Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Contents lists available at ScienceDirect Journal of Loss Prevention in the Process Industries journal homepage: http://www.elsevier.com/locate/jlp Severity of emergency natural gas distribution pipeline incidents: Application of an integrated spatio-temporal approach fused with text mining Xiaobing Li a, *, Praveena Penmetsa a, Jun Liu b, Alexander Hainen b, Shashi Nambisan c a b c Alabama Transportation Institute, The University of Alabama, Tuscaloosa, AL, 35487, USA Department of Civil, Construction, and Environmental Engineering, The University of Alabama, Tuscaloosa, AL, 35487, USA Transportation Research Center, Department of Civil and Environmental Engineering and Construction, The University of Nevada, Las Vegas, NV, 89154, USA A R T I C L E I N F O A B S T R A C T Keywords: Natural gas Distribution pipeline incidents Severity GTWOLR Text mining The transportation of natural gas often relies on pipelines which require constant monitoring and regular maintenance to prevent spills or leaks Pipeline incidents could pose a huge adverse impact on people, the environment, and society Numerous efforts have been invested to identify contributing factors to pipeline in cidents so that countermeasures could be developed to proactively prevent some incidents and reduce incident severities or impacts However, the countermeasures may need to vary for different incidents due to the potential heterogeneity between incidents, and such heterogeneity is likely related to the geology, weather, and built environment which vary across space and time domain The objective of this study is to revisit the correlates of pipeline incidents, focusing on the spatial and temporal patterns of the correlations between natural gas pipeline incident severity and contributing factors This study leveraged an integrated spatio-temporal modeling approach, namely the Geographically and Temporally Weighted Ordered Logistic Regression (GTWOLR) to model the natural gas pipeline incident report data (2010–2019) from the U.S Pipeline and Hazardous Material Safety Administration Text mining was performed to extract additional information from the narratives in re ports Results show several factors have significant spatiotemporally varying correlations with the pipeline incident severity, and these factors include excavation damage, gas explosion, iron pipes, longer incident response time, and longer pipe lifetime Findings from this study are valuable for pipeline operators, end-users, responders to jointly develop localized strategies to maintain the natural gas distribution system More impli cations are discussed in the paper Introduction Natural gas is the earth’s cleanest fossil fuel and is colorless and odorless in its natural state Natural gas offers 91% high energy effi ciency and low cost of capital investment compared to conventional coal plants The American Gas Association (AGA) reported that since 1990, emissions from the natural gas distribution system have declined by 73%, and CO2 emissions have hit 25-year lows due to high natural gas efficiency and the growth of renewable energy use (AGA, 2020) How ever, if natural gas distribution pipelines aren’t well maintained and operated by end-users or contractors, accidental loss of containment or pipeline failures occur due to risk factors related to human errors or equipment failures, posing huge adverse effects on people, environment, and society The U.S Pipeline and Hazardous Material Safety Adminis tration (PHMSA) reported that the 10-year average (2010–2019) num ber of incidents related to the natural gas distribution system was 110, with peak 140 incidents reported in 2019 These incidents had led to a total of 106 fatalities and 510 injuries in the last decade In addition, the reported annual total cost associated with these incidents was $0.26 billion from 2010 to 2019, with a peak total cost of $2.27 billion re ported in 2018 Thus, natural gas pipeline safety is the top priority for delivering the best energy service to customers through 2.6 million miles of pipeline nationwide (AGA, 2020) More than $86 million were invested daily in enhancing the safety of natural gas distribution and transmission sys tems (AGA, 2020) Fig depicts the natural gas production and delivery * Corresponding author E-mail addresses: xli158@ua.edu (X Li), ppenmetsa@ua.edu (P Penmetsa), jliu@eng.ua.edu (J Liu), ahainen@eng.ua.edu (A Hainen), shashi@unlv.edu (S Nambisan) https://doi.org/10.1016/j.jlp.2020.104383 Received 13 July 2020; Received in revised form 30 October 2020; Accepted 23 December 2020 Available online 30 December 2020 0950-4230/© 2020 Elsevier Ltd All rights reserved X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 system A holistic approach to improving pipeline safety has been pro posed by AGA to identify, prevent, and remedy safety hazards Such a Pipeline Safety Management System (PSMS) manages pipeline safety through continuous monitoring and improvement based on the “Plan- Do-Check-Act” cycle It involves leadership and management, process, compliance, procedure, and many other organizational changes With respect to the holistic safety improvement approach, this study intends to add extra safety knowledge by investigating the contributing factors related to the severity of natural gas distribution pipeline incidents Understanding the contributing factors is key to developing safety countermeasures that can help prevent future natural gas distribution pipeline incidents or alleviate the associated severity of these incidents PHMSA has reported seven apparent causes of the incidents including corrosion failure, natural force damage, excavation damage, other outside force damage, pipe/weld/joint failure, equipment failure, incorrect operation Some of these causes are not in control by humans (e.g., flooding natural force damage), however, some of these causes can be avoided by pipeline operators or end-users (e.g., excavation damage, incorrect operations) Additionally, these incidents usually present spatial and temporal patterns interacting with the diverse local socioeconomic, cultural, and geographic contexts It would be interesting to see whether these contributing factors have spatially and temporally varying correlations with incident severity With all that said, this study aims to investigate the contributing factors, especially the apparent causes that correlate with natural gas pipeline incident severity by applying a simultaneous Geographically and Temporally Weighted Or dered Logistic Regression (GTWOLR) approach To achieve the research goal, this study collected natural gas pipeline incidents from PHMSA that occurred between 2010 and 2019 The integrated GTWOLR approach can capture the complexities embedded within the incident data by accounting for the unobserved heterogeneity that might relate to the diversified spatial and temporal contexts To