Content uploaded by Andy Mauromoustakos
Author content
All content in this area was uploaded by Andy Mauromoustakos on May 29, 2020
Content may be subject to copyright.
_____________________________________________________________________________________________________
*Corresponding author: Email: amauro@uark.edu;
Original Research Article
Journal of Basic and Applied Research International
26(2): 39-52, 2020
ISSN: 2395-3438 (P), ISSN: 2395-3446 (O)
FURTHER PREDICTIVE STATISTICAL ANALYSIS OF A
GFSI SURVEY OF INTERNATIONAL FOOD PROCESSORS
ANDY MAUROMOUSTAKOS
1*
, PHILIP G. CRANDALL
2
, KEVIN C. THOMPSON
1
AND CORLISS A. O’BRYAN
2
1
Agricultural Statistics Laboratory, University of Arkansas, Fayetteville, AR 72701, USA.
2
Department of Food Science, University of Arkansas, Fayetteville, AR 72704, USA.
AUTHORS’ CONTRIBUTIONS
This work was carried out in collaboration among all authors. All authors read and approved the final
manuscript.
Received: 18 March 2020
Accepted: 23 May 2020
Published: 29 May 2020
__________________________________________________________________________________
ABSTRACT
Food safety and quality regulations that are agreed on by the buyer and seller are the cornerstone of food safety
and quality. Earlier, we reported the top-level survey results of 2,300 food manufacturers as to their costs and
benefits of becoming compliant with the benchmarked Global Food Safety Initiative (GFSI) schemes. In this
paper, we provide a more in-depth statistical analysis with rationale for using these analyses. The Bradley Terry
probability model showed that “existing customer requirements” were 4.6 times more important than “potential
new customer requirements.” A Multiple Correspondence Analysis associated the reasons for certification using
the following customer demographics: the GFSI scheme (British Retail Consortium [BRC] or Safe Quality
Foods [SQF]), the region of the world (Europe/EU or North America/NA) and the food safety risk (high,
medium or low). EU companies certified by BRC were primarily driven by wanting to improve their business
reputation while those certified by the SQF scheme, mostly in NA, were primarily driven by requirements of an
existing customer. A Generalized Linear Model with the multinomial ANOVA and a cumulative logistic link
functioned to estimate the amount of agreement on four important questions about Key Performance Indicators,
KPI’s sales/revenues, numbers of customers, employees, and suppliers. A Quantile Regression was used to
analyze the length of time companies spent becoming GFSI Certified, with 80% of the companies achieving this
goal in less than 1 year. Finally, a Logistic Regression and tree-based models used 24 perceived benefits to
predict if certification would be beneficial for a new company considering this option.
Keywords: Food safety; quality; statistical analytics; large data analysis; analytic survey data; generalized
regression models; Bradley-Terry probability model; multi-dimensional preference analysis;
multiple correspondence analysis and predictive modeling.
1. INTRODUCTION
One of the preeminent, international food safety and
quality standards, the Global Food Safety Initiative
(GFSI) continues to be a hot topic with world-wide
interest. An on-line search of the term, GFSI, resulted
in 573,000 hits [1]. Our initial publication provided
the top-notes of food companies' opinions following
their initial GFSI implementation [2]. Overall, GFSI
was perceived as being beneficial or very beneficial
by 90% of these newly certified suppliers. Nearly
three-quarters of these certified food manufacturers
would choose to go through the certification process
again, even if they were not required to do so by one
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
40
of their current retail customers. They believed GFSI
was foundational in starting to do business with new
customers, reducing the number of third-party food
safety audits, and/or improving their own or their
suppliers’ food safety programs. Six out of ten
respondents reported significant capital and staff time
were required to become GFSI compliant.
Consumers, governmental regulators, and the food
industry all share the common desire and are working
diligently to increase the safety of our food supply.
However, despite the millions of dollars spent on
basic microbiological research, promulgating and
enforcing ever-expanding regulations and round-the-
clock food safety testing by the food industry to
assure the absence of pathogenic microorganisms,
more than half of US consumers are concerned about
the safety of the food they serve their families.
According to a poll, 57% of consumers surveyed are
concerned or very concerned about the safety of their
food, and 22% reported getting sick from something
they ate in the past year [3]. Echoing these concerns,
a 2015 FAO survey lists food safety and selecting
food for health benefits as major drivers of consumer
purchasing decisions [4]. Rohrich [5] calculated the
number of meals potentially associated with food
borne illnesses based on CDC risk assessment data.
Of the more than 600 million meals eaten every day in
the USA, 24,000 meals are potentially infected with
pathogenic bacteria.
The global marketplace for food products and raw
ingredients continues to expand. World-wide exports
of food, excluding live animals, reached $1.1 trillion
in 2013, up more than three-fold from $342 billion in
2001 [6]. There have been notable breakdowns in the
necessary world-wide trust relationships with food
suppliers whose foods were intentionally adulterated
with horse meat [4] or whose milk was intentionally
tainted with melamine [7]. The wide-spread, negative
publicity surrounding these and other headline
grabbing incidents has continued to fuel consumers’
skepticism that regulators and the food industry are
able to ensure the safety of their food supply.
