Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Journal of Safety Research 76 (2021) 154–165

Contents lists available at ScienceDirect

Journal of Safety Research


journal homepage: www.elsevier.com/locate/jsr

Risk factors associated with truck-involved fatal crash severity:


Analyzing their impact for different groups of truck drivers
Yalong Yuan a,b,c, Min Yang a,b,⇑, Yanyong Guo a, Soora Rasouli c, Zuoxian Gan d, Yifeng Ren a
a
School of Transportation, Southeast University, Jiangsu Key Laboratory of Urban ITS, Southeast University, Jiangsu Province Collaborative Innovation Center of Modern Urban
Traffic Technologies, Southeast University, PR China
b
School of Transportation, Southeast University, Jiangsu Key Laboratory of Urban ITS, Southeast University, Jiangsu Province Collaborative Innovation Center of Modern Urban
Traffic Technologies, Southeast University, 2 Sipailou, Nanjing 210096, PR China
c
Urban Planning Group, Eindhoven University of Technology, PO Box 513, 5600 MB Eindhoven, The Netherlands
d
School of Transportation, Dalian Maritime University, PR China

a r t i c l e i n f o a b s t r a c t

Article history: Introduction: Fatal crashes that include at least one fatality of an occupant within 30 days of the crash
Received 2 May 2020 cause large numbers of injured persons and property losses, especially when a truck is involved.
Received in revised form 21 July 2020 Method: To better understand the underlying effects of truck-driver-related characteristics in fatal
Accepted 15 December 2020
crashes, a five-year (from 2012 to 2016) dataset from the Fatality Analysis Reporting System (FARS)
Available online 31 December 2020
was used for analysis. Based on demographic attributes, driving violation behavior, crash histories, and
conviction records of truck drivers, a latent class clustering analysis was applied to classify truck drivers
Keywords:
into three groups, namely, ‘‘middle-aged and elderly drivers with low risk of driving violations and high
Truck-involved fatal crash
Risk factors
historical crash records,” ‘‘drivers with high risk of driving violations and high historical crash records,”
Crash severity and ‘‘middle-aged drivers with no driving violations and conviction records.” Next, equivalent fatalities
Latent class clustering were used to scale fatal crash severities into three levels. Subsequently, a partial proportional odds
Partial proportional odds model (PPO) model for each driver group was developed to identify the risk factors associated with the crash
severity. Results’ Conclusions: The model estimation results showed that the risk factors, as well as their
impacts on different driver groups, were different. Adverse weather conditions, rural areas, curved align-
ments, tractor-trailer units, heavier weights and various collision manners were significantly associated
with the crash severities in all driver groups, whereas driving violation behaviors such as driving under
the influence of alcohol or drugs, fatigue, or carelessness were significantly associated with the high-risk
group only, and fewer risk factors and minor marginal effects were identified for the low-risk groups.
Practical Applications: Corresponding countermeasures for specific truck driver groups are proposed.
And drivers with high risk of driving violations and high historical crash records should be more
concerned.
Ó 2020 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction the cargo demand and truck volume increase (Dong, Richards,
Huang, & Jiang, 2015). Therefore, research is urgently needed to
Trucks (vehicles with a gross weight rating greater than 10,000 analyze and understand truck-involved fatal crashes as corre-
pounds) are a vital component of freight logistics and economic sponding safety countermeasures should be developed.
well-being. However, they are overrepresented in terms of both For fatal crashes involving trucks, previous studies have mainly
fatal crash frequency and serious property damage compared with focused on driver fatalities (Brodie, Lyndal, & Elias, 2009;
passenger cars (Uddin & Huynh, 2017; Yuan et al., 2019; Zaloshnja Häkkänen & Summala, 2001; Lombardi, Horrey, & Courtney,
& Miller, 2004). In the United States between 2012 and 2016, an 2017; Weaver & Bédard, 2013). However, Slovic, Lichtenstein,
average of 3,591 fatal crashes involved large trucks, leading to an and Fischhoff (1984) indicated that the social impact of fatal acci-
annual average of 4,049 people killed (NIOSH, 2016). Furthermore, dents cannot be accurately measured by only considering the fatal-
the number of truck-involved fatal crashes will likely increase as ities, and the detailed characteristics of casualties should be
analyzed comprehensively. In addition, Zhu and Srinivasan
(2011) indicated that heterogeneity occurs in crashes with the
⇑ Corresponding author.
same injury severity, such as fatal accidents. Moreover, drivers,
E-mail address: yangmin@seu.edu.cn (M. Yang).