the best of authors’ knowledge, such an integrated spatio-temporal modeling approach has not been applied to natural gas distribution pipeline incident severity analysis The study is expected to contribute to natural gas distribution pipeline safety improvement by leveraging the space- and timereferenced incident data The outcomes are expected to help opera tors, customers, regulators, emergency responders, excavators, and policymakers to develop and implement localized safety improvement strategies and continue improving the industry’s longstanding record of providing natural gas service safely and efficiently both regionally and nationwide Literature review The number of pipeline incidents was increasing over time from 1968 to 2009 (Siler-Evans et al., 2014) People have shown great interest to analyze the trends, causes, and consequences of natural gas pipeline incidents In the last several decades, natural gas distribution pipeline safety has been investigated intensively, and safety risk factors were observed to contribute to the natural gas distribution pipeline incidents In general, these contributing factors can be categorized into two groups: one group is related to human activities or system errors, and the other group is related to natural force damage that is not directly linked to human activities or system failures Human activities such as digging, drilling, traffic accident, construction, excavation by operators or a 3rd party, rupture tasks, ignition, fire, and explosion are the potential threats to pipeline safety (Montiel et al., 1996; Simonoff et al., 2010; Li et al., 2016) Additionally, human errors or pipeline system characteristics or failures may lead to even severer natural gas distribution pipeline in cidents These factors include unlicensed operations, misoperation, imperfect inspection system, unclear facility marks, outdated pipeline information, failure of on-site protection, misuse of equipment, absence of barriers, incomplete warnings, lack of training, loss of containment, mechanical failure, joint failures, pipe material, pipe age, and pipeline pressure re-distribution (Jo and Ahn, 2002; Liu, 2003; Han and Weng, 2010; Cunha, 2012; Rui and Wang, 2013; Naik and Kiran, 2018; Bur gherr and Hirschberg, 2005; Montiel et al., 1996; Li et al., 2016) Natural force damage or environmental characteristics of the incidents are also investigated in previous studies The damage may be related to land slide, corrosion, earthquakes, hurricanes, mudslide, frost heave, rock falls, rockslides, soil erosion, volcanic eruption, flooding, storm, high winds, heavy rainfall, hail, cyclone, hot or cold temperature, drought or wildfire (Han and Weng, 2010; Cunha, 2012; Naik and Kiran, 2018; Girgin and Krausmann, 2014; Burgherr and Hirschberg, 2005; Zheng et al., 2012) The consequences of natural gas pipeline incidents were also inves tigated Environmental pollution, fatalities, injuries, hospitalization, property damage, gas release, and evacuation of people from the affected area are the typical consequences derived from these incidents (Ramírez-Camacho et al., 2017) Besides, there are also costs associated with the consequences To study the natural gas pipeline accident risks or the consequences of the incidents, researchers or safety engineers have applied various methods For example, Fuzzy hierarchical model (Li et al., 2016), weighted least squares and logistic regression models Fig Natural gas production and delivery system (Source: U.S Energy Information Administration) X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 (Simonoff et al., 2010), finite element model (Zheng et al., 2012), quantitative risk assessment (QRA) (Zhou et al., 2014), hazard model (Jo and Ahn, 2002) were utilized to study the accident risks, while hazard-based models (Hainen et al., 2020) were used to study the duration of interruptions caused by the incidents However, the severity of natural gas distribution pipeline incidents was barely mentioned or analyzed in those studies and currently, there is no standard to define the severity of these incidents This study wants to fill in this research gap by developing a rigorous way to categorize the severity of these incidents Additionally, few studies were specifically designed to un cover the varying correlations with incident severity across space and time In a recent study done by Naik and Kiran (2018), the spatial and temporal patterns of U.S pipeline incidents were investigated (Naik and Kiran, 2018) However, the authors did not attempt to uncover the correlations between incident severity and it associated factors They only presented the spatial and temporal patterns for factors of interest (e.g., temperature, precipitation, financial loss) Therefore, it would be interesting to see how natural gas distribution pipeline incident severity correlates with the contributing factors, and how the correlations vary across the U.S in the last decade pipeline incident reports If the incident involved fatalities or injuries, the economic value of a statistical life (VSL) generated by the U.S Department of Transportation is utilized to estimate the cost associated with fatalities and injuries According to their VSL estimates, VSLs of 2007, 2013, 2014, 2015, and 2016 were $5.8 million, $9.1 million, $9.2 million, $9.4 million, and $9.6 million, respectively (Moran and Monje, 2016) A linear interpolation method was used to generate the VSLs for years 2010, 2011, 2012, 2017, 2018, and 2019 (Friedman, 1962) They were estimated to be were $7.5 million, $8.0 million, $8.6 million, $9.8 million, $10.0 million, and $10.2 million, respectively In addition, the cost associated with injuries is estimated using the fraction of VSL The pipeline incident reports did not capture the injury severity levels of a person, a rational 0.105 fraction of VSL was utilized for injury cost estimation (Friedman, 1962) The total cost is classified by categories as minor ($0–0.01 million), moderate ($0.01–0.1 million), serious ($0.1–0.5 million), severe ($0.5–7.5 million), and critical (over $7.5 million) We also identified observations with missing information in data pre-processing, and they were removed from the final sample To further extract information from the reports, the narratives in text format were also analyzed based on the text mining technique (Magh rebi et al., 2015; Moran and Monje, 2016; Das et al., 2016) (see Fig 2) In text mining, the Latent Dirichlet allocation (LDA) method is often used to find the topics based on the document frequency generated by tokenization Table shows the first topics generated from the nar ratives Particularly, we want to capture people’s concerns about these topics based on key words Word “pressure” was used 276 times in the narratives (see also word cloud in Fig 1), thus concern regarding pipeline pressure was selected as one of the incident severity covariates Table provides detailed