It has also become increasingly clear that to rebuild
consumers’ confidence in this global marketplace and
to meet these increasingly diverse global challenges,
global food safety standards need to be strengthened.
For nearly 20 years an international food safety
standard, the Global Food Safety Initiative (GFSI),
has been continuously updated to provide real-time,
science-based information to minimize risks from
foodborne pathogens, managed costs of establishing
and maintaining an effective food safety culture and
to improved consumer confidence in our food
production and retail industries. Today, GFSI certified
third party auditors work with more than 100,000
food manufacturing facilities in 160 countries to assist
them in validating their food safety management
practices.
GFSI efficacy continues to be of interest to food
manufacturers world-wide as evidenced by several
recent studies. King et al. [8] discussed mega-trends
that will challenge the future of food safety. They
provided the rational for the role Big Data can play in
optimizing food safety legislation to be proactive
rather than the current reactive environment. These
authors highlighted the disturbing trend of private
retailers’ individual food safety standards that are
counterproductive to global harmonization. Luning
et al. [9] who surveyed 100 EU food manufacturers as
to their food safety management practices and
analyzed the data with a hierarchical cluster analysis.
Their analysis showed that most firms’ management
adequately controlled for the level of food safety risk.
Researchers in China used secondary governmental
data to show that food safety certifications played a
positive role in their supply chain management and
the firm’s economic performance [10]. Chen, et al.
[11] surveyed 115 New Zealand food manufacturers
and reported a number of positive gains from
embracing third-party food safety management
schemes, including improved product traceability.
While traceability standards are beyond the scope of
this paper, they are certainly a supportive part of
many international regulations [12,13].
The objective of this current paper is to extend the
statistical analysis of the survey data from 828 food
manufacturers regarding the costs and benefits of
becoming compliant with one of the benchmarked
GFSI schemes. It is hoped that readers would consider
using these more advanced statistical techniques to
improve their ability to predict future behavior and to
meet key benchmarks for new companies which
might want to pursue GFSI certification. Techniques
such as the Bradley Terry probability model, a Multi-
Dimensional Preference Analysis, a Multiple
Correspondence Analysis, and a Quantile and Logistic
Regression are discussed.
2. METHODS AND MATERIALS
A survey of international food manufacturers about
their perceived benefits and the costs of complying
with an international food safety benchmark, GFSI,
was conducted. Survey data was collected on the key
demographics of survey respondents, the rationale for
becoming GFSI certified, the costs and benefits of
becoming certified, additional food safety training
necessitated, responsibilities required and an estimate
of the length of time spent in becoming certified.
This survey was initially conducted
in t
by a project steering group comprised of staff from
Sealed Air and McCallum Layton, the University of
Arkansas, Diversey Consulting, the Consumer Goods
Forum, and the officers of GFSI. Additional details
on the results of the survey, and a
copy of the survey,
are available at Crandall et al. [2].
In this paper, we make extensive use of the various
forms of the Generalized Regression Model with
binary, multinomial and categorical data and some
continuous data with the hope of choosing model
with good predictive power (higher Akaike’s
Information Criterion, AICc). We endeavor to provide
a detailed rationale for using the selected statistical
analysis and subsequent interpretation of the results
for this large data set. It is imperative
that researchers know the questions they want to
answer before beginning their analysis. Knowing
the questions ahead of time does two important
things: it gives the persons conducting the statistical
analysis the questions
that need data to be answered
and it prevents “p hacking,” which is the
inappropriate use of statistical analysis to run multiple
probability tests to calculate the probability that the
null hypothesis (of no effect) can be rejected. It is
quite possible
that, especially with very large data
sets, that there will be “statistically significant” p <
0.05 results that are simply due to random error. At
this level of “p” one would expect there to be
significance simply due to random error to occur less
than
5 times out of 100. See the perspective article by
Fig
Percentage of companies self-
reported as certified by GFSI schemes and producing foods of various risk categor
previously described [2]
. BRC, British Retail Consortium; SQF, Safe Quality Food, Other, other third
collapsed because of the low rate of response
Mauromoustakos et al.
; JOBARI,
41
in t
he fall of 2013
by a project steering group comprised of staff from
Sealed Air and McCallum Layton, the University of
Arkansas, Diversey Consulting, the Consumer Goods
Forum, and the officers of GFSI. Additional details
copy of the survey,
In this paper, we make extensive use of the various
forms of the Generalized Regression Model with
binary, multinomial and categorical data and some
continuous data with the hope of choosing model
s
with good predictive power (higher Akaike’s
Information Criterion, AICc). We endeavor to provide
a detailed rationale for using the selected statistical
analysis and subsequent interpretation of the results
for this large data set. It is imperative
that researchers know the questions they want to
answer before beginning their analysis. Knowing
the questions ahead of time does two important
things: it gives the persons conducting the statistical
that need data to be answered
and it prevents “p hacking,” which is the
inappropriate use of statistical analysis to run multiple
probability tests to calculate the probability that the
null hypothesis (of no effect) can be rejected. It is
that, especially with very large data
sets, that there will be “statistically significant” p <
0.05 results that are simply due to random error. At
this level of “p” one would expect there to be
significance simply due to random error to occur less
5 times out of 100. See the perspective article by
Head, Holman, Lanfear, Kahn, & Jennions
additional information.
