You are on page 1of 8

Cognitive Factors Influencing Safety Behavior at Height:

A Multimethod Exploratory Study


Yang Miang Goh 1 and Nur Faddilah Binte Sa’adon 2

Abstract: Despite efforts in recent years, the construction industry remains one of the top contributors for workplace fatalities in many
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

countries. One of the key concerns in the industry is the management of workers’ safety behavior. This paper aims to explore the cognitive
factors influencing the unsafe behavior of not anchoring a safety harness when working at height. In addition, multiple stepwise linear
regression, artificial neural network, and decision tree techniques were applied in the study to assess their usefulness in evaluating survey
data of safety cognitive factors. The theory of planned behavior (TPB) was adopted to model the cognitive factors influencing the unsafe
behavior of scaffolders. The TPB postulates that attitude, perceived behavioral control, and subjective norms affect the intention of workers,
which ultimately affects intentional behavior. The unsafe act of not anchoring harnesses while working on a scaffold was selected as the focal
behavior based on observations and interviews with safety supervisors. Supervisors also provided their opinions on the underlying reasons for
the unsafe act. A questionnaire was then developed based on the site observations, interviews, and literature review. Subsequently, 40 migrant
workers from Bangladesh, India, and China were surveyed. Stepwise multiple linear regression, neural network, and decision tree analyses
were implemented. The analyses revealed that subjective norm was the key variable influencing a worker’s decision to anchor the safety
harness. The significance of subjective norm was probably affected by the national culture of the migrant workers. In addition, the analyses
showed that the relationships between the variables were probably nonlinear, thus neural network and decision tree are suitable techniques.
The exploratory study provides the basis for design of an in-depth study on the cognitive factors influencing safety behavior and it expands
the choice of analyses techniques. DOI: 10.1061/(ASCE)CO.1943-7862.0000972. © 2015 American Society of Civil Engineers.
Author keywords: Cognitive analysis; Safety behavior; Theory of planned behavior; Construction safety; Neural network; Decision tree;
Data mining; Migrant worker; Working at height; Labor and personnel issues.

Introduction construction companies implemented behavioral safety programs


to improve workers’ safety-related behaviors, and many studies
Fall-from-height (FFH) accidents have been a leading cause of have demonstrated the effectiveness of the approach [e.g., Lingard
injury in the construction industry in many countries including and Rowlinson (1997)]. However, studies on the factors influenc-
Singapore, New Zealand, Hong Kong, Taiwan, Kuwait, the United ing safety-related behaviors in the construction industry, especially
States, and Israel (Bentley et al. 2006; Chan et al. 2008; Cheng et al. high-risk trades such as scaffolding and formwork installation,
2010; Kartam et al. 1998; Lipscomb et al. 2004; Workplace Safety are not as prevalent as other industries (Cameron and Duff 2007).
and Health Council 2012; Yanai et al. 1999). Not only do FFH ac- The construction industry needs to understand the cognitive factors
cidents cause human suffering, substantial economic losses associ- influencing workers’ behaviors so as to design better behavioral
ated with fall injuries were also reported worldwide (Lockhart et al. interventions.
2005). For example, in the United States, the annual direct cost of The recent cognitive-behavioral study by Zhang and Fang
occupational injuries due to falls has been estimated to be in excess (2013) identified that Chinese scaffolders decide not to use safety
of $6 billion (Courtney et al. 2001). harness because of “inconvenience and discomfort of using safety
Management of worker behavior is a critical component in harnesses, underestimating the risk of not using safety harnesses,
preventing FFH. Despite having the necessary safety equipment, negative pressures from gangmasters, foremen, and safety officers,
procedures, rules, and training, workers frequently choose to vio- and lack of safety lines.” The study provided useful guidance for
late safety rules, procedures, and training and not use protective design of behavioral interventions, but the survey was focused on
equipment for reasons such as comfort, convenience, productivity, Chinese scaffolders in China. Thus, this paper aims to explore the
and group norms (Hale and Borys 2013). In response, many cognitive factors influencing migrant workers in Singapore. Like
Singapore, many countries are employing construction labor from
1
Assistant Professor, Safety and Resilience Research Unit (SaRRU), overseas and it had been shown that migrant workers are a vulner-
Dept. of Building, School of Design and Environment, National Univ. able group in terms of workplace safety (Ahonen et al. 2007;
of Singapore, 4 Architecture Dr., Singapore 117566 (corresponding Guldenmund et al. 2013; Hare et al. 2013). In addition, this study
author). E-mail: bdggym@nus.edu.sg also evaluated the potential of using data mining techniques,
2
Property Executive, Property Facility Services Pte. Ltd.; formerly, namely, artificial neural network and decision tree, in studying
Dept. of Building, School of Design and Environment, National Univ. cognitive factors influencing safety behaviors. In contrast to linear
of Singapore, 4 Architecture Dr., Singapore 117566.
regression analysis, neural network (Samarasinghe 2007) and de-
Note. This manuscript was submitted on July 13, 2014; approved on
December 4, 2014; published online on January 7, 2015. Discussion period cision tree (Witten 2011) does not require assumptions about the
open until June 7, 2015; separate discussions must be submitted for indi- statistical characteristics of the population and these data mining
vidual papers. This paper is part of the Journal of Construction Engineer- techniques are suited for analyzing nonlinear problems. Despite
ing and Management, © ASCE, ISSN 0733-9364/04015003(8)/$25.00. its potential, there is insufficient interest in applying neural network