descriptions of incident severity and selected variables of natural gas distribution pipeline incidents Methodology 3.1 Data pre-processing This study collected the PHMSA natural gas distribution pipeline incident data from 2010 to 2019 In this study, the severity of these incidents is measured based on the total cost related to fatalities, in juries, property damage, and gas released The latter two types of cost can be directly obtained from the PHMSA natural gas distribution Fig Frequent words discussed in narratives from the natural gas distribution pipeline incident reports X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Table Topics based on the Narratives from the Incident Reports Topic Topic Topic Topic Topic Topic Topic gas service fire main meter leak line inch damaged natural steel gas fire service report inch valve approximately damage pressure response department gas fire line department hours main meter valve natural reported customers service line incident main leak fire valve investigation regulator gas pressure gas main incident leak damage approximately report distribution area damaged contractor gas service main incident leak pipe damage inch pressure street water main service fire incident approximately gas investigation inch repair street pressure Statistics show that approximately 50% of the incidents caused damages costing $0.1 million or more Exaction damage, outside force damage, and incorrect operations contributed apparently to about 70% of the incidents Additionally, root causes analyzed based on the Common Ground Alliance Damage Information Reporting Tool (CGA-DIRT) further indicates that excavation, locating, and one-call notification practices were not sufficient, which led to over 27% of the incidents Descriptions and distributions of other incident associated factors are also detailed in Table In addition to general descriptions of natural gas distribution pipe line incident covariates, this study also looks into the spatial and tem poral patterns across the nation in the last decade Fig depicts the geographical distributions of the natural gas distribution pipeline in cidents It indicates that more incidents occurred in or nearby metro politan cities such as San Francisco, Los Angeles, New York, and Chicago Fig presents the temporal patterns of distribution pipeline incidents by months of the year in the last decade The trendline of the incident numbers does not reveal a significant change over these years, even though the total number of incidents increased by 8.3% from 2010 to 2019 This study aims to discover the non-stationary correlations between incident severity and the covariates, thus a spatio-temporal approach will be discussed in the next section Table Descriptive statistics of selected variables Variable description (Total N = 1094) Frequency Percentage Severity of emergency natural gas distribution pipeline incidents measured by total cost Minor Moderate Serious Severe Critical Corrosion failure Natural force damage Excavation damage Other outside force damage Pipe, weld, or joint failure Equipment failure Incorrect operation Other incident cause Others Excavation practices not sufficient Locating practices not sufficient One-call notification practices not sufficient Yes 150 402 355 114 73 25 81 330 355 13.71% 36.75% 32.45% 10.42% 6.67% 2.29% 7.40% 30.16% 32.45% 75 6.86% 49 77 102 795 158 4.48% 7.04% 9.32% 72.67% 14.44% 73 6.67% 68 6.22% 15 1.37% Other reasons Intentional release of gas Unintentional release of gas Yes Steel Plastic Iron Copper Others Yes 104 13 9.51% 1.19% 977 89.31% 257 509 336 53 12 184 24 23.49% 46.53% 30.71% 4.84% 1.10% 16.82% 2.19% 0–5 6–10 11–30 31–60 Over h Unknown 0–10 years 11–20 years 21–50 years Over 50 years Others No concern Minor concern Major concern 258 93 413 234 48 48 166 137 372 197 222 818 149 127 23.58% 8.50% 37.75% 21.39% 4.39% 4.39% 15.17% 12.52% 34.00% 18.01% 20.29% 74.77% 13.62% 11.61% Apparent causes of the incident CGA-DIRT root cause Intentional outside force damage Incident resulted from Gas explosion Material involved in the incident Controller(s) or control room issues Response time to incident site Lifetime of gas distribution pipes Concerns with pipeline pressure 3.2 Analysis methodology 3.2.1 GTWOLR The response variable in this study is incident severity, which is measured based on the total cost It has been re-coded as a 5-level cat egorical variable, where ordinal logistic regression models or ordered probit models can be applied to estimate the probabilities of incident severity at each level (Abdel-Aty, 2003; Chen and Jovanis, 2000; Islam and Hernandez, 2013; Khattak et al., 2003; Khattak and Targa, 2004; Lemp et al., 2011; Xu et al., 2019; Zhu and Srinivasan, 2011) These conventional models can be modified by incorporating the randomized associations to improve prediction accuracy as stated in the previous studies (Islam et al., 2014; Xu et al., 2019) As mentioned, natural gas pipeline incidents are dependent on geographical locations, therefore the Geographically Weighted Regression (GWR) models can also be applied, which were frequently used in previous studies (Liu et al., 2019b, 2020) Since incidents were also time-referenced, the GWR models can be extended to incorporate time information in the modeling process using a similar concept to GWR Thus, this study introduces an integrated GTWOLR approach to disentangle the complex non-stationary relationships between incident severity and the associ ated factors to account for the unobserved heterogeneity that may relate to space and time The conventional ordinal logistic regression (OLR) is written as (StataCorp, 2019), X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Fig Nationwide spatial distribution of natural gas distribution pipeline incidents Fig Temporal patterns of natural gas distribution pipeline incidents ( ) /[ ( )] pij = Pr(yi = j) = Pr κj− < β1 x1i + β2 x2i + … + βk xki + ui < κj = 1 + exp − κj + β1 x1i + β2 x2i + … + βk xki + ui /[ ( )] − 1 + exp − κj− + β1 x1i + β2 x2i + … + βk xki + ui where pij represents the probability of an outcome yij = j; i is the index for incident observations; j is the index levels of incident severity, and k represent total severity levels; β1 , …, βk are the estimated coefficients; pij = Pr(yi = j) = (1) correlations will vary substantially across space and time The overall GTWOLR model is written in Equation (2), )] /[ ( )] /[ ( + exp − κj (gi , li , ti ) + β(gi , li , ti )xi − 1 + exp − κj− (gi , li , ti ) + β(gi , li , ti )xi κ1 , …, κk− are the cutpoints at jth level of incident severity with staring κ0 taken as − ∞ and κk taken as + ∞; uj is assumed to be a logistically distributed error term This study first estimates the global OLR model by taking all the crash observations Estimated coefficients of the covariates represent stationary global relationships between incident severity and the cova riates To further evaluate