AUTCON-02217; No of Pages 9

Automation in Construction xxx (2017) xxx–xxx

Contents lists available at ScienceDirect

Automation in Construction

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

Investigation of the causality patterns of non-helmet use behavior of

construction workers

Heng Li a, Xiaoying Li a, Xiaochun Luo a, Joanna Siebert b,⁎

a
Department of Building and Real Estate, The Hong Kong Polytechnic University, Hong Kong
b
Department of Computing, The Hong Kong Polytechnic University, Hong Kong

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

Article history: Not wearing a safety helmet can result in serious injuries on construction sites, but the cause of such unsafe be-
Received 22 February 2017 havior is not yet well understood. This paper describes a practical field study to find the factors that influence
Accepted 28 February 2017 non-helmet use on construction sites using a real-time tracking system (the Eye on Project or EOP), which is de-
Available online xxxx
veloped by our research team to provide an objective record of helmet use; and which overcomes the deficiencies
of traditional recording methods (i.e. self-reporting). The application of association rules then combines desig-
Keywords:
Non-helmet use
nated risk factors and finds the causality patterns of non-helmet use. Based on the analyses, this paper develops
Association rules a method of evaluating workers' risk level involving the intensity of multiple risk factors.
Risk assessment © 2017 Elsevier B.V. All rights reserved.
Unsafe behavior
Construction safety

1. Introduction
The construction industry is considered to be one of the most haz-
ardous and dangerous industries; there is a high accident rate within
the field [21]. According to the Health and Safety Executive [22], con-
struction workers suffer 10% of major injuries and 31% of fatal injuries.
At same time, out of almost 4000 worker fatalities in private industry
in a calendar year 2013, 796 or 20.3% were in the construction industry
[36].
The International Labor Organization [24] and the Bureau of Labor
Statistics [8] also published reports confirming a high fatality rate in
the construction industry. Sousa et al. [40] stated that compared with
workers in other industries, the construction workers are under a
higher probability suffer from potential injuries and even die at work.
In China alone, the State Administration of Work Safety reported that
at least 2197 deaths occurs on construction site and almost 90% of the
construction accidents are caused by human factors(i.e. unsafe behav-
ior) 2014,the average death on construction site from 2005 to 2104
also reach at a formidable amount which is almost 2600 per year [23].
To enhance construction safety, a series of measures has been put
forward to guide workers to wear helmets by both the law and con-
struction managers. Kelm et al. [26] stated that employers should in-
crease helmet use in three ways: (1) education and training, (2)
incentives, and (3) enforcement. Despite large efforts and prevention
⁎ Corresponding author.
E-mail addresses: heng.li@polyu.edu.hk (H. Li), xiaoying.li@connect.polyu.hk (X. Li),
eric.xiaochun.luo@polyu.edu.hk (X. Luo), Joanna.Siebert@polyu.edu.hk (J. Siebert).
techniques for accidents, many studies highlight the remaining factors
that reduce safety on sites.
Head injuries are one of the most severe kinds of injury possible, as
the head is both fragile and prone to collision [33]. On construction
sites, traumatic brain injury is usually caused by falls and trench/scaffold
collapse. Therefore, helmet is an important personal protect equipment
(PPE) on construction site. Despite the great importance of helmet use,
data collection of helmet use on construction sites is still in a relatively
early stage. Previous studies always focus on the helmet use behavior it-
self and its contributing factors. Individual factors, such as gender, age,
work experience and time of day have an obvious impact on unsafe be-
havior. Onsite data has always been collected after the occurrence of in-
cidents and, until now, research has been forced to rely on data analysis
and feedback. Current hazard-identification technology has obvious
limitations [9]. Although modern eye–tracking technology can measure
and analyze the eye position and movement of workers onsite [18], pre-
vious studies have failed to create a technology to supervise non-helmet
use. This technology would be especially useful since helmet misuse/ne-
glect is one of the most common forms of equipment misuse that can
occur on construction sites, and can therefore be easily compared with
other unsafe behaviors (i.e. loitering around dangerous areas). This
paper provides such a technology: a real-time tracking system known
as the “Eye on Project” (EOP), to provide an objective record of helmet
use in construction sites. This improved hazard-identification technolo-
gy can automatically monitor and record helmet misuse.
Another considerable improvement is made in this paper: previous
research was mostly based on the reporting of accidents or near-acci-
dents, which usually just consisted of a description of the individuals

http://dx.doi.org/10.1016/j.autcon.2017.02.006
0926-5805/© 2017 Elsevier B.V. All rights reserved.

Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
2 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx
involved and the accident itself [32]. In addition, the frequency of inci-
dents was also often used as data. The real-time tracking safety helmet
designed by the researchers can identify and store the data by monitor-
ing real-time non-helmet use behaviors, and then to improve the cur-
rent hysteresis analysis based on accident reports.
This paper reports a field experiment investigating workers' com-
mon characteristics of non-helmet use and then assessing workers'
risk levels on construction sites. Association rules were used as a data
analysis method to find the obscure combination of risk factors hidden
in the data collected. Many data mining techniques have been used for
safety analysis in recent years [13,19,20]. Meanwhile, a risk assessment
matrix system was obtained by the above association rules with levels
classified by characteristics. The significance of this paper is to propose
a more intuitional way to find and demonstrate risk contributors.
The paper is organized into four sections: (1) review of previous re-
search on helmet use including the importance of helmet use, as well as
current methods and technologies for helmet use inspection; (2) an ex-
perimental design comprising of the Eye on Project (EOP) system and a
method of defining and evaluating workers' risk levels based on the
common characteristics of non-helmet use; (3) experimental settings,
procedure and results; and (4) discussion of the findings summarized
from the previous results, the limitations of the current research and fu-
ture research avenues.
2. Literature review
2.1. Individual factors influencing unsafe behaviors on construction sites
Since non-helmet use is a common unsafe behavior in site condi-
tions, it is important to refer to the contributory factors influencing un-
safe behaviors. Previous studies classified the key factors leading to
unsafe behaviors on construction sites. According to previous studies,
one of the most significant factors influencing unsafe behaviors is indi-
vidual characteristics [25]. Seven items were identified from previous
studies, namely gender, age, work experience, time of day, attitude
and motivation, psychological distress and intended acts. The related
items of literature are listed in Table 1.
2.2. Importance of helmet use for construction workers
Helmets are an important way of reducing fatal injuries and evaluat-
ing workers' safety performance on construction sites. Head injuries are
a common source of trauma in the workplace worldwide, since the
human head is the part of body with the highest potential for serious in-
jury and even death [33]. Medical Online, for example, states that about
230,000 Americans suffer traumatic head injuries each year, with more
than a fifth dying. Similarly, the Center for Disease Control and Preven-
tion [10] estimates that head injuries account for almost half (49%) of
fatal injuries. Moreover, a survey focused on worksite accidents and in-
juries collected by the Bureau of Labor Statistics [8] demonstrated that
not wearing head protection results in almost 90% of all traumatic
brain injury. This is especially true in the case of the construction indus-
try, which has the highest rate of traumatic brain injury of all industries.
Therefore, on-site construction workers are required to wear a helmet
or a hardhat on site.
Table 1
Literature related to individual characteristics.
No Individual contributory factors Studies
1 Gender [4,7] [15,34]
2 Age [2,4,7] [12]
3 Work experience [3,14,42] [43]
4 Time of day [28]
5 Attitude and motivation [11,21]
6 Psychological distress [44]
7 Intended acts [39,42]
Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
Safety performance is a key factor influencing the decision to evalu-
ate construction workers' risk level and adopt further safety manage-
ment on those of high risk. On the other hand, non-helmet use is a
direct safety performance indicator for predicting the occurrence of in-
cidents. Safety helmets are one of the most widely used pieces of per-
sonal protective equipment (PPE) which help to mitigate the severity
of injuries to humans in hazardous conditions [31,41,44]. Government
departments have published a series of booklets to provide safety-relat-
ed regulations and rules to reduce safety helmet misuse. Employees in
the U.S., for instance, are protected by the Occupational Safety and
Health Administration [35], which has developed a series of regulations
to ensure that employers provide appropriate head protection and bear
the responsibility of supervising its use in conditions where objects
might fall from above and strike workers on the head. Over the past de-
cade, several studies concerning the evaluation of wearing PPEs as a
safety performance indicator have been reported, which vary in their
purpose and focus.
2.3. Current methods and technologies on helmet use inspection
Previous researchers have attempted to figure out the major causes
of helmet misuse on construction sites [45]. However, it is hard to obtain
data about helmet misuse (i.e. time spent without wearing a helmet, pe-
riods of the day where it is more likely to occur, and correlation between
personality and helmet misuse). Due to the hysteresis of accident pre-
diction and warning functions in the construction sector, the research
is always based on accident reports instead of real-time data. For exam-
ple, statistics of the lack of PPE use during daily construction processes
have always emphasized self-reports of construction staff by construc-
tion site managers [39]. Incident reporting systems (IRSs) are also wide-
ly used for post hoc analysis, which provides proactive analysis for
safety management [38]. The current analysis is mainly based on
reporting that cannot visually display workers' unsafe behaviors and
their updated risk. Therefore, current studies may fail to provide effec-
tive safety analysis for a complex industry such as construction.
Because of the danger present on construction sites, the use of safety
protection equipment, such as PPE, has gained much of attention from
researchers. The current method of worker supervision using PPEs is
simply visual surveillance by supervisors or construction managers.
However, this method is ineffective and time-consuming, since such
surveillance is executed only at scheduled times. In the past few de-
cades, studies have moved from this manual method of supervision to
the use of advanced remote sensing, which negates the need for
human interference entirely. Kelm et al. [26], for example, use various
existing commercial automated identification (ID) and information
technologies (IT) to design a mobile RFID to check the use of PPE by
workmen. Barro-Torres et al. [5] have introduced an advanced cyber-
physical system (CPS) to check in real time whether a PPE is worn by
workers based on an architecture composed of a wireless local area net-
work and a body area network. Since the helmet is a widely used piece
of personal protective equipment (PPE) on construction sites that can
directly reduce the risk of head injury or prevent workers from injury
from falling items [16,41,44], proactive research is urgently needed to
promote the workers' safety [27].
Therefore, it is necessary to develop methods and technologies to in-
spect helmet use based on real-time and visual data. Meanwhile, perfor-
mance should be not only evaluated in terms of the behavior itself, but
also from the workers' common characteristics of non-helmet use. This
study aims to investigate the evaluation of non-helmet use in both fre-
quency and duration.
3. System framework and methodology
In this section, the EOP is used as the non-helmet-use behavior in-
spection system in an experiment. The features and operational proce-
dure are introduced in detail. Then, a method of evaluating workers'
 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx
risk levels is proposed based on the system's features. The framework of
the system implementation procedures is demonstrated in Fig. 1.
3.1. Factors designated in the study
In Step 1, the paper mainly focuses on four risk factors: gender, age,
experience and time of day. “Sex-linked differences” are always found
related to occupational accidents and the accident rates of men and
women workers show that women suffer fewer accidents than men
[34]. Age is regarded as a significant contributor to construction injuries.
Cheng et al. [12] state that the most commonly injured group are people
aged 35–44 years (31%), while it is apparent that workers older than 55
and younger than 24 are the most likely to be involved in fatal accidents.
Given that experience represents the workers' own proficiency, [3]
shows that poor work experience is a major factor affecting the fatal ac-
cident rate on construction sites. Meanwhile, López et al. [28] analyze
the severity of occupational accidents suffered by construction workers
at different hours of the day and illustrate that the interval of time be-
tween 13:00 and 17:00 has a significantly higher rate of severe and
fatal accidents than any other period; this is termed the “lunch effect”.
3.2. Non-helmet-use behavior inspection system: EOP
3.2.1. System overview
In Step 2, a non-helmet-use behavior inspection system called the
Eye on Project (EOP) was developed and applied. The EOP was mainly
designed to collect the time data of workers' helmet use (including fre-
quency and duration) and analyze their behavior patterns when they
are occupied in their own jobs. The real-time data collection and
uploading system was built with Bluetooth technology. As noted in
the literature review above, radio frequency (RF)-based technologies,
Fig. 1. Framework of the implementation procedure.
Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
3
such as RFID, are popularly used to record and upload data currently col-
lected about safety helmets. However, these technologies are too bulky
to carry everywhere during construction. Moreover, they consume a
large amount of power, which makes them impractical for long-term
usage. The tradeoff between different characteristics such as range, ac-
curacy, ease of deployment on construction sites, costs of purchase,
use and maintenance must also be considered [30].
The EOP developed in this study is divided into two parts: (1) a
Bluetooth endnote device and (2) a data collector with a low-power
Bluetooth protocol. The Bluetooth device with a coin cell battery is at-
tached to the inside shell of the safety helmet and connected to a sili-
cone single-point detector placed on the sweat pads of the safety
helmet. Meanwhile, to protect workers from suffering potential discom-
fort, the Bluetooth device has been produced to be very small and thin.
Therefore, the device is still suitable for workers' daily on-site work
compared with the other technologies mentioned earlier. Although a
Bluetooth device is attached to the inside shell of the helmet, it will
not affect physical health. Moreover, the privacy of workers is also con-
sidered. Detection of safety helmet use occurs when the detector is
touched. In this way, the device can capture the period the helmet is
worn and store the data in its memory for regular transmission to an ex-
ternal location. The Bluetooth device attached to the safety helmet is
shown in Fig. 2.
Fig. 3 represents the framework of the EOP. The data collector can be
fixed in any place and is mainly used to establish an automatic connec-
tion with the device, read the data stored in the device and then trans-
mit the data to a server by GPRS. Under normal circumstances, the auto-
connection between the device and collector can be established due to
the wide communication coverage of low consumption Bluetooth. The
process moves in circles. When data appears in the memory of the de-
vice within scanning range, the collector reads and uploads it. It then
4 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx
Fig. 2. Structure of the safety helmet.
re-scans continuously until the next broadcast within the scanning
range. To negate the effects of false touches, an exception is made
when a connection of only a few seconds is sensed.
3.2.2. System operational procedure
A typical operational procedure of EOP consists of three stages: prep-
aration, running and data analysis. In preparation, (1) the data collector
is placed in a fixed area for workers passing on time (such as the con-
struction site entrance), and the data transmission distance from end-
note to collector is 30 m (if workers' active areas exceed this range, it
is necessary to deploy more than one collector in the site); (2) the
workers' personal information (including ID, age, gender and work ex-
perience) are established and updated, the characteristics can be differ-
entiated by workers' ID; and (3) the relationship between endnote and
endnote carriers is established through the association of worker-ID and
helmet-ID, which can realize the combination between workers' per-
sonal characteristics and their performance on helmet use.
The data collection was realized by the record of the collector phys-
ical ID, the helmet physical ID and the time of helmet on and off with
EOP. Part of the helmet use records from EOP is shown in Fig. 4. When
running EOP, we can obtain the time data of helmet use by inspecting
record by record. A single helmet record contains the data collector ID,
Fig. 3. Framework of the EOP.
Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
Helmet ID, duration of use and duration of non-helmet use for one
time. The duration of non-helmet use is equal to the difference between
the {Take-off Time} for this record and the {Take-on Time} for the next
record during their working time. Thus, the frequency and duration of
the workers' non-helmet use can be acquired during the period of
data collection.
Finally, when helmet use records and worker information have been
collected, the workers can be categorized into risk levels according to
the frequency and duration of non-helmet use and by using association
rules to explore the characteristics and contributors involved. More de-
tails of this procedure are given in the sections below. These character-
istics are the foundation of the worker's risk level assessment during the
construction period. This study provides an example of exploring the
use of this data.
3.3. A method for assessing workers' risk level
The data analysis of non-helmet use can be divided into two steps:
(1) using the association rules to analyze the characteristics of workers
who neglect helmet use; and (2) using the risk assessment matrix to
evaluate the risk levels of workers; workers who have higher risk
being the target during construction work.
3.3.1. Association rule
Association rules were proposed to underline groups of correlated
variables that typically occur together defined on transaction. Currently,
the association rule is widely used to delve into relationships of vari-
ables from big databases, and to explore potential associations [13]. In-
dependent variables are combined stochastically. The strongest rule is
identified by using the malleable association rules. The antecedent is
the input controllable variable and the consequent is the variable
which are supposed to be predicted, the relation always meet the
form “If antecedent, then consequent” [29],
In this paper, an algorithm by Agrawal et al. [1] is adopted and used
to analyze the collected data. In furtherance of finding the valid associ-
ation rules in a transactional dataset, a specified minimum support and
specified minimum confidence are supposed to be designed. The rela-
tionship between the variables will be strong if they meet the threshold
of minimum support or minimum confidence.
The details of the support-confidence framework following the orig-
inal definition by Agrawal can be shown as follows. Let Ι ¼ fi1; i2; i3;…; im g
be a set of binary attributes (also named items). Let D ¼ ft 1; t 2; t 3; …; t m
 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx 5