https://doi.org/10.1016/j.jsr.2020.12.012
0022-4375/Ó 2020 National Safety Council and Elsevier Ltd. All rights reserved.
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

as the operators of trucks, have an important responsibility in the facilities at intersections, traffic flows in different periods, and
context of crashes and the safety of other occupants involved in the lighting conditions in work zones were significantly related to
accidents. Studies have indicated that 90% of crashes can be attrib- crash injury severity.
uted mainly to drivers (NHTSA, 2010). Therefore, it is necessary to Regarding collision-related risk factors, some studies have indi-
study the roles of the relevant characteristics of drivers involved in cated that specific collision types, such as head-on, rear-end, and
crashes in detail. Previous studies have mainly classified drivers sideswipe collisions, and the number of vehicles involved in a crash
into male or female, young or old, and novice or experienced have significant effects on the crash-based injury severity. Chen
groups according to a single attribute, such as the driver’s sex, and Chen (2011), Xie et al. (2012), and Islam, Jones, and Dye
age, or driving experience (Kim, Ulfarsson, Kim, & Shankar, 2012; (2014) studied the injury severity of drivers in single- and multive-
Lombardi et al., 2017; Morgan & Mannering, 2011; Yuan et al., hicle crashes involving trucks. The results showed that risk factors,
2019a,b). However, truck drivers are agents with many attributes, such as driving under the influence of alcohol, use of safety belts,
such as specific demographics, driving habits, crash histories, and weight and body type of trucks, and the curviness of roadway seg-
conviction records. A single attribute, such as age or gender, is ments, have significantly different impacts on the injury severity in
not sufficient as a proxy for different driver types, and this single- and multi-vehicle crashes. Wang and Abdel-Aty (2008), Pai
approach may result in an incomplete or biased evaluation of dri- and Saleh (2008), and Guo, Li, Liu, and Wu (2019) studied specific
vers due to the heterogeneity that exists among different age or manners of collision in crashes. The results indicated that the crash
gender groups. For example, one study found that young drivers characteristics for different collision manners, such as head-on,
were more likely to be involved in crashes than middle-aged dri- sideswipe, and rear-head collisions, are significantly different and
vers (Massie, 1995), but a certain proportion of young drivers that attention should be given to the manner of collision in the
demonstrated good driving habits and maintenance of a low crash context of crashes.
involvement rate. To minimize the heterogeneity among different For vehicle-related risk factors, Chen and Chen (2011) found
driver groups, it is necessary to comprehensively consider driver- that a single-unit truck is safer for drivers than non-single-unit
related characteristics. We can also speculate about who is a dan- trucks in single-vehicle crashes, but single-unit trucks tend to
gerous or safe driver based on more detailed profiles of different increase driver injury severity in multi-vehicle crashes. They also
diver groups. Moreover, implementing corresponding countermea- found that truck tire defects are more dangerous than brake
sures for specific driver groups is of great significance for safety defects in terms of driver safety. Kim et al. (2013) found that
managers. new vehicles can reduce injury severity in single-vehicle crashes
The objective of this study was to identify various risk factors for male drivers and that older drivers who drive older vehicles
associated with fatal crash severities and to analyze their impact tend to suffer more severe injuries. However, Khorashadi et al.
on different truck driver groups by incorporating a comprehensive (2005) found that multiunit trucks and large tractors have higher
list of variables to classify the drivers. The remainder of the paper risks than single-unit large trucks due to the chance of experienc-
is organized as follows. First, a literature review is presented. Sec- ing severe injuries in all types of crashes. Zhu and Srinivasan
ond, data collection and the classification of truck drivers are per- (2011) indicated that tractor trailers transporting heavy cargo
formed. Subsequently, the methodology, which focuses on a partial can lead to high injury severity relative to that for single-unit
proportional odds (PPO) model, is introduced. Next, the model trucks hauling less heavy cargo. It should be noted that vehicle-
results are provided and analyzed. Finally, the discussion is pre- related characteristics influence driver injury severity differently
sented, and conclusions are drawn. based on the overall crash level.
Among driver-related risk factors, the attributes of drivers,
2. Literature review including the relevant demographic information, driving behavior,
and history of conviction, have been studied in depth. Morgan and
Truck-involved crash analysis has attracted worldwide atten- Mannering (2011), Kim et al. (2012), and Lombardi et al. (2017)
tion (Brodie et al., 2009; Häkkänen & Summala, 2001; Lombardi studied the effects of age and sex on the injury severity of drivers,
et al., 2017; Slovic et al., 1984; Weaver & Bédard, 2013). Previous and a comparative analysis was performed among different age
studies have mainly identified the key risk factors from driver-, and gender groups of drivers. The results showed that the likeli-
vehicle-, roadway-, environment-, and collision-related perspec- hood of being involved in a fatal crash was higher for older drivers
tives (Chang & Mannering, 1999; Dong et al., 2013; Lemp, than for other age groups. Geneviève et al. (2002), Li, Baker, Qiang,
Kockelman, & Unnikrishnan, 2011; Zhu & Srinivasan, 2011; Grabowski, and McCarthy (2005), Alver, Demirel, and Mutlu
Chang & Chien, 2013; Yang, Guo, & Xu, 2019). Results have shown (2014), and Das, Sun, Wang, and Leboeuf (2015) analyzed the
that various risk factors, including driving behaviors, road loca- effects of driver histories, including traffic violations and crash
tions, weather conditions, the manner of collisions, and the num- records, on risky driving behavior and the likelihood of being
ber of vehicles involved, are significantly related to crash severity involved in a future crash. They indicated that the crash records
(Khorashadi, Niemeier, Shankar, & Mannering, 2005; Xie, Zhao, & of drivers can be better used to predict risky driving behavior than
Huynh, 2012; Zhu & Srinivasan, 2011). the history of traffic violations, and a high traffic violation involve-
For roadway- and environment-related risk factors, previous ment rate had a significant positive effect on risky driving behav-
studies have found that crashes that occurred in different time ior. In addition, Mercer (1987), Bédard, Guyatt, Stones, and
periods, area types, and roadway segments have distinct character- Hirdes (2002), Behnood, Roshandeh, and Mannering (2014), and
istics. Notably, Khorashadi et al. (2005) studied differences in truck Sullman, Stephens, and Pajo (2016) studied the effects of driving
driver injury severity in urban and rural areas. Uddin and Huynh violation behavior, such as drunk driving and wearing a seat belt,
(2017) analyzed crash injury severity under different lighting con- on crash severity. The results showed that both driving drunk
ditions in rural and urban areas. The results showed significant dif- and not wearing a seat belt can lead to serious driver injuries in
ferences in risk factors under different combinations of lighting crashes.
conditions and area types. Dong, Clarke, Richards, and Huang In summary, previous studies have mainly focused on safety
(2014), Pahukula et al. (2014), Wei, Wang, and Zhang (2016), and analyses of nonfatal crashes, and there is also a lack of specific
Osman, Paleti, Mishra, and Golias (2016) investigated the risk fac- studies on fatal crash severities, especially for truck-involved ones.
tors contributing to crash frequency and injury severity at intersec- In addition, the multiple attributes of truck drivers such as viola-
tions and work zones in different time periods. They found that the tion behaviors and conviction records, which greatly affect driver
155
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

