A hybrid BN-HFACS model for predicting safety performance in construction projects

Safety Science 101 (2018) 332–343

Contents lists available at ScienceDirect

Safety Science journal homepage: www.elsevier.com/locate/safety

A hybrid BN-HFACS model for predicting safety performance in construction projects

MARK

⁎

Nini Xiaa, Patrick X.W. Zoub,c, Xing Liua, Xueqing Wanga, , Runhe Zhua a b c

College of Management and Economics, Tianjin University, No. 92 Nankai District, Tianjin 300072, PR China College of Management and Economics, Tianjin Chengjian University, No. 26 Jinjing Road, Xiqing District, Tianjin 300384, PR China Department of Civil and Construction Engineering & Centre for Sustainable Infrastructure, Swinburne University of Technology, Hawthorn, Victoria 3122, Australia

A R T I C L E I N F O

A B S T R A C T

Keywords: Safety performance Construction project Human factor Bayesian networks Safety risk assessment

Lacking a holistic framework for analyzing risk factors would result in the inaccurate assessment of safety performance and poor safety management. This research aims to establish a Bayesian-network (BN)-HFACS hybrid model to proactively predict safety performance in construction projects. First, a causation framework for analyzing the underlying factors influencing construction safety performance was established using the Human Factors Analysis and Classification System (HFACS). This causation framework incorporates 18 risk factors from organizational, environmental and human aspects that are categorized into five levels: L1: “unsafe acts of workers,” L2: “preconditions for unsafe acts,” L3: “unsafe supervision and monitoring,” L4: “adverse organizational influences,” and L5: “adverse environmental influences.” The relationships between these factors and project safety performance were then hypothesized in the BN-HFACS model, and validated by data collected with questionnaires. The proposed model was applied to a subway project with AgenaRisk software. This application demonstrated the model’s capabilities in systematically identifying risk factors, predicting the probabilities of safety states in project level and in the five specific cause levels, and diagnosing the most sensitive risk factor. This research contributes to safety assessment and management by modifying the original HFACS for the causation analysis of construction safety performance, and by establishing a BN model for quantifying the total influences of the risk factors at five distinct levels on project safety performance. The integration of HFACS and BNs may be instructive in other contexts where diverse safety risk factors are involved in a system and safety prediction of the system is necessary.

1. Introduction The construction industry worldwide is constantly prone to injuries and accidents (Fang and Wu, 2013; Wang et al., 2016; Zou and Sunindijo, 2015, 2013). This causes great damage to the well-being of the workers and to the profitability and reputation of project parties (Zou et al., 2014). Concerning these losses, safety performance is a critical measure of project success alongside the traditional “iron triangle” view of time, cost, and quality (Alzahrani and Emsley, 2013). The best way to reach good safety performance is to mitigate or minimize risks before they occur (Fung et al., 2010). Therefore, safety performance in a construction project should be predicted and managed proactively to prevent accidents, and thus to improve the chances of project success. An accurate prediction depends on reliable and systematic analyses of the sources of risks, i.e., risk factors that can threaten safety performance. In the construction context, extensive attention has been paid

⁎

to fragmented factors, including human factors such as safety behavior (Guo et al., 2016), safety attitude (Langford et al., 2000), and risk tolerance (Wang et al., 2016), as well as organizational and environmental factors such as senior management commitment (Zou and Sunindijo, 2013) and frontline supervision (Fang et al., 2015). These studies have made valuable contributions by examining the individual effects of those factors on safety performance. However, how project safety performance can be predicted based on the collective effects of those factors remains unaddressed. Safety performance can be determined by a number of organizational, environmental and human factors (Hale et al., 2012; Zhou et al., 2014). Furthermore, existing studies have mainly focused on the unsafe behavior of frontline workers, while less attention was paid to the underlying causes behind that behavior, such as organizational factors (Fung et al., 2010; Zou and Sunindijo, 2013). Unsafe acts, as active failures, are hard to foresee, while latent variables like senior management commitment can be identified and then mitigated before an accident or incident occurs

Corresponding author. E-mail addresses: [email protected] (N. Xia), [email protected] (P.X.W. Zou), [email protected] (X. Liu), [email protected] (X. Wang), [email protected] (R. Zhu).

http://dx.doi.org/10.1016/j.ssci.2017.09.025 Received 3 April 2017; Received in revised form 24 August 2017; Accepted 26 September 2017 0925-7535/ © 2017 Published by Elsevier Ltd.

Safety Science 101 (2018) 332–343

N. Xia et al.

(Love et al., 2016). To summarize, in order to predict safety performance, together with mitigating risk factors especially adverse latent variables, a holistic causation analysis framework for safety performance is a prerequisite while it seems to lack in the current literature. The present study, thus, aims to establish a hybrid BN-HFACS model for predicting safety performance in construction projects. First, as discussed above, the underlying factors influencing safety performance, i.e., safety risk factors in construction projects are holistically explored using the Human Factors Analysis and Classification System (HFACS) (Shappell and Wiegmann, 2000). The HFACS focuses on the underlying causes of accidents, especially the management and organization aspects, and is widely used for safety and accident analysis in various areas such as maritime, railway, and construction (Akyuz, 2017; Chen et al., 2013; Garrett and Teizer, 2009; Zhan et al., 2017). In this study, the original HFACS is modified to fit the specific context of construction projects. By integrating BNs thereafter, the model established is capable of describing the relationships between safety risk factors and project safety performance, and in predicting the probabilities of project safety states and failure, and in diagnosing the most sensitive factor causing project safety failure. BNs and HFACS have been used separately in previous construction safety research, but they are integrated for the first time in this study. Such a combination will advance our full understanding of the underlying causes of construction project safety failure, and of the interrelationships among the risk sources and their total effects on project safety performance. In the following section, main characteristics, current applications and gaps concerning HFACS and BNs, and advancement that the present study can make to these two theories are presented.

Fig. 1. Example of a Bayesian Network.

the HFACS framework that currently remain at identifying the core risk factors of accidents (Akyuz, 2017; Soner and Asan, 2015; Zhan et al., 2017). 2.2. Bayesian Networks (BNs) BNs, also called belief networks, Bayesian belief networks, influence diagrams, or causality diagrams, are directed acyclic graphical models that represent a set of random variables (i.e., nodes) and their conditional dependent or influencing relationships (McCabe et al., 1998). Fig. 1 describes a simple example of a BN. The nodes acting as independent variables are also called parent nodes (root nodes), while nodes acting as dependent variables are also called child nodes (intermediate and leaf nodes). BNs are especially well suited to risk assessment and reliability problems (Francis et al., 2014), including the construction field. Luu et al. (2009) incorporated a BN into a schedule risk model to quantify the probability of project delays and proved its effectiveness. Nasir et al. (2003) also focused on schedule risk assessment, while Khodakarami and Abdi (2014) focused on cost risk. Concerning construction safety, Leu and Chang (2013) established a BN-based model for assessing safety risk in steel projects by combining a fault tree approach, whereas Zhang et al. (2013, 2014), and Wu et al. (2015) put forward a BN model for providing decision support relating to accident prevention and safety control in the event of tunnel-induced damage during construction. Focusing on individual safety behavior, Jitwasinkul et al. (2016) established a BN model for identifying the most critical organizational factors to enhance safe behavior. Although these applications were in different settings, it can be concluded that BNs are capable of risk and safety analysis, including prediction, diagnosis, decision-making, and the provision of insights into relationships among variables. In addition, BNs have been seldom utilized to predict safety performance of construction projects, despite the wide application in the construction setting. This study, thus, aims to introduce BNs as a suitable method for predicting safety states and failure in construction projects, and for diagnosing the causes of safety failure in a proactive manner.

