Content uploaded by Emad Mohamed
Author content
All content in this area was uploaded by Emad Mohamed on Jun 14, 2019
Content may be subject to copyright.
CON113- 1 -
7
th CSCE/ASCE/CRC International Construction Speciality Conference
7
e SCGC/ASCE/CRC International Conférence Spécialisée sur la Construction
Laval
(Greater Montreal), Canada
June 12–15, 2019/ Juin 12–15, 2019
LEADING INDICATORS FOR SAFETY MANAGEMENT: UNDERSTANDING
THE IMPACT OF PROJECT PERFORMANCE DATA ON SAFETY
PERFORMANCE
Emad Mohamed1, Parinaz Jafari1, Shih-Chung Kang1, Estacio Pereira1, and Simaan
AbouRizk1,2
1Department of Civil and Environmental Engineering, University of Alberta, Edmonton, AB, Canada
2abourizk@ualberta.ca
Abstract: The construction industry continues to experience an elevated number of accidents and
fatalities, rendering safety a major concern for many construction companies. To develop more effective,
proactive strategies capable of reducing future accidents, safety performance must be monitored and
assessed prior to incident occurrence. Safety leading indicators can be used to proactively assess safety
performance, provide insights into the effectiveness of an organization’s safety practices, and offer
guidance on how to improve. Although useful, an agreed-upon set of leading indicators for proactively
assessing safety performance has yet to be established in the literature. This research aims to investigate
and test the feasibility of using project-related data together with safety-related data to more accurately
assess proactive safety performance in industrial construction projects. Data utilized in this study were
obtained from a large contractor in North America, pulled from eight industrial construction projects over a
period of two years. Databases from different departments are matched and integrated into a single
dataset. Correlation and feature selection techniques were used to identify the variables with the greatest
impact on safety performance. Results of this study indicate that project performance data were
associated with safety performance, demonstrating that project data, in addition to traditional safety
leading indicators, can be used to build a safety management system to more effectively monitor safety in
a project. Additionally, this study has shed light on the project performance metrics that could be collected
by safety leaders to better predict safety performance on construction sites.
1 INTRODUCTION
Effective management of safety, health, and the environment are essential to the success of any
construction company. Indeed, safety performance is one of the main measures of the success in a
project (Mohammadi, Tavakolan, and Khosravi 2018). The construction industry is dynamic in nature;
safety-management techniques, therefore, must regularly be altered to satisfy industry requirements
(Akroush and El-adaway 2017). Safety indicators play an important role in providing information on
organizational safety performance, and recognizing these safety indicators is a motivating factor for
stakeholders to increase organizational potential for safety (Reiman and Pietikäinen 2012).
Traditionally, safety performance has been measured by lagging indicators, or ‘after the loss’
measurements such as accident rates, which can only be measured after the occurrence of an accident
(Grabowski et al. 2007, Akroush and El-adaway 2017, Hinze, et al. 2013). Used primarily by insurance
companies, owners, and companies to compare performances or to examine performance trends over
CON113- 2 -
cost-time, lagging indicators can be used neither to predict the safety performance of a construction
company nor to predict the current level of risk of a particular construction project (Hinze et al., 2013).
Indeed, monitoring classic measures such as total recordable injury rate (TRIR) and the experience
modification rate (EMR) has not, as of yet, reduced injury rates to achieve optimal improvement in safety
performance. This is primarily because it is both difficult and unlikely to identify deficiencies and flaws
prior to incident occurrence using lagging indicators (Hinze et al. 2012). Although there is prevalent use of
lagging indicators, their effectiveness in anticipating safety performance and proactively reducing
occurrence of accidents is under scrutiny (Akroush and El-adaway 2017). Many safety professionals and
researchers agree that lagging indicators may not provide the necessary insights for taking corrective
action to avoid future accidents (Grabowski et al. 2007, Hinze et al. 2013, Kjellén 2009).
