International Journal of Industrial Ergonomics 66 (2018) 101e109

Contents lists available at ScienceDirect

International Journal of Industrial Ergonomics

journal homepage: www.elsevier.com/locate/ergon

A statistical study of occupational accidents in the manufacturing

industry in Turkey
Bulent Altunkaynak
Department of Statistics, Gazi University, 06500, Ankara, Turkey

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

Article history: One of the most important field of economy is manufacturing. The highest number of occupational
Received 29 August 2016 accidents occurs in manufacturing. In Turkey, more than half of these accidents occur in this field (56% in
Received in revised form 2012). This study aims to identify the contributory factors of such accidents and examine the relationship
17 January 2018
between Lost Work Days (LWD), Post-Accident State (PAS) and Type of Accidents (TA) using data mining,
Accepted 20 February 2018
correspondence analysis and chi-square test. The analyses have shown the most frequently encountered
TA is “pressing of the body or members between two objects” and its predictors are Sector, Size, and Age,
Experience, Hours Day. This type of accident is most likely to occur in sector 25 (manufacture of fabri-
Keywords:
Manufacturing industry
cated metal product, except machinery and equipment), with a higher likelihood for the businesses with
Accident statistics fewer than 500 employees and in the employees aged less than 40. As a result, the incapability of work is
Accident causes found to be as a period of 30e59 days. Consequently, it is recommended that higher premium ratios
Data mining should be imposed to urge the businesses to take measures and supervision practices by the social se-
curity supervisors should be more frequent.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction occupational accidents, exceeding the 1235 deaths in 2013. In the

Occupational accidents is one of the leading challenges most
sectors encounter in economy. Most laborers are faced with
numerous risks and the number of occupational accidents is higher
than reported in developed and developing countries (Ha € m€ €inen
ala
et al., 2006). 6300 people die every day as a result of occupational
accidents or work-related diseases e more than 2.3 million deaths
per year. Annually, 317 million accidents most of which bring about
prolonged absences from work occur. (ILO, 2015). Turkey in
particular ranks the third after El Salvador and Algeria in fatal ac-
cidents in the world and is at the top of the list of occupational
accidents in Europe. The rate of the fatal occupational accidents is
20.5 employees per 100,000 population in Turkey, but it is two
employees per 100,000 population in countries such as Norway,
Sweden, and Denmark (ILO, 2014). According to the Ministry of
Labor and Social Security, 172 daily occupational accidents occur in
Turkey, with four deaths and six workers becoming incapable of
work permanently (Ministry of Labor and Social Security, 2013).
According to the Council of Workers’ Health and Workplace Safety,
1270 workers lost their lives in the first eight months of 2014 due to
E-mail address: bulenta@gazi.edu.tr.
last decade, 6428 workers have lost their lives due to occupational
accidents. The number of the workers having lost their lives at work
between 2000 and 2014 is over 15000 (Occupational Health and
Safety Council, 2014).
The major reasons why the occupational accidents are so com-
mon in Turkey are the lack of effective policies, that of strict su-
pervision, the 2% premium set without considering the number of
the occupational accidents per company per year and with no
reference to the category and severity of the hazard (Altunkaynak
and Ozturk, 2016). However, countries such as Germany, France,
Belgium, and Italy use a premium system that incorporates rewards
and sanctions in the determination of the rate of premiums for
occupational injuries and diseases (Eurogip, 2004). Moreover, the
formal statistics concerning unregistered employment rate, the
number of occupational accidents and diseases are not reliable and
thought to reflect the less of what the actual situation is (Karadeniz,
2012).
The cost of occupational accidents is huge. At the country level,
the economic burden of poor occupational safety and health
practices is estimated at 4 percent of global Gross Domestic Product
each year (ILO, 2015). At the organization level, there emerges a
significant negative influence of occupational accidents on financial
performance of the company (Argiles-Bosch et al., 2014). Further, as