2. Theoretical background and research gapsm 2.1. Human Factors Analysis and Classification System (HFACS) In most cases, accidents result from various factors which can be classified (Reason, 1990; Shappell and Wiegmann, 2000; Wiegmann and Shappell, 2001). Among existing classification methods, one particularly appealing approach is the Human Factors Analysis and Classification System (HFACS) originally designed within the aviation accident setting (Shappell and Wiegmann, 2000). HFACS describes four levels of failure: unsafe acts, preconditions for unsafe acts, unsafe supervision, and organizational influences, with the first level (unsafe acts) being closest to the accident itself. Since its initial development, HFACS has been demonstrated to be a valid tool for human error analyses in various fields, such as railways (Zhan et al., 2017), mining (Patterson and Shappell, 2010), maritime shipping (Chauvin et al., 2013), health-care (Diller et al., 2014), and aviation (Daramola, 2014). Despite its validity and wide-ranging applicability, the framework was less commonly applied in the context of the construction industry (Garrett and Teizer, 2009). This may be attributable to the inadequate focus on the underlying causes of safety failure and accidents in this industry, as argued previously. HFACS is valuable in providing a systematic analysis of accident causes. However, it considers merely the sequential influences from the higher levels on the lower levels. In reality, influencing relationships may exist among the cause levels that are not neighboring. For example, safety failures at the organizational factor level can lead directly to unsafe acts by employees (Hofmann and Stetzer, 1996). Given this, this study develops a more deliberate influencing network. Furthermore, the extant application of HFACS in the construction setting appears to remain mainly on a qualitative basis. Therefore, the present study utilizes Bayesian networks (BNs) to construct an interrelated network among the risk factors at different levels in HFACS, and between the risk factors and project safety performance. Based on this, a BN-HFACS model for quantitative prediction of safety states and failure in construction projects can be built. This quantitative application for performance prediction can add knowledge to quantitative analyses of

3. Research framework In order to depict clearly how HFACS and BNs can be combined in carrying out safety performance prediction, a systematic framework is proposed in Fig.2. Phase 1 is concerned with establishing a general BNHFACS model for safety performance prediction, including establishing an HFACS framework for describing and categorizing accident causes in construction projects (Step 1), and developing an influencing network among the variables in the HFACS with BNs (Step 2). In Phase 2, this model can be applied to a specific project to examine the feasibility and capability of the BN-HFACS model in real cases (Step 3). In applications, modifications should be made to the general BN-HFACS established in Phase 1 to reflect specific characteristics of different projects. Step 1. The first step is to establish a modified HFACS framework for analyzing safety risk factors specific to construction projects. To this end, the original HFASC framework was revised based on a literature review of the relevant knowledge of HFACS and safety and accident management in construction. We also consulted six experts who had been working in construction safety management for more than 15 years. These experts had participated in a large number of accident investigations in the construction industry. Details of Step 1 are provided in Section 4. 333

Safety Science 101 (2018) 332–343

N. Xia et al.

Fig. 2. Research framework.

Step 2. After the initial HFACS framework was modified to the construction context, the second step is to visualize the BN-HFACS model (Luu et al., 2009) – in particular, to establish the influencing relationships among different variables involved in the established HFACS framework. In the BN-HFACS model, each variable in the HFACS framework was treated as a node. For the influencing relationships among the nodes, it was hypothesized that (1) factors at each level (e.g., “inadequate supervision” in level 3) would influence the overall safety state at that level (e.g., L3: “unsafe supervision and monitoring”); (2) the safety states at higher levels would influence all lower one(s), for example, L4: “organization influence” would influence the safety states of all three lower levels, including L3: “unsafe supervision and monitoring,” L2: “preconditions for unsafe behavior,” and L1: “unsafe acts of workers”; and (3) the safety state of each level would directly influence the project safety performance. The above hypothesized influencing relationships were then examined using empirical data collected from a total of 142 practitioners in China’s construction industry. Before assessing the influencing relationships hypothesized, we investigated the feasibility of those factors (i.e., the criticality of each factor) in China’s construction industry. It was assessed on a 5-point Likert-type scale, from 1 representing very low and 5 representing very high. Similarly, the degree of influence of one node upon another is also assessed on a 5-point Likert-type scale. The average scores of the experts’ opinions on both factor criticality and path strength were rounded into a five-point format (from 1 = “very low” to 5 = “very high”). Out of those 142 sets of questionnaires, 78 sets were valid, reaching a valid response rate of 54.9%. This rate was considerably higher than the normal for questionnaire surveys in the construction industry (Akintoye, 2000). The respondents’ backgrounds are described in Table 1. Results of Step 2 are presented in Section 5. Step 3. In Step 3, the established BN-HFACS model was applied to a real case, validating the capability of the model in assisting practitioners to predict project safety performance. When a BN model is used for prediction, prior-probabilities of root nodes in the model should be given. In traditional BNs, expert judgments are used for assigning priorprobabilities, namely, experts should provide probabilistic parameters of root nodes in a BN-model (Khodakarami and Abdi, 2014; McCabe et al., 1998). However, the requirement for probabilistic parameters

Table 1 Respondents’ Background. Item

Frequency

Percent (%)

Project type Building (commercial, industrial, and residential) Infrastructure (subway, highway, and airport)

32 46

41.0 59.0

Work experience (years) ≤3 4–6 ≥7

13 47 18

16.7 60.3 23.1

Role Client Contractor Supervisor

23 20 35

29.5 25.6 44.9

Position Project manager Safety director Safety supervisor Civil engineer

17 19 32 10

21.8 24.4 41.0 12.8

acts as a major barrier to BNs application, because construction experts working at a practical level in the industry normally have limited or poor knowledge of probability theory (McCabe et al., 1998; Taroun, 2014). To overcome this barrier, and to improve the practicability of the established model, this research used the ranked nodes/paths method. Fenton et al. (2007) first proposed the ranked nodes method, in which scale judgment is adopted rather than precise data; as a result, the amount of evaluation data dramatically decreased. Like ranked nodes (Neil et al., 2005; Fenton et al., 2007), the ranked paths method also provides a good solution to experts’ inadequate knowledge in probability theory and can relieve the exhaustion of eliciting a large number of prior-probabilities (Xia et al., 2017). With the ranked nodes/ path method, two types of parameters should be assessed by experts (Fenton et al., 2007; Xia et al., 2017): (1) the criticality of root nodes to safety failure, and (2) the degree of influence of one node upon another. In this research, both factor criticality and path strength were assessed on a 5-point Likert-type scale, and the average scores of the experts’ 334

Safety Science 101 (2018) 332–343

N. Xia et al.

Fig. 3. Modified HFACS for the causation analysis of construction safety performance.

opinions were also rounded into a 5-point format (from 1 = “very low” to 5 = “very high”). With these two types of data input into the AgenaRisk software and the parameters defined in the software, postprobabilities of child nodes can be rapidly calculated in AgenaRisk (2016). Details of this case application are described in Section 6.