© ASCE 04015003-1 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


and other soft computing methods in safety studies (Ciarapica and researchers [e.g., Cheng et al. (2009), Chua et al. (1997), and
Giacchetta 2009). Thus, this exploratory study implemented step- Goh and Chua (2013)] have successfully applied neural network
wise multiple linear regression, neural network, and decision tree to techniques on complex and nonlinear problems. However, the ap-
assess the potential of the data mining techniques. plication of neural network technique in construction safety is still
uncommon.
In this study, the generalized regression neural network (GRNN)
Background is used. Fig. 2 shows the generic structure of a GRNN with two
input numeric variables (equivalent to independent variables in re-
Theory of Planned Behavior gression). One neuron is used for each input variable. Similarly,
each neuron on the pattern layer represents a training case. The
This study adopted the theory of planned behavior (TPB) (Ajzen neurons in the pattern layer use a function to determine the distance
1991) (Fig. 1) in modeling the cognitive factors influencing the
(or level of similarity) between the input case and the training case.
behaviors of scaffolders. The TPB had been widely applied in
The distance function has smoothing parameters that scale the dif-
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

many areas such as traffic safety, health interventions, adolescent


ferent inputs (Specht 1991) and the training process seeks to adjust
behavior, food safety, and information security [e.g., Gerend and
these smoothing parameters to reduce the prediction error of the
Shepherd (2012), Heirman and Walrave (2012), Ifinedo (2012),
network. The summation layer has two neurons (denominator and
Milton and Mullan (2012), and Parker et al. (1992)]. The TPB
numerator neurons) that take the values from the pattern layer and
postulates that planned behaviors are significantly influenced by
does a dot product. The denominator neuron sums the weight val-
intention, subjective norm (SN), attitude, and perceived behavioral
control (PBC). ues of the neurons in the hidden layer and the numerator neuron
Intention refers to the readiness to perform a particular behavior adds the weight values multiplied by the actual dependent (or out-
despite the difficulty in doing so. TPB postulates that in a planned put) values. Finally, the output neuron divides the value generated
behavior, where the individual can opt to perform or not perform by the numerator neuron by the value produced by the denominator
the behavior, the intention or motivation of the individual is the neuron.
key determinant of the actual behavior. In addition, intention is
predicted by three cognitive attributes, namely, attitude, subjective Decision Tree
norm, and perceived behavioral control. Attitude towards the
behavior refers to an individual’s judgment of whether the individ- A decision tree specifies a sequence of decisions needed to arrive at
ual has a favorable or unfavorable evaluation or appraisal of the a recommendation or prediction (see Fig. 3 for an example) (Witten
behavior. Subjective norm refers to perceived social pressure or ex- 2011). The approach adopts a divide and conquer strategy in which
pectations to perform the behavior or not. PBC is a reflection of the an attribute is selected and the training data set is then split into two
individual’s perception of his/her ability to perform the behavior,
while taking into account the resources and opportunities availabil-
ity to the individual. Output
Neuron

Summation Layer
Neural Network
Neural network is an established artificial intelligence technique
that mimics biological nervous systems, which have adaptive learn- Pattern Layer
ing properties. Each neural network is a mathematical model made
Input
up of nodes (or neurons) that are connected through a network.
A learning algorithm is implemented to adjust the weights of
the connections between the neurons so that the network is able Fig. 2. Structure of a generalized regression neural network with two
to minimize prediction errors. Many construction management inputs

>=3

Contract sum

>=3
<3

Competency
Low Risk of supervisor

<4 >=4

High Risk Low Risk

Fig. 1. Theory of planned behavior Fig. 3. Hypothetical decision tree

© ASCE 04015003-2 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


subsets repeatedly (for numerical variables) until the remaining observations, it was decided that this exploratory study will focus
cases can no longer be divided. The distribution of the class var- on failure to anchor safety harness.
iable (or dependent variable) is then used to provide the recommen- During the site visits, five supervisors were interviewed to
dation. The criterion used to split data sets is usually based on the understand the underlying reasons for the failure to anchor har-
attribute that provides the greatest information gain (bits). Once a nesses. The supervisors have an average of 8 years of experience
tree is generated, the tree will have to be pruned (or simplified) so in the construction industry and all of them have attended safety-
that it is readable and not too specific to the data set (Williams related trainings. All five supervisors agreed that failure to anchor
2011). As an illustration of how a decision can be used, assume a safety harnesses is one of the most common unsafe acts on site. In
survey of 10 attributes was used to capture expert assessment of the addition, it was felt that the main reason for the unsafe act was the
safety and health risk of a series of projects; the results are used to poor safety attitude of scaffolders. The supervisors highlighted that
generate the decision tree in Fig. 3. Based on the tree, if personnel most scaffolders have adequate safety training, but they tend to be
shortage is low (<3), safety and health risk is low. On the other complacent and do not implement the safety measures that they
hand, if personnel shortage is high (≥ 3), contract sum is high (≥ 3) have learned in their training. Some of the supervisors suggested
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