https://doi.org/10.1016/j.ergon.2018.02.012
0169-8141/© 2018 Elsevier B.V. All rights reserved.
102 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109
an outcome concerning organizational culture, workplace acci-
dents cause a perceived lack of influence and a distrust of man-
agement (Barling et al., 2003). These observations point to the
growing importance of workplace health and safety.
On the other hand, it is quite remarkable to note that 50% of the
occupational accidents can easily be prevented, 48% are prevent-
able with a systematic effort, but only 2% of the accidents cannot be
prevented (Disbudak, 2013). It is necessary to analyze occupational
accidents by type and causality to produce solutions for prevention.
This also has a positive effect on the selection and application of the
preventive measures.
A perusal of the literature demonstrates that occupational acci-
dents happen more frequently in sectors such as mining, industry,
construction and agriculture and more fatal accidents take place in
these sectors. Studies conducted in different occupational sectors
include mining (Maiti et al., 2009; Sanmiquel et al., 2015), industry
(Nenonen, 2011; Carrillo-Castrillo et al., 2012; Hedlund, 2013), con-
struction (Haslam et al., 2005; Hola and Szostak, 2014; Irumba, 2014)
and agriculture (Kumar and Dewangan, 2009; Robert et al., 2015).
Among these sectors, the manufacturing industry is one of the
most dangerous branches in light of the frequency of occupational
accidents (Nenonen, 2011). In 2007, the number of accidents causing
absence more than 3 days at manufacturing companies in EU-15
countries and Norway added up to 942,000 and 667 fatal accidents
were recorded in the manufacturing industry (Eurostat, 2009). In
Turkey, more than half of the occupational accidents occur in
manufacturing industry. However, there are very few studies inves-
tigating the accidents in manufacturing industry as a whole. Limited
number of current studies focus on the accidents in sub-sectors such
_
as metal and furniture industries in Turkey (Ilhan et al., 2013; Akalp
et al., 2015). However, there are some studies reporting occupa-
tional accidents in the manufacturing sector. As these studies deal
with only a small part of the manufacturing sector, they cannot be
used to compare the sectors by the number of accidents.
The manufacturing sector is heterogeneous, considering the
subsectors involved. The types of accidents with lost work days
may be variable by sector. However, the social security system does
not take this into account and imposes the same premium ratio
(2%) for all the sectors and companies. As a result, firms feel not
obliged to take measures to prevent occupational accidents.
Identifying the type of accidents and the outcomes of these ac-
cidents could help to find out which sector experiences relatively
more occupational accidents. However, the type of accidents may
not only be determined by sector but also by factors such as em-
ployees and the workplace conditions. As a result, if these factors can
be identified, the sources of the accidents could be ascertained. For
example, the accidents may be caused by inexperience on the part of
the employees, inadequate training, and poor working conditions.
Consequently, the current study has the following objectives:
To identify
1. the variables affecting the type of accidents in the
manufacturing sector
2. the most frequent type of accident
3. the outcomes of the most frequent type of accident
4. the antecedents of the most frequent type of accident
5. the variable levels that maximize the degree of probability of the
type of accidents
2. Material and methodology
Data mining techniques have been used in the present study to
meet the objectives of the study, even though the aims of data
mining, and some of its methods, overlap with those of classical
statistics. A great number of statistical models are available for the
statistician to explain relationships in a data set or to make pre-
dictions. Among these are cluster analysis, discriminant analysis
and nonparametric regression that can be used in data mining
problems. As a result, data mining seems, to the statistician, to be
only one of the branches of statistics (Hosking et al., 1997). How-
ever, the data mining problems and the methods employed have
their own characteristics. First, data sets can be far larger than it is
usually used in statistics. Second, differences of emphasis exist in
the approach to modeling: compared with statistics, data mining
pays less attention to the large-sample asymptotic properties of its
inferences and more to the general philosophy of “learning”,
including consideration of the complexity of models and of the
computations that they need. Some modeling techniques, such as
rule-based methods (decision tree, association rules etc.), are
difficult to fit into the classical statistical framework, and others,
such as neural networks, have an extensive methodology and ter-
minology that has developed largely independently of input from
statisticians. On the other hand, analysis of simulated data indicates
that, in general, methods based on data mining are more sensitive
in detecting associations among many categorical variables than
other statistics methods (Sanmiquel et al., 2015).
Powerful statistical tools, such as variable selection, decision
trees, Bayesian networks, association rules among other data
mining techniques are available for data analysis. In addition, some
of the classical statistical techniques, such as Chi-square, Z test, and
Correspondence Analysis can be effectively used in examining
uncomplicated relationships (e.g., examining the relationship be-
tween two variables) (Hosking et al., 1997).
This study has used data mining techniques and classical sta-
tistical techniques in combination for different purposes. First, us-
ing variable selection, we identified which variables had significant
effects on the type of accident. Thanks to variable selection, it is
possible to obtain better classification results using fewer variables.
At this stage, there are different evaluators and search methods
suggested in the literature. However, being heuristic, these
methods may yield conflicting results. Thus, it is necessary to try
more than one method simultaneously for the results to be reliable
in order to obtain similar results (Guyon and Elisseeff, 2003). This is
very important in terms of consistency of the results. Next, we used
decision trees to determine in which classes of the selected variables
for classification emerged what type of accidents. Decision trees are
functional if the response variable is categorical (e.g., type of acci-
dent in the current study) (Agrawal et al., 1993). There are a number
of algorithms like OneR, KNN, ID3, J48, NaiveBayes in building de-
cision trees when the response variables are categorical. Each of
these algorithms should be tried and the algorithm with the
highest accuracy for classification should be selected. In addition,
we used Bayesian networks to calculate probability for the nodes of
decision trees. Bayesian networks are used in association rules to
determine the rules of highest probability as well. As for classical
statistics, we used chi-square to test the relationship between the
two categorical variables (e.g., Type of Accidents and Lost Work
Days). In addition, when the relationship is found to be significant, a
more detailed explanation is necessary. Therefore, when we found
a significant relationship between variables, we used Z-test to
compare the proportions between classes. Finally, we preferred
correspondence analysis, whereby the relationship between classes
can be illustrated in a more visual way, to graphically show the
relationship between classes.
2.1. Study population
There are 39,327 instances of occupational accidents recorded in
the Turkish manufacturing sector in 2012. This data is concerned
 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109 103

