of detection (at high probabilities) would
be when used to estimate an upper limit for the limit of detection.
We used a number of scenarios. Each scenario consisted of an average probability
of a positive response (0.1, 0.05, 0.01) with beta-distributed between-laboratory
variation such that the probability of a positive response at the 95
th
percentile of
laboratories was at double the average (0.2, 0.1, 0.02). The protocol was applied to
each scenario using a range of numbers of laboratories (5, 10, 15, 20, 30, 50, 100)
and replicate analyses within each laboratory (5, 10, 20, 50, 100). Parameters
describing the scenarios are given in Table 16.

The simulation algorithm was
For each number of laboratories (L=5, 10, 15, 20, 30, 50, 100), number of
replicates per laboratory (n=5, 10, 20, 50, 100), and scenario ( =0.1, 0.05,
0.01)

20

Generate a probability of detection for each laboratory p
i
(i=1 to L) by
selecting a random value from the beta distribution.
Generate a number of positive and negative responses from n Bernouli
trials per laboratory with probability p
i
(i=1 to L).
Analyze the responses using the draft protocol
Report the estimated probability of a positive response for the upper
95
th
percentile of laboratories.

10000 estimates of probability of a positive response for the upper 95
th
percentile of
laboratories were produced for each scenario and study design, which were
summarized by their mean (solid line) , 5
th
and 95
th
percentile (dotted lines). (Figures
9 to 11).
The general pattern is that increasing the number of laboratories and the number of
replicate analyses reduces the size of the variation associated estimates. Also,
estimates of the probability of a positive response for the upper 95
th
percentile of
laboratories becomes increasingly conservative where low numbers of replicate
analyses are used. For example, for the scenario =0.2, 95
th
percentile of
laboratories at p=0.4, average of the estimates of the 95
th
percentile of laboratories
lies at approximately 0.60 for 5 replicates per laboratory and at 0.46 for 20 replicates
per laboratory. This is because low numbers of replicates add additional variation to
estimates of the probability of detection.
This means that a finding that a method gives fit for purpose results is likely to be
safe even for studies based on lower numbers replicate analyses and laboratories if
we use the estimated upper limit, at 95% confidence, to assess fitness for purpose of
a methods false positive probability and the lower limit, at 95% confidence, to
estimate limit of detection. Larger studies (more replicates and laboratories) should
be used to give more assurance that methods which do have fit for purpose
performance will be found to produce fit for purpose results. For example, if we
decide that validation studies should be conservative, but should give results which
demonstrate that a method with an average false response probability (positive at
zero, negative at the limit of detection) of 0.01 with the upper 95
th
percentile of
laboratories having a false response probability of 0.02 should, on average be

21

expected to provide an estimate of the false response probability at the upper 95
th

percentile of laboratories of 0.05, then the 20 replicate analyses per laboratory
analyzed across as many laboratories as can be managed (Figure 12). In general
examining performance where the true value is close to zero or one requires the use
of a larger number of replicates within each laboratory, and then the use of as many
laboratories as can be afforded given the use of approximately 20 replicate samples
per laboratory.

Conclusions
We have presented a draft protocol for the validation of qualitative analytical
methods which can be applied to the validation of methods by collaborative trial or
single-laboratory validation. The analysis of results using the draft protocol is based
on the estimation of the average probability of a positive response and the observed
reproducibility standard deviation. Because a single estimate of the probability of
detection is produced in each laboratory (or in each group in a single laboratory
study) when replicate analyses are undertaken under repeatability conditions an
analysis of variance such as that used in the IUPAC/ISO/AOAC harmonized protocol
for the validation of quantitative methods is not useful. An estimate of the
reproducibility standard deviation is gained by calculating the mean and the standard
deviation of the probabilities of detection across laboratories (or groups for single
laboratory studies). Then a simple plot of confidence intervals for the probability of
detection across laboratories based on the beta distribution against the
concentration of analyte is used to provide an estimate of the range of limits of
detection and false positive probability that we can expect to see when the validated
method is applied in different laboratories.
The draft protocol has been applied to a number of trials based on the analysis of
samples at three to seven levels of analyte concentration, using five to 20 replicate
analyses per laboratory, in 10 to 18 laboratories, for PCR based detection of DNA in
animal feed, ELISA based detection of food allergens and the detection of
salmonella in ground beef.
A simulation study showed that the draft protocol tends to perform conservatively.
This means if this protocol is applied to a study based on a low number of