and competency of supervisor is very high (≥ 4), then the risk of the the migrant workers are used to the lower level of safety standards
project is low. The algorithm has only selected three attributes to in their home countries. The supervisors opined that scaffolders
classify the data set. The selected attributes are more likely to be tend to be focused on getting the task done and frequently neglect
more influential in determining the risk level of projects. However, safety measures such as anchoring their safety harnesses. In gen-
most decision trees can only predict nominal class. eral, the supervisors felt poor safety attitude of individual scaffold-
ers was the main reason for the failure to anchor safety harnesses.

On-Site Observations and Interviews


Questionnaire Survey
In view of the exploratory nature of this study, a range of data was
collected to understand the critical safety behavior of scaffolders A questionnaire survey was developed to determine the cognitive
and the perceived underlying factors influencing behaviors. The factors (as highlighted in the TPB) influencing the unsafe act of not
study started with observations of two construction sites to deter- anchoring safety harnesses when it is feasible to do so. The selected
mine common unsafe behaviors of scaffolders. The observations approach is aligned with many psychological studies that used a
were unstructured so that the researchers can explore common un- questionnaire survey to study safety-related behaviors [e.g., Ross
safe behaviors in an uninhabited fashion. The observations were et al. (2011) and Sniehotta (2009)]. Besides demographic ques-
conducted on a mix of erection, dismantling, and normal use of tions, questions were developed to measure each of the key vari-
scaffolds. Workers were informed of the presence of the researchers ables highlighted in the TPB, i.e., behavior, intention, subjective
and the purpose of the research. Seven 2-h observation sessions norm, attitude, and PBC.
were conducted over 2 weeks. As a whole, 50 workers were ob- In the questionnaire, respondents were asked to rate 33 state-
served. Table 1 summarizes the types of unsafe acts and their ments based on a five-point Likert scale with 1 being strongly dis-
corresponding frequencies. agree and 5 being strongly agree. Even though the questionnaire
Even though failure to put on safety harnesses was relatively developed by Zhang and Fang (2013) was focused on failure to
uncommon among the scaffolders observed, many scaffolders did use safety harness, their questionnaire was based on TPB and had
not connect their safety harnesses onto suitable anchors while much relevance for this study. However, the questions had to be
working at height. It was noticed that some of the scaffolders would phrased differently so as to suit the background of the workers and
immediately anchor their harnesses when their supervisors were the local context and account for the difference in the target unsafe
nearby. This indicates that the scaffolders were able to anchor their behavior. A pilot study was conducted by using the draft question-
safety harness, but decided not to do so until their supervisors ap- naire to survey eight respondents so as to improve the clarity of the
pear. For the sites visited, and many sites in Singapore, scaffolders instructions and assess the time needed for the survey. Several im-
were required to attach their harness whenever they are working on provements were made to the questionnaire to ensure that the state-
a scaffold and scaffolders are allowed to anchor their harnesses ments were easy to understand. In addition, the questionnaire was
to the scaffolds [this is aligned with the advice given in the U.K. modified based on the interviews with the supervisors. Even though
National Access and Scaffolding Confederation (2011)]. In addi- the survey covered 33 statements, this paper is focused on 11 of
tion, it was observed that 12 (44%) of the observed unsafe acts were them (Table 2). Question 11 is assessed by both the supervisor of
related to fall arrest system (including safety harness). Based on the the scaffolders and the scaffolder. A Student t-test showed that the
difference in the assessments by the scaffolders and their supervisor
were statistically insignificant. However, inspection of the data
Table 1. Observed Unsafe Acts of Scaffolders showed that the assessments by the scaffolders were either equal
Number Observation Frequency Percent to or higher than the assessment by the supervisors. Thus, to reduce
the effect of self-reporting bias, the supervisor’s assessment was
1 Failure to anchor safety harness 6 22
2 Improper use of fall arrest system 4 15
used in the analysis.
3 Failure to ensure proper housekeeping 4 15 Forty scaffolders from three construction sites in Singapore
4 Misuse of tools and equipment 4 15 completed the survey with the assistance of interpreters. In view of
5 Stepping on guardrail to access 3 11 the exploratory nature of this study, the scaffolders were recruited
unreachable areas based on convenience sampling. Out of the 40 scaffolders, 27 were
6 Failure to put on safety harness 2 7 from Bangladesh, 11 from India, and two from China. The inter-
7 Taking unsafe shortcuts to access scaffold 2 7 preters were site personnel that were able to communicate in
8 Distracted when performing task 2 7 English and the native language of the responding worker. Each
(e.g., talking vigorously) survey took approximately 25 min. The respondents had working
Total 27 100
experience of 3 to 10 years (mean ¼ 4.98 years).