with those employees with sick leave for one day or more. Data was trainees and Number of convicts are considered as only one attri-
obtained from the database on accidents of the Social Security bute because, when combined, they show the size. The attributes
Institution. The companies have to report any accident to the Social are explained in detail below.
Security Institution by filling out a form named “Occupational Ac- Frequency distributions to categorize the variables have been
cident Report Form”c used, which results in nonconventional intervals. The reason for
There are 22 subsectors in the manufacturing industry this is that if we had used such conventional intervals as 50 and
(Eurostat, 2008), and some of these sectors, such as tobacco, less ¼ small for firm size and 0e1 ¼ year for experience, the num-
pharmaceutical products and preparations, have a fewer number of ber of businesses in some intervals would have been too few.
accidents. As a result, the study considers those sub-sectors with Age (A): Age of the injured worker at the time of the accident.
1000 and more accidents because 96% of the total number of ac- Six classes are formed: [16, 24], [25, 29], [30, 34], [35, 39], [40, 44],
cidents in the manufacturing sector occurred in these sectors. [45, /].
Therefore, the study population is composed of 37,735 instances. Experience (E): Experience in this type of work (in years) of the
injured employee. Six classes are formed: [0, 5], [6, 10], [11, 15], [16,
20], [21, 25], [26, /].
2.2. Variables
Size (S): Number of employees in the company. The classes are:
[0, 500], [501, 1000], [1001, 1500], [1501, 2000], [2001, 2500], [2501,
The accident report consists of 39 attributes listed in Table 1.
3000], [3001, /].
This report can be downloaded from http://www.sgk.gov.tr/wps/
Work Hours (WH): How many hours the employee worked
wcm/connect/a899e80b-4f5c-4b8c-9a96-cfcf9215ea12/SGK-032.
before the accident. Six groups have been formed: (0, 1], (1, 4], (4,
xls?MOD¼AJPERES. As seen in the table, the information as to
8], (8, 10], (10, 12], (12,/].
Company ID, address, company name, and the like cannot be
Hour Day (HD): Hour of the day at the time of the accident. Four
counted as accident attributes because it does not provide identi-
classes are: (0, 6], (6, 12], (12, 18], (18, 24].
fying information regarding the accident. Although 39 variables
Gender (G): 1 ¼ Male; 2 ¼ Female.
were collected, only 12 of them were selected by using such criteria
Status at Workplace (SW): 1 ¼ Own-account worker;
as our previous experience and other results published on this
2 ¼ Worker by salary or daily wage; 3 ¼ Employer; 4 ¼ Unpaid
topic. Jeong (1997) investigated some characteristics in the occu-
family worker.
pational accidents of manufacturing industry and concluded that
Month Year (MA): Month of the year when injury happened
the company size and worker experience influenced the number of
1 ¼ January, …12 ¼ December.
accidents. In his 1999 study, he considered company size, injured
Sector (SE): 12 sub-sectors with annual 1000 and more acci-
person's age, work experience, accident time, activity at time of
dents according to NACE classification were selected for the study.
accident, accident type, injury type, injured part of body, and
(Eurostat, 2008). 10 ¼ Manufacture of food products; 13 ¼ Manu-
agency of accident as accident attributes. He found in both studies
facture of textiles; 16 ¼ Manufacture of wood and of products of
that the most frequent type of accident was pressing of the body or
wood and cork, except furniture; manufacture of articles of straw
members between two objects. Sanmiquel et al. (2015) listed the
and plaiting materials; 22 ¼ Manufacture of rubber and plastic
accident attributes as age of the injured worker, experience in this
products; 23 ¼ Manufacture of other non-metallic mineral prod-
kind of work (in months) of the injured employee, Number of
ucts; 24 ¼ Manufacture of basic metals; 25 ¼ Manufacture of
employees of the company, type of employment contract, existence
fabricated metal products, except machinery and equipment;
of a previous cause before the accident, place where the injury
27 ¼ Manufacture of electrical equipment; 28 ¼ Manufacture of
occurred, type of physical activity that the worker was making at
machinery and equipment n.e.c.; 29 ¼ Manufacture of motor ve-
the time of the accident, preventive organization in the company,
hicles, trailers and semi-trailers; 31 ¼ Manufacture of furniture;
accident risk, day of the week that injury happened, hour of the day
33 ¼ Repair and installation of machinery and equipment.
when injury happened, how many hours the employee worked
Type of Accident (TA): The cause of the accident. Six categories
before the accident and contractual status.
are 1 ¼ Accident caused by machinery or motor vehicle; 2 ¼ Acci-
In the current study, the accident attributes are shown with an
dent due to a sharp piercing device; 3 ¼ Pressing of the body or
asterisk (*). Number of men workers, Number of women workers,
members between two objects; 4 ¼ Falling from high places or
Number of child workers, Number of disabled workers, Number of

Table 1
Attributes in the accident report.

Company information Accident information Injured worker information

Company ID City ID
The name of tax office District Status at workplace *
The ID of tax office Date * Date of recruitment *
Company name Time * Date of birth *
Company address Type of accident * Place of birth
Company telephone Where, exactly, did it happen? Gender *
Company fax Number of injured person Marital status
Sector * Names of witnesses (if any) Education level
Number of men workers * Address
Number of women workers * Job
Number of child workers * Task
Number of disabled workers * Work hours *
Number of trainees * Lost work days *
Number of convicts * Post-accident state *
Working hours Nature of injury
Part of body affected

*The accident attributes considered in the analysis.

104 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109

falling on the same level; 5 ¼ Struck by falling objects; 6 ¼ others.

