Finding causal paths between safety management system factors and accident precursors 1

11 Understanding the causal relationships between safety management system (SMS) factors and 12 accident precursors helps construction organizations identify which factors require improvement 13 upon observing an accident precursor. Previous research has not clearly distinguished between 14 SMS factors and accident precursors. This background examines the relationships between SMS 15 factors and accident precursors using empirical data. Specifically, five structural equation models 16 (SEMs) are developed to map causal paths between SMS factors and accident precursors. Each of 17 the SEMs helps identify what specific SMS factors would have a significant influence on the 18 occurrence of a particular type of accident precursor. These models can thus help describe what 19 specific SMS factors would need to be improved when a certain type of accident precursors 20 appears on site. The SEM results show in particular that the occurrence of accident precursors can 21 be attributed largely to adverse project conditions such as project schedule pressure, reworks, and 22 change orders. Construction organizations may capitalize on these findings by prioritizing safety 23 management resources to address specific observed accident precursors in a more informed and 24 targeted manner. 25


INTRODUCTION 26
To investigate the complex associations between the condition of SMS factors and the occurrence 118 of specific types of accident precursors in a quantitative manner, a structural equation modeling 119 (SEM)-based approach was used in this research. Specifically, the research was conducted in the 120 following two stages: (1) defining constructs and collecting empirical data for each measure of 121 SMS factors and accident precursors, and (2) constructing and testing SEMs to connect each type 122 of accident precursor with SMS factors. The data analysis stage was further divided into two steps: 123 (1) Confirmatory Factor Analysis (CFA), and (2) Structural Equation Modeling (SEM) and 124 analysis, as outlined by Hair et al. (2014). 125

Measures and Data Collection 126
Based on a comprehensive review of the construction safety management literature, a total of 28 127 SMS factors (Table 1) and 24 accident precursors (Table 2) were selected for inclusion in the 128 questionnaire. As indicated in Tables 1 and 2 After data collection, the data were pre-processed so that all variables could be interpreted such 157 that a higher value means a more undesirable state, whether or not the measure is related to a SMS 158 factor or accident precursor. 159 A link to the online survey questionnaire form was distributed to key contact individuals of 15 160 major construction companies in Alberta, Canada, who were asked to circulate the questionnaire 161 link to site managers, safety managers, and other construction practitioners in their companies. 162 Survey participation was voluntary, anonymous, and confidential. Respondents were asked to 163 respond to items based on their experience from their current or most recent project to reflect a 164 single project. A total of 102 responses were received, of which 6 were removed due to 165 incompleteness; therefore, 96 responses were used in the analysis stage. While the majority (60%) 166 of the respondents were currently working on an industrial construction project, 31% were in the 167 heavy construction sector, 6% in the building industry, and 3% in the other construction sectors of 168 the construction industry. Of those respondents, 24% were also health, safety, and environment 169 (HSE) managers, 25% were project managers, 21% were superintendents, 19% were other safety 170 staff members, and 11% had other managerial positions in the construction industry. The 171 respondents were predominantly from Alberta, Canada. 172

Data Analysis and Modeling 173
The data analysis process of this research was guided by the widely adopted SEM process 174 suggested by Hair et al. (2014). In the process, a confirmatory factor analysis (CFA) is first 175 performed to confirm that the small number of predetermined constructs (i.e., "Groups;" see 176 Tables 2 and 3)