© ASCE 04015003-3 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


Table 2. Statements Used in Survey Questionnaire
Question
number Statement TPB variable
1 It is important to anchor the safety harness when working on scaffolds Attitude
2 I am responsible for anchoring my safety harness Attitude
3 Accident is not a matter of luck Attitude
4 I am constantly reminded by my colleagues to hook my safety harness onto suitable anchors Subjective norm
5 My colleagues are committed to follow all safety rules Subjective norm
6 My supervisor will reprimand me if I do not hook my safety harness to suitable anchors Subjective norm
7 I will fall from height if I do not hook my safety harness onto suitable anchors when working on scaffold Perceived behavioral control
8 It is easy to hook safety harness onto suitable anchors Perceived behavioral control
9 I plan to hook safety harness onto suitable anchors when working on scaffold Intention
10 I plan to remind others to hook their safety harness onto suitable anchors when working on scaffold Intention
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

11 I always hook safety harness onto suitable anchors when working on scaffolding Behavior

Table 3. Descriptive Statistics for Each Question


Question number Mean Standard deviation
1 4.95 0.221
2 4.83 0.385
3 4.08 0.997
4 4.45 0.597
5 4.5 0.599
6 4.3 0.687
7 4.53 0.554
8 4.08 0.694
9 4.73 0.554
10 4.5 0.641
11 4.4 0.545

Table 4. Pearson Bivariate Correlation Matrix between Independent


Variables
Variables PBC Attitude SN Intention Fig. 4. Residual versus fit plot (intention as dependent variable and SN,
attitude, and PBC as independent variables)
PBC 1 0.296 0.177 0.210
Attitude — 1 −0.062 −0.126
SN — — 1 0.443a
Intention — — — 1 Table 5. Results of Stepwise Regression Analysis of Factors Predicting
a
Correlation is significant at the 0.01 level (two-tailed). Intention
Step b Standard error β t Significant
The results in Table 3 show that the supervisors and individuals 1 (Constant) 2.484 0.702 — 3.537 0.001
generally feel that the respondents were anchoring their safety har- SN 0.482 0.158 0.443 3.049 0.004
nesses. Table 4 indicates that the independent variables are essen- Note: R2 ¼ 0.197 for Step 1.
tially uncorrelated. SN has significant correlation with intention,
but the correlation is only moderate. Table 4 also shows that multi-
collinearity was not an issue. In stepwise regression, independent variables are entered or
removed based on the F-statistic (in relation to user-defined thresh-
olds) at each step (all the F-statistics for the following regression
Analyses models are significant with p ≤ 0.01). The F-statistic is a reflec-
tion of the level of correlation between the independent variable
and dependent variable. The stepwise regression is one of the
Stepwise Multiple Linear Regression most widely used variable selection techniques (Montgomery and
Based on the TPB, attitude, SN, and PBC can affect intention of Runger 1999). The residual versus fit plot in Fig. 4 seems to
workers. In turn, intention and PBC influences the actual behavior. suggest that the assumption of linearity is acceptable. However,
Thus, stepwise multiple linear regression was applied in two stages. a closer examination shows that the points are more dispersed be-
For the first stage, the intention of scaffolders was assigned as low the zero line and do not spread randomly across the chart. This
the dependent variable and SN, attitude, and PBC were assigned points to possible nonlinearity issues.
as independent variables. Subsequently, intention and perceived Table 5 summarizes the regression model for predicting inten-
behavioral control were used as independent variables to predict tion based on SN, attitude, and PBC. The model indicates that SN is
behavior. Because it is also possible for SN, attitude, and PBC to the only independent variable entered into the model (intention ¼
have a direct relationship with behavior, a third regression analysis 2.484 þ 0.482 × SN). Due to the high uncertainty in human cog-
was also conducted using SN, attitude, and PBC as independent nition and behavior and the small sample size (Newman and
variables and behavior as dependent variable. Newman 2000), it is not surprising that the R2 is relatively low

© ASCE 04015003-4 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


Table 6. Results of Stepwise Regression Analysis of Factors Predicting
Behavior
Step b Standard error β t Significant
1 (Constant) 2.554 0.685 — 3.726 0.001
Intention 0.400 0.148 0.403 2.712 0.010
Note: R2 ¼ 0.162 for Step 1.