Fig. 1 shows the distribution of frequencies for this response.
Lost Work Days (LWD): Number of lost work days due to the
accident. The classes are: [1, 9], [10, 19], [20, 29], [30, 59], [60, /].
There were no given information on LWD for 141 accident in-
stances. Thus, LWD is set to be 37,594. Fig. 2 shows the distribution
of frequencies for this response.
Post-Accident State (PAS): The health status of the injured in
the post-accident period when the person starts work or dies.
1 ¼ healthy; 2 ¼ disabled 3 ¼ dead. Fig. 3 shows the distribution of
frequencies for this response variable.
9 of them were considered to be predictive variables and 3 of
them were response variables.
Fig. 1 shows that the most frequently encountered type of ac-
cident was pressing of the body or members between two objects
(22.39%). The accidents in others ranked second (18.79%), and ac- Fig. 2. Distribution of the Lost Work Days (n ¼ 37; 735).
cident caused by machinery or motor vehicle (18.20%) ranked third.
Fig. 2 illustrates the percentage of the number of the workers in
[1,9] lost work days (23.53%), [10,19] 24.54%, and 60 days or more
17.52%. Fig. 3 shows death rate due to accidents in the
manufacturing sector was 0.42%, rate of disability 2.82%, the
healthy 96.76%.

2.3. Statistical Procedure

The present study aims to identify the contributory factors to

the Type of Accidents and the factors which the type of accidents
brings about. The procedure followed to this end is fivefold:

1. To identify which type of accidents are more likely to occur

under which conditions
2. To apply the required methods to test the validity of results
3. To identify the effects of the type of accidents on lost work days
and Post-Accident State considering the type of accidents as
predictive variable
4. To test the differences between the classes of type of accidents
[1e6] and the classes of lost work days [1e5], and the differ-
ences between the classes of type of accidents [1e6] and the
classes of Post-Accident State [1e3]
5. To demonstrate the use of CA for the correspondences between
these classes
Weka v. 3-7-13 and R-project v. 3.2.2 software have been used
for the statistical analyses of the study. In the first step, using Weka,
we applied the data mining techniques of Attribute Selection, De-
cision Trees and Association Rules to identify the predictive
Fig. 1. Distribution of the Type of Accident (n ¼ 37; 735).
Fig. 3. Distribution of the Post-Accident State (n ¼ 37; 735).
variables that could be effective on Type of Accident (Bouckaert
et al., 2014). A similar application was carried out in mining
sector by Sanmiquel et al. (2015). In the second step, we considered
the Type of Accident as predictive of Lost Work Days and Post-Acci-
dent State. Using R-project, we applied chi-square to identify the
relationship between these variables (Venables et al., 2015), and
using z-test, we compared the rates of each class of lost work days
by all classes of type of accident and the rates of each class of Post-
Accident State by all classes of type of accident. P-values were
calculated using Benferroni Method. Using correspondence anal-
ysis, we analyzed the relationships between the classes of lost work
days and the classes of type of accident, and the relationships be-
tween the classes of Post-Accident State and the classes of type of
accident (Fig. 4).
In this study, we used the following procedures for statistical
analyses. Firstly, the predictive variables were selected using Select
Attributes in Weka and using the methods shown in Table 2.
Attribute Evaluator and Search Method forms the attribute selec-
tion process. Multiple techniques are used in each section. The
Attribute Evaluator is the technique by which each attribute in the
dataset is evaluated in the context of the response variable. The
Search Method is the technique to try or navigate different com-
binations of attributes in the dataset in order to arrive on a short list
of chosen features. Some Attribute Evaluator techniques require the
use of specific Search Methods. For example, the InfoGainAt-
tributeEval technique used in this study can only be used with a
Ranker Search Method that evaluates each attribute and lists the
results in a rank order.
 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109 105

Fig. 4. Attributes and statistical procedures.

Table 2 space of attribute subsets starting from the empty set of attributes.
Attribute evaluators, search methods and attribute selection modes for Type of The attribute evaluators InfoGainAttributeEval, ChiSquaredAt-
Accidents.
tributeEval, ClassifierAttributeEval, CfsSubsetEval, ClassifierSubsetEval
Attribute evaluators Search methods Attribute selection modes combined with Ranker, GreedyStepwise, GeneticSearch, BestFirst and
InfoGainAttributeEval Ranker 3, 5 and 10 cross-val. ExhaustiveSearch search methods have been used for Type of Acci-
ChiSquaredAttributeEval Full training set dent response. For the validity of evaluators, cross-validation (the
ClassifierAttributeEval number of folds n ¼ 3, 5 and 10 in this case) and full training were
CfsSubsetEval GreedyStepwise
used.
ClassifierSubsetEval GeneticSearch
BestFirst A data set can contain a large number of predictors. Some pre-
ExhaustiveSearch dictors are useful for predicting the response variable, and others
are not. Importance level is an indication of which predictors are
most useful for predicting the response variable. Importance level
2.3.1. Attribute evaluators is calculated through these steps:
InfoGainAttributeEval: Evaluates the worth of an attribute by
measuring the information gain with respect to the class. 1) The attribute evaluator is calculated for each predictor.
InfoGain(Class, Attribute) ¼ H(Class) - H(Class j Attribute). 2) The value for the percentage attribute evaluator is found for
The information gain (also called entropy) of each attribute can each predictor.
be calculated for the output variable. Entry values vary from 0 (no
information) to 1 (maximum information). Those attributes that For example, let the InfoGain value for the predictor i be IGi . The
contribute more information will have a higher information gain percentage InfoGain value is calculated for the predictor i PIGi using
value and can be selected. However, the attributes which do not this formula:
add much information will have a lower score and can be removed
(Kononenko and Hong, 1997).
.X 
ChiSquaredAttributeEval: Evaluates the worth of an attribute by PIGi ¼ 100 IGi IGi
computing the value of the chi-squared statistic with respect to the
class.
ClassifierAttributeEval: Evaluates the worth of an attribute by
using a user-specified classifier. 3) Finally, the importance level for each predictor is found by
CfsSubsetEval: Evaluates the worth of a subset of attributes by calculating the mean of the values for the percentage attribute
considering the individual predictive ability of each feature along evaluators (Kononenko and Hong, 1997).
with the degree of redundancy between them. Subsets of features
that are highly correlated with the class while having low inter- The analyses explained in Table 2 resulted in the selection of 5
correlation are preferred (Hall, 1998). predictive variables out of 9, the importance level of which are
ClassifierSubsetEval: Evaluates attribute subsets on training data presented in Fig. 5. The first 5 variables had an importance level of
or a separate hold out testing set. The 'merit' of a set of attributes is 92%; therefore, we formed association rules using these 5 variables.
estimated using a classifier. Secondly, a decision tree based on J48 algorithm is used for