4.1. L1: Unsafe acts Workers are the last defense in situations of risk or accidents; their unsafe acts are the dominant cause of accidents (Didla et al., 2009). Unsafe acts are divided into errors and violations (Reason, 2000). Errors refer to unintended activities that fail to meet their desired outcomes. It is inevitable for human beings to make errors, such as slips, lapses, and mistakes. The original HFACS includes skill-based errors, decision errors, and perceptual errors. Skill-based errors are referred to as those errors resulting from memory slips and/or attention misallocation when individuals conduct repetitive work. In onsite construction projects, most workers (e.g., steel fixers) normally carry out one type of work for many years and they are extremely familiar with the work; thus, memory and/or attention failures may occur and further result in skill-based errors. More importantly, skill-based errors frequently occur because of the inadequate skills of individuals. For example, a construction procedure may involve a series of complex steps and workers may arrange those steps in the wrong order by

4. Modifying the HFACS for construction context As HFACS has been seldom applied in the general context of construction projects (Garrett and Teizer, 2009), its original structure and classifications need to be adjusted according to the specific features of construction projects. The main tactics modifying the HFACS framework include adding or removing items and adding additional layers to the original framework (Zhan et al., 2017). The present study then deleted, adjusted certain items, and added some items and an additional layer (Level 5: Environmental influences). The modified HFACS is shown in Fig. 3, with main changes and reasons for the changes presented below. 335

Safety Science 101 (2018) 332–343

N. Xia et al.

two are risk factors associated with human behavior. Adverse mental conditions include distraction and mental fatigue, whereas adverse physiological conditions refer to illness, injury, physical fatigue, or an otherwise impaired physiological state. These two kinds of impaired states mainly result from stressors such as sleep loss. In the construction industry, construction workers normally work under high pressure and challenging demands (Hoffmeister et al., 2014). This demanding environment is likely to place great pressure on workers, which would then impair their mental and/or physical states. Once in an adverse state, either mental or physical, there is a chance of unsafe acts taking place. The original HFACS framework considers adverse mental and physical states resulting from external environmental factors. By contrast, personal readiness in the original HFACS framework accounts for instances in which people engage in their job while disregarding crew rest requirements, violating alcohol restrictions, selfmedicating, etc. Namely, personal readiness is internally caused by the workers themselves rather than external factors. This research accordingly integrated personal readiness into the adverse mental and physiological condition categories that may be caused by both external and internal factors. Construction projects, especially large-scale ones, involve large numbers of workers cooperating with each other (Sousa and Einstein, 2012). As a consequence, it is especially important to guarantee effective safety/risk communication and coordination in teamwork among workers and between managers and workers. All these can contribute to accident prevention. In the original HFACS framework, communication and coordination were important aspects of crew resource management. Due to the particular importance of these aspects in the construction context, this research separated them away from the original crew resource management factor and into a single factor. The physical and mental limitations in the original HFACS were removed since they are inherent in the nature of individuals. In contrast, safety awareness was recognized as an important precondition of unsafe acts. Compared to traditional emphases on time, cost, and quality, safety receives a lower priority and poor safety awareness in construction personnel has long been recognized (Tam et al., 2004; Shin et al., 2014). It is possible that weak safety awareness may result in unsafe acts.

mistake. Moreover, construction projects, especially large-scale ones, frequently involve complex equipment and hydrogeological conditions, in which case new techniques and materials will be introduced. However, construction workers may not grasp the newly adopted techniques proficiently, possibly resulting in skill-based errors. In this research, both decision and perceptual errors were placed in a single category in the modified HFACS, because they are both generally cognitive biases regarding actual states in a construction process. The best phrase to describe decision errors is “honest mistakes.” That is, although decisions are made in pursuit of good outcomes, they can prove to be inappropriate or wrong in the context of an actual situation. A decision error occurs when information, knowledge, or experience is lacking (Diller et al., 2014). When an individual’s perceptions of the world differ from reality, perceptual errors often occur. In other words, errors occur when the individual sensory input is degraded or incomplete due to objective factors. For example, construction workers may not be able to see warnings about a foundation pit when carrying out a task in darkness, and thus risky acts may occur. Complexities both of equipment and of the environment in large-scale construction processes make construction projects vulnerable to both decision errors and perceptual errors. Violations are defined as workers’ intended deviations from the rules and/or regulations related to safety, such as not wearing safety helmets, chatting during the working period, and disregarding standard technical processes (Reason, 2000). Generally, the frequency of violations is lower than the frequency of errors, but the severity of violations is mostly high (Fogarty and Shaw, 2010). In the original HFACS framework, violations are divided into habitual violations and exceptional violations. Habitual violations mainly occur over a very long period of either neglecting or not knowing the regulations or appropriate skills. The experts reflected that habitual violations had occurred frequently in almost every project in which they had participated and that they can easily cause incidents. They further argued that many supervisors tolerated the act of not wearing safety helmets. In a worse case, in certain projects, safety helmets were not provided. This indicates that it is important to create conditions that reduce or avoid the possibility of habitual violations. In contrast, exceptional violations refer to irregular and accidental violations that are hard to predict and rarely occur (Wiegmann and Shappell, 2001). As they are always unexpected, it is impractical to cope with them proactively. Accordingly, the present study excluded exceptional violations. To summarize, L1: “unsafe acts of workers” finally included three factors: L1R1: “skill-based errors,” L1R2: “decision and perceptual errors,” and L1R3: “habitual violations.”

4.3. L3: Unsafe supervision The behavior of frontline supervisors will affect their workers’ attitude and behavior towards safety (Fang et al., 2015; Kouabenan et al., 2015). In the original HFACS model, unsafe supervision is divided into four categories: inadequate supervision, planned inappropriate operations, failure to correct problems, and supervisory violations. In fact, the first three aspects all referred to inadequate supervision and mainly occurred at different stages. Specifically, planned inappropriate operations, inadequate supervision, and failure to correct problems refer respectively to before, during, and after conducting a supervision task. Therefore, these three aspects were combined into one item denoted L3R1: “inadequate supervision.” One interviewee from a client said that supervisory violations occurred constantly in construction projects. Professionals in the construction industry have been criticized for not adhering to their corresponding obligations (Vee and Skitmore, 2003). Frequent violating behaviors of supervisors in construction projects include authorizing processes that are not up to standard, failing to enforce rules and regulations, authorizing workers to undertake dangerous construction work, and conducting supervision without qualifications. Thus, “supervisory violations” is the second factor at the third level and is labeled L3R2. Supervision refers to inspection by human beings, while continuous monitoring means using automatic instruments to monitor the safety state of project conditions onsite. The present research added in this factor because automatic monitoring is critical for accident prevention in construction projects, due to their dynamic and complex features

4.2. L2: Precondition of unsafe acts It has been widely acknowledged that unsafe acts of employees contribute to a high proportion of accidents (Didla et al., 2009; Heinrich et al., 1950). However, it is of great importance to investigate the preconditions of unsafe acts which lead directly to them. This is the first step in conducting a causation analysis of accidents. Given the characteristics of construction projects, modifications were made to the original HFACS. In this research, risk sources at level 2 were categorized into two sub-groups: “work conditions” and “worker and personnel conditions.” “Work conditions” refers to L2R1: “physical environment” regarding the living and working environment (e.g., room layout, construction dust and noise, and excessive clutter), and to L2R2: “technological conditions,” including equipment conditions, material conditions, and technical measures. Conditions and practices of operators in the original HFACS were combined into one category named “worker and personnel conditions,” comprising L2R3: “adverse mental conditions,” L2R4: “adverse physiological conditions,” L2R5: “ineffective safety/risk communication and coordination,” and L2R6: “poor safety awareness.” The first two factors refer to impersonal substandard conditions, whereas the latter 336

Safety Science 101 (2018) 332–343

N. Xia et al.

L5R1: “insufficient project stakeholder coordination,” L5R2: “poor social and industrial environment,” and L5R3: “inappropriate legislation and enforcement.” This additional level follows the attempts of previous research such as Patterson and Shappell (2010) and Chen et al. (2013). Construction projects involve and can be affected by a large number of stakeholders (Mok et al., 2015). Therefore, project stakeholder coordination was included as an external influence. For example, poor relationships and cooperation among project stakeholders can incur severe conflicts, improper actions, and destructive reactions (Klijn and Teisman, 2003). Furthermore, the social and industrial environment concerning safety issues directly affects people’s attitudes and habits towards risk management. Especially in developing countries like China, slack attention to safety management and negligence of social impacts can lead to inefficient management commitment, supervisory practices, and subsequent adverse consequences (Van Os et al., 2015). Finally, legislation and enforcement are also critical factors for determining a good safety climate and practices in an industry. Previous studies have argued that incomplete legislation and poor enforcement are key causes for injuries and accidents in the construction industry (Hale et al., 2012; Yu et al., 2014).