Fig. 4. Example of helmet use records from EOP.

g be a set of variable transactions (also called the transaction database),
where each transaction Τ is a set of items such that Τ ⊆ Ι. Therefore, every
transaction in D contains a subset of the items in Ι. At meanwhile, each
Association rule consists of two different sets of items (also named item
sets) Χ and Y, where Χ is called the antecedent or left-hand-side (LHS)
and Y is the consequent or right-hand-side (RHS). The association rule
is the implication that Χ ⇒ Y where Χ , Y ⊆Ι and Χ ∩ Y ≠∅. The significance
of association rules is usually measured by three indices: support, lift
and confidence [6]. The probability that Χ ⇒ Y holds in D is defined
as support:
Support ðΧ⇒Y Þ ¼ Support ðΧ∪Y; D:Þ
The confidence of a rule is the probability which a rule can be true.
Confidence can be defined as follows:
ConfidenceðΧ⇒Y Þ ¼ Support ðΧ∪Y; DÞ∕Support ðΧ; DÞ
The lift lift(Χ ⇒ Y) is the ration of the observed support value to ex-
pected support value.
Lift ðΧ⇒Y Þ ¼ Support ðΧ∪YÞ∕ðSupport ðΧ ÞSupport ðY ÞÞ
(severity). However, this risk assessment method is always focused on
the risk of an activity, which represents the amount of injuries that is ex-
pected to occur as the result of a potential accident associated with an ac-
tivity. In present research, we intend to analyze the characteristics of
workers that lead to non-helmet use and assess their risk level based on
the helmet use data. Therefore, the risk of non-helmet use means the
number of non-helmet use in terms of frequency and duration.
Two-dimensional risk assessment matrices have been widely used
to define different levels of risk. The color codes in this matrix corre-
spond to various levels of risk (low, medium and high). However, this
matrix has been criticized for being too subjective and qualitative. In
this research, assigning numerical values for frequency and duration
based on the association rules can provide quantitative risk values for
the matrix.
4. Pilot studies
4.1. Overview of the experimental construction
The non-helmet use behavioral data was collected for the Shenzhen
project “The Peninsula” applied with the Eye on Project (EOP). The