2.3.2. Search methods

Search methods are used to identify the most irrelevant fea-
tures. This technique optimizes the time complexity (Onik et al.,
2015). Several search techniques can be used under a specific
attribute evaluator. Ranker worked under the Info gain Attribute
Evaluator. Ranker ranks attributes by their individual evaluations.
Greedy Stepwise performs a greedy forward or backward search
through the space of attribute subsets. The search starts with all
attributes or from an arbitrary point in space. It stops when the
deletion of any remaining attributes results in a decrease in the
evaluation. It also produces a ranked list of attributes and records
the order in which the attributes are selected. Genetic Search per-
forms a search using the simple genetic algorithm described in
Goldberg (1989). Best First Searches the space of attribute subsets
by greedy hill climbing augmented with a backtracking facility.
Exhaustive Search performs an exhaustive search through the Fig. 5. Ranking of the best predictor variables for Type of Accidents.
106 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109
classification. The C4.5 algorithm which is developed by Quinlan
(1993) for building decision trees is implemented in Weka as a
classifier called as J48. Using this classifier, we established a deci-
sion tree with a size of 4634 nodes and 3753 leaves. The correct
classification percentage of the decision tree was 71%. We selected
this tree because the other trees we tried had worse percentages.
Finally, the rules were obtained using Association Rules regarding
type of accident by the class of the predictive variables and selected
the best 10 of the rules (Table 3). The confidence or predictive ac-
curacy (at the end of each row) indicates the number of instances
for which all the conditions are true (the coverage), divided by the
number of instances for which the conditions in the antecedent are
true. Thus, for instance, the first row in Table 3 has to be interpreted
as: Sector ¼ 25, Size ¼ 1, Age ¼ 1, and Hour Day ¼ 3 0 Type of Acci-
dent ¼ 3, with a confidence level of 0.992. However, there is no
standard test to choose a specific rule. As the confidence level ap-
proaches to 1, the rule gets stronger. In this study, the rules that
have 95% or more confidence level are used.
3. Results and discussion
3.1. Evaluating the contributing variables to type of accident -
attribute selection
We have concluded that Work Hours, Month Year, Gender and
Status at Workplace have no significant effect on the type of acci-
dent, using the techniques in Table 2, while Sector, Size, Age, Hour
Day and Experience have significant effects on the type of accident.
As expected, Sector is found to be an important predictive var-
iable of Type of Accident. Just as the danger classes of sectors are
different, so are types of accidents (Soaita, 2014). This indicates that
it is impractical to use a fixed premium ratio for all sectors as it is
practiced in Turkey (2%) and it is imperative that stricter measures
need to be taken in those sectors with more economic loss by the
type of accidents. We recommend that the sectors with more loss of
work days and risks of accidents should be identified and higher
premium ratios should be imposed on them to force them to take
measures to prevent occupational accidents. However, this involves
the risk of understating the number of accidents on the part of
those firms, since reporting the accidents is the responsibility of the
firm in Turkey. As a result, a stricter control mechanism is
necessary.
Strong correlations exist between occupational accidents and
size (Jeong, 1997, 1999; Hunting and Weeks, 1993; Saari, 2005),
which is the same as it is in Turkey. This may stem from the fact the
control mechanisms depend on the size of the firms in Turkey. That
is, if the firm has fewer employees than 10, no control mechanism is
required by law. On the other hand, it is not true to claim that size is
the reason for the number of occupational accidents. It seems
necessary to consider size in combination with the other factors
(see association rules).
Table 3
The ten best rules for Type of Accident output variable.
Rule Sector Size Age
Rule 1 25 1 1
Rule 2 25 1 2
Rule 3 25 7 3
Rule 4 25 7 3
Rule 5 24 1 1
Rule 6 25 1 1
Rule 7 24 1 2
Rule 8 31 1 2
Rule 9 31 1 1
Rule 10 25 1 3
As for age, it has an effect on the levels of physical activity and
attention of the workers (Swaen et al., 2003; Safe Work Australia,
2012). In the manufacturing sector, accidents resulting in serious
injury occur at older ages (Jeong, 1999). On the other hand, inex-
perience and overconfidence in the young employees may be one
reason for accidents.
Experience at work increases with age. On the whole, age affects
risky behavior, which could be seen in the rate of accidents at work
(Unsar and Sut, 2009; Mehrdad et al., 2014). The frequency distri-
bution of experience may prove that young workers (<35 years old)
face occupational accidents because of their lack of experience
(Alizadeh et al., 2015). Hour Day has a significant effect on the type
of accident, with more accidents occurring during daytime than
during nightshifts (06:00e18:00) (Macedo and Silva, 2005; Wig-
glesworth E., 2006; Alizadeh et al., 2015).
3.2. The contributing variables to the most frequently encountered
type of accident e association rules
The association rules indicate which accident types co-occur
with which variables handled in the study. The results of the as-
sociation rules show that the major contributory factors to
“pressing of the body or members between two objects” are Sector,
Size, and Age, which are included in all the rules. The probability of
the emergence of this type of accident is the highest in sector 25,
which is one of the most dangerous branches of the manufacturing
sector. This type of accident is more likely to occur in the businesses
of sector 25 with fewer than 500 employees and in the employees
aged less than 40.
3.3. The relationships between Type of Accident and Lost Work Days
and Post-Accident State e chi-square and CA
As a result of the study, we found a significant relationship be-
tween Type of Accident and Lost Work Days (Table 4; c220 ¼ 637:22;
P < .001). The comparison results of the columns in Table 5, for
example, show that the rate of LWD ¼ 1 ([1, 9]) when TA ¼ 6
(others) is 30.2%, which is significantly higher than the other Type of
Accident classes. An examination of crosstabs points out that the
relationship between Type of Accident and Lost Work Days is sta-
tistically significant. It is seen that Type of Accident does not occur at
the same rate as each class of Lost Work Days. For example, TA ¼ 2
(accident due to a sharp piercing device) in lost work days [10, 19]
and [20, 29] are 27.7% and 19.6%, respectively, which is higher than
the rates for other types of accidents. TA ¼ 3 (pressing of the body
or members between two objects) in lost work days [30, 59] is
higher than the rates for other types of accidents. When lost work
days are 60 days and more, the rate of TA ¼ 4 (falling from high
places or falling on the same level) is higher (22.8%). The closest
classes to LWD ¼ 5 ([60, /]) are TA ¼ 4 (Falling from high places or
falling on the same level) and TA ¼ 1 (accident caused by machinery
Experience Hour Day Type of Accident Confidence level
e 3 3 0.992
2 e 3 0.988
3 e 3 0.987
1 2 3 0.980
1 2 4 0.975
1 e 3 0.968
2 2 4 0.962
e 3 3 0.961
e 3 3 0.958
e 3 3 0.950
 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109 107