2.1 Key Demographics
We grouped 828 respondents according to
party GFSI certification scheme they used, their
region of the world where their plant(s) were located
and the risk
category of the foods they manufactured
(Fig. 1).
2.2 Analysis of Rank Order Data
An important question asked the food manufac
respondents to rank, in order of importance, their
top three reasons for becoming GFSI certified.
Preliminary surveys and interviews with suppliers
were used to generate eight specific reasons used in
the questionnaire.
The Bradley Terry
probability model was used to
compare all pairs of reasons for becoming GFSI
certified. For each reason, a probability, and an odds
ratio, along with their confidence interval were
calculated for choosing this reason over the other
reasons. The Bradley Ter
ry model also allowed the
researchers to focus the comparison on two specific
reasons. Companies which did not rank both reasons
under consideration as one of their top three reasons
were excluded from that round of calculations. This
process allowed for
the elimination of lower, non
significant reasons from the analysis.
Fig
. 1. Demographics of respondents
reported as certified by GFSI schemes and producing foods of various risk categor
. BRC, British Retail Consortium; SQF, Safe Quality Food, Other, other third
-
party schemes were
collapsed because of the low rate of response
; JOBARI,
26(2): 39-52, 2020
Head, Holman, Lanfear, Kahn, & Jennions
[14] for
We grouped 828 respondents according to
which third
party GFSI certification scheme they used, their
region of the world where their plant(s) were located
category of the foods they manufactured
2.2 Analysis of Rank Order Data
An important question asked the food manufac
turing
respondents to rank, in order of importance, their
top three reasons for becoming GFSI certified.
Preliminary surveys and interviews with suppliers
were used to generate eight specific reasons used in
probability model was used to
compare all pairs of reasons for becoming GFSI
certified. For each reason, a probability, and an odds
ratio, along with their confidence interval were
calculated for choosing this reason over the other
ry model also allowed the
researchers to focus the comparison on two specific
reasons. Companies which did not rank both reasons
under consideration as one of their top three reasons
were excluded from that round of calculations. This
the elimination of lower, non
-
reported as certified by GFSI schemes and producing foods of various risk categor
ies as
party schemes were
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
42
A Multidimensional Preference Analysis (MDPREF)
was used to graphically group the rankings of the
reasons into clusters based on the scoring method. A
Multiple Correspondence Analysis (MCA) was also
used to graphically illustrate an estimate of the
association among the most important reasons
compared to the companies’ demographics: scheme,
region, and risk classifications to see if there were
reason by demographic interactions.
The MDPREF and MCA examples used a Heisman
weighting method that gives more weight to the
primary reason (one-half of a point), less weight to the
second reason (one-third of a point), even less weight
to the third reason (one-sixth of a point), and no
weight to the remaining reasons. See Crandall et al.
[15] for a discussion alternative data weighting
methods.
2.3 Analysis of Ordinal Data
A set of four questions compared the Key
Performance Indicators (KPI’s) of the company
before and after GFSI certification. Responses were
given either on a 3-point ordinal scale or as actual
numeric values before and after GFSI certification.
Responses returned as the numeric before/after values
were converted to the 3-point scale. The responses
were analyzed separately in a generalized linear
model using a multinomial distribution and a
cumulative logistic link function. A stepwise method
was used to select the best subset of the three
demographic items (scheme, region, risk) to be
included in the final model. For each significant
effect, the primary contrasts of interest were
investigated: British Retail Consortium (BRC) versus
Safe Quality Food (SQF) for the scheme, North
America (NA) versus Europe (EU) for the region, and
high versus low for the food safety risk category of
food manufactured.
2.4 Analysis of Highly Skewed Continuous
Data
Data on the question, what was the time in months
your company needed to attain certification, was the
only data collected on a continuous scale. This item
was analyzed using Quantile Regression to account
for the highly skewed nature of the data. Companies
certified under multiple schemes were removed from
this analysis. The 50
th
and 80
th
percentiles were fitted
to the best subset of the three demographic items
(scheme, region, risk).
2.5 Predictive Analytics
The primary focus question of the survey was whether
certification was beneficial. Responses were given on
a 5-point ordinal scale. For prediction purposes, the
results were collapsed into a binary response:
beneficial vs not beneficial (output variable). The
3.5% of the responses in the neutral level were
eliminated. A set of 24 questions asked the
respondents for their level of agreement with
statements in four general categories: their intentions,
regulatory compliance, benefits, and costs resulting
from becoming GFSI certified. For prediction
purposes, these data were analyzed and reported as
responses related to intentions, compliance, other
benefits and costs and were treated as predictors
(inputs).
Various predictive modeling methods were used to
relate the response of whether GFSI certification was
beneficial or not to the set of the intentions,
compliance, benefits, and costs items. These models
might be useful to the managers of the various GFSI
certification schemes that marketers use to target
potential future retail clients who may be considering
asking their suppliers to become GFSI certified.
The model comparison platform in JMP Pro was
utilized to compare these models. All analysis
presented in this paper used SAS/Stat ®, Version 9.4
and the univariate and multivariate platforms of JMP
Pro ®, Version 14 [16].