The construction industry is moving away from lagging indicators and toward leading safety indicators as
an alternative method of measuring safety performance (Akroush and El-adaway 2017). Due to their
ability to facilitate preventative decision making, leading indicators have become the preferred method for
assessing construction safety (Versteeg 2018). Leading indicators in current use are based on case
studies, content analysis of completed projects, and safety experts’ knowledge (Guo and Yiu 2016), with
many leading safety indicators reported in literature focusing only on safety-related data. However,
multiple researchers have demonstrated the potential associations between quality (Wanberg et al. 2013)
and schedule (Han et al. 2014) performance and safety. Despite collecting rich stores of performance
data from different departments (such as cost, quality, and schedule) construction companies do not fully
utilize this existing information to develop safety leading indicators in practice.
The objective of this paper is to investigate and test the feasibility of using existing, yet under-utilized,
project-related data together with safety-related data to more accurately assess proactive safety
performance in industrial construction projects. First detailing the development process of finding leading
indicators, the study investigates whether or not integrated project-performance data can be used within
quantitative development methodologies to identify safety useful leading indicators.
2 LITERATURE REVIEW
To take corrective actions, safety indicators must be both detectable before an event and be linked in a
causal pathway leading to the event (Hale 2009). A leading indicator is a measure of attitudes, behaviours,
practices, or conditions that influence construction safety performance (Hinze et al. 2012, Guo and Yiu
2016). In addition to informing practitioners of the current safety performance, threshold values for leading
indicators (below which corrective actions should be taken) can be established to guide risk mitigation
practices to reinstate above-level performance (Akroush and El-adaway 2017). The careful selection,
measurement of, and response to leading indicators of safety performance in the construction industry all
have helped to improve construction-site safety and organization (Ng et al. 2012, Hinze et al. 2012).
Leading indicators can be classified as passive or active (Hinze et al. 2012). Passive indicators are a set
of strategies and actions that are set up prior to the beginning of the project and cannot be adjusted once
the project has started (Hinze et al. 2012, Akroush and El-adaway 2017). In contrast, active indicators
can be measured and adjusted dynamically during the construction phase, allowing for the real-time
implementation of risk mitigation practices (Akroush and El-adaway 2017, Hinze et al. 2012).
2.1 Development of safety leading indicators in construction
Multiple researchers have defined many base-criteria for selecting leading indicators in construction.
Criteria established by Hale (2009) include validity, reliability, sensitivity, representativeness, openness to
bias, and cost-effectiveness. Guo and Yiu (2016) categorized essential attributes of safety leading
indicators into two dimensions: a scientific dimension and a managerial dimension. From the scientific
perspective, indicators should have strong scientific and conceptual bases, be developed from safety
models, reflect causes of accidents, be sensitive to changes in safety conditions, and allow for early
warning. From the managerial perspective, indicators should be compatible with practical safety
management, drive appropriate behaviour, be easily observable, and be cost-effective in terms of
collection. Akroush and El-adaway (2017) added to this list by identifying four additional criteria that
CON113- 3 -
selected leading indicators should satisfy: indicators should be complete, consistent, and reliable in
covering critical assumptions of safety; they should be unbiased, and not susceptible to or influenced by
manipulation; they must be easily measured and quantifiable on a numerical scale; and finally, they
should significantly correlate to a reduction in number of incidents.
Qualitative and quantitative methods have been used to develop safety leading indicators in the
construction industry. Qualitative models, aimed at identifying leading indicators and then assessing both
their effectiveness and their correlation to safety performance (Akroush and El-adaway 2017), include (1)
questionnaires, interviews, accident investigations and focus groups; (2) safety audits built by the
organization to monitor and measure safety performance factors; (3) perception surveys asking
employees, supervisors, and top management about their perceptions regarding the corporate and safety
climate in the organization; (4) behavioural observation to identify unsafe behaviours and promote safer
attitudes through necessary training; (5) case studies and brainstorming sessions by research teams and
experts in the field, and data extraction from industry databases; and (6) the Delphi method: a structured
communication technique involving a panel of experts giving initial estimates, and revising those
estimates after in-depth discussion. A number of passive and leading indicators have been identified
using qualitative methods (Hallowell et al. 2013, Hinze et al. 2012, Akroush and El-adaway 2017).