Table 4
Crosstabs for type of accident and lost work days.

Lost Work Days Type of Accident Total

1 2 3 4 5 6

1 Count 1451a 1235b 1712a 867a.b 1455b 2125c 8845

2 % 21.3 24.2 20.3 21.9 23.4 30.2 23.5
3 Count 1587a.b 1418c 1878b 911a.b 1578a.c 1852c 9224
4 % 23.3 27.7 22.2 23.0 25.4 26.3 24.5
5 Count 1160a.b 1002c 1497b.c 647a.b 1105b.c 1117a 6528
Total % 17.0 19.6 17.7 16.3 17.8 15.9 17.4
Count 1200a 798a.b 1717c 632a.b 1017a.b 1048b 6412
% 17.6 15.6 20.3 16.0 16.4 14.9 17.1
Count 1422a.b 660c 1642b 902a 1057d 902c 6585
% 20.9 12.9 19.4 22.8 17.0 12.8 17.5
Count 6820 5113 8446 3959 6212 7044 37594
% 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Each subscript letter denotes a subset of Lost Work Days categories whose column proportions do not differ significantly from each other at the .05 level.

Table 5
Crosstabs for type of accident and post-accident state.

Post-Accident State Type of Accident Total

1 2 3 4 5 6

1 Count 6451a 5023b 8262b 3844c 6071b.d 6861c.d 36512

2 % 93.9 98.2 97.8 96.4 97.5 96.8 96.8
3 Count 364a 90b 180b.c 114c 139b.c 179b.c 1066
Total % 5.3 1.8 2.1 2.9 2.2 2.5 2.8
Count 52a 0b 8b.c 30a 16c 51a 157
% 0.8 0.0 0.1 0.8 0.3 0.7 0.4
Count 6867 5113 8450 3988 6226 7091 37735
% 100.0 100.0 100.0 100.0 100.0 100.0 100.0

Each subscript letter denotes a subset of Result of Accident categories whose column proportions do not differ significantly from each other at the .05 level.

or motor vehicle). These types of accidents have the highest rate of
lost work days. Consequently, they should be examined closely for
prevention.
The relationship between Type of Accident and PAS has been
found to be statistically significant (Table 5; c210 ¼ 293:83;
P < .001). It is seen that Type of Accident does not occur at the same
rate as each class of Post-Accident State. Considering the accidents
resulting in death, the rates for TA ¼ 1 (accident caused by ma-
chinery or motor vehicle), TA ¼ 4 (falling from high places or falling
on the same level) and TA ¼ 6 (others) are 0.8%, 0.8%, and 0.7%,