Table 7. Regression Model for Predicting Behavior: Intention Excluded


Standard
Step b error β t Significant
1 (Constant) 2.031 0.676 — 3.005 0.005
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

SN 0.536 0.152 0.497 3.527 0.001


2 (Constant) −0.238 1.048 — −0.227 0.821
SN 0.560 0.141 0.518 3.967 0.000
Attitude 0.469 0.174 0.353 2.701 0.010
Fig. 5. Residual versus fit plot (behavior as dependent variable and
Note: R2 ¼ 0.247 for Step 1; ΔR2 ¼ 0.124 for Step 2 (ps < 0.05). intention and PBC as independent variables)

(0.197). The model implies that among the TPB variables, subjective
norm was the only significant predictor of intention. The residual ver-
sus fit plot in Fig. 4 seems to suggest that the assumption of linearity is
acceptable. However, a closer examination shows that the points are
more dispersed below the zero line and do not spread randomly across
the chart. This points to possible nonlinearity issues.
In accordance with the TPB, a stepwise multiple regression was
used to assess the importance of intention and PBC in influencing
the scaffolders’ decision to anchor their safety harness. With refer-
ence to Table 6, Intention was included in the regression model
(behavior ¼ 2.554 þ 0.4 × intention) and PBC was excluded. As
can be observed from Table 6, the R2 was also low. Table 7 shows
the results for the regression analysis to predict behavior with
attitude, SN, and PBC as the independent variables. The final
model (Step 2), behavior ¼ −0.238 þ 0.56 · SN þ 0.469 · attitude
(Table 7) shows that SN and attitude were the significant variables
influencing behavior of scaffolders and PBC was excluded. The
final model has a relatively higher R2 (0.371). In addition, both
Figs. 5 and 6 show that the relationships of the regression analyses Fig. 6. Residual versus fit plot (behavior as dependent variable and SN,
are probably nonlinear. attitude, and PBC as independent variables)
Thus as a whole, the three regression analyses indicated that
among the TPB variables, PBC is not significant. On the other
hand, SN is a relatively influential variable. Furthermore, the useful for small data sets, such as herein. In the traditional parti-
residual analyses suggested that the relationship between the var- tioning or holdout approach, a fixed percentage of the training set is
iables was probably nonlinear. used for testing. However, in small data sets, the small number of
cases for testing is very limited and the testing results will not be
credible. In contrast, the cross-validation method employs all the
Neural Network Analysis
available data for testing.
The neural network analyses were conducted using NeuralTools 6 NeuralTools (Palisade 2014) produces an indicator called var-
(Palisade 2014). Because the variables are numeric, the GRNN iable impact to measure the relative importance of each indepen-
(Specht 1991) was adopted. Unlike the multilayer feed-forward dent variable in predicting the dependent variable. Even though
network (MLFN), the GRNN does not require arbitrary assignment variable impact is not meant to give a firm assessment of the im-
of the number of hidden layers and the number of neurons on each portance of each variable, the indicator does provide useful guid-
layer [e.g., Goh and Chua (2013)]. The 10-fold cross-validation ance on relative importance. Based on Palisade (2013), variable
technique (Samarasinghe 2007; Witten 2011) was used to test the impact is calculated by focusing on each independent variable,
credibility of the network. In the 10-fold cross validation, the train- X i sequentially, where i ¼ 1; : : : ; m and m is the number of inde-
ing set was randomly split into 10 subsets and within each subset, pendent variables. The algorithm then steps through each training
10% of the training set were randomly assigned for testing and case j, where j ¼ 1; : : : ; n and n is the number of training cases.
withdrawn from the training process. Each subset of testing cases For each training case j, the value of the ith independent variable
was then used to assess the root-mean-square error of the corre- X ij , is varied, but the other independent variables are kept un-
sponding network. The process was repeated 10 times and the aver- changed. Each specific value of X ij within the training set is used
age of the root-mean-square error for the 10 testing sets was used to to derive the corresponding dependent variable Y ij derived from the
estimate the root-mean-square error of the network trained using neural network. The difference between the maximum and mini-
the full training set. The cross-validation technique is especially mum value of the dependent variable Y ij , Δij , is then calculated

© ASCE 04015003-5 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


Table 8. RMS Errors and Impact Analysis Table 9. Percent Bad Prediction for Decision Trees
Independent Dependent Variable RMS Dependent Selected Percent bad
Network variable variable impact error Tree Independent variable variable variable prediction
1 Attitude, PBC, SN Intention SN (100%) 0.4620 1 Attitude, PBC, SN Intention class SN 50
2 Intention, PBC Behavior PBC (50%), 0.5103 2 Intention, PBC Behavior class Intention 32.5
intention (50%) 3 Attitude, PBC, SN Behavior class SN 25
3 Attitude, PBC, SN Behavior SN (42%), 0.3948
attitude (36%),
PBC (22%)
Centre 2014)]. Because the interviews indicate that supervisors
expected the scaffolders to anchor their harnesses, the high power
distance score may mean that the scaffolders are heavily influenced
for each training case j and independent variable i. By averaging by their perception of the expectation of their supervisors. In ad-
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