Qualitative methods are questioned by some researchers (Guo and Yiu 2016, Guo et al. 2017) owing to
the fact that the indicators are not selected based on a conceptual framework that provides theoretical
guidance on developing leading indicators. Furthermore, (Hinze et al. 2013) pointed out that safety
practitioners face challenges in developing indicators that fit well into existing safety programs. A few
construction companies have successfully implemented the monitoring of safety leading indicators;
however, there is little published information concerning successfully applied, specific leading indicators
(Hinze et al. 2013).
Limitations associated with qualitative methods, such as subjective nature, have prompted researchers to
develop quantitative methods for the identification of leading indicators. Guo and Yiu (2016) developed a
conceptual framework for developing leading indicators in construction where they clarified the concept of
a leading indicator; a model was then used to conceptualize the safety conditions. Despite being an
important first step, their approach lacked quantitative validation, as their only means of validation was
expert judgment. Guo et al. (2017) proposed a pressure-state-practice model as a theoretical basis for
developing leading indicators in construction. Their model represented the safety level of a construction
project as a dynamic phenomenon characterized by interrelationships between safety state, safety
practice, and pressures. Similar to the early neural network model proposed by Goh and Chua (2013), the
model only considered safety aspects of the project. Later, Poh, Ubeynarayana, and Goh (2018)
proposed a machine-learning approach for developing leading indicators that incorporated other aspects
of projects, including project delay and percentage completion. In spite of the model’s novelty, data
regarding quality and cost performance were not examined, and the procedures other contractors should
use to collect and combine such data to apply the methodology were not clarified.
3 RESEARCH METHODOLOGY
The research methodology for investigating the association and impact of project performance data on
safety performance is described. This methodology presents a general framework that can be applied by
any construction company. Collection, processing, and preparation of input data, as well as data analysis
and investigation, are described in Figure 1.
CON113- 4 -
Safety
Department
Quality
Department
Cost
Department
Schedule
Department
• Match data and
combine
Data Integration
• Data Wrangling
• Data Cleaning
Data Processing
• Correlation
• Feature selection
Data Analysis
• Hypothesis testing
• Confirmed features
Output
Figure 1: Research methodology
3.1 Data Collection and Understanding
In industrial construction projects, project performance data (including safety, quality, schedule, and
costs) are collected separately by each designated department. Data are collected periodically (usually
every two weeks) according to the specific conditions of each project. Each department collects only
related data. For example, the cost department tracks the financial records and indicators of cost
performance, such as the cost performance index and money spent in each reporting period. The
scheduling department records the schedule performance index and assesses whether or not the project
has progressed as planned. The quality department tracks of the number of change orders. The safety
department keeps track of the number of working-hours spent in each reporting period and tracks
whether or not incidents have occurred.
3.2 Data Preparation and Processing
In order to process the data, it must first be integrated into one centralized dataset. A common challenge
with centralizing the data is the differences in reporting periods used by the various departments within an
organization; in these instances, recording dates may be used to match and combine project data. Once
centralized, the dataset is cleaned to identify missing data and remove outliers.
3.3 Data Analysis and Investigation
The final step is to test the hypothesis that project performance data can be used to develop safety
leading indicators. This step is conducted by correlation testing and the Boruta feature selection function
in R (R Team 2013) to determine if variables are feasible (i.e., important) or not.
4 CASE STUDY
The dataset used in this research was collected from a large construction company in Alberta, Canada.
The company engages in many types of projects such as building, industrial, and infrastructure projects.
Data spanned eight industrial projects over a period of two consecutive years, from 2016 to 2017. Project-
performance-related data within the dataset consists of cost, schedule, quality, and safety performance
data collected from different departments.
4.1 Data Collection and Understanding
Table 1 describes the collected features from each department and defines each attribute.