and compare the model performance among the global and local models, this study will estimate a local GTWOLR model and perform non-stationarity tests to examine whether the local (2) where pij represents the overall probabilities at all levels of incident severity; κj (gi , li , ti ) and β(gi , li , ti ) are the estimated localized cutpoints and coefficients for ith observation; (gi , li ) together denote the longitude and latitude coordinates for ith incident location; ti denotes time infor mation of ith incident (date is used in this study) The detailed GTWOLR modeling process has four major steps: Step Selection of a spatio-temporal target center and weight all observations with respect to this target center Each observation has a X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 chance to be chosen as the target center Once the target center is selected, all other observations are assigned with a spatial weight based on their geographical distance to the target center The Bi-square Kernel Weighting function is usually utilized for spatial weighting purposes (Fotheringham et al., 1998) It’s formulated as, )2 ] [ ( din wSi (gi , li ) = − dmax there is a unit change in independent variables while other variables are kept constant at their means Such marginal effects can be calculated using Equation (6) (StataCorp, 2019), ) ( exp − κj (gi , li , ti ) + β(gi , li , ti )xi ∂Pr(yi = j) ∂[β(gi , li , ti )xi ] =[ )]2 × ( ∂xim ∂xim + exp − κj (gi , li , ti ) + β(gi , li , ti )xi (3) =[ wSi (gi , li ) where is the spatial weight for nth observation with regards to the target center at ith observation geo-coded as (gi , li ); din is the geographical distance from nth observation to the target center; dmax is the maximum geographical distance from the furthest observation to the target center Similar procedures can be followed to calculate the tem poral weight for all the observations It’s the time difference between the current observation and the target center Its formulation is, [ ( )2 ] tin wTi (ti ) = − (4) tmax (6) where yi is the outcome of the response variable for the ith observation with incident severity levels (j = 1, 2, 3, 4, or 5); xim is the corre sponding variable for marginal effect, and m denotes the index for the total number of M selected variables in the regression models; βm is a (1 ×M) vector for each one of the covariates 3.2.2 Delta test and model performance comparison Like the random parameter models (Hainen et al., 2020; Islam et al., 2014; Xu et al., 2019), the randomized coefficients should be significant based on the distributions (e.g., normal) Thus, it’s necessary to perform statistical tests to validate the significance of the spatio-temporal vari ations of the estimated coefficients A simple analysis of variance (ANOVA) test could be applied, but with regard to spatial stratified heterogeneity, advanced tests are always preferred such as the q-statistic (Wang et al., 2016a, 2016b), which measures both total sum of squares and within sum of squares for groups However, it can’t be used to test spatial and temporal heterogeneity simultaneously This study utilizes a non-stationary significance test, the Delta test, to reveal the significant varying correlations with incident severity The Delta test compares the variations of localized estimated coefficients from the GTWOLR model with global estimated coefficients from the conventional OLR model The variations of the locally estimated coefficients can be calculated based on the maximum and minimum values To avoid the extreme dispersion of the local coefficients, the interquartile range Delta between the first quantile (25th percentile) and the third quantile (75th percentile) is used for the non-stationary test Variables that pass the Delta test are considered as significant spatio-temporal variables, meaning substantial variations of correlations with incident severity across space and time The Delta test is formulated as, where wTi (ti ) is the temporal weight for nth observation with respect to the target center at ith observation labeled (ti ); tin is the temporal dis tance from nth observation to the target center; dmax is the maximum temporal distance of the furthest observation with regard to the target center The final spatio-temporal weight is calculated based on both spatial and temporal weights calculated for each observation in the entire sample with respect to the target center (Liu et al., 2019a) To capture the spatial and temporal effects simultaneously on model esti mations, the spatio-temporal weight is chosen as the final weight in the modeling process The spatio-temporal weight for the target center is equal to Equation (5) formulates the spatio-temporal weight for each observation, [ ( ( )2 ] )2 ] [ din tin S T wST (g , l , t ) = w (g , l ) × w (t ) = − × − i i i i i i i i i dmax tmax ) ( exp − κj (gi , li , ti ) + β(gi , li , ti )xi ( )]2 βm + exp − κj (gi , li , ti ) + β(gi , li , ti )xi (5) Step Local sample selection Unlike conventional regression modeling approaches, GTWOLR will subtract a local sample for local model estimations The local sample selection is based on the spatiotemporal weights derived from Step Observations with larger weights are highly likely to be selected in the local sample because they are neighbors with both small spatial and temporal distance to the target center The optimal bandwidth for the local sample size can be deter mined based on model performance indicators such as the Akaike In formation Criterion (AIC) (Bozdogan, 1987) Delta = βfq − βtq { Test Results= (7) Passed Non, stationary test, if Delta > 1.96SE &max|z| > 1.96 Failed Non, stationary test, if otherwise (8) zi = β(gi , li , ti ) / SE[β(gi , li , ti )] Step Re-weighting of local sample observations Once the local sample is determined based on Step 2, Step procedures can be per formed for the observations in a local sample to get the spatio-temporal weights In the same way, the observations in a local sample are not handled equally but weighted based on Equation (5) (9) where Delta is the test values for each variable; βfq is the 1st quantile of the corresponding variable coefficients from the GTWOLR model; βtq denotes the 3rd quantile of the corresponding variable coefficients from the GTWOLR model; SE is the standard error of the corresponding var iables from the conventional OLR model; |z| is the z-value of the local variable coefficients from the GTWOLR model The model performance can be measured by many other statistical tests The most often used tests for logistic models are log-likelihood, Pdeudo-R2, and AIC (Liu et al., 2019a; Bozdogan, 1987) Models with larger log-likelihoods and Pseudo-R2, smaller AICs are preferred, meaning the model can explain more information about the data with greater goodness-of-fit The equations of these performance indicators are (StataCorp, 2019): Step Weighted OLR In the last step, GTWOLR will estimate local