Fig. 6. Biplots for Lost Work Days and Type of Accidents (a), Biplots for Type of Accidents and Post-Accident State (b).
108 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109
respectively, which are higher than the rates for other Type of Ac-
cident. As for the accidents resulting in disability, the rate of TA ¼ 1
(accident caused by machinery or motor vehicle) is significantly
higher than the rates of the other Type of Accident.
As a result of the correspondence analysis, considering the most
frequent type of accident (pressing of the body or members be-
tween two objects), the incapability of work is found to be as a
period of 30e59 days (Fig. 6a Region I). In the case of “falling from
high places or falling on the same level,” the incapability of work is
found to be as a period of 60 or more days, which is more than the
period in other types of accidents (Fig. 6a Region III). These two
types of accidents are predictable by the other variables in associ-
ation rules analysis, as a result of which, for instance, the “pressing
of the body or members between two objects” is more likely to
occur in sector 25, while the “falling from high places or falling on
the same level” is more likely in sector 24 (Manufacture of basic
metals). Gulhan et al. (2012) have found similar results, concluding
that the primary reason for these types of accidents is employee
carelessness.
4. Conclusions
37,735 instances of occupational accidents recorded in the
Turkish manufacturing sector in 2012 have been analyzed in this
study with the objective to identify the main factors causing
manufacturing injuries by using statistical techniques. The study
shows that the variables Sector, Size, Age, Experience and Hour Day
are in the occurrence of most accidents. In addition, Type of Accident
as a predictor variable both for Lost Work Days and Post-Accident
State as outputs have served to measure the severity of the acci-
dents. In sector 24 and sector 25, more precautions to prevent ac-
cidents seem necessary as they have the highest number of
accidents. Therefore, it is recommended that higher premium ratios
be imposed to urge the businesses to take measures, and supervi-
sion practices by the social security supervisors should be more
frequent.
This study has covered the manufacturing sector in Turkey,
which has the biggest share in national income and the highest
number of occupational accidents. Similar studies with the meth-
odology followed in the present study could be performed for the
other sectors.
The limitations of the study include, firstly, the incom-
prehensiveness of the data due to the unregistered nature of
Turkish economy and employment. Secondly, there is no variable to
determine whether the accident is caused by human error or poor
safety conditions. Thirdly, there are not enough additional variables
concerning physical conditions to analyze the accidents thoroughly.
Finally, the data for 2014 and 2015 are still unavailable.
Acknowledgments
The authors would like to thank the anonymous reviewers and
editors for their valuable comments that greatly helped to improve
the quality of the paper.
We also would like to thank the Social Security Institution for
allowing us to use the Turkish general database of accident. We
€
would like to thank Pınar Oztürk for her comments on the technical
definitions of occupational accidents and her interpretations as an
expert in social security.
References
Agrawal, R., Imielinski, T., Swami, A., 1993. Database mining: a performance
perspective. IEEE Trans. Knowl. Data Eng. 5, 914e925.
Akalp, G., Aytac, S., Yamankaradeniz, N., Cankaya, O., Gokce, A., Tufekci, U., 2015.
Perceived safety culture and occupational risk factors among women in metal
industries: a study in Turkey. Procedia Manufacturing 3, 4956e4963.
Alizadeh, S.S., Mortazavi, S.B., Sepehri, M.M., 2015. Analysis of occupational accident
fatalities and injuries among male group in Iran between 2008 and 2012. Iran.
Red Crescent Med. J. 17, 1e6.
Altunkaynak, B., Ozturk, P., 2016. How to minimize occupational accidents in
Turkey: the Bonus-Malus System. Journal of Social Insurance 9, 1e22.
Argiles-Bosch, J.M., Marti, J., Monllau, T., Garcia-Blandon, J., Urgell, T., 2014.
Empirical analysis of the incidence of accidents in the workplace on firms'
financial performance. Saf. Sci. 70, 123e132.
Barling, J., Kelloway, E.K., Iverson, R.D., 2003. Accidental outcomes: attitudinal
consequences of workplace injuries. J. Occup. Health Psychol. 8, 74e85.
Bouckaert, R.R., Frank, E., Hall, M., Kirkby, R., Reutemann, P., Seewald, A., Scuse, D.,
2014. WEKA Manual for Version 3-7-12.
Carrillo-Castrillo, J.A., Rubio-Romero, J.C., Onieva, L., 2012. Causation of severe and
fatal accidents in the manufacturing sector. Int. J. Occup. Saf. Ergon. 19,
423e434.
Disbudak, Z., 2013. The Situations in Which Nurses Are Injured with Sharp Tools
and Needlesticks and the Methods They Followed after the Injury. Gaziantep
University, Graduate School of Health Sciences. Master Thesis.
Eurogip, 2004. Costs and Funding of Occupational Diseases in Europe. Eurogip,
Paris.
Eurostat, 2008. NACE Rev. 2 statistical classification of economic activities in the
european community. In: Eurostat Methodologies and Working Papers,
ISSN1977e0375.
Eurostat, 2009. Number of Accidents at Work by Economic Activity and Size of