3. RESULTS AND DISCUSSION
3.1 Demographics
Of the 15,000 questionnaires sent out, we used data
cleaning to attain 828 usable responses corresponding
to a 5.5% response rate. The percentage responses
were well distributed with 52% of the respondents
from North America (NA), 34% from Europe (EU),
10% from the Pacific, and 4% as having multiple
plants in multiple regions, which we termed as
Global. The vast majority, 90%, of the 2,300
companies responding to the survey, were GFSI
certified, using either the British Retail Consortium
(BRC) 51% or Safe Quality Food (SQF) 39%
schemes. A mosaic plot summary capturing the most
important demographics summary shows both a
strongly significant association between scheme and
region (P-value<.0001) and lack of association
between risk and region depending on location
(Fig. 1).
3.2 What was the Importance of the Reasons
for Becoming GFSI Certified
In the first paper we simply reported on the top three
reasons from this large data set of food manufacturers
to become GFSI certified by tallying the reasons that
received the most votes [2].
This gave us a good
starting point in understanding the sentime
respondents, but in this paper two more advanced
statistical techniques were used to further our
understanding of this important question. The
results of the Bradley Terry model are given as
probabilities, odds
along with their 95%
intervals
in Table 1. From the pooled results, by
and-
away the most statistically significant reason for
becoming GFSI certified was that an “existing
customer required (our company) to become (GFSI)
certified.” This reason was 4.61
times
(3.84, 5.54
) than the second most prevalent reason.
To see a direct comparison of the first to the second
reason, look at the bottom of the table
. There was an
82% probability that reason number one was selected
first compared t
o the second or third reason. When
comparing the second to the third reason
Fig. 2. Multidimensional preference analysis on reasons using Heisman scores
Mauromoustakos et al.
; JOBARI,
43
This gave us a good
starting point in understanding the sentime
nts of the
respondents, but in this paper two more advanced
statistical techniques were used to further our
understanding of this important question. The
results of the Bradley Terry model are given as
along with their 95%
confidence
in Table 1. From the pooled results, by
-far-
away the most statistically significant reason for
becoming GFSI certified was that an “existing
customer required (our company) to become (GFSI)
times
more likely
) than the second most prevalent reason.
To see a direct comparison of the first to the second
. There was an
82% probability that reason number one was selected
o the second or third reason. When
comparing the second to the third reason
-they
were evenly split with a 52% (basically a 50:50)
probability.
To dig deeper into this important question, a
Multidimensional
Preference Analysis (MDPREF)
analysis, based on the Heisman weighted scores, was
used to group the reasons graphically (three
destination points on the figure) for becoming GFSI
certified and the reasons companies cited for
certification (arrows) (Fig. 2). Fig. 2 provides an easy
way
for the reader to visulize the comparison of the
dependant variable (reasons for becoming certified)
with the independent variable, the demographics of
the company responding. A major advantage of using
MDPREF
analysis was that it allowed the reader to
gr
asp the interrelationships among a group of reasons
and in this case, the number of companies associated
with that reason.
Fig. 2. Multidimensional preference analysis on reasons using Heisman scores
(n=828)
; JOBARI,
26(2): 39-52, 2020
were evenly split with a 52% (basically a 50:50)
To dig deeper into this important question, a
Preference Analysis (MDPREF)
analysis, based on the Heisman weighted scores, was
used to group the reasons graphically (three
destination points on the figure) for becoming GFSI
certified and the reasons companies cited for
certification (arrows) (Fig. 2). Fig. 2 provides an easy
for the reader to visulize the comparison of the
dependant variable (reasons for becoming certified)
with the independent variable, the demographics of
the company responding. A major advantage of using
analysis was that it allowed the reader to
asp the interrelationships among a group of reasons
and in this case, the number of companies associated
(n=828)
Table 1. Bradley-
Terry analysis of main reasons for suppliers becoming GFSI certified
Reason
An existing customer requirement
A potential new customer requirement
Improve our food safety program
Improve business reputation generally
Improve our operational efficiency
Existing vs new
customer requirement
New customer vs food safety improvement
a
Numbers not followed by the same superscript letter are significantly different at the 95% confidence level
Fig. 3. Multiple Correspondence Analysis (MCA) that associates the main reasons for certification with
the three demographic delineators: Scheme, region, and risk
Dimension 1. Reasons for becoming GFSI certified. Existing customer = an existing customer re
certified; New customers = a potential new customer requires GFSI certification; business reputation= hope that GFSI will
improve business reputation in general; food safety = GFSI will improve the overall food safety program.
Dimension
2. Schemes: BRC, British Retail Consortium; SQF Safe Quality Food; Other third
Region: North America, Europe and Global (plants in more than one region)
Risk: Level
of risk of foods being
Mauromoustakos et al.
; JOBARI,
44
Terry analysis of main reasons for suppliers becoming GFSI certified
Probability Odds
0.91
a
(0.89,
0.92)
10.12
a
(8.36,
A potential new customer requirement
0.69
b
(0.65, 0.72) 2.19
b
(1.84,
0.65
b
(0.60, 0.69) 1.85
b
(1.51,
Improve business reputation generally
0.58
c
(0.52, 0.63) 1.36
c
(1.10,
0.39
d
(0.28, 0.51) 0.64
d
(0.39,
customer requirement
0.82
a
(0.79, 0.85) 4.61
a
(3.84,
New customer vs food safety improvement
0.54
b
(0.49,
0.59)
1.18
b
(0.96,
Numbers not followed by the same superscript letter are significantly different at the 95% confidence level
Fig. 3. Multiple Correspondence Analysis (MCA) that associates the main reasons for certification with
the three demographic delineators: Scheme, region, and risk
Dimension 1. Reasons for becoming GFSI certified. Existing customer = an existing customer requirement to become
certified; New customers = a potential new customer requires GFSI certification; business reputation= hope that GFSI will
improve business reputation in general; food safety = GFSI will improve the overall food safety program.