the Δij for all training cases, the mean of Δij (Δ̄i ) for independent dition, because all scaffolders feel that they always anchor their
variable i is then calculated. The process is repeated for other in- harnesses, this forms peer pressure on the other scaffolders, espe-
dependent variables (2; : : : ; m) to obtain the mean difference for cially if the level of individualism is low. Furthermore, because
each independent variable. Thus, the variable impact for indepen- individualism is likely to be low in this group of workers, the effect
dent variable i is of an individual’s attitude and perceived behavioral control, as re-
flected in the analysis, is likely to be low. Thus, the national cul-
Δ̄i
VI i ¼ ð1Þ tures of the migrant workers and the location of work could have
Δ̄1 þ : : : þ Δ̄m played a part in making subjective norm the key variable influenc-
ing assessment of consistency of safety behavior. However, this
As in the case of regression, three networks were trained and
exploratory study did not collect information on the site and organi-
tested (Table 8). Besides variable impact, Table 8 shows the RMS
zational culture in which the scaffolders worked. As shown in past
error for each of the networks trained. Because the Likert scale used
studies (Zohar 2010), safety climate at group and organizational
in the survey is an integer scale of 1 to 5, a RMS error of 0.5 or less
levels are important and these factors should be considered in future
was deemed to be acceptable. It can be observed that Network 3 has
studies.
the lowest RMS error. In alignment with the regression analysis, the
Based on the residual versus fit plots of the stepwise multiple
GRNN analyses indicate subjective norm as the independent var-
linear regression models, the relationships between the variables
iable with the greatest influence on behavior and intention.
were probably nonlinear. In addition, regression analysis requires
more stringent checks on the statistical characteristics of the pop-
Decision Tree Analysis ulation to assure that the assumptions are valid. Artificial neural
To facilitate the application of the decision tree technique, the de- network and decision tree are better suited to handle nonlinear
pendent variables, intention and behavior had to be converted into relationships and they do not require any assumptions about the
nominal data type. Thus, for both intention and behavior, a value of statistical characteristics of the population being investigated. How-
5 is classified as A and a value less than 5 is classified as B. The ever, a neural network is very much a black box and it is hard to
analyses used the J48 classifier (Witten 2011), which is a modifi- understand the details of the network. In contrast, a decision tree
cation of the widely implemented C4.5 (Quinlan 1993). In accor- usually provides a simple and straightforward solution. In this
dance with the previous analyses, three trees were developed study, all the decision trees only have one node. This could be be-
(Table 9) and all the trees were tested using 10-fold cross valida- cause of the small number of variables and sample size. Further-
tion. All three trees only have one node. Tree 1 has a high percent more, a decision tree analysis require a nominal class to act as the
bad prediction (percentage of predictions that were wrong), but dependent variable. Arbitrary classification of the dependent vari-
Trees 2 and 3 had acceptable prediction accuracy. In addition, it able may affect the results of the decision tree. Nevertheless, both
was noted that as in previous analyses, subjective norm surfaced as data mining approaches have shown that they are useful approaches
the most influential independent variable. for understanding cognitive factors of safety-related behaviors.
During the interviews, the supervisors opined that scaffolders
fail to anchor their harnesses due to poor safety attitude. However,
Discussions the study showed that workers were more influenced by subjective
norm than attitude. The supervisors underappreciate their own in-
The analyses consistently identified SN as the most influential var- fluence on the workers and the importance of group norms. Further
iable in determining supervisors’ ratings of migrant scaffolders’ studies of the contribution of supervisory actions to unsafe behavior
consistency in anchoring their safety harnesses. It is not unaccept- should be conducted.
able that only one variable (SN) is identified as significant because Future studies can use this exploratory study as a basis to gen-
the TPB only stipulates possible relationships between the variables erate hypotheses and select research methods. Studies on cognitive
and does not demand inclusion of all the variables for all situations factors should take into account nationality of the subjects being
(Ajzen 1991). Different behavior and population can yield differ- studied. For migrant workers from Asia, particularly southern Asia,
ent results. In the context of this study, the respondents are from it is necessary to measure and probe the factors influencing sub-
Bangladesh, China, and India. In addition, they are working in jective norm in more detail. A wider range of behaviors can also
Singapore. All four countries are known to have high scores (77–80) be studied to assess if the type of behavior played a significant role
for power distance dimension (degree of acceptance of unequal in influencing the relative importance of the underlying cognitive
distribution of power) and low or moderate scores (20–48) for indi- factors. Because worker behavior is affected by organization struc-
vidualism dimension [degree of preference for a “loosely-knit ture, processes, and culture (Guldenmund 2007), a wider range of
social framework in which individuals are expected to take care factors can also be studied. In addition, many safety-related behav-
of only themselves and their immediate families” (The Hofstede iors are decided on the spot depending on the decision environment

