Professional Documents
Culture Documents
Food Control
journal homepage: www.elsevier.com/locate/foodcont
a r t i c l e i n f o
a b s t r a c t
Article history:
Received 28 October 2013
Received in revised form
21 March 2014
Accepted 29 March 2014
Available online 12 April 2014
Variables sampling plans for microbial safety are usually based on the log transformation of the observed
counts. We propose a new variables plan for lognormal data using the angular transformation. In a
comparison with the classic approach, this new method shows more stringency and allows the use of
smaller sample sizes to obtain the same level of consumer protection. This transformation is robust when
the underlying distribution departs from the lognormal distribution as well as in the presence of
contamination. A description of the new plan and the software codes are provided.
2014 Elsevier Ltd. All rights reserved.
Keywords:
Food safety
Acceptance sampling
Lognormal distribution
Robust plan
Sinhearcsinh transformation
1. Introduction
Sampling inspection plans are used to assess the tness for use
of batches of products providing protection to the consumers.
Sampling inspection cannot be avoided due to the high cost associated with laboratory testing and the destructive nature of the
microbiological tests makes 100% inspection impossible. A single
sample of size n is usually inspected or tested for the specied
microbiological criteria for conformance. The conformance criteria
are often regulated for potentially dangerous pathogens e.g.
Regulation (EC) No 2073/2005, (Anonymous, 2005). In addition to
variables plans, attribute plans are also used for food safety inspection. In twoeclass attribute plans, items are classied as conforming or not while the variables plans deal with parameters such
as the mean concentration. The International Commission on
Microbiological Specications for Foods in ICMSF (2002) as well as
the Codex Alimentarius Commission (CAC) in CAC (2004) recommend both attribute and variables plans for food quality assessment. The performance of simple attributes plans in the food safety
context is widely studied, see for instance, Hildebrandt, Bhmer,
and Dahms (1995), Legan, Vandeven, Dahms, and Cole (2001),
Dahms (2004), Wilrich and Weiss (2009). However, the variables
plans are preferred because the whole information from the sample
250
diseases. Often only few units of each lot are tested because of the
budget constraints on laboratory analysis. A small sample size may
fail to detect high levels of fraction nonconforming compromising
protection to the consumer. The paper presents a new decision
criterion with a better performance than the classic approach as
well as robust to distributional uncertainties. The proposed variables plan provides the same level of protection to the producer
with fewer units of the batch to be tested, but with a stringent level
of protection to the consumer.
h
i
Hx sinh d sinh1 x
Z1
my Y
Sy
(1)
(2)
p
where sinhx ex ex =2 and sinh1 x log x x2 1
are the hyperbolic sine and its inverse. The hyperbolic functions are
the counterpart in the hyperbola of the classic trigonometric
functions. This transformation allows to control the skewness and
kurtosis with the parameters d and . The skewness and kurtosis are
measures of the asymmetry and the peakedness of the distribution
respectively. We used a particular combination of parameters,
d 0.1 and 0. Let V H(x) and mv H(m), then an analogous test
statistic to Eq. (1) for the new approach can be dened.
Z2
mv V
Sv
(3)
This plan will be referred as the new method from now on.
2
3
4
5
10
15
20
30
40
50
60
a 0.01
a 0.05
k1
k2
k1
k2
0.97
1.22
1.39
1.51
1.85
2.02
2.14
2.28
2.37
2.43
2.48
1.03
1.34
1.54
1.71
2.16
2.41
2.57
2.79
2.93
3.02
3.10
1.41
1.63
1.77
1.87
2.15
2.29
2.38
2.49
2.56
2.61
2.64
1.63
1.90
2.08
2.22
2.61
2.82
2.95
3.13
3.23
3.31
3.37
3. Results
Critical values for an AQL 0.1%, different sample sizes and
producers risk probabilities (a) are shown in Table 1. The kevalues
for the Z1 statistic given in Table 1 are the same as can be found in
the variables sampling plan literature. See the Appendix section for
critical distances for other set of AQL values.
An evaluation of the performance of the new method in comparison to the traditional approach is shown in Fig. 1. A reasonable
sample size of size (n 10) and two producers risk levels (a 0.01,
0.05) were considered. Following the approach of Wilrich and
Weiss (2009), this gure shows two Xeaxes: one for the proportion nonconforming (p) and the other for the associated process
level in log10 (CFU/g). It can be seen from these gures that the new
approach reduces at a consumers risk of 10% the rejectable fraction
nonconforming in about 3% while maintaining the producers risk
at 1%. Alternatively the new approach will reduce the number of
test to be done. For instance, the same or more protection is obtained with nine and eight units at a 0.01 and 0.05 in comparison
with that of obtained with 10 units using by the classic approach.
Figs. A1 and A2 in the Appendix section show OC curves for various
sample sizes recommended in ICMSF (2002) and other popular
combinations of AQL and a.
251
Fig. 1. Comparison of Operating Characteristic (OC) curves for n 10, AQL 0.1% and different values of producers risk. The OC curves of the log and sinhearcsinh transformations
are shown in solid and dashed lines respectively. The new approach offers better consumer protection by lowering the consumers risk at poor quality levels.
252
Fig. 2. Comparison of Operating Characteristic (OC) curves at a false positive misclassication error of 1% for n 10, AQL 0.1% and different values of producers risk. The OC curves
of the log and sinhearcsinh transformations are shown in heavy solid and dashed lines respectively.
40,000
4.602
1.385
69,000
4.839
1.480
81,000
4.908
1.509
200,000
5.301
1.679
350,000
5.544
1.791
253
Fig. 3. Effect in the OC curves when the true distribution is gamma (displayed in thicker line width). The difference in the LQL at a b risk for the Z2 statistic is much smaller than that
of Z1.