22

laboratories and replicate analyses and estimates of limit of detection and false
positive probability are sufficiently low, then the conclusion that the method is fit for
purpose is likely to be safe. Increasing study size, in particular by increasing the
number of replicate analyses per laboratory to 20, gives a better chance that good-
enough, methods will give results that lead to a favorable assessment.
Hence, we believe that this protocol strikes the right balance between the three
competing goals for a standard method: to give correct answers; have a broad scope
of application; and be accessible to a wide range of users.
Aknowledgements
We would like to thank Dr Stphane Pietrevalle for his excellent advice. This
research has been supported by funding from IUPAC (IUPAC project 2005-024-2-
600 Establishment of guidelines for the validation of qualitative and semi quantitative
(screening) methods by collaborative trial: a harmonized protocol) and the EU 6
th

Framework Network of Excellence Project MoniQA (Food-CT-2006-036337,
Monitoring and Quality Assurance in the Food Supply Chain
http://www.moniqa.org/). Furthermore we are grateful to Gilber Berben, Ollivier
Fumire and Ana Boix for sharing with us the results from a collaborative study for
the validation of a PCR method for the detection of meat and bone meal in
feedingstuffs conducted within the EU 6
th
Framework project Safeed-PAP (FOOD-
CT-2006-036221, Detection of presence of species-specific processed animal
proteins in animal feed http://safeedpap.feedsafety.org).

Appendix 1: Calculating the inverse beta distribution function
Example
The value of the inverse beta distribution at at the 95% percentile with shape
parameters v=10, w=2 is 0.9667.
Excel and OpenOffice
In excel (Excel 2003 onwards) [15] the value of the inverse beta distribution function,
with shape parameters v and w, at probability x is given by
Betainv(x,v,w)

23

For example if cell A1 contains the probability at which the function is to be
evaluated (x), cell B1 contains the shape parameter v and cell C1 contains the shape
parameter w, then use
=BETAINV(A1,B1,C1)
If an error is generated then instead use
=1-BETAINV(A1,C1,B1)
The same function is also available in the Calc module of OpenOffice [
16
]
R
In R (2.8.1 onwards) [17] the value of the inverse beta distribution function, with
shape parameters v and w, at probability x is given by
Qbeta(x,w,w)
Numerical approximation
If you do not have access to software for calculating the inverse beta distribution an
approximation based on the normal distribution and some arithmetic can be used
(from page 945 of Abramovitz and Stegun[18]):
, ,

. 2

Where

1
2 1

1
2 1

5
6

2
3
,
h
2
1
2v 1

1
2w 1

an

3
6

k is the inverse standard normal distribution function for (1-x). E.g for calculating the
5
th
percentile use k=1.645 and use k=-1.645 for the 95
th
percentile.


24

Appendix 2: calculating the beta binomial distribution function and using it to
examine observed variation
The beta-binomial distribution can be used to estimate the maximum and minimum
expected number of positive results from n tests given that we have observed X
positive results in N tests as follows.
If we have observed X positive results in N tests then, assuming that the probability
of a positive results does not change, the natural log of the probability of observing i
positive results in n tests is given by
|,
n 1 i 1 n i 1 X 0.5 i
N X 0.5 n i N 1 n X 0.5
N X 0.5 N 1
And
|, n 1 i 1 n i 1 X 0.5 i
N X 0.5 n i N 1 n X 0.5
N X 0.5 N 1
Where is the log gamma function.
Hence where X out of N tests are positive then we can expect to see at least x
L

positive results out of n tests, where x
L
is the lowest integer for which
|,

0.05,
And no more than x
U
+1 positive results out of n tests, where x
U
is the highest integer
for which
|,