2. Schemes: BRC, British Retail Consortium; SQF Safe Quality Food; Other third
-
party schemes were collapsed
because of the low response rate.
Region: North America, Europe and Global (plants in more than one region)
of risk of foods being produced: high risk foods like raw meats, medium and low risk
; JOBARI,
26(2): 39-52, 2020
Terry analysis of main reasons for suppliers becoming GFSI certified
(n=828)
(8.36,
12.25)
(1.84,
2.62)
(1.51,
2.27)
(1.10,
1.69)
(0.39,
1.06)
(3.84,
5.54)
(0.96,
1.47)
Numbers not followed by the same superscript letter are significantly different at the 95% confidence level
Fig. 3. Multiple Correspondence Analysis (MCA) that associates the main reasons for certification with
quirement to become
certified; New customers = a potential new customer requires GFSI certification; business reputation= hope that GFSI will
improve business reputation in general; food safety = GFSI will improve the overall food safety program.
party schemes were collapsed
produced: high risk foods like raw meats, medium and low risk
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
45
Component 1 clearly showed that “an existing
customer requirement” was the primary reason for
becoming GFSI certified at 72.8%. The variation
among the companies making that ranking is shown
by the scatter of the vector arrows between
components. Together, both Components 1 and 2
accounted for more than 80% of the multivariate
reasons for companies to complete GFSI certification.
The arrows show clustering of companies with respect
to their prefered reason for becoming certified.
However, we were not able to further describe this
clustering of companies by testing the correlations
among the three company demographic items of:
scheme (BRC vs SQF), region (EU vs NA) or food
safety risk (high, medium or low). The MDPREF
analyses using the ranks and the other three weighted
systems produced similar graphs (not shown).
However, we believed there were additional insights
that could be garnered---so we explored another
statistical approach.
A Multiple Correspondence Analysis (MCA)
extended the Principal Component Analysis so that it
also graphically associated the reasons for companies
becoming GFSI certified with the three demographic
delinators: scheme, region, and risk (Fig. 3). Less
important reasons were collapsed into an “other
reasons” category for this analysis. Fig. 3 shows that
the BRC scheme was more closely associated with
European (EU) food manufacturers and the “improve
business reputation” and “potential new customer
requirement” reasons. The SQF scheme was most
closely associated with North American (NA) food
manufacturers and the “existing customer
requirement” reason. Additionally, high-risk food
products were most closely associated with
companies’ reason to “improve (their) food safety
program” in both geographic regions. With this MCA
analysis we have found that the use of a graphical
figure represented this data, allowing the reader to
make rapid associations. This reinforces the point of
needing to try both MDPREF and MCA to see which
produces a figure with the most information that is
easiest for the reader to interperet. Chen et. al. 2015
survey of New Zeland food producers similarly found
that the major reason for adopting food safety
management schemes was to satisfy the requirements
of a current customer.
3.3 Key Performance Indicators before and
after GFSI Certification
We used the Generalized Linear Model with the
multinomial ANOVA and a cumulative logistic link
function to estimate the amount of agreement on four
important questions about KPI’s sales/revenues,
numbers of customers, employees, and suppliers.
About one-fourth of the companies reported their
actual data before and after becoming certified and
that summary information is shown in the last column
of Table 2. There were several outliers that skewed
the initial analysis so only a percentage change in the
median is reported. It is important to point out that all
four of the KPIs improved substantially in the year
following becoming GFSI certified. Annual
sales/revenue for the companies reporting an
increase following certification was 52% with a
median 20 % increase. For the KPI, that was most
important to manufacturers in the EU, number of
customers increased for 43% of these companies
and they saw a median increase of 25% in their
number of customers. There was also a 17% median
increase in the number of employees for 40% for
these companies. The number of new suppliers
increased for 30% of the firms with a median
increase of 16%. Because of the additional analytics
we were able to present a much more concise and
compelling argument for GFSI certification to food
company managers than we presented in our
first paper [2], we included Fig. 4, heretofore, not
reported.
3.4 Time it Takes to Achieve Certification
Because the length of time (in months) for achieving
certification was non-normal and highly skewed, a
quantile regression analysis was used to estimate and
compare the 50 and 80% quantiles (Table 3). The
median value for the time it took to implement GSFI
was nine months with the mean of almost 10 months,
and a 95% confidence interval (CI) ranged from 9.5 to
10.4 months. A vast majority, 80% of the 543
respondents, succeeded in achieving GFSI
implementation of a single scheme in less than a year.