CON113- 5 -
Table 1: Description of the collected variables in each department
No. Variable/feature Description Department
1 Project ID Unique identifier for the project All
2 Contract type The specific type of contract used for the project All
3 Report date The date at which the variables are measured All
4 Contract change order
(CCO)
A change in the work or change in the contract sum
or the contract time at report time Cost
5 Outstanding CCO
A change in the work or change in the contract sum
or the contract time at report time that will change
original contract
Cost
6 CCO submitted to
date The cumulative amount of CCO at report time Cost
7 RFI The total number of Requests For Information at
report time Quality
8 RFI submitted to date The cumulative amount of RFI at report time Quality
9 Open RFI The number of unresolved RFI at report time Quality
10 Original budget The planned/ estimated cost of the project Cost
11 Re-baseline budget The modified budget of the project at the report time Cost
12 Approved changes The cost of the changes in the project at the report
time Cost
13 Revised contract
value
The sum of re-baseline budget and approved
changes Cost
14 Pending changes The cost of the changes in the project at the report
time waiting for approval Cost
15 Forecast at
completion (FAC) Forecasted value of the project at completion time Cost
16 Earned value The budgeted cost of work performed Cost
17 Incurred value The money spent for the work accomplished Cost
18 Outstanding change
%
The percentage of change in the original budget at
report time Cost
19 Field Surveillance
Report (FSR) A proactive quality surveillance technique Cost
20 FSR submitted to date The cumulative amount of FSR at report time Cost
CON113- 6 -
21 Open FSR The number of unresolved FSR at report time Cost
22 Work order CPI The ratio of the earned value to the incurred/actual
value at report time (Cost Performance Index) Cost
23 Work order %
complete
The progress of the project at report time as
percentage Schedule
24 Non-conformance
report (NCR)
The number of reports showing the quality deviation
at report time Quality
25 Open NCR The number of reports showing the quality deviation
which are not resolved Quality
26 Work order HPI
Evaluation of the accomplishment of the schedule
and budget of the activities executed at report
time(Human Performance Index)
Cost &
Schedule
27 Work order SPI The ratio of the earned value to the planned value at
report time (Schedule Performance Index) Schedule
28
(0-1) years’
experience direct
hours
The total numbers of hours spent on the project up to
the report time by workers that have experience less
than one year
Safety
29
(1-2) years’
experience direct
hours
The total numbers of hours spent on the project up to
the report time by workers that have experience
greater than one year and less than two years
Safety
30
(2-3) years’
experience direct
hours
The total numbers of hours spent on the project up to
the report time by workers that have experience
greater than two years and less than three years
Safety
31
(3-4) years’
experience direct
hours
The total numbers of hours spent on the project up to
the report time by workers that have experience
greater than three years and less than four years
Safety
32 +4 years’ experience
direct hours
The total numbers of hours spent on the project up to
the report time by workers that have experience
greater than four years
Safety
33 Foreman hours The total numbers of hours spent on the project up to
the report time by foreman Safety
34 Shift hours The total number of hours spent on the project by all
workers at the report time Safety
35 Exposure hours The cumulative amount of shift hours Safety
36 Incident The variable that shows if an incident happened on
the project at report time Safety
CON113- 7 -
4.2 Data Preparation and Processing
Raw data was processed, cleaned, and transformed into a proper format before analysis. Because data
may have been collected in an ad hoc manner, including empty fields in records or mistakes in data entry,
data preparation was given the utmost care (Soibelman and Kim 2002). Data were integrated into one
dataset; both project ID and report dates were then used to match different records from various datasets.
The collected variables were investigated, and features/variables that were not useful for the purpose of
this work were removed. For example, in the provided dataset the variables “Contract Type” and “Project
ID,” which are the same type for all the 8 projects (and, therefore, for all the records in the dataset) were
removed. Due to the similarity in meaning between the three budget variables (Original Budget, Re-
Baseline Budget, Revised Contract Value), only “Re-baseline Budget” was kept. Additionally all “To Date”
variables were removed, as they could be easily calculated from the corresponding non-cumulative
variable (e.g., CCO Submitted and CCO Submitted to Date).
Another common issue in all datasets, and in the case study dataset particularly, is missing values.
Missing values were removed or substituted by values that allowed them to be used in further analysis
while, at the same time, not adversely affecting dataset behaviour (Witten et al. 2016). In this particular
dataset, 50% of the total records for “0-1,” “1-2,” “2-3,” and “3-4 Years’ Experience Direct Hours” were
missing; therefore, these columns were removed due to the limited number of data points. After filtering
the variables (from Table 1), 23 features remained.