(Zou et al., 2007). In short, certain hidden risks are difficult for human beings to identify. Therefore, many investigators have devoted time and effort to establishing dynamic monitoring and warning systems (Ding et al., 2013; Ding and Zhou, 2013). Accordingly, L3: “unsafe supervision and monitoring” in the modified HFACS includes L3R1: “inadequate supervision,” L3R2: “supervisory violations,” and L3R3: “inadequate dynamic monitoring.” 4.4. L4: Organizational influences In HFACS, organizational influences refer to the internal management aspects of the party who is directly responsible for safety management (i.e., contractors in construction projects). In the modified HFACS, L4: “adverse organizational influences” contained L4R1: “poor management commitment,” L4R2: “inappropriate procedures and practices,” and L4R3: “inadequate investment in safety management.” In the original HFACS framework, organizational climate is one factor in this level. In the present research, we replaced organizational climate with safety climate, with a specific focus on the safety issue. By definition, safety climate refers to policies, procedures, and practices in an environment and serves as a frame of reference for guiding and directing appropriate and adaptive safety behavior in carrying out activities (Zohar, 2010). However, the interviewees commented that this concept sounded too broad and confusing and that it is difficult for construction practitioners to assess in practice. As a result, this aspect was divided into two aspects, namely management commitment, and procedures and practices, which are two important dimensions of safety climate (Zohar, 1980, 2000). Neal and Griffin (2004) defined management commitment concerning safety as “the extent to which management is perceived to place a high priority on safety and communicate and act on safety issues effectively” (p. 27). Many studies have attested to the positive association between higher management commitment and positive safety outcomes (e.g., Hofmann and Morgeson, 1999). Furthermore, the importance of management commitment to safety lies in its far-reaching influence on safety management strategies (Jitwasinkul et al., 2016). Namely, if supervisors and workers perceive a high commitment to safety from senior managers, they will be likely to engage in safer behavior. Otherwise, under the high stress of production, they may not have sufficient time to behave safely or may tend to take short-cuts to save time (Choudhry and Fang, 2008; Guo et al., 2016). In the interview, one supervisor said, “We understand that intense production pressure will pose adverse mental conditions on workers. But we can do little to alter this situation as project managers always push us to accomplish a project as early as possible.” This statement points to the importance of top managers’ valuing safety over other project objectives such as the schedule. The factor named “procedures and practices” refers to procedures and practices in a project that should be carefully designed for enhancing project safety performance. Resource management refers to the management, allocation, and maintenance of organizational resources, including human resource management (selection, training), monetary safety budgets, and equipment design (ergonomic specifications). Among these aspects, the investment in safety management was considered by the interviewees to be the most important factor. Similarly, Yu et al. (2014) found that the ratio of safety investment to the total project investment ranked top in terms of occurrence of accidents in metro construction projects. Therefore, we substituted “investment in safety management” for the original resource management factor.

5. Empirical findings of factor criticality and influencing relationships The respondents’ ratings on the criticality of the 18 identified safety risk factors are presented in Table 2. The factors were ranked according to the sample population mean values. Overall, fifteen among the total 18 factors were deemed critical (Mean ≥ 3.00 on a 5-point scale), which to some extent validated the feasibility of the safety risk factors identified using a combination of literature review and expert interviewing. The top five critical factors include L2R6: “poor safety awareness,” L3R2: “supervisory violations,” L1R3: “habitual violations,” L4R2: “inappropriate procedures and practices,” and L1R1: “skill-based errors.” First, it can be concluded that L1: “unsafe acts of workers” was a significant source of accidents because two factors at this level were ranked as the third and fifth. This finding confirms the argument about the dominant role of individual unsafe behavior in accidents and lasting efforts on mitigating risky behavior (Didla et al., 2009; Fang et al., 2015; Langford et al., 2000). More importantly, the respondents considered poor awareness of safety among project parties as the most critical factor for accidents. This is consistent with the previous stress on improving construction practitioners’ safety awareness (e.g., Tam et al., 2004), and thus justified the inclusion of this Table 2 Mean, standard deviation (SD), and rank of variables.

4.5. L5: Environmental influences In contrast to the original HFACS, this research included a new level, namely L5: “adverse environmental influences,” above the fourth level relating to organizational influences. This level refers to the influences from the environment external to the onsite project, including 337

Risk source (ID)

Mean

SD

Rank

Skill-based errors (L1R1) Decision and perceptual errors (L1R2) Habitual violations (L1R3) Adverse physical environment (L2R1) Adverse technological environment (L2R2) Adverse mental conditions (L2R3) Adverse physiological conditions (L2R4) Ineffective safety/risk communication and coordination (L2R5) Poor safety awareness (L2R6) Inadequate supervision (L3R1) Supervisory violations (L3R2) Inadequate dynamic monitoring (L3R3) Poor management commitment (L4R1) Inappropriate procedures and practices (L4R2) Inadequate investment in safety management (L4R3) Insufficient project stakeholder coordination (L5R1) Poor social and industrial environment (L5R2) Inappropriate legislation and enforcement (L5R3)

3.40 3.35 3.60 2.54 3.36 3.12 3.10 3.13

0.98 1.01 0.94 1.06 1.12 1.04 1.33 1.31

5 9 3 18 7 12 13 11

3.87 3.40 3.62 3.36 3.03 3.44 3.24 2.92 2.73 3.08

1.00 0.94 1.00 1.06 1.00 1.13 0.96 1.07 1.17 1.02

1 6 2 8 15 4 10 16 17 14

Safety Science 101 (2018) 332–343

N. Xia et al.

Fig. 4. The conceptual hybrid BN-HFACS model for predicting safety performance in construction projects.