A robust analysis utilizing a Quantile Regression for
the 50
th
and 80
th
quantiles for each of the two major
schemes was necessary to minimize the effects of
extremes in reported values. This approach to the
analysis found that both quantiles took, on average, a
significantly longer period, two months longer than
the mean. Even when we conditioned the data to only
contain reports from only companies manufacturing in
N. A., we still found the same difference in both of
the two quantiles of SQF, requiring two more months
in comparison to BRC for the 50 and 80% of the NA
responders. Surprisingly, there were no significant
differences in time to become certified among the
demographics among the three risk categories. We
had hypothesized that companies producing high-risk
foods would be vastly more familiar with the third-
party certification process than the producers of low-
risk foods for whom this may have been their first
experience preparing for formal third-party
certification.
Fig. 4. Cha
nges in business profile benchmarks since GFSI certification
Table 2.
Comparison of key performance indicators (KPI’s) before and after certification
Freq
Annual sales/revenue 704
Number of Customers 754
Number of Employees 770
Number of Suppliers
754
Table 3. Quantile regression
relating the time (in months) for achieving certification, n=543
Quantile Scheme
50%
a
BRC
SQF
80% BRC
SQF
a
Median values and confidence intervals.
Mauromoustakos et al.
; JOBARI,
46
nges in business profile benchmarks since GFSI certification
Comparison of key performance indicators (KPI’s) before and after certification
Freq
Percentage of respondents saying Freq
Percent median
change
Increased Same Decreased
52
(49,
56) 40 8 (6, 10) 155
20
43
(39,
47) 49 8 (6, 10) 165
25
40
(37,
44) 46 14
(12,
17) 251
17
30
(27
34)
59
11
( 9.
13)
180
16
relating the time (in months) for achieving certification, n=543
Estimate Lower 95%
Upper 95%
8.0 7.1
8.94
10.0 9.2
10.80
12.0 10.7
13.26
14.0 12.9
15.08
Median values and confidence intervals.
British Retail Consortium, BRC;
Safe Quality Food, SQF
; JOBARI,
26(2): 39-52, 2020
Comparison of key performance indicators (KPI’s) before and after certification
Percent median
change
20
25
17
16
relating the time (in months) for achieving certification, n=543
Upper 95%
8.94
10.80
13.26
15.08
Safe Quality Food, SQF
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
47
Table 4a. Prediction modeling summary
Validation Creator Entropy R square Generalized R square Mean -Log p RMSE Mean Abs Dev Misclassification Rate N AUC
Training Boosted Tree 0.5069 0.5876 0.1609 0.2152 0.1089 0.0624 577 0.97
Training Bootstrap Forest 0.4601 0.5411 0.1761 0.2286 0.1221 0.0780 577 0.97
Training
Ordinal Logistic
0.4894
0.5693
0.1643
0.2251
0.0985
0.0640
406
0.94
Validation Boosted Tree 0.2460 0.3146 0.2654 0.2830 0.1416 0.1036 222 0.85
Validation
Bootstrap Forest
0.2580
0.3286
0.2612
0.2808
0.1502
0.0991
222
0.85
Validation Ordinal Logistic 0.0755 0.1008 0.3075 0.2947 0.1262 0.1034 145 0.84
Table 4b. Ordinal logistic
Source Log worth P-value
Compliance Improved preparation for current and future regulatory change 4.6
<0.0000
Compliance
Increased market share
3.6
0.0002
Intentions
Staff more aware of food safety KPIs
3.5
0.0003
Costs Higher production cost 2.5
0.0030
Costs External help 1.3
0.0485
Benefits Less rework in production 1.0
0.0911
Intentions Enhanced ability to produce safe food 0.5
0.3410
Benefits
Reduced the working capital cost
0.0
0.9516
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
48
Table 4c. Boosted tree
Source Splits G
2
Portion
Benefits Improved the quality of the food produced 18 5790
0.1933
Compliance Improved preparation for current and future regulatory change 13 2841
0.0948
Intentions Staff more aware of food safety KPIs 8 2664
0.0889
Compliance Increased market share 13 2605
0.0869
Benefits
Less rework in production
7
2501
0.0835
Intentions Less time spent complying with 3rd party audits 14 2208
0.0737
Intentions Enhanced ability to produce safe food 12 1915
0.0639
Costs Capital investment 5 1583
0.0528
Intentions Food safety management system more effective 7 1482
0.0495
Benefits Reduced the working capital cost 4 1065
0.0355
Benefits
Net savings in insurance costs
8
1012
0.0338
Intentions Fewer 3rd party food safety audits 7 948
0.0317
Costs Staff time investment 6 921
0.0307
Costs Changes to system 5 909
0.0304
Costs Higher production cost 4 612
0.0204
Compliance Compliance costs being offset by greater operating efficiencies 8 483
0.0161
Compliance
Improved regulatory compliance
4
194
0.0065
Benefits More consistency in operations and documentation 4 119
0.0040
Costs External help 1 76
0.0026
Benefits Improvements and time savings in internal audit 1 16
0.0005
Compliance Reduced costs of complying with safety regulations 1 16
0.0005
Compliance Reduction in corrective actions following food audits 0 0
0.0000
Intentions
Net savings of cost of 3rd party audits
0
0
0.0000
Benefits Savings in working capital 0 0
0.0000
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
49
Table 4d. Bootstrap forest
Source Splits G
2
Portion
Compliance Improved preparation for current and future regulatory change 25 12.5
0.1050
Intentions Staff more aware of food safety KPIs 24 11.7
0.0976
Intentions Food safety management system more effective 21 10.1
0.0844
Compliance Increased market share 27 9.3
0.0779
Benefits
Improved the quality of the food produced
18
8.9
0.0743
Intentions Enhanced ability to produce safe food 19 8.2
0.0687