4.3 Correlation
To investigate the degree of association, the Pearson correlation coefficient between all attributes was
calculated. The results of correlation analysis are summarized in Table 2. The attributes with the highest
positive and negative correlation with “Accident” were “Foreman Hours” (r = 0.50) and “Work Order HPI”
(r = -0.28), respectively. Evans (1996) suggested that a value of r between 0 and 0.19 represents a very
weak correlation; 0.20 and 0.39 represents a weak correlation, and 0.40 and 0.59 represents a moderate
correlation. Here, “Foreman Hours,” “Shift Hours,” “+4 Years’ Experience Direct Hours,” “RFI,” “Open
FFI,” and “Contract Change Order (CCO),” were determined to be moderately correlated with “Accident.”
In contrast, “Work Order HPI,” “Forecast at Completion (FAC),” and “Re-baseline Budget,” were weakly
correlation with “Accident.” Remaining attributes were very weakly correlated with “Accident.” The feature
selection process was also performed to provide further insight into the feasibility of the features.
Table 2: Correlation matrix
ReBaseB udget
ApprChanges
PendChanges
FAC
Earned
Incurred
Outstanding. Change
s.of.O riginal.Budget
WOCom plete
WOHPI
WOSPI
WOCPI
FM.Hour s
Shift.Hours
X4.yrs.Exp
RFI
Open.RFI
FSR
Open.FSR
NCR
Open.NCR
CCO
Outstanding. CCO
ACCIDENT
ReBaseB udget 1
ApprChanges 0.63 1
PendChanges 0.16 0.16 1
FAC 10.69 0.18 1
Earned 0.79 0.73 0.08 0. 82 1
Incurred 0.78 0.72 0.09 0. 82 1 1
Outstanding. Changes.of.O riginal.Budget -0.3 -0.23 0.55 -0.3 -0.23 -0.22 1
WOC omplete -0.11 0.21 -0 -0.07 0.34 0.34 0.16 1
WO HPI -0. 46 -0.35 -0.4 -0. 47 -0.38 -0.39 -0.2 -0.1 1
WOSPI -0.57 - 0.19 -0.1 -0.54 -0.06 -0.05 0.24 0.65 0. 16 1
WOCPI -0.09 - 0.09 -0.3 -0.11 -0.11 -0. 13 -0.5 -0.2 0.58 0.08 1
FM.Hours 0.74 0.42 0.23 0.75 0.73 0.74 -0.2 0.1 -0.53 -0.19 -0. 1 1
Shift.Hours 0.67 0.41 0.25 0. 69 0.76 0.77 -0.1 0.21 -0.49 -0.1 -0.2 0.94 1
X4.yrs.E xp 0.38 0.2 0.32 0.39 0.43 0.44 0.04 0.18 -0.55 - 0.03 -0.3 0.72 0.76 1
RFI 0.56 0.34 0.19 0.57 0.58 0.58 -0.2 -0.1 -0.29 -0.14 0.15 0.72 0. 72 0.43 1
Open.RFI 0.7 0.31 0.03 0.69 0.6 0.6 - 0.3 -0.1 -0. 29 - 0.36 -0 0.74 0.72 0.4 0.73 1
FSR 0.52 0.43 -0 0.53 0.66 0.65 -0. 1 0.19 -0.26 -0.02 -0.1 0.56 0.59 0.35 0.42 0.48 1
Open.FSR 0.67 0.48 0.06 0.68 0.86 0.87 -0.2 0.27 -0.32 -0.02 -0.1 0.75 0.78 0. 43 0.64 0.66 0.73 1
NCR 0.34 0.28 - 0.1 0.35 0.45 0.45 -0.2 0.21 -0.2 0.02 -0. 1 0.5 0. 53 0.37 0.26 0.32 0.57 0.53 1
Open.NCR 0.52 0.47 0.09 0.55 0.78 0.79 -0.2 0.42 -0.26 0.11 -0.1 0.66 0.73 0.41 0. 52 0.48 0.51 0.85 0.64 1
CCO 0.58 0.5 0.09 0. 6 0.59 0.58 - 0.1 0.05 -0.35 -0.2 -0.1 0.56 0.54 0.27 0.44 0.44 0.41 0. 49 0.28 0.43 1
Outstanding. CCO -0.11 -0.08 0.23 -0.1 -0.07 -0.07 0.69 0.19 -0.26 0.16 -0.4 -0.2 -0.18 -0.1 -0.2 -0.21 -0.1 -0.1 -0.1 -0. 1 0.13 1
ACCIDENT 0.39 0.03 0.06 0.37 0.29 0.28 -0.1 -0. 1 -0.28 -0.18 0.01 0.5 0.48 0.41 0.47 0.48 0.22 0. 26 0.15 0.12 0.43 -0.08 1
CON113- 8 -
4.4 Feature Selection
Feature subset selection is another important data preparation and testing step for checking and selecting
the most important predictive features (Poh et al. 2018). This technique aims to reduce the number of
features based on their importance and impact as irrelevant features in the dataset can negatively impact
the accuracy of the prediction model and cause unnecessary computational complexity. The feature
subset selection algorithm performs a subset search using the induction algorithm as part of the
evaluation function (Soibelman and Kim 2002).