relationships between incident severity and covariates for each obser vation in a local sample based on Equation (2) The local model will incorporate the spatio-temporal weights of all the locally sampled ob servations These above-mentioned steps will be repeated until all the observations are treated as the target center once in turn The GTWOLR approach will produce a local set of coefficients for all the covariates of each observation It reveals the localized varying cor relations of incident severity with the selected variables in correspon dence to space and time The GTWOLR modeling process is coded in the R environment For more details of this approach, please refer to the study done by (Liu et al., 2019a, b) The marginal effects can be esti mated based on the model outputs They are often used to reveal the changes of probabilities on different levels of the response variable when ⎧∑ n ∑ k ⎨ ln pij , if yi = j ln L = i=1 j=1 ⎩ 0, if otherwise (10) X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Table Global OLR model estimates for selected variables and the marginal effects Y = Natural Gas Distribution Pipeline Incident Severity Model Variables β S.E Minor Moderate Serious Severe Critical 0.652 1.132** 0.862** 0.506 − 0.029 − 0.062 0.322 − 0.246 − 0.675* − 0.130 0.421 0.486 0.389 0.424 0.464 0.436 0.425 0.347 0.383 0.382 − 9.46% − 14.40% − 11.83% − 7.63% 0.51% 1.08% − 5.09% 2.86% 8.91% 1.45% − 4.20% − 9.83% − 6.47% − 2.85% 0.07% 0.13% − 1.47% 2.57% 5.64% 1.42% 7.13% 11.02% 9.04% 5.65% − 0.34% − 0.72% 3.67% − 2.32% − 7.05% − 1.18% 3.72% 7.16% 5.17% 2.79% − 0.14% − 0.29% 1.69% − 1.69% − 4.22% − 0.91% 2.81% 6.05% 4.09% 2.05% − 0.09% − 0.20% 1.20% − 1.41% − 3.29% − 0.78% − 1.402*** 0.610 0.610*** 1.355*** − 0.530* − 0.633** − 1.208** − 0.186 0.739** 0.490** 0.452** 0.361 0.461 1.090*** 0.465** 0.283 0.452*** 0.495** 0.011 0.557*** − 0.136 0.540 0.654 0.228 0.147 0.281 0.306 0.605 0.317 0.358 0.236 0.220 0.234 0.340 0.339 0.203 0.216 0.172 0.198 0.164 0.190 0.564 23.03% − 8.07% − 8.07% − 11.66% 5.21% 6.45% 14.87% 1.61% − 6.63% − 6.24% − 5.82% − 4.77% − 5.93% − 11.59% − 5.69% − 3.66% − 5.55% − 6.00% − 0.13% − 5.51% 4.41% − 5.21% − 5.21% − 18.58% 6.46% 7.49% 11.16% 2.43% − 9.41% − 4.46% − 4.03% − 3.07% − 4.14% − 12.18% − 4.57% − 2.54% − 4.41% − 4.92% − 0.11% − 6.73% − 15.57% 6.55% 6.55% 8.86% − 4.03% − 5.05% − 11.50% − 1.14% 4.72% 5.06% 4.72% 3.85% 4.80% 8.87% 4.70% 3.00% 4.58% 4.95% 0.11% 4.24% − 6.96% 3.80% 3.80% 11.25% − 3.92% − 4.61% − 7.88% − 1.44% 5.61% 3.15% 2.88% 2.26% 2.95% 7.75% 3.08% 1.80% 2.99% 3.30% 0.08% 4.12% − 4.91% 2.93% 2.92% 10.14% − 3.71% − 4.28% − 6.66% − 1.47% 5.72% 2.48% 2.25% 1.73% 2.31% 7.14% 2.48% 1.40% 2.40% 2.68% 0.06% 3.88% Cutpoint κ2 1.899*** 0.567 3.640*** 0.574 Cutpoint κ4 4.775*** 0.583 Apparent cause of the incident (base: corrosion failure) CGA-DIRT root cause (base: others) Intentional outside force damage (yes) Incident resulted from (base: other reasons) Gas explosion (yes) Material involved in the incident (base: iron) Controller(s) or control room issues (yes) Response time to incident site (base: 6–10 min) Lifetime of gas distribution pipes (base: others) Perceptions of pipe pressure (base: no concern) Cutpoint κ1 Natural force damage Excavation damage Other outside force damage Pipe, weld, or joint failure Equipment failure Incorrect operation Other incident cause Excavation practices not sufficient Locating practices not sufficient One-call notification practices not sufficient Intentional release of gas Unintentional release of gas Steel Plastic Copper Others 0–5 11–30 31–60 Over h Unknown 0–10 years 11–20 years 21–50 years Over 50 years Min or concern Major concern Cutpoint κ3 Summary statistics Number of observations Log-likelihood at zero LL(0) 1094 − 1555.469 Log-likelihood at convergence LL(β) − 1469.988 Degree of freedom k Pseudo-R2 AIC 30 0.055 2999.976 Marginal Effects Note: “Coef.” indicates the estimated coefficients; “S.E.” indicates the standard error; “***“, “**“, and “*” represent the variable significance at 99%, 95%, and 90% significance levels separately / Pseudo R2 = − ln L ln Lnull (11) AIC = 2k-2lnL (12) likelihood at convergence = − 1284.68 vs − 1469.99) The OLR model results show that excavation damage and outside force damage could lead to higher incident severity These two types of apparent causes of incidents are related to human errors For example, when workers were performing drilling tasks on the ground for con struction, they could falsely damage the pipes under the ground if the pipes were not detected before the task Or a drunk motor vehicle driver crashed unto the pipes that were laid out above the ground Addition ally, the marginal effects reveal that excavation damage and outside force damage were associated with 18.3% and 24.4% increased likeli hood of a serious or higher incident severity (e.g., incident costing $0.1 million or more), respectively However, if the outside force damage was intentional (meaning the damage with done with care), the incident was found to associate with a 27.4% decreased likelihood of a serious or higher severity In other words, unintentional damage was more likely to correlate with a serious or higher severity Results show that the unintentional release of gas was associated with a 13.3% increased likelihood of serious or higher incident severity Other variables such as iron pipe, controller or control room issues, longer response time (e.g., 11–30 min), longer lifetime of pipes (e.g., 21–50 years), and pipe pres sure concerns were found to associate with increased incident severity Specifically, some of these variables are found to have varying Where ln L is the model log-likelihood; Lnull is the likelihood of the null model without any explanatory variables; k is the model degree of freedom A three-point percentage decrease in AIC values indicates a significant performance improvement when comparing model perfor mances (Liu et al., 2019a) Results 4.1 Model results overview Table presents the model estimates and the marginal effects of the conventional global OLR model Table presents the variations of the GTWOLR model estimations and the Delta test results The AIC of the GTWOLR model is 2629.36, which is about 12.4-points percentage decrease from the AIC = 2999.98 of the OLR mode This indicates that the GTWOLR model significantly improves the model performance by accounting for the unobserved heterogeneity related to space and time Other model performance indicators also lead to the same conclusion in favoring of the GTWOLR model (e.g., Pseudo-R2 = 0.174 vs 0.055, Log7 X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Table Local GTWOLR model estimates for selected variables and the delta test results Y = Natural Gas Distribution Pipeline Incident Severity GTWOLR Model Estimates Variables βMean Apparent cause of the incident (base: corrosion failure) CGA-DIRT root cause (base: others) Intentional outside force damage (yes) Incident resulted from (base: other reasons) Gas explosion (yes) Material involved in the incident (base: iron) Natural force damage Excavation damage Other outside force damage Pipe, weld, or joint failure Equipment failure Incorrect operation Other incident cause Excavation practices not sufficient Locating practices not sufficient One-call notification practices not sufficient Intentional release of gas Unintentional release of gas Steel Plastic Copper Others Controller(s) or control room issues (yes) Response time to incident site (base: 0–5 6–10 min) 11–30 31–60 Over h Unknown Lifetime of gas distribution pipes 0–10 years (base: others) 11–20 years 21–50 years Over 50 years Perceptions of pipe pressure (base: no Min or concern concern) Major concern Cutpoint κ1 βMin β - 1st quantile Delta Test Results β - 3rd quantile βMax Max|z| Delta values Pass/ Fail 0.498 0.903 0.629 0.218 − 0.580 − 0.429 − 0.033 − 0.333 − − − − − − − − 0.631 1.664 0.161 0.752 3.602 1.595 1.526 1.648 0.168 0.505 0.326 − 0.299 − 1.095 − 1.002 − 0.515 − 0.724 0.851 1.461 0.926 0.810 0.454 0.066 0.367 0.002 1.581 2.590 1.915 1.550 1.001 1.528 0.859 1.395 1.120 2.178 1.417 1.555 2.350 1.428 0.796 1.662 0.682 0.955 0.601 1.109 1.549 1.068 0.882 0.726 Fail Pass Fail Fail Pass Fail Fail Fail − 0.766 − 0.232 − 2.294 − 1.489 − 1.266 − 0.680 − 0.321 0.119 0.904 2.482 2.623 1.452 0.945 0.799 Pass Fail − 0.476 1.098 1.021 1.588 − 0.396 − 0.301 − 0.136 0.136 1.328 0.803 0.775 0.640 1.215 1.438 0.815 0.565 0.703 0.680 0.359 0.814 0.231 0.921 3.209 3.956 2.148 0.142 0.435 0.944 1.175 1.670 1.390 1.445 1.093 2.129 2.263 1.449 1.057 1.009 1.198 0.949 1.357 1.304 8.59E+07 4.98E+06 2.598 5.122 2.387 2.661 3.28E+06 1.343 1.645 2.278 2.469 1.403 2.369 5.66E+08 2.788 1.388 1.999 1.840 1.556 2.877 1.842 − 1.715 − 0.278 − 0.647 − 1.358 − 0.741 − − 0.790 − − 1.395 − − 0.204 − 0.876 0.477 − 0.459 − 0.343 − 0.240 − 0.421 − 0.448 − 0.312 − 0.492 − 0.448 − 0.126 − 0.652 − − 0.751 − 15.531 − 2.610 13.859 − 1.301 1.191 − 0.066 0.352 1.117 1.818 − 1.119 1.959 − 1.332 15.562 − 2.262 1.507 − 0.656 − 0.730 0.482 0.456 0.202 1.027 0.187 0.882 0.105 3.436 − 0.564 15.333 0.352 0.530 − 0.027 0.868 0.102 0.294 0.323 0.362 0.216 0.784 − 0.127 0.213 0.440 − 3.558 − 1.652 0.600 0.588 0.535 1.778 1.086 0.842 0.463 0.380 0.464 0.486 0.374 1.884 2.134 Pass Pass 0.472 Pass Pass Pass Fail 0.847 Pass Pass Fail Pass Pass Pass Fail Pass Fail Fail Pass Fail 2.399 1.087 0.723 1.030 2.126 0.791 Cutpoint κ2 1.581 − 0.726 0.799 2.221 4.044 1.979 1.422 Pass Cutpoint κ3 3.430 0.889 2.811 3.962 6.343 3.171 1.151 Pass Cutpoint κ4 4.577 2.173 3.985 5.103 7.855 4.108 1.119 Pass Summary statistics Number of observations Log-likelihood at zero LL(0) 1094 − 1555.469 Log-likelihood at convergence LL(β) − 1284.679 Degree of freedom k Pseudo-R2 AIC 30 0.174 2629.357 Note: “Coef.” indicates the estimated coefficients; “S.E.” indicates the standard error; “***“, “**“, and “*” represent the variable significance at 99%, 95%, and 90% significance levels separately correlations with incident severity across space and time (e.g., excava tion damage, intentional outside force damage, unintentional release of gas, plastic pipe, response time, pipe lifetime, the major concern of pipe pressure) The next section will discuss these varying correlations in detail facility markers should be placed wherever appropriate for operators to clearly identify natural gas distribution pipes both underground and above the ground The root cause of excavation damage was related to insufficient practices related to locating, excavation, and notification Fig (b) indicates that insufficient locating practices compared to other root causes (e.g., abandoned or deteriorated facilities, working outside of the scope of work - beyond the marked-out area, inaccurate marking, crews did not follow procedures, and mapping records error) did not seem to cause severer incidents But this is not always the case, insuf ficient locating practices would correlate with severer incidents in Oregon and Alaska Intentional outside force damage to the pipes (e.g., theft of transported commodity, theft of equipment) was found to associate with decreased incident severity because these actions could be detected if sudden changes were identified out of the normal func tioning system Or it could be the pipes were damaged for certain maintenance purposes (e.g., pipe joint fixing) and the damage was taken care of by the crews However, unintentional damage to gas releasing could lead to increased incident severity, especially for incidents that occurred in mid-western states such as Montana, Idaho, Nevada, Utah, 4.2 Discussions of key contributing factors Based on the Delta test results presented in Table 4, this study reveals that some of the contributing factors hold significant varying correla tions with natural gas distribution pipeline incident severity Fig ex hibits the correlations of those factors across different states in the U.S Spatially, regarding the apparent cause of natural gas distribution pipeline incidents, excavation damage done by either the operator, contractor or another third party could lead to even severer incidents costing $0.1 million or more in New York, Connecticut, Massachusetts, Washington D.C., and Kentucky compared to other states (see Fig (a)) Thus, pipe inspection using various tools (e.g., magnetic, ultrasonic, caliper) before the excavation work should be encouraged and pipeline X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Fig Spatial-varying correlations with incident severity for key contributing factors and Arizona (see Fig (d)) Interestingly, unintentional gas release was found to associate with decreased incident severity in some states (e.g., North Carolina) This might due to a faster emergency response to those incidents Statistics show that average response time for all incidents was 64 min, while response time for incidents in North Carolina was just 19 Other natural gas distribution pipeline incident associated contrib uting factors were also investigated Fig (e) reveals that if the gas explosion occurred in incidents, it was 30.25% more likely that those serious, severe, or critical incidents that would cost $0.1 million or more Materials of the pipelines (e.g., iron, steel, plastic, coppers) also matter in terms of the corrosion failures The results indicate that compared to iron, steel, plastic, and copper pipelines were associated with decreased incident severity due to better corrosion resistance of these materials AGA also reported that cast iron pipelines have declined by 61% since 1990, while the use of modern plastic pipelines has Fig Time-varying correlations with incident severity for key contributing factors X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 increased by 215% for improved natural gas service (AGA, 2020) Replacing iron pipelines are strongly recommended, especially for Washington, Montana, Nevada, and