factor in the modified HFACS framework. In addition, L3R2: “supervisory violations” and L4R2: “inappropriate procedures and practices” respectively from the third (supervision) and fourth (organization) levels were considered to be important in construction accidents. This may imply the necessity and value of considering the potential underlying causes of unsafe behavior, especially supervisors’ supervisory obligations and the establishment of appropriate safety procedures and practices. For example, it has been argued that some of the safety practices appear to be impracticable (Fogarty and Shaw, 2010). By contrast, “adverse physical environment (L2R1)” (rank = 18) and “poor social and industrial environment (L5R2)” (rank = 17) were considered relatively uncritical. In China, construction projects have witnessed repeated large-scale accidents in recent years (Zhou and Irizarry, 2016). Correspondingly, the government has released more stringent regulations on construction safety, which may contribute to improved physical working conditions and emphasis on safety throughout the entire industry. The results of examination of the hypothesized influencing relationships (Fig. 4) are depicted in Table 3. Concerning the overall influences of each level on final project safety performance, the influences of L1: “unsafe acts of workers” and L2: “preconditions for unsafe acts” were rated as the most important. Among the influencing relationships among different levels, the risk propagation L2 → L1: “preconditions for unsafe acts → unsafe acts of workers” was the most critical. At Level 1, “habitual violations” influenced the overall safety performance at this level the greatest. For Levels 2, 3, 4, and 5, the strongest influencing paths were L2R6 → L2: “poor safety awareness → preconditions for unsafe acts,” L3R2 → L3: “supervisory violations → unsafe supervision and monitoring,” L4R1 → L4: “poor management commitment → adverse organizational influences,” and L5R3 → L5: “inappropriate legislation and enforcement → adverse environmental influences,” respectively. Overall, the majority of the expected influences were confirmed

with an average value above 3.00 (“medium” influence degree); and the values of the seven exceptions were all nearly 3.00. To summarize, according to the questionnaire survey, it can be concluded that the hypothesized BN-HFACS model was supported concerning the feasibility of those identified safety risk factors, as well as the demonstrated influencing relationships among those factors. The general BN-HFACS model is described in Fig. 4, which can be adjusted for safety performance prediction in a specific construction project. 6. Application of the BN-HFACS model in a real case construction project Using a real-life case application, this section describes how the BNHFACS model developed by this research can assist project management personnel to predict safety performance of construction projects. The selected case is the subway station construction project of Line Six in Tianjin, China. A project management team from the China Railway Construction Bridge Engineering Bureau Group Co., LTD won the bid for this project and was responsible for its construction. With a budgeted investment of 370 million RMB, the entire project covered 12,000 square meters and 2350 m of a shield-driven tunnel. Starting on 12 November 2015, it was scheduled to be completed by approximately July 2017. Seven individuals were invited to participate in the case application: one general manager, one civil engineer, two project managers, one safety director, and two safety supervisors. All of them had been participating in the project from the beginning. During the application, one of the present authors presented the visualized model. The experts agreed with the factors included and the influencing relationships established in the model. Then they were asked to individually rate the criticality of each safety risk factor (root node). Responses were given on a designated questionnaire and the levels of risk criticality ranged 338

Safety Science 101 (2018) 332–343

N. Xia et al.

Table 3 The influencing relationships among variables. Category and Relationship (ID)

Mean

SD

Rank

Single-level influences on project safety performance Unsafe acts of workers → Project safety performance (L1 → PSP) Preconditions for unsafe acts → Project safety performance (L2 → PSP) Unsafe supervision and monitoring → Project safety performance (L3 → PSP) Adverse organizational influences → Project safety performance (L4 → PSP) Adverse environmental influences → Project safety performance (L5 → PSP)

4.21 3.81 3.71 3.37 2.92

0.99 0.91 0.91 0.91 1.06

1 3 6 13 33

Influences among levels Adverse environmental influences → Adverse organizational influences (L5 → L4) Adverse environmental influences → Unsafe supervision and monitoring (L5 → L3) Adverse environmental influences → Preconditions for unsafe acts (L5 → L2) Adverse environmental influences → Unsafe acts of workers (L5 → L1) Adverse organizational influences → Unsafe supervision and monitoring (L4 → L3) Adverse organizational influences → Preconditions for unsafe acts (L4 → L2) Adverse organizational influences → Unsafe acts of workers (L4 → L1) Unsafe supervision and monitoring → Preconditions for unsafe acts (L3 → L2) Unsafe supervision and monitoring → Unsafe acts of workers (L3 → L1) Preconditions for unsafe acts → Unsafe acts of workers (L2 → L1)

3.13 2.96 3.19 2.95 3.32 3.37 3.33 3.36 3.42 3.72

0.99 0.99 1.11 1.07 1.10 1.03 1.02 1.11 1.10 1.01

24 29 23 31 18 13 16 15 12 5

Level 1 Skill-based errors → Unsafe acts of workers (L1R1 → L1) Decision and perceptual Errors → Unsafe acts of workers (L1R2 → L1) Habitual violations → Unsafe acts of workers (L1R3 → L1)

3.44 3.27 3.76

1.25 1.15 1.06

11 20 4

Level 2 Adverse physical environment → Work Conditions (L2R1 → L2) Adverse technological environment → Work Conditions (L2R2 → L2) Adverse mental conditions → Worker and personnel conditions (L2R3 → L2) Adverse physiological conditions → Worker and personnel conditions (L2R4 → L2) Ineffective safety/risk communication and coordination → Worker and personnel conditions (L2R5 → L2) Poor safety awareness → Worker and personnel conditions (L2R6 → L2) Work Conditions → Preconditions for unsafe acts (L2R7 → L2) Worker and personnel conditions → Preconditions for unsafe acts (L2R8 → L2)

2.87 3.12 2.96 3.03 3.08 3.90 3.31 3.53

1.00 1.09 1.08 1.31 1.29 1.14 1.02 1.21

34 25 30 28 26 2 19 8

Level 3 Inadequate supervision → Unsafe supervision and monitoring (L3R1 → L3) Supervisory violations → Unsafe supervision and monitoring (L3R2 → L3) Inadequate dynamic monitoring → Unsafe supervision and monitoring (L3R3 → L3)

3.49 3.54 3.33

0.92 1.03 0.93

10 7 16

Level 4 Poor management commitment → Adverse organizational influences (L4R1 → L4) Inappropriate procedures and practices → Adverse organizational influences (L4R2 → L4) Inadequate investment on safety → Adverse organizational influences management (L4R3 → L4)

3.53 3.24 3.23

0.93 1.09 1.04

8 21 22

Level 5 Insufficient project stakeholder coordination → Adverse Environmental influences (L5R1 → L5) Poor social and industrial environment → Adverse Environmental influences (L5R2 → L5) Inappropriate legislation and enforcement → Adverse Environmental influences (L5R3 → L5)

2.86 2.94 3.06

1.14 0.99 0.98

35 32 27

(L5), at the medium level, was estimated at 75.66%. To summarize, the safety performance of the project and four cause levels out of five were predicted at the high-risk level. Thus, the project management team ought to conduct continuous monitoring of the criticality of risk factors and to assess the total influence of those risk factors on project safety performance. Besides the capability in predicting the probabilities of safety states and failure, through sensitivity analysis, the constructed model was able to identify the most sensitive factor in terms of safety failure within a targeted project. Fig. 6 shows the project safety performance was the most sensitive to L5: “Environmental influences.” This suggests that management personnel in the subway project should be cautious about risk factors resulting from the environmental influence level, especially the social and industrial environment and legislation and enforcement concerning safety issues in construction. Although these factors are out of the control of project management teams, they should be aware of these factors and implement strict safety management. Project management teams should by no means slacken their management of safety issues, especially in conditions where the outside environment pays little attention to construction safety.

from 1 (very low) to 5 (very high). Then, the participants were asked to individually assess the degree of each influencing relationship in the BN-HFACS network on the same questionnaire. The degrees also ranged from 1 (very low) to 5 (very high). The above process took approximately 45 min. Almost all the items reached the required level (inter-rater agreement (Rwg) > 0.7) proposed by James et al. (1984). The final levels of risk criticality and influencing relationships were obtained according to the expert average ratings. With these parameters input in the model (Fig. 5) constructed in AgenaRisk, safety performance prediction can be conducted. From the BN-HFACS model in Fig. 5, it can be seen that child nodes, post-probabilities of which require being predicted, include L1: “unsafe acts of workers,” L2: “preconditions for unsafe acts,” L2R8: “worker and personnel conditions,” L2R7: “work conditions,” L3: “unsafe supervision and monitoring,” L4: “adverse organizational influences,” L5: “environmental influences,” and PSP: “project safety performance.” Each variable in the BN-HFACS model was assigned five risk levels, namely “very low,” “low,” “medium,” “high,” and “very high.” Table 4 shows the post-probability distribution obtained. It was predicted that the probability of project safety performance is approximately 84.12% at the high-risk level, while it is 15.30% at medium risk level. Risk sources originated from Levels 1–4 were all at high-risk levels, with probabilities of 87.51%, 74.20%, 96.58%, and 96.02%, respectively, while “Environment influences” 339

Safety Science 101 (2018) 332–343

N. Xia et al.

Fig. 5. The BN-HFACS model built in AgenaRisk for predicting safety performance in the studied project.