Enterprise. Retrieved from. http://nui.epp.eurostat.ec.europa.eu/nui/show.do?
dataset¼hsw_aw_nnasz&lang¼en.
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization, and Machine
Learning. Addison-Wesley, Reading, MA.
Gulhan, B., Ilhan, M.N., Civil, E.F., 2012. Occupational accidents and affecting factors
of metal industry in a factor in Ankara. Turk. J. Pub. Health 10, 76e85.
Guyon, I., Elisseeff, A., 2003. An Introduction to variable and feature selection.
J. Mach. Learn. Res. 3, 1157e1182.
Hall, M.A., 1998. Correlation-based Feature Subset Selection for Machine Learning.
Hamilton, New Zealand.
Ha€m€al€
ainen, P., Takala, J., Saarela, K.L., 2006. Global estimates of occupational ac-
cidents. Saf. Sci. 44, 137e156.
Haslam, R., Hide, S., Gibb, A., Gyi, D., Pavitt, T., Atkinson, S., Duff, A., 2005.
Contributing factors in construction accidents. Appl. Ergon. 36, 401e415.
Hedlund, F.H., 2013. Recorded fatal and permanently disabling injuries in South
African manufacturing industry - overview, analysis and reflection. Saf. Sci. 55,
149e159.
Hola, B., Szostak, M., 2014. Analysis of the development of accident situations in the
construction industry. Procedia Eng. 91, 429e434.
Hosking, J.R.M., Pednault, E.P.D., Sudan, M., 1997. A statistical perspective on data
mining. Future Generat. Comput. Syst. 13, 117e134.
Hunting, K.L., Weeks, J.L., 1993. Transport injuries in small coal mines: an explor-
atory analysis. Am. J. Ind. Med. 22 (3), 391e406.
Ilhan, A., Kosar, G., Karapinar, A., Gedik, T., 2013. Analysis of underlying causes of the
occupational accidents and diseases in furniture manufacturing in Sakarya
province. Kastamonu University. J. For. Fac. 13, 202e210.
ILO, 2014. Occupational Injuries. Retrieved from: http://www.ilo.org/ilostat.
ILO, 2015. Retrieved from: http://www.ilo.org/global/topics/safety-and-health-at-
work/langeen/index.htm.
Irumba, R., 2014. Spatial analysis of construction accidents in Kampala, Uganda. Saf.
Sci. 64, 109e120.
Jeong, B.Y., 1997. Characteristics of occupational accidents in the manufacturing
industry of South Korea. Int. J. Ind. Ergon. 20, 301e306.
Jeong, B.Y., 1999. Comparisons of variables between fatal and nonfatal accidents in
manufacturing industry. Int. J. Ind. Ergon. 23, 565e572.
Karadeniz, O., 2012. Occupational accidents and diseases, and the deficiency of
social protection in Turkey and the world. Lab. Soc. 3, 15e75.
Kononenko, I., Hong, S.J., 1997. Attribute selection for modelling. Future Generat.
Comput. Syst. 13, 181e195.
Kumar, G.V., Dewangan, K.N., 2009. Agricultural accidents in north eastern region of
India. Saf. Sci. 47, 199e205.
Macedo, A.C., Silva, I.L., 2005. Analysis of occupational accidents in Portugal be-
tween 1992 and 2001. Saf. Sci. 43, 269e286.
Ministry of Labor and Social Security, 2013. Labour Statistics. Retrieved from: http://
www.csgb.gov.tr/csgbPortal/ShowProperty/WLP%20Repository/csgb/dosyalar/
istatistikler/CalismaHayati-2013.
Maiti, J., Khanzode, V., Ray, P., 2009. Severity analysis of Indian coal mine accidents:
a retrospective study for 100 years. Saf. Sci. 47, 1033e1042.
Mehrdad, R., Seifmanesh, S., Chavoshi, F., Aminian, O., Izadi, N., 2014. Epidemiology
of occupational accidents in Iran based on social security organization database.
Iran. Red Crescent Med. J. 16, 1e5.
Nenonen, S., 2011. Fatal workplace accidents in outsourced operations in the
manufacturing industry. Saf. Sci. 49, 1394e1403.
Occupational Health and Safety Council, 2014. In 2014, at Least 1886 Workers Lost
Their Lives. Retrieved from: http://www.guvenlicalisma.org/icerik/haber/
dosyalar/2014.pdf.
Onik, A.R., Haq, N.F., Alam, L., Mamun, T.I., 2015. An analytical comparison on filter
feature extraction method in data mining using J48 classifier. Int. J. Computer
Appl. 124, 1e8.
 B. Altunkaynak / International Journal of Industrial Ergonomics 66 (2018) 101e109
Quinlan, R., 1993. C4.5: Programs for Machine Learning. Morgan Kaufmann Pub-
lishers, New York.
Robert, K., Elisabeth, Q., Josef, B., 2015. Analysis of occupational accidents with
agricultural machinery in the period 2008e2010 in Austria. Saf. Sci. 72,
319e328.
Safe Work Australia, 2012. Australian Work-related Injury Experience by Sex and
Age 2009-10. Retrieved from: http://www.safeworkaustralia.gov.au/sites/SWA/
about/Publications/Documents/705/Australian_work-related_injury_
experience_by_sex_and_age_2009-10.PDF.
Saari, J., 2005. Small and medium enterprises. Technical conference: Prevention of
Occupational Hazards in SMEs. Asepeyo, Seville, pp. 11e20.
109
Sanmiquel, L., Rossell, J.M., Vintro, C., 2015. Study of Spanish mining accidents using
data mining techniques. Saf. Sci. 75, 49e55.
Soaita, D., 2014. Workplace accidents dynamics and structural changes in economy.
Procedia Technol. 12, 500e504.
Swaen, G., Amelsvoort, L.G.P.M., Bultmann, U., Kant, I., 2003. Fatigue as a risk factor
for being injured in an occupational accident: results from the Maastricht
Cohort Study. Occup. Environ. Med. 60, 88e92.
Unsar, S., Sut, N., 2009. General assessment of the occupational accidents that
occurred in Turkey between the years 2000 and 2005. Saf. Sci. 47, 614e619.
Venables, W.N., Smith, D.M., R Core Team, 2015. An Introduction to R, Notes on R: a
Programming Environment for Data Analysis and Graphics Version 3.2.2.