Utah states (see Fig (f)) Simi larly, pipelines that have long been used to provide the gas service should also be considered for replacement or at least some maintenance The results indicate that pipelines that had over 20 service years would be associated with a 10% increased likelihood of leading to a natural gas distribution pipeline incident that cost $0.1 million or more Fig (h) reveals that those pipelines in South Carolina, North Carolina, Iowa, and Minnesota were associated with even higher incident severity and should be given more maintenance service or replacements using modern pipes In terms of emergency incident response time, the longer response time (e.g., over h) was found to associated with a 10% increased likelihood of an incident that could cost $0.1 million or more Fig (g) indicates that such correlations were even stronger in South Carolina, North Carolina, North Dakota, New Hampshire, and Con necticut Interestingly, the short response time (e.g., 0–5 min) was also found associated with increased incident severity The statistics show that on average, 7.5% more fatalities and 75% more injuries were associated with incidents that triggered fast emergency incident response within Lastly, pipeline pressure was another issue for those incidents If the crew was talking more about the pipeline pres sure, the incidents were 12% more likely to be a serious or severer incident that cost $0.1 million or more Routine monitoring and ad justments of pipeline pressure are recommended for a better natural gas distribution pipeline safety In addition to spatially varying correlations with incident severity, significant temporal correlations are also revealed Fig depicts the overall trends of these correlations For those contributing factors that have a consistent positive correlation with incident severity (e.g., gas explosion, pipeline pressure), safety countermeasures such as routine natural gas pressures monitoring and adjustments of the pressure by controllers in the control room are encouraged As for pipeline mate rials, protected steel pipes, modern plastic pipes are recommended to replace the iron pipes that were associated with corrosion failures Efficient and coordinated emergency response to those natural gas dis tribution pipeline incidents is recommended Such a coordinated response system would help significantly reduce the cost associated with the incidents after 2017 investigate the contributing factors, especially human-related contrib uting factors that are correlated with the incident severity A total of 1094 natural gas distribution pipeline incidents were collected and a comprehensive cost was estimated for each incident based on the fa talities, injuries, property damage cost, and released gas cost It was then categorized into levels (minor moderate, serious, severe, and critical) based on the estimate comprehensive cost Text mining was utilized to extract useful information from incident report narratives, and GTWOLR was applied to examine the varying correlations between incident severity and the contributing factors Key findings of this study are: • GTWORL model outperforms the conventional OLR and significantly improves the model AIC performance indicator by 12%; • Human-related contributing factors excavation damage and outside force damage were associated with 18.3% and 24.4% increased likelihood of a serious or higher incident severity that cost $0.1 million or more; Factors such as unintentional gas release, iron pipes, longer emergency response, extended pipe lifetime service, and concerns with pipeline pressure were also found to positively asso ciate with a serious or greater severity of a natural gas distribution pipeline incident; • Improving excavation skills using advanced pipe detection tools, separating fire ignition source from natural gas pipes, replacing iron pipes or long-service pipes with modern plastic or protected steel, and establishment of a coordinated natural gas distribution pipeline incident response team are strongly recommended, especially for states with those particular issues The outcomes of this study are expected to attract the attention of the practitioners who work in the pipeline distribution system Human er rors are the main source to contribute to the natural gas distribution pipeline incidents besides the natural pipeline corrosion and the natural force damage (e.g., flooding and hurricanes) Thus, efforts should be put forward jointly by operators, managers, policymakers, customers, ex cavators, and emergency responders to develop safety countermeasures and compliance protocols in improving natural gas distribution pipeline safety for an efficient, clean and safe energy future Such a study method can be applied to other pipeline incidents In the future, we plan to extend our study to transmission pipelines, and incidents involving hazardous liquid Limitations Author statement The study quality is heavily dependent on the accuracy and validity of the natural gas dis distribution pipeline data explored in the models The final incident data presented to us were pre-processed by pro fessionals in PHMSA where possibly some important information is missing due to data publication regulations Thus, the accuracy and completeness of the incident data remain unclear Additionally, the incident reports may not truly represent the facts, especially when fa talities were involved, because incorrection pipeline operations are hard to obtain when those faults were made by the dead Thus, the accuracy of natural gas distribution pipeline incident data may impact the out comes of this study Furthermore, the study outcomes may also be affected by the model specifications (e.g., coding of variable categories) due to the small sample size The inclusion of more detailed information would enable a better model estimation Xiaobing Li: Study conceptualization and design, Data collection, Methodology, Visualization, Formal analysis and interpretation of re sults, Writing-Original draft preparation Praveena Penmetsa: Study conception and design, Data collection, Analysis and Interpretation of results, Writing- Reviewing and Editing Jun Liu: Study conception and design, Resources, Methodology, Analysis and Interpretation of results, Writing- Reviewing and Editing Alexander Hainen: Draft manuscript preparation, Writing- Reviewing and Editing Shashi Nambisan: Su pervision, Writing- Reviewing and Editing Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors Conclusions Declaration of competing interest Natural gas distribution pipeline incident statistics show that in the last decade, these incidents have caused a total of 107 life loss and 510 injured These incidents have caused society productivity loss, waste of energy, and many other costs related to health care In total, the average cost of a natural gas distribution pipeline incident was estimated to be $1.9 million Motivated