Fig. 4. Effect in the OC curves when the true distribution is contaminated lognormal (displayed in thicker line width). The Z2 statistic shows a much smaller reduction in LQL than Z1.
254
8. Conclusions
Effect of the parameters in the sampling performance
The performance of variables sampling plans using logetransformed data was compared with that obtained using the sinhearcsinh transformation, for right skewed distributions used for
Fig. A1. Comparison of OC curves at a producers risk (a) of 0.01 for different combinations of sample size and AQL. The common cause situation is assumed to be the lognormal
distribution with m 0 and s 1, both in log scale.
255
Fig. A2. Comparison of OC curves at a producers risk (a) of 0.05 for different combinations of sample size and AQL. The common cause situation was modelled in the lognormal
distribution using m 0 and s 1, both in log scale.
Table A1
Monte Carlo estimates of the critical distance factor (k) for three values of producers
risk and AQL 0.01.
Table A2
Calculated estimates of the critical distance factor (k) for three values of producers
risk and AQL 0.0001.
2
3
4
5
10
15
20
30
40
50
60
a 0.01
a 0.02
a 0.05
k1
k2
k1
k2
k1
k2
0.56
0.78
0.92
1.03
1.31
1.46
1.55
1.67
1.74
1.79
1.83
0.54
0.78
0.95
1.08
1.44
1.63
1.76
1.93
2.03
2.11
2.17
0.71
0.91
1.04
1.14
1.41
1.54
1.63
1.73
1.80
1.85
1.89
0.71
0.94
1.10
1.22
1.57
1.75
1.87
2.02
2.12
2.19
2.25
0.95
1.13
1.25
1.33
1.56
1.68
1.75
1.84
1.90
1.94
1.97
1.02
1.22
1.37
1.48
1.79
1.95
2.05
2.18
2.27
2.33
2.38
2
3
4
5
10
15
20
30
40
50
60
a 0.01
a 0.02
a 0.05
k1
k2
k1
k2
k1
k2
1.26
1.56
1.75
1.89
2.28
2.48
2.61
2.77
2.88
2.95
3.01
1.41
1.78
2.03
2.22
2.77
3.06
3.26
3.51
3.68
3.80
3.89
1.43
1.72
1.90
2.04
2.41
2.60
2.72
2.87
2.96
3.03
3.08
1.66
2.02
2.27
2.44
2.97
3.25
3.43
3.67
3.82
3.93
4.01
1.76
2.02
2.18
2.30
2.63
2.79
2.89
3.02
3.10
3.16
3.20
2.13
2.46
2.68
2.85
3.31
3.56
3.71
3.92
4.05
4.14
4.21
256
Software code
The R code shown below computes the acceptability constants
k1 and k2 for a given combination of a, AQL and n.
Step-by-step guide
The sampling design for the new approach can be developed by
using the following guide:
Dene the producers risk (a) and the number of samples to be
drawn (n).
For a given regulatory limit (m), compute the associated AQL as
the right tail area in the lognormal distribution. If m is not
established, dene an AQL and obtain m as the quantile of the
lognormal distribution.
Compute the statistic Z2.
For a given n, AQL and a, obtain k from Table 1, A1 or A2. For
other combinations, obtain k using the R snippet.
Apply the decision criterion. If Z2 k accept the lot; otherwise
reject.
Distribution matching
Fig. A3 shows the three distributions matched through their
mode and density. The lognormal distribution is more skewed that
the gamma and Weibull models.
\mathcal
{LN}\left
( \mu,\sigma
\right )
f x=s; m
1 exp
p
xs 2p
m
s
Z1 my Y =Sy
Z2 mv V =Sv
h
i
Hx sinh d sinh1 x
m
P
Y
Yi =n
q
P
Yi Y2 =n 1
Sy
e1
n
q
k
AQL
LQL
a
b
m2
lnx
2s2
Probability density
function
Logscale
Shape parameter
Test statistic for the
normal distribution
Statistic of the
sinhearcsinh transformation
sinhearcsinh transformation
Microbiological limit
Sample mean
sample standard deviation
Type I measurement error
Number of samples
Quantile function
Critical distance
Acceptable Quality Limit
Limiting Quality Level
Producers risk
Consumers risk
References
Fig. A3. Lognormal probability density function with m 0 and s 1 in solid line
matched with the gamma (c 1.5,b 0.75) and Weibull (k 1.3, l 1.14) distributions
through the mode and the density. The gamma and Weibull distribution are in dashed
and dotdashed line.
Albin, S. L. (1990). The lognormal distribution for modeling quality data when the
mean is near zero. Journal of Quality Technology, 22, 105e110.
Anonymous. (2005). Commission regulation (EC) No 2073/2005 of 15 November
2005 on microbiological criteria for foodstuffs. Ofcial Journal of the European
Union, L 338(1).
CAC. (2004). General guidelines on sampling. Codex Alimentarius Commission
Accessed 15.10.13.
Dahms, S. (2004). Microbiological sampling plans: statistical aspects. Mitteilungen
aus Lebensmitteluntersuchung und Hygiene, 95, 32e44.
Fowlkes, E. B. (1979). Some methods for studying the mixture of two normal
(lognormal) distributions. Journal of the American Statistical Association, 74,
561e575.
Gonzales-Barron, U., & Butler, F. (2011). A comparison between the discrete Poissongamma and Poisson-lognormal distributions to characterise microbial counts in
foods. Food Control, 22, 1279e1286.
257