7.1. Theoretical implication

Table 4 Post-probability distribution of safety performance of the studied project. Risk sources

Very low Unsafe acts of workers (L1) Preconditions for unsafe acts (L2) Work conditions (L2R7) Worker and personnel conditions (L2R8) Unsafe supervision and monitoring (L3) Adverse organizational influences (L4) Environmental influences (L5) Project safety performance (PSP)

This research contributed to accident causation analysis and then safety performance prediction. Systematic examination of accident causes seems to be lacking with inadequate focus on the underlying causes or antecedents of safety performance (Fung et al., 2010; Zou and Sunindijo, 2013). In the proposed HFACS, accident causes were grouped into five levels: L1: “unsafe acts of workers,” L2: “preconditions of unsafe acts,” L3: “unsafe supervision and monitoring,” L4: “adverse organizational influences,” and L5: “adverse environmental influences.” It is especially argued that the new level, i.e., L5: “adverse environmental influences” should be considered. Construction projects can be affected by diverse stakeholders and social and economic factors (Mok et al., 2015). Thus, factors from the environment outside of the project can play a critical role in shaping the safety performance of construction projects. Specific factors concerning environmental influences include project stakeholder coordination, social and industrial environment, and legislation and enforcement. The modified HFACS framework provided an overall picture of the underlying causes of safety performance in construction projects. The incorporation of BNs with the HFACS thereafter facilitates forecasting safety performance in construction projects, thereby contributing to safety performance measurement, especially, to the stream of establishing leading indicators to predict safety performance in the future. Traditional research was to develop lagging indicators, such as total recordable incident rate, to measure safety performance, while recent research emerges on leading factors, such as safety leadership and strategic management (Awolusi and Marks, 2017). But few studies have analyzed those factors in a systematic way that can be multi-levels and multi-facets, and thus predicted the collective effects of those factors on safety performance. The model proposed addressed this limitation by considering the influences of multiple leading factors from

Risk levels (percent) Low

3.21

Medium

High

12.43 25.65

87.51 74.20

93.58

3.21 83.32

13.68

96.58

3.27

96.02

3.44

75.66 15.30

Very high

24.34 84.12

Note: Only the probabilities of intermediate and leaf nodes (i.e., child nodes) were predicted.

7. Discussion This research has established a hybrid BN-HFACS model for predicting safety performance in construction projects. Theoretical analyses and an empirical survey demonstrated this model, in which (1) a total of 18 safety risk factors were identified and classified into five levels, and (2) the influencing relationships between the risk factors and project safety performance were established. Finally, a real-case application showed the model’s capabilities in safety assessment and cause diagnosis.

340

Safety Science 101 (2018) 332–343

N. Xia et al.

Fig. 6. Sensitivity analysis of the project safety failure.

five levels and aspects, and by quantifying their collective influences on project safety performance. This research provides a new model (a hybrid BN-HFACS) for safety performance prediction. The present study also contributed to both BNs and HFACS theories. BNs have been rarely used for predicting project safety performance; this research expanded its application. The demonstration of the construction-specific HFACS framework validated the HFACS’s ability in accident analyses in the construction context. The HFACS is mainly used for qualitative analyses of accident causes, and current quantitative analyses merely focus on identifying the core safety risk factors. Hence, this research contributed to the HFACS by expanding its quantitative application for predicting safety performance. Another contribution to the HFACS was that we explored the influencing relationships among the cause levels that are not neighboring in the framework.

projects. Furthermore, the BN-HFACS model based on this HFACS can help practitioners to predict the probabilities of project safety states and specific states at each cause level, and to diagnose the most sensitive risk factor contributing to project safety failure. This is meaningful, as practitioners were found to lack systematic approaches towards safety and risk assessment in the construction domain (Fung et al., 2010), especially for overall projects (Taroun, 2014). More importantly, as construction practitioners frequently find it difficult to use complex quantitative techniques for risk assessment (Taroun, 2014), this research provides an approachable visualization of the BNs-based model, an easily understood ranked nodes/paths method for eliciting quantitative prior-probabilities, and a software (i.e., AgenaRisk) for calculating post-priorities rapidly. During the applied case, the involved participants reflected that the model and the calculating processes in AgenaRisk could be easily understood. Finally, as the variables in a Bayesian network can be added or deleted without affecting the remainder of the network, the model proposed by the present study can be modified and then applied to different projects where safety risks may vary.

7.2. Practical implication For safety and accident management in the construction industry, practitioners have long been focused on preventing and/or reducing unsafe behavior of workers; applied measures include material incentives, goal setting, performance feedback, and positive or negative management reinforcement (Fang et al., 2015; Hermann et al., 2010). However, these measures lack an in-depth intervention to the root causes of unsafe acts. The modified HFACS suggests that construction practitioners should not only focus on preventing the unsafe behavior of frontline workers, but also on seeking interventions by thoroughly examining the underlying causes of accidents on site. In particular, besides unsafe acts as the first level of the causes of accidents, the modified HFACS suggests that attention should be paid to underlying factors at other four levels: L2: “preconditions for unsafe behavior,” L3: “unsafe supervision and monitoring,” L4: “adverse organization influences,” and L5: “adverse environmental influences.” This modified HFACS provides practitioners with a framework to thoroughly identify human, organizational, and environmental causes of accidents in construction

7.3. Limitations and future research directions Despite the above implications for theory and practice, this research has several limitations. First, the generalization of the proposed BNHFACS model should be cautious. As historical data regarding construction accidents were not available for the present study, this research modified the HFACS based on related studies and expert interviews, and the influencing relationships were demonstrated through a questionnaire survey on construction practitioners in China. We do not intend to provide a standardized, exhaustive accident causation framework and model, but to provide an initial model which can help analyze future accidents and predict project safety performance. It, thus, suggests that future research should modify and then improve the 341