by these high incident costs and the lack of adequate regulation, inspection, and enforcement, we intend to The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper 10 X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Acknowledgements Lemp, J.D., Kockelman, K.M., Unnikrishnan, A., 2011 Analysis of large truck crash severity using heteroskedastic ordered probit models Accid Anal Prev 43, 370–380 Li, J., Zhang, H., Han, Y., Wang, B., 2016 Study on failure of third-party damage for urban gas pipeline based on fuzzy comprehensive evaluation PloS One 11 Liu, H., 2003 Pipeline Engineering CRC Press Liu, J., Hainen, A., Li, X., Nie, Q., Nambisan, S., 2019a Pedestrian injury severity in motor vehicle crashes: an integrated spatio-temporal modeling approach Accid Anal Prev 132, 105272 Liu, J., Khattak, A.J., Li, X., Fu, X., 2019b A spatial analysis of the ownership of alternative fuel and hybrid vehicles Transport Res Transport Environ 77, 106–119 Liu, J., Nambisan, S., Li, X., Fu, X., 2020 Are young Americans carless across the United States? A spatial analysis Transport Res Transport Environ 78, 102197 Maghrebi, M., Abbasi, A., Rashidi, T.H., Waller, S.T., 2015 Complementing travel diary surveys with twitter data: application of text mining techniques on activity location, type and time In: IEEE 18th International Conference on Intelligent Transportation Systems, 2015 IEEE, pp 208–213 Montiel, H., Vílchez, J.A., Arnaldos, J., Casal, J., 1996 Historical analysis of accidents in the transportation of natural gas J Hazard Mater 51, 77–92 Moran, M J & Monje, C 2016 Guidance on treatment of the economic value of a statistical life (vsl) in us department of transportation analyses-2016 adjustment memorandum to secretarial officers and modal administrators, August, Naik, D.L., Kiran, R., 2018 Data mining and equi-accident zones for US pipeline accidents J Pipeline Syst Eng Pract 9, 04018019 Ramírez-Camacho, J.G., Carbone, F., Pastor, E., Bubbico, R., Casal, J., 2017 Assessing the consequences of pipeline accidents to support land-use planning Saf Sci 97, 34–42 Rui, Z., Wang, X., 2013 A comprehensive analysis of natural gas distribution pipeline incidents Int J Oil Gas Coal Technol 6, 528–548 Siler-Evans, K., Hanson, A., Sunday, C., Leonard, N., Tumminello, M., 2014 Analysis of pipeline accidents in the United States from 1968 to 2009 Int J Critic Infrastruct protect 7, 257–269 Simonoff, J.S., Restrepo, C.E., Zimmerman, R., 2010 Risk management of cost consequences in natural gas transmission and distribution infrastructures J Loss Prev Process Ind 23, 269–279 StataCorp, 2019 Stata Statistical Software: Release, vol 16 StataCorp LLC, College Station, TX Wang, J.-F., Zhang, T.-L., Fu, B.-J., 2016a A measure of spatial stratified heterogeneity Ecol Indicat 67, 250–256 Wang, X., Liu, J., Khattak, A.J., Clarke, D., 2016b Non-crossing rail-trespassing crashes in the past decade: a spatial approach to analyzing injury severity Saf Sci 82, 44–55 Xu, J., Wali, B., Li, X., Yang, J., 2019 Injury severity and contributing driver actions in passenger vehicle–truck collisions Int J Environ Res Publ Health 16, 3542 Zheng, J., Zhang, B., Liu, P., Wu, L., 2012 Failure analysis and safety evaluation of buried pipeline due to deflection of landslide process Eng Fail Anal 25, 156–168 Zhou, Y., Hu, G., Li, J., Diao, C., 2014 Risk assessment along the gas pipelines and its application in urban planning Land Use Pol 38, 233–238 Zhu, X., Srinivasan, S., 2011 Modeling occupant-level injury severity: an application to large-truck crashes Accid Anal Prev 43, 1427–1437 The authors would like to thank the PHMSA for offering the publicly available incident data on its official website The authors are respon sible for the fact, views, and accuracy of the information presented in this study References Abdel-Aty, M., 2003 Analysis of driver injury severity levels at multiple locations using ordered probit models J Saf Res 34, 597–603 AGA, 2020 2020 AGA playbook (AGA) http://playbook.aga.org/: American Gas Ass ociation (Accessed 23 May 2020) Bozdogan, H., 1987 Model selection and Akaike’s information criterion (AIC): the general theory and its analytical extensions Psychometrika 52, 345–370 Burgherr, P., Hirschberg, S., 2005 Comparative Assessment of Natural Gas Accident Risks Paul Scherrer Institut Chen, W.-H., Jovanis, P.P., 2000 Method for identifying factors contributing to driverinjury severity in traffic crashes Transport Res Rec 1717, 1–9 Cunha, S.B., 2012 Comparison and analysis of pipeline failure statistics In: 9th International Pipeline Conference, 2012 American Society of Mechanical Engineers Digital Collection, pp 521–530 Das, S., Sun, X., Dutta, A., 2016 Text mining and topic modeling of compendiums of papers from transportation research board annual meetings Transport Res Rec 2552, 48–56 Fotheringham, A.S., Charlton, M.E., Brunsdon, C., 1998 Geographically weighted regression: a natural evolution of the expansion method for spatial data analysis Environ Plann 30, 1905–1927 Friedman, M., 1962 The interpolation of time series by related series J Am Stat Assoc 57, 729–757 Girgin, S., Krausmann, E., 2014 Analysis of Pipeline Accidents Induced by Natural Hazards, JRC88410 Finalreport JRC Technical Reports (European Union) Hainen, A.M., Harbin, K.B., Dye, D., Lindly, J.K., 2020 Duration analysis of emergency shutdown incidents regarding hazardous liquid pipelines J Perform Constr Facil 34, 04020040 Han, Z., Weng, W., 2010 An integrated quantitative risk analysis method for natural gas pipeline network J Loss Prev Process Ind 23, 428–436 Islam, M., Hernandez, S., 2013 Large truck–involved crashes: exploratory injury severity analysis J Transport Eng 139, 596–604 Islam, S., Jones, S.L., Dye, D., 2014 Comprehensive analysis of single-and multi-vehicle large truck at-fault crashes on rural and urban roadways in Alabama Accid Anal Prev 67, 148–158 Jo, Y.-D., Ahn, B.J., 2002 Analysis of hazard areas associated with high-pressure naturalgas pipelines J Loss Prev Process Ind 15, 179–188 Khattak, A.J., Schneider, R.J., Targa, F., 2003 Risk factors in large truck rollovers and injury severity: analysis of single-vehicle collisions, 40 Transportation Research Record Khattak, A.J., Targa, F., 2004 Injury severity and total harm in truck-involved work zone crashes Transport Res Rec 1877, 106–116 11 ... 2019), X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 Fig Nationwide spatial distribution of natural gas distribution pipeline incidents Fig Temporal patterns of natural... gas production and delivery system (Source: U.S Energy Information Administration) X Li et al Journal of Loss Prevention in the Process Industries 69 (2021) 104383 (Simonoff et al., 2010), finite... associated with fatalities and injuries According to their VSL estimates, VSLs of 2007, 2013, 2014, 2015, and 2016 were $5.8 million, $9.1 million, $9.2 million, $9.4 million, and $9.6 million, respectively