Table 2
Furthermore, the three indices above demonstrate the strength of Descriptive statistics of non-helmet use data from EOP.
such an association rule and the stronger association can be represented
by a higher index [35]. Characteristic Level Description Count Percentage

To set a threshold of such an association rule, the minimum support Gender male GEN1 14 72.36%
σ and minimum confidence δ has been set respectively. The association female GEN2 5 27.64%
Age b24 AGE1 1 5.26%
rule must meet the requirement:
24–34 AGE2 3 15.79%
35–44 AGE3 9 47.37%
Support ðΧ⇒Y Þ≥σ 45–54 AGE4 5 26.32%
≥55 AGE5 1 5.26%
and Work experience b1 EXP1 2 8.94%
1–5 EXP2 3 16.67%
ConfidenceðΧ⇒Y Þ ≥δ 6–10 EXP3 6 30.49%
N10 EXP4 8 43.90%
Time of day for frequency [9:00, 12:00] TIM 1 35 7.32%
of non-helmet use (12:00, 15:00] TIM2 345 71.36%
3.3.2. Risk assessment (15:00, 18:00] TIM3 103 21.32%
Risk assessment has traditionally involved quantifying the risk of an Time of day for duration [9:00, 12:00] TIM1 204 13.61%
incident based on two or more variables, such as the probability of a of non-helmet use (12:00, 15:00] TIM2 841 56.10%
(15:00, 18:00] TIM3 454 30.29%
risk (frequency) and the impact or consequence of the risk occurring

Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
6 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx

Table 3
Overlapped rules from the frequency and duration of non-helmet use.

Rule ID Consequent Antecedent Frequency Duration

Support (%) Confidence (%) Lift Support (%) Confidence (%) Lift

1 EXP1 GEN2 26.91 72.55 1.40 28.69 56.33 1.03

2 EXP1 GEN2 and TIM2 21.17 81.24 1.56 21.16 56.38 1.03
3 TIM2 GEN2 and AGE4 19.84 90.19 1.13 12.27 72.56 1.06
4 TIM2 AGE3 and EXP1 18.85 84.09 1.06 19.68 69.74 1.02
5 EXP1 GEN2 and AGE4 and TIM2 17.90 90.01 1.73 8.90 59.20 1.09
6 GEN1 AGE5 and EXP1 9.63 100.00 1.37 3.25 100.00 1.40
7 EXP2 TIM3 and AGE4 7.35 53.51 1.29 9.12 58.25 1.53
8 GEN1 TIM3 and AGE4 7.35 73.51 1.01 9.12 71.36 1.00
9 GEN1 AGE2 6.28 100.00 1.37 11.07 87.00 1.22
10 EXP2 TIM3 and AGE4 and GEN1 5.40 50.00 1.21 6.51 50.34 1.32
11 EXP2 TIM3 and GEN2 4.57 74.78 1.80 5.91 55.81 1.46
12 AGE3 GEN2 and EXP2 and TIM2 3.97 55.00 1.83 9.17 52.90 1.34
13 GEN1 EXP3 3.91 100.00 1.37 6.77 99.02 1.39
14 GEN1 AGE2 and EXP1 3.42 100.00 1.37 5.11 100.00 1.40
15 EXP1 TIM1 and GEN1 3.36 81.07 1.56 3.96 64.80 1.19
16 AGE4 TIM1 and EXP1 and GEN1 2.72 62.04 1.17 2.57 55.17 1.53
17 GEN1 EXP3 and TIM2 2.70 100.00 1.37 5.16 98.71 1.38
18 EXP1 TIM1 and AGE3 2.07 82.69 1.59 2.08 68.09 1.25
19 GEN1 EXP3 and AGE3 2.05 100.00 1.37 5.20 100.00 1.40
20 GEN1 TIM1 and AGE4 1.97 100.00 1.37 2.72 72.36 1.02
21 EXP1 TIM1 and AGE4 and GEN1 1.97 85.86 1.65 1.97 71.91 1.32
22 GEN1 AGE2 and EXP1 and TIM2 1.89 100.00 1.37 2.50 100.00 1.40
23 GEN2 TIM1 and AGE3 and EXP1 1.71 68.61 2.55 1.42 60.94 2.12
24 GEN1 AGE2 and TIM3 and EXP2 1.53 100.00 1.37 2.52 72.81 1.02
25 GEN1 AGE2 and TIM3 and EXP1 1.53 100.00 1.37 2.61 100.00 1.40
26 GEN1 EXP3 and AGE3 and TIM2 1.27 100.00 1.37 4.07 100.00 1.40
27 AGE3 EXP3 and TIM3 1.21 63.93 2.12 1.62 69.86 1.78
28 GEN1 EXP3 and TIM3 1.21 100.00 1.37 1.62 100.00 1.40
29 AGE3 EXP3 and TIM3 and GEN1 1.21 63.93 2.12 1.62 69.86 1.78
30 AGE3 TIM1 and GEN2 1.17 100.00 3.32 1.62 53.43 1.36
31 EXP1 TIM1 and GEN2 1.17 100.00 1.92 1.62 100.00 1.83
32 AGE3 TIM1 and GEN2 and EXP1 1.17 100.00 3.32 1.62 53.43 1.36