Safety Science 101 (2018) 332–343

N. Xia et al.

causes in cargo ships. Saf. Sci. 92, 17–25. Alzahrani, J.I., Emsley, M.W., 2013. The impact of contractors’ attributes on construction project success: a post construction evaluation. Int. J. Project Manage. 31 (2), 313–322. Awolusi, I.G., Marks, E.D., 2017. Safety activity analysis framework to evaluate safety performance in construction. J. Construct. Eng. Manage. 143 (3), 05016022. Chauvin, C., Lardjane, S., Morel, G., Clostermann, J.-P., Langard, B., 2013. Human and organisational factors in maritime accidents: analysis of collisions at sea using the HFACS. Accid. Anal. Prev. 59, 26–37. Chen, S.T., Wall, A., Davies, P., Yang, Z., Wang, J., Chou, Y.H., 2013. A Human and Organizational Factors (HOFs) analysis method for marine casualties using HFACSMaritime Accidents (HFACS-MA). Saf. Sci. 60, 105–114. Choudhry, R.M., Fang, D., 2008. Why operatives engage in unsafe work behavior: investigating factors on construction sites. Saf. Sci. 46 (4), 566–584. Daramola, A.Y., 2014. An investigation of air accidents in Nigeria using the Human Factors Analysis and Classification System (HFACS) framework. J. Air Transp. Manage. 35, 39–50. Didla, S., Mearns, K., Flin, R., 2009. Safety citizenship behaviour: a proactive approach to risk management. J. Risk Res. 12, 475–483. Diller, T., Helmrich, G., Dunning, S., Cox, S., Buchanan, A., Shappell, S., 2014. The Human Factors Analysis Classification System (HFACS) applied to health care. Am. J. Med. Qual.: Off. J. Am. College Med. Qual. 29 (3), 181–190. Ding, L.Y., Zhou, C., 2013. Development of web-based system for safety risk early warning in urban metro construction. Autom. Construct. 34, 45–55. Ding, L.Y., Zhou, C., Deng, Q.X., Luo, H.B., Ye, X.W., Ni, Y.Q., Guo, P., 2013. Real-time safety early warning system for cross passage construction in Yangtze Riverbed Metro Tunnel based on the internet of things. Autom. Construct. 36, 25–37. Fang, D., Wu, C., Wu, H., 2015. Impact of the supervisor on worker safety behavior in construction projects. J. Manage. Eng. 31 (6), 04015001. Fang, D., Wu, H., 2013. Development of a Safety Culture Interaction (SCI) model for construction projects. Saf. Sci. 57, 138–149. Fenton, N.E., Neil, M., Caballero, J.G., 2007. Using ranked nodes to model qualitative judgments in Bayesian networks. IEEE Trans. Knowl. Data Eng. 19 (10), 1420–1432. Fogarty, G.J., Shaw, A., 2010. Safety climate and the theory of planned behavior: towards the prediction of unsafe behavior. Accid. Anal. Prev. 42 (5), 1455–1459. Francis, R.A., Guikema, S.D., Henneman, L., 2014. Bayesian Belief Networks for predicting drinking water distribution system pipe breaks. Reliab. Eng. Syst. Saf. 130, 1–11. Fung, I.W.H., Tam, V.W.Y., Lo, T.Y., Lu, L.L.H., 2010. Developing a Risk Assessment Model for construction safety. Int. J. Project Manage. 28 (6), 593–600. Garrett, J.W., Teizer, J., 2009. Human factors analysis classification system relating to human error awareness taxonomy in construction safety. J. Construct. Eng. Manage. 135 (8), 754–763. Guo, B.H.W., Yiu, T.W., González, V.A., 2016. Predicting safety behavior in the construction industry: development and test of an integrative model. Saf. Sci. 84, 1–11. Hale, A., Walker, D., Walters, N., Bolt, H., 2012. Developing the understanding of underlying causes of construction fatal accidents. Saf. Sci. 50 (10), 2020–2027. Heinrich, H.W., Petersen, D., Roos, N., 1950. Industrial Accident Prevention. McGrawHill, New York. Hermann, J.A., Ibarra, G.V., Hopkins, B.L., 2010. A safety program that integrated behavior-based safety and traditional safety methods and its effects on injury rates of manufacturing workers. J. Org. Behav. Manage. 30 (1), 6–25. Hoffmeister, K., Gibbons, A.M., Johnson, S.K., Cigularov, K.P., Chen, P.Y., Rosecrance, J.C., 2014. The differential effects of transformational leadership facets on employee safety. Saf. Sci. 62, 68–78. Hofmann, D.A., Morgeson, F.P., 1999. Safety-related behavior as a social exchange: the role of perceived organizational support and leader-member exchange. J. Appl. Psychol. 84 (2), 286–296. Hofmann, D.A., Stetzer, A., 1996. A cross-level investigation of factors influencing unsafe behaviors and accidents. Pers. Psychol. 49 (2), 307–339. James, L.R., Demaree, R.G., Wolf, G., 1984. Estimating within-group interrater reliability with and without response bias. J. Appl. Psychol. 69 (1), 85–98. Jitwasinkul, B., Hadikusumo, B.H.W., Memon, A.Q., 2016. A Bayesian Belief Network model of organizational factors for improving safe work behaviors in Thai construction industry. Saf. Sci. 82, 264–273. Khodakarami, V., Abdi, A., 2014. Project cost risk analysis: a Bayesian networks approach for modeling dependencies between cost items. Int. J. Project Manage. 32 (7), 1233–1245. Klijn, E.-H., Teisman, G.R., 2003. Institutional and strategic barriers to public—private partnership: an analysis of dutch cases. Publ. Money Manage. 23 (3), 137–146. Kouabenan, D.R., Ngueutsa, R., Mbaye, S., 2015. Safety climate, perceived risk, and involvement in safety management. Saf. Sci. 77, 72–79. Langford, D., Rowlinson, S., Sawacha, E., 2000. Safety behaviour and safety management: its influence on the attitudes of workers in the UK construction industry. Eng. Construct. Archit. Manage. (Wiley-Blackwell) 7 (2), 133–140. Leu, S.-S., Chang, C.-M., 2013. Bayesian-network-based safety risk assessment for steel construction projects. Accid. Anal. Prev. 54, 122–133. Love, P.E.D., Teo, P., Morrison, J., Grove, M., 2016. Quality and safety in construction: creating a no-harm environment. J. Construct. Eng. Manage. 142 (8), 05016006. Luu, V.T., Kim, S.-Y., Tuan, N.V., Ogunlana, S.O., 2009. Quantifying schedule risk in construction projects using Bayesian belief networks. Int. J. Project Manage. 27 (1), 39–50. McCabe, B., AbouRizk, S.M., Goebel, R., 1998. Belief networks for construction performance diagnostics. J. Comput. Civil Eng. 12 (2), 93–100. Mok, K.Y., Shen, G.Q., Yang, J., 2015. Stakeholder management studies in mega construction projects: a review and future directions. Int. J. Project Manage. 33 (2),

HFACS framework and the BN-HFACS model based on causation analyses of actual accident records and surveys with a larger sample. Second, the hypothesized BN-HFACS model considered the influence relationships between risk factors and their belonged cause level, among distinct cause levels, and between the five distinct cause levels and final project safety performance. However, the model did not consider the influence relationships among risk factors within the same level and across different levels. Another issue related to the influence relationships is whether the lower levels can influence the higher levels. Thus, it merits future studies on the establishment of more accurate influence relationships. Finally, expert judgment was utilized to elicit prior-probabilities of root nodes in the BN-HFACS model. Although this method is mostly used, it inevitably involves cognitive biases (Zhang and Thai, 2016). Future research may develop suitable ways to mitigate these biases. 8. Conclusion High rates of accidents exist in construction projects. Prediction of safety performance in construction projects is critical to safety management and project success. As the proverb goes, “we cannot manage what we cannot measure.” However, there seems to be a lack of reliable approaches to predict safety performance in construction projects. This research thus established a BN-HFACS model to quantitatively predict the probability distribution of safety performance in construction projects. First, an HFACS framework was modified for analyzing factors contributing to construction safety performance. It comprised a total of 18 key risk factors that were classified at five cause levels, i.e., L1: “unsafe acts of workers,” L2: “preconditions for unsafe acts,” L3: “unsafe supervision and monitoring,” L4: “adverse organizational influences,” and L5: “adverse environmental influences.” The established HFACS, together with codes of the variables included, can help project practitioners to identify potential safety risks more easily and holistically than has previously been possible. A BN-HFACS model was subsequently established using this modified HFACS, which was demonstrated to be capable of helping practitioners to predict the probabilities of safety states in project level and in the five specific cause levels, and to diagnose the most sensitive factor. The established HFACS contributes to the causation analysis of safety performance in construction projects by providing a systematic framework comprising diverse causes at five levels, as previous studies have mainly analyzed accident causes in a fragmented manner. Furthermore, the combination of HFACS and BNs makes HFACS for quantitative prediction of project safety performance possible and extends the application of BNs in the safety domain. Together, the proposed HFACS-BNs model for predicting project safety performance contributes to the research on safety performance prediction. The setting of this research is construction projects, but it can also act as a model for safety prediction using BNs and HFACS in other settings, especially where diverse risk factors are involved and may exert an integrated influence on project safety performance. Acknowledgments This research is fully supported by the National Natural Science Foundation of China (Grant Nos. 71231006, 71772136, and 71722004). Many thanks to all the participants who have participated in the empirical surveys. Besides, all the present authors appreciate the editors and anonymous referees for their constructive feedback. References Agenarisk, 2016. . Akintoye, A., 2000. Analysis of factors influencing project cost estimating practice. Construct. Manage. Econ. 18 (1), 77–89. Akyuz, E., 2017. A marine accident analysing model to evaluate potential operational