Confirmatory Factor Analysis 200
Because the measurements used in this research are self-reported and collected through the same 201 questionnaire during the same period of time, a common method variance (a variance that is 202 attributed to the measurement method rather than the constructs of interest) could cause systematic 203 measurement errors. To ensure that the data is not substantially influenced by a common method 204 variance, the Harman's single factor test was applied. The result suggests that 23.54% of the 205 dataset variance could be explained by one latent factor, which is much lower than the 50% 206 threshold for common method variance (Podsakoff et al, 2003). 207 The CFA was conducted, and the results of the analysis on the SMS factors are summarized in 208 Table 3. To examine the factor models' reliability, the internal consistency of the measures for 209 each group was tested. Items with a factor loading of greater than 0.6 were accepted to be 210 unidimensional (Hair et al. 2014). The following SMS factors had a factor loading less than 0.6 211 and, therefore, were excluded from the factor models: Emergency  In addition, the convergent validity-the degree to which indicator variables correlate and share 216 variance with each other-was tested using the Average Variance Extracted (AVE) metric. 217 According to Fornell and Larcker (1981), it is recommended that AVE be 50% or greater. In addition, 218 the Composite Reliability (CR) test was used to evaluate the convergent validity of reflective 219 constructs. According to Hair et al (2014), CR has a threshold value of 0.7. The following factors 220 (Table 3)   Consistency, Convergent Validity, and CR-were also applied to the accident precursor factors. 230 All accident precursor factors also satisfied these criteria, and the factor models were therefore 231 deemed acceptable. 232

Hypotheses for Structural Models 233
Based on the CFA results, five SEMs were hypothesized: one for each accident precursor factor. 234 Each model was designed to examine the associations between one type of accident precursor and 235 the SMS factors. According to Ullman and Bentler (2003), the first phase in a SEM analysis is the 236 specification of a model. Although the factor analysis for each construct can be built based on 237 exploratory or confirmatory approaches, the researcher should hypothesize the causal paths and 238 directionality between the variables in the model specification (Gunzler and Morris 2015). That is, 239 a researcher is more likely to use SEM to determine whether a certain model is valid, rather than 240 using SEM to "find" a suitable model. In this research, the hypothesized relationships for each 241 structural model were constructed based on the research findings reported in the construction 242 safety management literature. The hypotheses tested in the structural models are summarized in 243 Table 5. 244

Final Causal Path Models between SMS Factors and Accident Precursors 245
The structural models based on the hypotheses were built using AMOS 24. The internal validity 246 test-the discriminant validity between the factors-was analysed to verify if each construct is 247 truly distinct from the others so as to avoid the issue of multicollinearity. According to Hair et al 248  Table 6. 257 The final model for the worker-related precursors (WOR) is illustrated in Figure 1 (Model 1). 258 Worker-related precursors (WOR) were found to be significantly affected by adverse project 259 conditions (ADV). Although the standardized coefficient (0.44) of the causal link from worker 260 behavior improvement efforts (BEHAV) to worker-related precursors (WOR) was higher than that 261 of the adverse project conditions (ADV) (0.42), the significance of this relationship was not 262 supported by the bootstrapping test (p > 0.05). As a note, the positive value of the coefficient 263 between BEHAV and WOR means that worker behavior improvement efforts can reduce worker-264 related precursors since all data were pre-processed such that a high value means an undesirable 265 state regardless of whether the variable is a SMS factor or an accident precursor. Similarly, the 266 causal link from resources for safety management (RES) to worker-related precursors (WOR) was 267 not supported by the test. The final model suggests that commitment to safety (COM) can 268 significantly affect resources for safety management (RES) as well as worker behavior 269 improvement efforts (BEHAV). 270 The model for Work team-related precursors (TEAM) is illustrated in Figure 2  Resources for safety management (RES), and also by adverse project conditions (ADV); however, 287 it did not support the hypothesis that site organization-related precursors (SITE) would be affected 288 by risk assessment and control efforts (RISK). As with Model 3, Model 4 confirms a strong 289 relationship between the following variables: between commitment to safety (COM) and project 290 administration for safety (ADMIN); and, between project administration for safety (ADMIN) and 291 risk assessment and control efforts (RISK); and lastly, between commitment to safety (COM) and 292 Resources for safety management (RES). 293 Finally, Figure 5 illustrates Model 5, the model for the Materials and equipment-related precursors 294 (MATEQ). Model 5 did not support the hypothesis that materials and equipment-related precursors 295 (MATEQ) are affected by risk assessment and control efforts (RISK), resources for safety 296 management (RES) or worker behaviour improvement efforts (BEHAV). However, the model 297 does support the hypothesis about the influence of the adverse project conditions (ADV) on the 298 accident precursors. As was the case in the previous models, strong relationships were observed 299 between commitment to safety (COM) and project administration for safety (ADMIN); and 300 between project administration for safety (ADMIN) and risk assessment and control efforts 301 The five structural models presented in this paper imply that the occurrence of accident precursors 304 is systemic. The models also suggest that each of the accident precursors may be linked with one 305 or two specific upstream SMS factors. Specifically, Model 1 suggests that the accident precursors 306 tight budget, schedule, or major rework. Workers and operators also may start to 'cut corners' in 360 using heavy equipment and tools ignoring best practices for safety performance. Contractors would, 361 therefore, need to manage project conditions such as time, changes and rework, effectively to 362 prevent accident precursors related to inadequate construction materials and equipment usage. 363 One notable finding of this study is the significant influence that adverse project conditions such 364 as tight contract schedule, a large number of change orders and reworks can have on the occurrence 365 of most types of accident precursors. The models demonstrate that even when a SMS is 366 implemented, adverse project conditions can still cause the occurrence of accident precursors. This 367 finding indicates the importance of a holistic approach to safety management. The mere 368 implementation of several safety improvement programs/practices may not be powerful enough 369 on its own to offset the impact of adverse project conditions. Therefore, SMSs should be integrated 370 into the larger project administration and planning framework, including project design, project 371 planning, human resources, change management, and quality assurance to ensure their 372 effectiveness in improving safety performance. 373