project is an on-going 44,481.2 m2-floor space housing estate with a
construction period from May 2015 to December 2017 (32 months). A
total of 43 workers participated in the second and third construction
phases of the project from the 23 November 2015 to 21 January 2016.
19 workers worked on-site without head protection at least once in
the experiment, resulting in 483 incidents recorded by EOP. Table 2
summarizes the categories of workers' characteristics in the
experiment.
4.2. Non-helmet use workers' characteristics
To create association rules for the characteristics and contributory
factors of non-helmet use in the EOP database, the researchers carried
out an a priori algorithm with 9 sets of data. According to the principle
of association rules, the first was to set the minimum value of support,
confidence and lift based on the actual situations. Workers' characteris-
tics were expressed by rules as the combinations of different character-
istics. In this investigated project, the daily lunch rest period was 12:00–
13:00, and the helmet and non-helmet use data during this period has
already been excluded in the data analysis section.
However, the resulting rules for frequency and duration of non-hel-
met use were diverse and independent, which produced two evaluated
results from the two dimensions of frequency and duration. Next, we se-
lected rules that simultaneously exhibit the frequency and duration re-
sults. In this study, a total of 32 rules were obtained as shown in Table 3.
frequency and duration can represent the probability of repetitive
non-helmet use. Thus, the matrices of frequency and duration descrip-
tors are the same, as shown in Table 4. In Table 4, the green cells show
the unlikely probability levels, yellow cells correspond to occasional
levels of probability, while the orange cells are the rules that occur
sometimes with likely probability levels and the red cells represent
the most frequent levels of probability. According to Tables 4, 32 rules
are classified into frequent, likely, occasional and unlikely levels from
the frequency and duration of non-helmet use.
We combined the 32 rules in Table 5 from the frequency and dura-
tion dimensions of non-helmet use. Table 6 shows the risk levels and
color codes based on the frequency and duration, which represent the
impact of the four factors (i.e., age, experience, gender, and time-of-
day) on workers' non-helmet use. Risk is classified into extreme, high,
moderate and low levels. The green cells show the risk levels, yellow
cells correspond to moderate levels of risk, while orange cells denote
Table 4
Frequency or duration descriptors for non-helmet use.
Support (%)
[1,2] [2,4] [4,7] [7,100]
Lift [1.4,100] Occasional Likely Frequent Frequent

4.3. Workers' risk level assessment [1.2,1.4] Unlikely Occasional Likely Frequent

To obtain a more accurate result from the association rules, we se-

[1.1,1.2] Unlikely Unlikely Occasional Likely
lected two indices: support and lift, to overcome the inaccuracies of a
single index. In this case study, support and lift intervals from the two
dimensions (frequency and duration) are set according to the same reg- [1.0,1.1] Unlikely Unlikely Unlikely Occasional
ulation. The support and lift intervals both range from 1 to 100. Both

Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx 7

Table 5 However, age was found to be a critical factor determining the sever-
Thirty-two rules classified by frequency levels. ity of an accident [4]; younger workers were more vulnerable to acci-
Level Rule ID for frequency Rule ID for duration dents and injuries due to such psychological and physical symptoms
Frequent 1; 2; 5; 6; 7; 11 6; 7; 9; 10; 11; 12
as worse understanding of safety programs [11], safety attitude [37],
Probable 3; 9; 10; 12; 15; 18 13; 14; 16; 17; 19 and the distance to managers [2]. Meanwhile, other studies have identi-
Occasional 4; 8; 13; 14; 17; 19; 21; 23; 27; 1; 2; 3; 4; 5; 8; 18; 22; 23; 25; 26; fied (lack of) work experience as a key contributory factor to occupa-
29; 30; 31; 32 27; 28; 29; 31; tional accidents [14,42], with the conclusion that work experience is
Remote 16; 20; 22; 24; 25; 26; 28 15; 20; 21; 24; 30; 32
linked to duration of job tenure and has a negative correlation with un-
safe behavior [43].
The current data collected from EOP are limited and may not be very
high risk and red extreme levels of risk. For example, the rule that is effective for discovering any universal laws. However, this pilot study is
classified as a extreme level of risk is the combination of the frequency mainly focuses on providing an approach to investigate the non-hel-
level and a duration level. High risk level events are those with associa- met-use behavior causality patterns of construction workers and the
tion rules classified as both frequent and unlikely. The 32 rules are proposed approach's successfully identification of the combination of
reclassified by the risk levels and listed in Table 7 according to the clas- simplex individual factors based on the current data is a significant im-
sification of risks in Table 6. Based on this result, the workers who ne- provement on previous studies.
glect to use helmets can easily be found.
In the extreme-risk section, it is apparent that {EXP1} is the conse-
5. Discussion
quence of the highest proportion of non-helmet use. {GEN2} and
{EXP1} have a stronger relationship than other rules, i.e. 1, 2 and 5.
This is the first comprehensive study to assess worker's risk levels
The extreme level of risk indicates that workers with these characteris-
based on their characteristics of non-helmet use during their work on
tics are more prone to experience construction accidents. The high risk
site. Organizations and industries can implement the proposed method
category provides another consequence where being a male ({GEN1})
to assess risks in the workplace by quantitative means. An important
worker is the biggest contributory factor to helmet misuse. Workers
consideration concerns the use of the results of the two dimensions of
with less than five years of work experience ({EXP1}) and work be-
frequency and duration to explore the characteristics of workers who
tween 9:00 am and 12 am ({TIM1}) are factors of approximately equal
neglect helmet use easily and to make decisions that improve their safe-
impact, while {TIM1} accounts for a high average support value, which
ty to prevent the occurrence of fatal injuries. With the help of these ma-
can be found at rule 3 and rule 4. Workers with a high level of risk per-
trices, the common characteristics of workers with poor safety
form an influential role on the whole work team, as they tend to affect
performance levels can be centralized.
their colleagues' behavior.
This study demonstrates the necessity for using real-time data in-
There are several comprehensive factors that mutually influence the
stead of logging data after an accident occurs (e.g. incident report).
moderate level, such as less than five years of work experience
This needs to target workers who are at extreme risk of non-helmet
({EXP1}), workers aged from 35 to 44 years old ({AGE3}), male workers
use during certain times to encourage them in appropriate helmets
({GEN1}); and {AGE3} leadership of all the impacts. Moderate level
use and participate in safety training programs.
workers are stable and less prone to taking dangerous actions. These
In this study, a data mining approach was proposed to investigate
workers are considered minimal risk actors. The low level is composed
non-helmet-use behavior causality characteristics and contributory fac-
of a singular component of male workers ({GEN1}). This is due to the
tors. Previous studies focus on an individual's characteristics and view
skill (or lack thereof) of these workers in the pilot study.
the found characteristics as the decisive reasons. Trying to find a single
Compared with the workers that have the patterns of combination
cause of dangerous behavior on-site is unlikely to succeed, as the situa-
individual factors leading to unsafe behavior, the characteristic of 24
tion is far too complex and multifaceted. Therefore, this paper has dealt
workers who are following the rule and wearing their helmets are con-
with the situation comprehensively in a manner that acknowledges
verging on the following aspects: below 35 years old ({AGE1}, {AGE2})
multiple factors. This paper is mainly focused on the mining association
and with over 5 years work experience ({EXP3}, {EXP4}). Moreover,
rule between sets of workers' characteristics of non-helmet use. The
those working after 3 pm tend to pay more attention to helmet use.
progressive methods proposed can automatically obtain important
The practical results of this study differ from previous findings. For
characteristics and contributory factors with the records of non-helmet
example, the finding that female workers are at a higher non-helmet
use. By using the two evaluation indexes as support and lift, more accu-
use risk level combined with other characteristics (such as b 5 years'
rate information of the combination of workers' characteristics can be
work experience). On the contrary, the previous findings always show
obtained to analyze the potential causes of serious accidents.
that male workers have higher fatal occupational injury rates than fe-
The test has shown the potential of the developed technology in im-
male workers, which is consistent with similar studies elsewhere [7,
proving construction safety culture and morale. Specifically, the use of
15,17].
helmet icons and symbols might affect how a construction worker
with diversified backgrounds (e.g., illiterate, foreign nationals) can be
educated or trained more effectively and efficiently. Therefore, the pro-
posed inspection system can also be adopted well within a construction
Table 6
Risk assessment matrix for non-helmet use rules. site through a simplified and mobile approach.