© ASCE 04015003-6 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


at that point of time. A questionnaire survey may not reflect the J. Constr. Eng. Manage., 10.1061/(ASCE)0733-9364(1997)123:3(214),
effect of the dynamic decision environment. Future studies have 214–222.
to take this limitation into consideration and triangulate the survey Ciarapica, F. E., and Giacchetta, G. (2009). “Classification and prediction
data with observations and interviews. of occupational injury risk using soft computing techniques: An Italian
study.” Saf. Sci., 47(1), 36–49.
The study adopted convenience sampling to ensure that there
Courtney, T. K., Sorock, G. S., Manning, D. P., Collins, J. W., and Holbein-
were sufficient data points for meaningful analyses. It is believed Jenny, M. A. (2001). “Occupational slip, trip and fall-related injuries-
that the profile of the workers were generally representative of can the contribution of slipperiness be isolated?” Ergonomics, 44(13),
the scaffolder population in Singapore, but the sampling approach 1118–1137.
remains a threat to internal and external validity. Further studies Gerend, M. A., and Shepherd, J. E. (2012). “Predicting human papilloma-
should devise a stratified sampling strategy taking into account virus vaccine uptake in young adult women: Comparing the health
variables such as nationality, years of experience, and trainings belief model and theory of planned behavior.” Ann. Behav. Med., 44(2),
attended. 171–180.
Goh, Y. M., and Chua, D. (2013). “Neural network analysis of construction
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

safety management systems: A case study in Singapore.” Constr.


Manage. Econ., 31(5), 460–470.
Conclusions Guldenmund, F., Cleal, B., and Mearns, K. (2013). “An exploratory study
of migrant workers and safety in three European countries.” Saf. Sci.,
This exploratory study investigated the cognitive factors that influ- 52, 92–99.
ence scaffolders’ decision not to anchor safety harnesses. The TPB Guldenmund, F. W. (2007). “The use of questionnaires in safety culture
was adopted to model the cognitive factors and the study aimed to research—An evaluation.” Saf. Sci., 45(6), 723–743.
determine which of the TPB constructs, namely, attitude, subjective Hale, A., and Borys, D. (2013). “Working to rule, or working safely? Part 1:
norms, perceived behavioral control, and intention, are potentially A state of the art review.” Saf. Sci., 55, 207–221.
significant predictor of safety behavior of scaffolders. In addition, Hare, B., Cameron, I., Real, K. J., and Maloney, W. F. (2013). “Exploratory
the study compared the usefulness of using multiple stepwise linear case study of pictorial aids for communicating health and safety for mi-
regression, artificial neural network, and decision tree to study cog- grant construction workers.” J. Constr. Eng. Manage., 10.1061/(ASCE)
nitive factors influencing safety behavior. CO.1943-7862.0000658, 818–825.
The analyses showed that for the 40 migrant workers that par- Heirman, W., and Walrave, M. (2012). “Predicting adolescent perpetration
ticipated in the study, subjective norm had the greatest influence in cyberbullying: An application of the theory of planned behavior.”
Psicothema, 24(4), 614–620.
on the workers’ decision to anchor their harness or not. In addition,
Hofstede Centre. (2014). “Country comparison.” 〈http://geert-hofstede
the relationships between the TPB constructs cannot be satisfacto- .com/countries.html〉 (Jul. 9, 2014).
rily modeled using linear regression, which suggests that the rela- Ifinedo, P. (2012). “Understanding information systems security policy
tionship between the constructs is probably nonlinear. The study compliance: An integration of the theory of planned behavior and
showed that data mining approaches such as artificial neural net- the protection motivation theory.” Comput. Secur., 31(1), 83–95.
work and decision tree have the potential to mine useful knowledge Kartam, N. A. K., Nabil, A., and Bouz, R. G. (1998). “Fatalities and
on the cognitive factors influencing safety behavior. Due to the ex- injuries in the Kuwaiti construction industry.” Accid. Anal. Prev., 30(6),
ploratory nature of this study, it is not possible to confirm the actual 805–814.
significance of the TPB constructs. However, this study provided Lingard, H., and Rowlinson, S. (1997). “Behavior-based safety manage-
the basis for devising the hypotheses for more in-depth study on ment in Hong Kong’s construction industry.” J. Saf. Res., 28(4),
the relative strengths of the TPB constructs in influencing construc- 243–256.
Lipscomb, H. J., Glazner, J., Bondy, J., Lezotte, D., and Guarini, K. (2004).
tion workers’ safety behaviors. In addition, this study provided the
“Analysis of text from injury reports improves understanding of con-
foundation for selecting analytical methods in future studies. struction falls.” J. Occup. Environ. Med., 46(11), 1166–1173.
Lockhart, T. E., Grönqvist, R., and Chang, W.-R. (2005). “Prevention of
fall-related accidents.” Saf. Sci., 43(7), 355–357.
References Milton, A. C., and Mullan, B. A. (2012). “An application of the theory of
planned behavior—A randomized controlled food safety pilot interven-
Ahonen, E. Q., Benavides, F. G., and Benach, J. (2007). “Immigrant pop- tion for young adults.” Health Psychol., 31(2), 250–259.
ulations, work and health—A systematic literature review.” Scand. J. Montgomery, D. C., and Runger, G. C. (1999). Applied statistics and
Work Environ. Health, 33(2), 96–104. probability for engineers, Wiley, New York.
Ajzen, I. (1991). “The theory of planned behavior.” Organ. Behav. Hum. National Access and Scaffolding Confederation. (2011). “SG4:10 prevent-
Decis. Processes, 50(2), 179–211. ing falls in scaffolding.” 〈http://www.nasc.org.uk/sg4_10_launch_1〉
Bentley, T. A., et al. (2006). “Investigating risk factors for slips, trips and (Oct. 1, 2014).
falls in New Zealand residential construction using incident-centred and NeuralTools 6 [Computer software]. New York, Palisade.
incident-independent methods.” Ergonomics, 49(1), 62–77. Newman, I., and Newman, C. (2000). “A discussion of low r-squares:
Cameron, I., and Duff, R. (2007). “A critical review of safety initiatives Concerns and uses.” Educ. Res. Q., 24(2), 3.
using goal setting and feedback.” Constr. Manage. Econ., 25(5), Palisade. (2013). “Calculation and use of variable impacts.” 〈http://kb
495–508. .palisade.com/index.php?pg=kb.page&id=225〉 (Jul. 14, 2014).
Chan, P. C., et al. (2008). “Work at height fatalities in the repair, main- Palisade, C. (2014). “NeuralTools 6 software, sophisticated neural net-
tenance, alteration, and addition works.” J. Constr. Eng. Manage., works for Excel.” 〈http://kb.palisade.com/index.php?pg=kb.chapter&
10.1061/(ASCE)0733-9364(2008)134:7(527), 527–535. id=47〉 (Jul. 14, 2014).
Cheng, C. W., Lin, C. C., and Leu, S. S. (2010). “Use of association rules Parker, D., Manstead, A. S. R., Stradling, S. G., Reason, J. T., and Baxter,
to explore cause-effect relationships in occupational accidents in the J. S. (1992). “Intention to commit driving violations—An application of
Taiwan construction industry.” Saf. Sci., 48(4), 436–444. the theory of planned behavior.” J. Appl. Psychol., 77(1), 94–101.
Cheng, M. Y., Tsai, H. C., and Liu, C. L. (2009). “Artificial intelligence Quinlan, J. R. (1993). C4. 5: Programs for machine learning, Morgan
approaches to achieve strategic control over project cash flows.” Autom. Kaufmann, New York.
Constr., 18(4), 386–393. Ross, L. T., Ross, T. P., Farber, S., Davidson, C., Trevino, M., and Hawkins,
Chua, D. K. H., Kog, Y. C., Loh, P. K., and Jaselskis, E. J. (1997). “Model A. (2011). “The theory of planned behavior and helmet use among
for construction budget performance—Neural network approach.” college students.” Am. J. Health Behav., 35(5), 581–590.