342

Safety Science 101 (2018) 332–343

N. Xia et al.

(HFACS). Aviat. Space Environ. Med. 72 (11), 1006–1016. Wu, X., Liu, H., Zhang, L., Skibniewski, M.J., Deng, Q., Teng, J., 2015. A dynamic Bayesian network based approach to safety decision support in tunnel construction. Reliab. Eng. Syst. Saf. 134, 157–168. Xia, N.N., Wang, X.Q., Wang, Y., Yang, Q.B., Liu, X., 2017. Lifecycle cost risk analysis for infrastructure projects with modified Bayesian networks. J. Eng. Des. Technol. 15 (1), 79–103. Yu, Q.Z., Ding, L.Y., Zhou, C., Luo, H.B., 2014. Analysis of factors influencing safety management for metro construction in China. Accid. Anal. Prev. 68, 131–138. Zhan, Q., Zheng, W., Zhao, B., 2017. A hybrid human and organizational analysis method for railway accidents based on HFACS-Railway Accidents (HFACS-RAs). Saf. Sci. 91, 232–250. Zhang, G., Thai, V.V., 2016. Expert elicitation and Bayesian Network modeling for shipping accidents: a literature review. Saf. Sci. 87, 53–62. Zhang, L., Wu, X., Ding, L., Skibniewski, M.J., Yan, Y., 2013. Decision support analysis for safety control in complex project environments based on Bayesian Networks. Expert Syst. Appl. 40 (11), 4273–4282. Zhang, L., Wu, X., Skibniewski, M.J., Zhong, J., Lu, Y., 2014. Bayesian-network-based safety risk analysis in construction projects. Reliab. Eng. Syst. Saf. 131, 29–39. Zhou, J.-L., Bai, Z.-H., Sun, Z.-Y., 2014. A hybrid approach for safety assessment in highrisk hydropower-construction-project work systems. Saf. Sci. 64, 163–172. Zhou, Z., Irizarry, J., 2016. Integrated framework of modified accident energy release model and network theory to explore the full complexity of the hangzhou subway construction collapse. J. Manage. Eng. 32 (5), 1–9. Zohar, D., 1980. Safety climate in industrial organizations: theoretical and applied implications. J. Appl. Psychol. 65 (1), 96–102. Zohar, D., 2000. A group-level model of safety climate: testing the effects of group climate on microaccidents in manufacturing jobs. J. Appl. Psychol. 85 (4), 587–596. Zohar, D., 2010. Thirty years of safety climate research: reflections and future directions. Accid. Anal. Prev. 42 (5), 1517–1522. Zou, P.X.W., Sunindijo, R.Y., 2013. Skills for managing safety risk, implementing safety task, and developing positive safety climate in construction project. Autom. Construct. 34, 92–100. Zou, P.X.W., Sunindijo, R.Y., 2015. Strategic Safety Management in Construction and Engineering. Wiley-Blackwell 250p. Zou, P.X.W., Sunindijo, R.Y., Dainty, A.R.J., 2014. A mixed methods research design for bridging the gap between research and practice in construction safety. Saf. Sci. 70, 316–326. Zou, P.X.W., Zhang, G., Wang, J., 2007. Understanding the key risks in construction projects in China. Int. J. Project Manage. 25 (6), 601–614.

446–457. Nasir, D., McCabe, B., Hartono, L., 2003. Evaluating Risk in Construction-Schedule Model (ERIC–S): construction schedule risk model. J. Construct. Eng. Manage. 129 (5), 518–527. Neal, A., Griffin, M., 2004. Safety climate and safety at work. In: Barling, J., Frone, M.R. (Eds.), The Psychology of Workplace Safety. Psychological Association, United States, pp. 15–34. Neil, M., Fenton, N., Tailor, M., 2005. Using bayesian networks to model expected and unexpected operational losses. Risk Anal. 25 (4), 963–972. Patterson, J.M., Shappell, S.A., 2010. Operator error and system deficiencies: analysis of 508 mining incidents and accidents from Queensland, Australia using HFACS. Accid. Anal. Prev. 42 (4), 1379–1385. Reason, J., 1990. Human Error. Cambridge University Press, New York. Reason, J., 2000. Human error: models and management. BMJ (Clinical Research Ed.) 320 (7237), 768–770. Shappell, S.A., Wiegmann, D.A., 2000. The human factors analysis and classification system–HFACS [microform]: final report/Scott A. Shappell, Douglas A. Wiegmann. Washington, DC: Office of Aviation Medicine, U.S. Dept. of Transportation, Federal Aviation Administration; Springfield, Va.: This document is available to the public through the National Technical Information Service, [2000]. Shin, M., Lee, H.-S., Park, M., Moon, M., Han, S., 2014. A system dynamics approach for modeling construction workers’ safety attitudes and behaviors. Accid. Anal. Prev. 68, 95–105. Soner, O., Asan, U., 2015. Use of HFACS-FCM in fire prevention modelling on board ships. Saf. Sci. 77, 25–41. Sousa, R.L., Einstein, H.H., 2012. Risk analysis during tunnel construction using Bayesian Networks: Porto Metro case study. Tunnel. Undergr. Space Technol. 27 (1), 86–100. Tam, C.M., Zeng, S.X., Deng, Z.M., 2004. Identifying elements of poor construction safety management in China. Saf. Sci. 42 (7), 569–586. Taroun, A., 2014. Towards a better modelling and assessment of construction risk: insights from a literature review. Int. J. Project Manage. 32 (1), 101–115. Van Os, A., Van Berkel, F., De Gilder, D., Van Dyck, C., Groenewegen, P., 2015. Project risk as identity threat: explaining the development and consequences of risk discourse in an infrastructure project. Int. J. Project Manage. 33 (4), 877–888. Vee, C., Skitmore, C., 2003. Professional ethics in the construction industry. Eng. Construct. Archit. Manage. 10 (2), 117–127. Wang, J., Zou, P.X.W., Li, P.P., 2016. Critical factors and paths influencing construction worker’ safety risk tolerances. Accid. Anal. Prev. 93, 267–279. Wiegmann, D.A., Shappell, S.A., 2001. Human error analysis of commercial aviation accidents: application of the Human Factors Analysis and Classification system

343