Benefits Less rework in production 17 6.2
0.0520
Compliance Reduction in corrective actions following food audits 17 4.7
0.0397
Benefits Reduced the working capital cost 19 4.6
0.0384
Benefits Improvements and time savings in internal audit 21 4.3
0.0360
Intentions
Fewer 3rd party food safety audits
22
3.7
0.0311
Benefits More consistency in operations and documentation 14 3.5
0.0289
Intentions Less time spent complying with 3rd party audits 14 3.3
0.0280
Benefits Net savings in insurance costs 20 3.2
0.0268
Compliance Improved regulatory compliance 16 3.1
0.0262
Costs External help 18 3.0
0.0252
Intentions
Net savings of cost of 3rd party audits
18
2.8
0.0236
Costs Higher production cost 17 2.8
0.0234
Costs Staff time investment 17 2.7
0.0227
Costs Capital investment 15 2.5
0.0210
Compliance Reduced costs of complying with safety regulations 13 2.5
0.0208
Benefits Savings in working capital 17 2.2
0.0187
Compliance
Compliance costs being offset by greater operating efficiencies
16
2.0
0.0169
Costs Changes to system 15 1.5
0.0126
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
50
3.5 Predictive Factors that Lead to Beneficial
Certification
A logistic regression was used to predict whether the
perceived impacts (intentions, regulatory compliance
and costs and benefits opinions) can be used to predict
whether GFSI certification was/would be beneficial.
The assumption was that we could use a subset of the
results as inputs to the regression to predict the
likelihood that those respondents would choose to be
certified again. The inputs to the models were the 24
perceived impacts (those with answers on the 1-5
numeric scale opinions) as the input predictor
variables. A 20% subset of the data was partitioned
at random to be used as the validation data set,
which allowed us to use the remaining 80% as the
training to develop the regression model.
Three models were used to build the prediction
rules for the probability of GFSI being beneficial, a
logistic regression and two classification trees
techniques.
The model comparison platform in JMP Pro
summarized all the performance criteria of all three
models for both the training and the validation sets of
data (Table 4a). All three models performed “equally
well.” They all had an Area Under the Curve (AUC
>=0.94) and low, acceptable misclassification rates of
only 6-8% for the training data. As expected, when
the model was applied to the validation data, the AUC
was smaller (approximately 0.85) and the
misclassification rates increased (approximately 10%)
(Table 4a).
Stepwise Ordinal Logistic Regression (with the
minimum Akaike’s Information Criterion, AICc
stopping rule) selected a “best” subset of the model
effects without including additional variables that
only minimally improved the fit of the model. The
top four of the eight variables selected were highly
significant with logWorth values >1.4. LogWorth
values (-log10(P-value)) provided a useful way to
study p-values graphically in these cases. Those four
primary key drivers for this logistic regression
prediction model were: “improved preparation for
regulatory changes,” “increase market share,” “staff
awareness of KPI’s,” and “higher production costs”
(Table 4b).
The other candidate prediction models we used were
partition trees type of models using bagging
(Bootstrap Forrest) and boosting (Boosted Tree) that
try to improve on the original classification tree
techniques that used binary partitioning (Tables 4c
and 4d). It is important to note that top predictor
variable from the logistic regression did not appear in
the top 4 predictors in the Boosted Tree that used
many more inputs. However, the other top 3 of the 4
inputs agreed with the key contributors as shown by
their Column Contribution summaries. Lastly, the
Bootstrap Forrest model that resampled the
observations and the variables tended to allow
all the variables to enter for some of 100 simple
trees. Only two of its top 4 agree with those of the
logistic regression including their top input
contributor.
4. CONCLUSIONS
By sharing our experience of statistical analysis of a
large data set, we have tried to highlight more
powerful statistical analysis techniques that readers
can use to analyze their own survey data. The use of
Bradley Terry Model was able to do a paired wise
comparison that showed there was an 82% greater
probability that reason number one (an existing
customer required CFSI certification) was selected
first compared to the second or third reason.
Following this with the MCA analysis and
corresponding plots we provided graphs tha would
provide greater insight that would be quickly and
easily understood in an oral presentation.
The Generalized Linear Model with the multinomial
ANOVA and a cumulative logistic link function was
able to handle companies’ responses on the
improvements on their 4 KPI’s in the year following
GFSI certification. The model was able to
differentiate the association of each of the
important KPI’s with the scheme, region, and risk
categories. These associations may be further
described by investigating specific contrasts of
interest, such as investigating significant
differences between the BRC vs SQF schemes, North
America vs Europe regions, and high vs low-risk
products.
With the only continuous variable, the length of
time to achieve certification, the use of quantile
regression allowed us to minimize the effects of
extreme outliers. Using both the 50
th
and 80
th
quantiles confirmed the difference in time to
become certified between the two major GFSI
schemes.