CONCLUSIONS 374
This study has developed five structural models to explain causal links between SMS factors and five 375 types of observable accident precursors on construction sites. This research used empirical data on 376 SMS factors and accident precursors collected from experienced site safety managers, and analyzed 377 the data using an established and rigorous SEM analysis process. The results of the SEMs enhance our 378 understanding of the relationships between SMS factors and accident precursors by (1) demonstrating 379 that adverse project conditions should be controlled, concomitantly, with traditional safety programs 380 to avoid the occurrence of incident precursors and 2) identifying SMS factors of interest for each 381 particular type of accident precursors. 382 The contributions of this research would be three-fold. First, from a practical perspective, the final 383 structural models can be used to address specific observable accident precursors in a more 384 informed and proactive manner. This evidence-based, focused approach is expected to enhance 385 the value for money of safety management resources by prioritizing measures and interventions 386 most relevant to specific conditions. Second, this research contributes to the understanding of the 387 complex cause-and-effect relationship between SMS factors and incident precursors. The results 388 reinforce that improving the SMS using a comprehensive approach (considering factors such as 389 performance and design) can reduce the occurrence of incident precursors and, consequently, 390 allow for a proactive approach for improving safety performance. Third, the models' results also 391 contribute to engineering management practice by corroborating or suggesting approaches to 392 enhance safety management onsite. The results reinforce that resources available for safety, and 393 implementation of safety programs to control unsafe behavior or to enhance risk assessments and 394 control on site, highly depend on organizational commitment to safety. The results also suggest 395 that merely enhancing traditional safety management programs to reduce the likelihood of accident 396 precursors may not be sufficient on its own. Therefore, organizations should adopt a holistic 397 approach in all project phases to avoid incidents 398 The findings of this study should be interpreted in consideration of the following limitations. The 399 SEM was built based on a sample size of 96 participants, which may be on the lower side for the 400 SEM analysis. Therefore, it is possible that the models developed in this research were influenced 401 by the biases that the respondents could have. It is recommended that the models are viewed as 402 most reflective of the circumstances in which they were gathered: Alberta, Canada. While this 403 research has used a bootstrapping method to enhance the reliability of the models by introducing 404 random sampling within the analysis process, further studies based on a larger sample size would 405 enable further reinforcement of the findings from this research to a greater degree of confidence. 406 Also, because the respondents were recruited from various types of construction projects, further 407 research may be warranted to identify project-specific SMS factors and accident precursors. 408 Additionally, efforts can be invested to test un-confirmed relationships. The cross-sectional design 409 of the current study can lead one only to infer causality, rather than prove causality.   Availability of skilled workers ADV6 The level of required worker compensation rate