0.95
For example if 150 out of 300 tests have given a positive result and there are 10
tests per laboratory then the probability of seining no positives in a set of results from
a laboratory is

25

0|150,300 10 1 0 1 10 0 1 150 0.5
0

300 150 0.5 10 0 300 1 20 150 0.5

300 150 0.5 300 1 0.0011289
And
x |150,300
|150,300

0 0.001129 0.001129
1 0.010652 0.011781
2 0.045817 0.057597
3 0.118299 0.175896
4 0.203055 0.378951
5 0.242098 0.621049
6 0.203055 0.824104
7 0.118299 0.942403
8 0.045817 0.988219
9 0.010652 0.998871
10 0.001129 1.000000

Hence a 90% confidence interval for the number of positive responses per laboratory
lies between 2/10 and 8/10 positive responses if the probability of a positive
response does not vary between laboratories.
This approach uses the log gamma function which is available in many software
packages including the following.
Excel and OpenOffice
In excel (Excel 2003 onwards) [15] the value of the log gamma function of x is given
by
Gammaln(x)
For example if cell A1 contains the value at which the function is to be evaluated (x,
then use
=GAMMALN(A1)
The same function is also available in the Calc module of OpenOffice [16]

26

R
In R (2.8.1 onwards) [17] the value of the log gamma function of x is given by
lgamma(x)

References


1 Lebesi, D., Dimakou, C., Alldrick, A.J., and Oreopoulou, V.,( 2010), QAS, 2(4),
173181
2 Horwitz, W., (1995), Pure Appl Chem 67, 331-343
3 EURACHEM (2000) Quantifying Uncertainty in Analytical Measurement, 2nd
Ed., http://www.eurachem.org/
4 Feinberg, M., Boulanger, B., Dew, W., & Hubert, P. (2004) Anal. Bioanal.
Chem. 380, 502514
5 Rose M, Poms R, Macarthur R, Ppping B, Ulberth F, QAS, Accepted for
publication
6 McCullagh P, Nelder J.A., 1989, Generalized Linear Models, Second edition,
Chapman and Hall, London
7 Langton SD, Chevennement R, Nagelkeke N, Lombard B (2002) Int J Food
Microbiol, 79:171181
8 van der Voet H, van Raamsdonk WD (2004) Int J Food Microbiol 95, 231234
9 Wilrich C, Wilrich P-T, (2009), JAOAC Int. 92(6), 1763-1772
10 Wilrich, P-T, (2010) Accred Qual Assur (2010) 15, 439444
11 Wehling, P., Labudde, R.A., Brunelle S.L., Nelson, M.T., (2011), JAOAC Int.,
94(1), 335-347
12 LaBudde, R.A. (2009) Coverage Accuracy for Binomial Proportion 95%
Confidence Intervals for 12 to 100 Replicates, TR297, Least Cost
Formulations, Ltd, Virginia Beach, VA,

27


http://www.lcfltd.com/Documents/tr297%20coverage%20accuracy%20binomi
al%20proportions.pdf
13 Brown L.D., Cai T.T., DasGupta A., (2001), Stat. Sci., 16(2), 101-133
14 van Hengel, A.J., Capelletti C., Brohee, M., Anklam E., (2006), JAOAC Int.
89(2), 462-468
15 Microsoft Excel online Help, http://office.microsoft.com/en-us/excel-
help/betainv-HP005209001.aspx?CTT=1
16 Open Office, http://www.OpenOffice.org
17 The R project for statistical computing, http://www.r-project.org/
18 Abramowitz, M., Stegun, I., (1972), Handbook of Mathematical Functions,
10th edition