The last part of the analysis demonstrated how to
utilize this large dataset to perform predictive
analytics. We attempted to build models that
addressed which and how the perceived impacts
(intentions, regulatory compliance and costs and
benefits opinions) could be utilized to predict if GFSI
would be beneficial to a new company considering
compliance. These models agreed that key
compliance reasons were to improve preparation for
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
51
current and future regulatory changes along with the
primary intention to make staff more aware of food
safety KPIs which are the most important for
important predictors for predicting beneficial GFSI.
It is hoped that the examples shown in the paper will
inspire the readers to use additional statistical analysis
of their own survey data sets.
DECLARATION
Partial funding for the statistical analysis was
provided by Walmart, Food Safety & Health and New
Diversey helped provide funds to collect the survey
data.
COMPETING INTERESTS
Authors have declared that no competing interests
exist.
REFERENCES
1. Anonymous. Google search of GFSI, Global
Food Safety Initiative; 2020.
Available:https://www.google.com/search?q=g
lobal+food+safety+initiative&rlz=1C1GCEU_
enUS819US820&oq=Global+Food+Safety+Ini
tiative&aqs=chrome.0.0l6.5238j0j8&sourceid=
chrome&ie=UTF-8
[Accessed 27 February 2020]
2. Crandall PG, Mauromoustakos A, O’Bryan
CA, Thompson KC, Yiannas F, Bridges NK,
Francois C. Impact of the global food safety
initiative on food safety worldwide: Statistical
analysis of a survey of international food
processors. Journal of Food Protection. 2017;
80:1613-1622.
3. Hensley S. Poll: Americans' concern about
food safety drops. NPR. 2011;4:19.
[PM ET. Site visited May 2018]
Available:https://www.npr.org/sections/health-
shots/2011/09/08/140291032/poll-americans-
concern-about-food-safety-drops
4. Lappo A, Bjorndal T, Fernandez-Polanco J,
Lem A. Consumers’ concerns and
external drivers in food marks. FAO
Fisheries and Aquaculture Circular No. 1102;
2015.
Available:http://www.fao.org/3/a-i4939e.pdf
5. Rohrich JD. The blog, Is our food safe?
Huffpost visited May 2018; 2017.
Available:https://www.huffingtonpost.com/jen
ny-dewey-rohrich/food-safety_b_8160054.html
6. Knoema. World exports and imports of
agricultural products; 2016.
Available:http://knoema.com/cduhihd/world-
exports-and-imports-of-agricultural-products
[Accessed 28 January 2016]
7. Branigan T. Chinese figures show fivefold
rise in babies sick from contaminated milk;
2008.
Available:https://web.archive.org/web/2008120
5093042/http://www.guardian.co.uk/world/200
8/dec/02/china
8. King T, Cole M, Farber J, Eisenbrand G,
Zabaras D, Fox E, Hill J. Food safety for
food security: Relationship between global
megatrends and developments in food safety.
Trends in Food Science and Technology.
2017;68:160-175.
9. Luning P, Kirezieva K, Hagelaar G, Rovira J,
Uyttedaele M, Jacxsens L. Performance
assessment of food safety management
systems in animal-based food companies in
view of their context characteristics: A
European study. Food Control. 2015;49:
11-22.
10. Song H, Turson R, Ganguly A, Yu K.
Evaluating the effects of supply chain quality
management on food firms’ performance. The
mediating role of food certification and
reputation. International Journal of Operations
& Production Management. 2107;37(10):1541-
1562.
DOI: 10.1108/IJOPM-11-2015-0666
11. Chen E, Flint S, Perry P, Perry M, Lau R.
Implementation of non-regulatory food safety
management schemes in New Zealand: A
survey of the food and beverage industry. Food
Control. 2015;47:569-576.
12. Stranieri S, Cavaliere A, Banterle A. Voluntary
traceability standards and the role of economic
incentives. British Food Journal. 2016;118(5):
1025-1040.
Available:https://doi.org/10.1108/BFJ-04-
2015-0151
13. Crandall P, O’Bryan C, Babu D, Jarvis N,
Davis M, Buser M, Adam B, Marcy J, Ricke S.
Whole-chain traceability, is it possible to trace
your hamburger to a particular steer, a U. S.
perspective. Meat Science. 2013;95(2):137-
144.
14. Head ML, Holman L, Lanfear R, Kahn AT,
Jennions MD. Perspective the extent and
consequences of p-hacking in science. PLOS
Biology; 2015.
DOI: 10.1371/journal.pbio.1002106 March 13,
2015
Mauromoustakos et al.; JOBARI, 26(2): 39-52, 2020
52
15. Crandall PG, Van Loo EJ, O’Bryan
CA, Mauromoustakos A, Yiannas F,
Dyenson N, Berdnik I. Companies’
opinions and acceptance of global food
safety initiative benchmarks after
implementation. J. Food Prot. 2012;75:1660-
1672.
Available:https://scholar.google.com/scholar?ci
tes=15393974083928811703&as_sdt=1005&s
ciodt=0,4&hl=en
16. Sall J, Stephens ML, Lehman A, Loring S.
JMP ®, Start statistics: A guide to statistics
and data analysis using JMP®. Sixth Edition;
2017.
__________________________________________________________________________________________
© Copyright International Knowledge Press. All rights reserved.