As opposed to previous feature selection techniques, which are limited by assumptions of normality, the
Boruta feature selection function in R (R Team 2013) runs a random forest classifier on the dataset and
ranks the features in a step-wise manner. Boruta was applied to the 23 features remaining after the data
cleaning and preparation step. The results of the feature selection method are summarized in Figure 2.
Box plots are used to show the distribution of a feature’s importance over a Boruta run, and colors are
used to indicate importance outputs, where green indicates an important feature, red indicates a feature
that is not important, and yellow indicates a tentative feature. A total of 10 features (indicated in green)
were identified as important, from which the “Shift Hours” feature was determined to be the most
important among the features analyzed. The correlation results are in agreement with the feature
selection results where the nine variables determined to be moderately or weakly correlated were
classified as the most important in the feature selection process. Only one variable with a very weak
correlation (“Approved Changes”) was considered important in the feature selection. This research study
is focused on proving that combined project performance metrics can be utilized as inputs for machine
learning algorithms when evaluating project safety performance. More details on applying machine
learning algorithms using selected features was provided by Jafari et al (2019) where the authors
extended their work by investigating different models and measuring their performance.
Figure 2: Boruta feature selection results
CON113- 9 -
5 CONCLUSIONS
In current practice, construction companies collect rich stores of data related to various aspects of project
performance such as quality, safety, cost, and schedule. Yet, it is uncommon for these data to be used to
predict safety performance and accident occurrence. This research study has demonstrated that existing
project performance data can be used as safety leading indicators and to build machine-learning
algorithms for predicting safety performance; of the ten variables identified as important by the Boruta
feature selection, seven were from non-safety-associated departments. Notably, while schedule and
quality performance variables were associated with accident occurrence, variables conventionally
associated with cost performance represented 40% of the most important variables identified using the
Boruta feature selection.
Although currently limited, an increasing number of studies are suggesting that project performance data
can be used to develop safety leading indicators. This study has expanded upon previous work and
applied a Boruta feature selection procedure to demonstrate not only that project performance variables
are important for predicting accident occurrence but that they are feasible for use to build machine-
learning algorithms for predicting safety risk. As these data are already being collected by various
departments for alternative purposes, using these existing data can reduce the time, efforts, and
resources required to monitor and track safety leading indicators.
While the specific variables identified as important in the current study (foreman hours, shift hours, years’
experience direct hours, re-baseline budget, forecasts at completion (FAC), contract change orders
(CCO), approved changes, and work order HPIs, requests for information (RFI), and open RFI) may
motivate practitioners to monitor these metrics in practice, it is important to note that, as a result of
differences in safety cultures, types of activities, and scope of projects, these variables may not be
associated with accident occurrence in all construction organizations. Additionally, the data set used in
the current study was limited by its small size and by the large number (up to 50%) of missing data points.
Future research is recommended to address these limitations.