© ASCE 04015003-7 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.


Samarasinghe, S. (2007). Neural networks for applied sciences and Workplace Safety and Health Council. (2012). “Workplace safety and
engineering, Taylor & Francis, Boca Raton, FL. health report 2011.” 〈https://www.wshc.sg/wps/themes/html/upload/
Sniehotta, F. (2009). “An experimental test of the theory of planned behav- announcement/file/WSH%20Stats%20Report%202011.pdf〉 (Jul. 4,
ior.” Appl. Psychol. Health Well-Being, 1(2), 257–270. 2012).
Specht, D. F. (1991). “A general regression neural network.” IEEE Trans. Yanai, O., Goldin, L., Kugel, C., and Hiss, J. (1999). “Occupational fatal-
Neural Networks, 2(6), 568–576. ities in Israel.” J. Clin. Forensic Med., 6(3), 129–132.
Williams, G. (2011). Data mining with rattle and R: The art of ex- Zhang, M., and Fang, D. (2013). “A cognitive analysis of why Chinese
cavating data for knowledge discovery (use R!), Springer, Canberra, scaffolders do not use safety harnesses in construction.” Constr.
Australia. Manage. Econ., 31(3), 207–222.
Witten, I. H. (2011). Data mining: Practical machine learning tools and Zohar, D. (2010). “Thirty years of safety climate research: Reflections and
techniques, Morgan Kaufmann, Burlington, MA. future directions.” Accid. Anal. Prev., 42(5), 1517–1522.
Downloaded from ascelibrary.org by Tokyo Univ Seisan Gijutsu on 05/14/15. Copyright ASCE. For personal use only; all rights reserved.

© ASCE 04015003-8 J. Constr. Eng. Manage.

J. Constr. Eng. Manage.

You might also like