Frequency
6. Conclusion
Frequent Likely Occasional Unlikely
This paper has investigated the contributory factors influencing non-
Duration Frequent Extreme Extreme High Moderate
helmet use on construction sites and provides an association-rules
Likely Extreme Extreme High Moderate based approach for figuring out the relationship between the contribut-
ing factors and non-helmet use behaviors. The hidden relationships of
Occasional Extreme High Moderate Low
the non-helmet use and contributing factors has been identified. The
Unlikely High Moderate Low Low discovered causality patterns where extracted from Statistic Package
STATISTICA DATA MINER V8.0.

Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
8 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx

Table 7
Thirty-two rules classified by risk levels.

Risk Rule ID Consequent Antecedent Risk Rule ID Consequent Antecedent

Extreme 1 EXP1 GEN2 Moderate 15 EXP1 TIM1 and GEN1

2 EXP1 GEN2 and TIM2 16 AGE4 TIM1 andEXP1 and GEN1
5 EXP1 GEN2 and AGE4 and TIM2 21 EXP1 TIM1 and AGE4 and GEN1
6 GEN1 AGE5 andEXP1 23 GEN2 TIM1 and AGE3 andEXP1
7 EXP2 TIM3 and AGE4 27 AGE3 EXP3 and TIM3
9 GEN1 AGE2 29 AGE3 EXP3 and TIM3 and GEN1
10 EXP2 TIM3 and AGE4 and GEN1 30 AGE3 TIM1 and GEN2
11 EXP2 TIM3 and GEN2 31 EXP1 TIM1 and GEN2
12 AGE3 GEN2 andEXP2 and TIM2 32 AGE3 TIM1 and GEN2 andEXP1
High 3 TIM2 GEN2 and AGE4 Low 20 GEN1 TIM1 and AGE4
4 TIM2 AGE3 andEXP1 22 GEN1 AGE2 andEXP1 and TIM2
8 GEN1 TIM3 and AGE4 24 GEN1 AGE2 and TIM3 andEXP2
13 GEN1 EXP3 25 GEN1 AGE2 and TIM3 andEXP1
14 GEN1 AGE2 andEXP1 26 GEN1 EXP3 and AGE3 and TIM2
17 GEN1 EXP3 and TIM2 28 GEN1 EXP3 and TIM3
18 EXP1 TIM1 and AGE3
19 GEN1 EXP3 and AGE3