ACKNOWLEDGEMENTS
This research was made possible by the financial support of a Collaborative Research and Development
Grant (CRDPJ 492657) from the Natural Sciences and Engineering Research Council of Canada. The
authors would also like to thank Elisia Snyder for her assistance with manuscript editing.
REFERENCES
Akroush, N.S. and El-Adaway, I.H., 2017. Utilizing construction leading safety indicators: Case study of
tennessee. Journal of Management in Engineering, 33(5): 06017002.
Evans, J.D., 1996. Straightforward statistics for the behavioral sciences. Thomson Brooks/Cole
Publishing Co.
Goh, Y.M. and Chua, D., 2013. Neural network analysis of construction safety management systems: a
case study in Singapore. Construction Management and Economics, 31(5): 460-470.
Grabowski, M., Ayyalasomayajula, P., Merrick, J., Harrald, J.R. and Roberts, K., 2007. Leading indicators
of safety in virtual organizations. Safety Science, 45(10):1013-1043.
Guo, B.H. and Yiu, T.W., 2015. Developing leading indicators to monitor the safety conditions of
construction projects. Journal of Management in Engineering, 32(1): 04015016.
Guo, B.H., Yiu, T.W., González, V.A. and Goh, Y.M., 2016. Using a pressure-state-practice model to
develop safety leading indicators for construction projects. Journal of Construction Engineering and
Management, 143(2): 04016092.
Hale, A., 2009. Why safety performance indicators?. Safety Science, 4(47): 479-480.
Han, S., Saba, F., Lee, S., Mohamed, Y. and Peña-Mora, F., 2014. Toward an understanding of the
impact of production pressure on safety performance in construction operations. Accident analysis &
prevention, 68: 106-116.
CON113- 10 -
Hinze, J., Hallowell, M.R., Gibbons, B., Green, L., Trickel, S. and Wulf, D., 2012. Measuring safety
performance with active safety leading indicators. Construction Industry Institute, CII Research
Summary.
Hinze, J., Thurman, S. and Wehle, A., 2013. Leading indicators of construction safety
performance. Safety science, 51(1): 23-28.
Jafari, P., Mohamed, E., Pereira, E., Kang, S. and AbouRizk, S., 2019. Leading safety indicators:
application of machine learning for safety performance measurement. In Proceedings for the 36th
international symposium on Automation and Robotics in Construction, ISARC, Banff, AB, Canada
[Accepted].
Kjellén, U., 2009. The safety measurement problem revisited. Safety Science, 4(47): 486-489.
Mohammadi, A., Tavakolan, M. and Khosravi, Y., 2018. Factors influencing safety performance on
construction projects: A review. Safety science, 109: 382-397.
Ng, K., Laurlund, A., Howell, G. and Lancos, G., 2012. lean safety: using leading indicators of safety
Incidents to improve construction safety. In Proceedings for the 20th Annual Conference of the
International Group for Lean Construction: are We Near a Tipping Point , San Diego, CA, USA, p. 173.
Poh, C.Q., Ubeynarayana, C.U. and Goh, Y.M., 2018. Safety leading indicators for construction sites: A
machine learning approach. Automation in Construction, 93: 375-386.
Reiman, T. and Pietikäinen, E., 2012. Leading indicators of system safety–monitoring and driving the
organizational safety potential. Safety science, 50(10): 1993-2000.
Soibelman, L. and Kim, H., 2002. Data preparation process for construction knowledge generation
through knowledge discovery in databases. Journal of Computing in Civil Engineering, 16(1): 39-48.
Team, R.C., 2013. R: A language and environment for statistical computing.
Versteeg, K., 2018. Utilizing Construction Safety Leading and Lagging Indicators to Measure Project
Safety Performance: a case study (Master's thesis, University of Waterloo).
Wanberg, J., Harper, C., Hallowell, M.R. and Rajendran, S., 2013. Relationship between construction
safety and quality performance. Journal of Construction Engineering and Management, 139(10):
04013003.
Witten, I.H., Frank, E., Hall, M.A. and Pal, C.J., 2016. Data Mining: Practical machine learning tools and
techniques. Morgan Kaufmann.