All the findings are facilitated by the real-time tracking system (EOP)
that helps to establish a risk assessment matrix between leading factors
and non-helmet use behavior and advise construction managers or
workers for purpose of preventing the causality patterns in future.
This proposed approach not only provides a brand-new method for
identifying factors related to unsafe behavior on construction sites, but
also developing more efficient and accurate risk assessment strategies.
The final analysis of data mining results will be used to provide and stip-
ulate safety rules in construction site.
Currently, the real-time Eye on Project (EOP) still has some limita-
tions. First, it acts as a limited identification function that recognizes
data concerning age, gender, individual experience, time and non-hel-
met use, however other potential factors (such as real-time location)
may be hindered. Meanwhile, the current association rule mining pro-
cess can be improved to discover unknown relationships, which also in-
dicates the need to improve the current data collection method of EOP
in the next stage. Furthermore, other sections of the presented work
have shown the potential for providing workers with real-time feed-
back on their helmet misuse. The current small sample size provided
in the pilot study is also limited in its ability to reveal any generalizable
laws relating to helmet use and non-use.
For future studies, more construction programs need to be involved
to reduce the limitations of the current singular case study; the factors
and association rules at the organizational level that also affect non-hel-
met use should be considered; more integrated individual factors and
other contributory factors (i.e. site condition, supervision, project man-
agement) need to be gradually taken into account; and a set of universal
rules that can provide guidance in advance and avoid the occurrence of
unsafe behavior in construction work is needed.
Funding sources
The work was supported by the Research Grants Council of Hong
Kong grant titled “Proactively Monitoring Construction Progress by In-
tegrating 3D Laser-scanning and BIM” (PolyU 152093/14E).
Acknowledgements
Many thanks go to Shuyi Huang for her support in the experiment
and Xintao Yang for his work in the development of the hardware.
References
[1] R. Agrawal, T. Imieliński, A. Swami, Mining association rules between sets of items in
large databases, ACM SIGMOD Rec. 22 (2) (1993) 207–216, http://dx.doi.org/10.
1145/170036.170072.
[2] R. Alsamadani, M.R. Hallowell, A. Javernick-Will, J. Cabello, Relationships among lan-
guage proficiency, communication patterns, and safety performance in small work
crews in the United States, J. Constr. Eng. Manag. 139 (9) (2013) 1125–1134,
http://dx.doi.org/10.1061/(asce)co.1943-7862.0000724.
[3] A.L. Arquillos, J.C.R. Romero, A. Gibb, Analysis of construction accidents in Spain,
2003–2008, J. Saf. Res. 43 (5) (2012) 381–388, http://dx.doi.org/10.1016/j.jsr.
2012.07.005.
[4] N. Azadeh-Fard, A. Schuh, E. Rashedi, J.A. Camelio, Risk assessment of occupational
injuries using accident severity grade, Saf. Sci. 76 (2015) 160–167, http://dx.doi.
org/10.1016/j.ssci.2015.03.002.
[5] S. Barro-Torres, T.M. Fernández-Caramés, H.J. Pérez-Iglesias, C.J. Escudero, Real-time
personal protective equipment monitoring system, Comput. Commun. 36 (1)
(2012) 42–50, http://dx.doi.org/10.1016/j.comcom.2012.01.005.
[6] S. Brin, R. Motwani, C. Silverstein, Beyond market baskets: generalizing association
rules to correlations, ACM SIGMOD Record, 26, ACM 1997, pp. 265–276.
[7] S.M. Buckley, D.J. Chalmers, J.D. Langley, Falls from buildings and other fixed struc-
tures in New Zealand, Saf. Sci. 21 (3) (1996) 247–254, http://dx.doi.org/10.1016/
0925-7535(95)00068-2.
[8] Bureau of Labor Statistics, Revisions to the 2009 Census of Fatal Occupational Inju-
ries (CFOI) counts, http://www.bls.gov/iif/oshwc/cfoi/cfoi_revised09.pdf 2009
(accessed June 3, 2016).
[9] G. Carter, S.D. Smith, Safety hazard identification on construction projects, J. Constr.
Eng. Manag. 132 (2) (2006) 197–205, http://dx.doi.org/10.1061/(asce)0733-
9364(2006)132:2(197).
[10] Center for Disease Control and Prevention, Chronic diseases: the leading causes of
death and disability in the United States, http://www.cdc.gov/chronicdisease/over-
view/ 2011 (accessed May 24, 2016).
[11] Q. Chen, R. Jin, A comparison of subgroup construction workers' perceptions of a
safety program, Saf. Sci. 74 (2015) 15–26, http://dx.doi.org/10.1016/j.ssci.2014.11.
021.
[12] C.-W. Cheng, S.-S. Leu, Y.-M. Cheng, T.-C. Wu, C.-C. Lin, Applying data mining tech-
niques to explore factors contributing to occupational injuries in Taiwan's construc-
tion industry, Accid. Anal. Prev. 48 (2012) 214–222, http://dx.doi.org/10.1016/j.aap.
2011.04.014.
[13] C.-W. Cheng, C.-C. Lin, S.-S. Leu, Use of association rules to explore cause–effect re-
lationships in occupational accidents in the Taiwan construction industry, Saf. Sci.
48 (4) (2010) 436–444.
[14] C.-F. Chi, T.-C. Chang, K.-H. Hung, Significant industry–source of injury–accident
type for occupational fatalities in Taiwan, Int. J. Ind. Ergon. 34 (2) (2004) 77–91,
http://dx.doi.org/10.1016/j.ergon.2004.03.002.
[15] C.-F. Chi, T.-C. Chang, H.-I. Ting, Accident patterns and prevention measures for fatal
occupational falls in the construction industry, Appl. Ergon. 36 (4) (2005) 391–400,
http://dx.doi.org/10.1016/j.apergo.2004.09.011.
[16] K. Cloute, A. Mitchell, P. Yates, Traumatic brain injury and the construction of iden-
tity: a discursive approach, Neuropsychol. Rehabil. 18 (5–6) (2008) 651–670, http://
dx.doi.org/10.1080/09602010701306989.
[17] A.E. Dillingham, Sex differences in labor market injury risk, Ind. Relat. 20 (1) (1981)
117–122, http://dx.doi.org/10.1111/j.1468-232x.1981.tb00187.x.
[18] R.-J. Dzeng, C.-T. Lin, Y.-C. Fang, Using eye-tracker to compare search patterns be-
tween experienced and novice workers for site hazard identification, Saf. Sci. 82
(2016) 56–67, http://dx.doi.org/10.1016/j.ssci.2015.08.008.
[19] K. Geurts, I. Thomas, G. Wets, Understanding spatial concentrations of road acci-
dents using frequent item sets, Accid. Anal. Prev. 37 (4) (2005) 787–799, http://
dx.doi.org/10.1016/j.aap.2005.03.023.
[20] K. Geurts, G. Wets, T. Brijs, K. Vanhoof, Profiling of high-frequency accident locations
by use of association rules, Transp. Res. Rec. 2003 (1840) 123–130, http://dx.doi.
org/10.3141/1840-14.
[21] T.C. Haupt, A study of management attitudes to a performance approach to con-
struction worker safety, J. Constr. Res. 4 (01) (2003) 87–100, http://dx.doi.org/10.
1142/S1609945103000327.

Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
 H. Li et al. / Automation in Construction xxx (2017) xxx–xxx
[22] Health and Safety Executive, Health and safety in construction sector in Great Brit-
ain, 2014/15, http://www.hse.gov.uk/statistics/industry/construction/construction.
pdf 2015 (accessed May 12, 2015).
[23] C.C.I.P. House, China's work safety yearbook, http://www.chinasafety.gov.cn/
newpage/ 2014 (accessed April 26, 2016).
[24] International Labor Organization, Delivering decent work in Europe and Central
Asia, http://www.ilo.org/wcmsp5/groups/public/—europe/—ro-geneva/documents/
meetingdocument/wcms_359771.pdf 2012 (accessed June 1, 2016).
[25] Z. Ismail, S. Doostdar, Z. Harun, Factors influencing the implementation of a safety
management system for construction sites, Saf. Sci. 50 (3) (2012) 418–423, http://
dx.doi.org/10.1016/j.ssci.2011.10.001.
[26] A. Kelm, L. Laußat, A. Meins-Becker, D. Platz, M.J. Khazaee, A.M. Costin, M. Helmus, J.
Teizer, Mobile passive Radio Frequency Identification (RFID) portal for automated
and rapid control of Personal Protective Equipment (PPE) on construction sites,
Autom. Constr. 36 (2013) 38–52, http://dx.doi.org/10.1016/j.autcon.2013.08.009.
[27] P. Kines, L.P. Andersen, S. Spangenberg, K.L. Mikkelsen, J. Dyreborg, D. Zohar, Im-
proving construction site safety through leader-based verbal safety communication,
J. Saf. Res. 41 (5) (2010) 399–406, http://dx.doi.org/10.1016/j.jsr.2010.06.005.
[28] M.A.C. López, D.O. Ritzel, I.F. González, O.J.G. Alcántara, Occupational accidents with
ladders in Spain: risk factors, J. Saf. Res. 42 (5) (2011) 391–398, http://dx.doi.org/10.
1016/j.jsr.2011.08.003.
[29] D.T. Larose, Discovering Knowledge in Data: an Introduction to Data Mining, John
Wiley & Sons, 2014.
[30] H. Li, G. Chan, T. Huang, M. Skitmore, T.Y. Tao, E. Luo, J. Chung, X. Chan, Y. Li, Chirp-
spread-spectrum-based real time location system for construction safety manage-
ment: a case study, Autom. Constr. 55 (2015) 58–65, http://dx.doi.org/10.1016/j.
autcon.2015.03.024.
[31] Y.-H. Lin, C.-Y. Chen, T.-W. Wang, Fatal occupational falls in the Taiwan construction
industry, J. Chin. Inst. Ind. Eng. 28 (8) (2011) 586–596, http://dx.doi.org/10.1109/
iccie.2010.5668168.
[32] A.-K. Lindberg, S.O. Hansson, C. Rollenhagen, Learning from accidents–what more
do we need to know? Saf. Sci. 48 (6) (2010) 714–721.
[33] J. Long, J. Yang, Z. Lei, D. Liang, Simulation-based assessment for construction hel-
mets, Comput. Methods Biomech. Biomed. Eng. 18 (1) (2015) 24–37, http://dx.
doi.org/10.1080/10255842.2013.774382.
[34] K. Messing, J. Courville, M. Boucher, L. Dumais, A.M. Seifert, Can safety risks of blue-
collar jobs be compared by gender? Saf. Sci. 18 (2) (1994) 95–112, http://dx.doi.org/
10.1016/0925-7535(94)90019-1.
Please cite this article as: H. Li, et al., Investigation of the causality patterns of non-helmet use behavior of construction workers, Automation in
Construction (2017), http://dx.doi.org/10.1016/j.autcon.2017.02.006
9
[35] Occupational Safety and Health Administration, Protecting roofing workers, https://
www.osha.gov/Publications/OSHA3755.pdf 2015 (accessed April 16, 2016).
[36] Occupational Safety and Health Administration, Worker injuries, illnesses and fatal-
ities, www.bls.gov/iif/cfoi_revised14.htm 2014 (accessed May 2, 2016).
[37] E. Öney-Yazıcı, M.F. Dulaimi, Understanding designing for construction safety: the
interaction between confidence and attitude of designers and safety culture, Archit.
Eng. Des. Manag. 11 (5) (2015) 325–337, http://dx.doi.org/10.1080/17452007.2014.
895697.
[38] T.A. Saurin, C.T. Formoso, R. Reck, B.M. Beck da Silva Etges, J.L.D. Ribeiro, Findings
from the analysis of incident-reporting systems of construction companies, J. Constr.
Eng. Manag. 141 (9) (2015) 05015007, http://dx.doi.org/10.1061/(asce)co.1943-
7862.0000988.
[39] F. Sherratt, M. Crapper, L. Foster-Smith, S. Walsh, Safety and volunteer construction
workers, Constr. Manag. Econ. 33 (5–6) (2015) 361–374, http://dx.doi.org/10.1080/
01446193.2015.1024269.
[40] V. Sousa, N.M. Almeida, L.A. Dias, Risk-based management of occupational safety
and health in the construction industry–part 1: background knowledge, Saf. Sci.
66 (2014) 75–86.
[41] H. Wagner, A.J. Kim, L. Gordon, Relationship between personal protective equip-
ment, self-efficacy, and job satisfaction of women in the building trades, J. Constr.
Eng. Manag. 139 (10) (2013) 04013005, http://dx.doi.org/10.1061/(asce)co.1943-
7862.0000739.
[42] K. Watanuki, K. Kojima, Knowledge acquisition and job training for advanced tech-
nical skills using immersive virtual environment, J. Adv. Mech. Des. Syst. Manuf. 1
(1) (2007) 48–57, http://dx.doi.org/10.1299/jamdsm.1.48.
[43] C. William McConnell, G. Gloeckner, J. Gilley, Predictors of work injuries: a quantita-
tive exploration of level of English proficiency as a predictor of work injuries in the
construction industry, Int. J. Constr. Educ. Res. 2 (1) (2006) 3–28, http://dx.doi.org/
10.1080/15503980500519585.
[44] Q. Williams, M. Ochsner, E. Marshall, L. Kimmel, C. Martino, The impact of a peer-led
participatory health and safety training program for Latino day laborers in construc-
tion, J. Saf. Res. 41 (3) (2010) 253–261, http://dx.doi.org/10.1016/j.jsr.2010.02.009.
[45] T.-W. Wong, Occupational injuries among construction workers in Hong Kong,
Occup. Med. 44 (5) (1994) 247–252.