classification results, must be considered. To fill in these gaps, the butions by maximally internally homogeneous and externally
classification of truck drivers, considering driver-related multiple heterogeneous (Abenoza, Cats, & Susilo, 2017; Chang & Yeh,
attributes, and underlying effects of different driver groups on fatal 2007; Feng et al., 2016; Guo & Fang, 2013; Moustaki &
crash severities, are analyzed in this study. Papageorgiou, 2005).
From the probabilistic perspective in LCC, every cluster has a
different underlying probability distribution from which its data
3. Data preparation and the classification of truck drivers
elements are generated (Depaire, Wets, & Vanhoof, 2008). When
the distribution functions are known, the problem of finding the
3.1. Data preparation
clusters reduces to a parameter estimation problem. Given a data
sample of N cases, measured with a set of observed variables
In the present study, data on truck-involved fatal crashes from
ðy1 ; y2 ;    ; ym Þ, the prior probability pðzÞ for cluster C z with
2012 to 2016 were retrieved and pooled from the Fatality Analysis
z ¼ 1; 2;    ; K and the conditional multivariate probability density
Reporting System (FARS) dataset (ftp://ftp.nhtsa.dot.gov/fars/). The
pðyj =C z ; hz Þ, where hz is the unknown parameter vector, and the
FARS files contain records of fatal motor-vehicle crashes within the
mixture probability density for the whole data set can be
United States and Puerto Rico, and each record includes at least one
expressed as in Eq. (1):
fatality of a vehicle occupant or non-occupant within 30 days of
the crash. Because the risk factors associated with fatal crash XK Y
m

severity are the main interest of this paper, only crash records pðY=hÞ ¼ z¼1
½pðC z Þ pðyj =C z ; hz Þ ð1Þ
j¼1
including environment, roadway, vehicle, driver, and collision
characteristics of interest were extracted. At the truck driver level, Following the maximum likelihood approach, the unknown
drivers’ age, gender, previous crash and traffic conviction histories, parameter vector is often estimated by means of the
and driving violation behaviors in the crash were considered in the expectation–maximization algorithm. After estimation of the
analysis. Ultimately, 15,506 valid records were obtained from the parameter vector h, the underlying statistical model assigns a set
original 17,955 total records. To facilitate the study of categorical of posterior probabilities pik belonging to cluster C k in each case i.
and nominal variables, environment-, roadway-, vehicle-, Compared with traditional clustering techniques, such as K-
collision-, and driver-related factors were transformed into binary means or hierarchical clustering algorithms, LCC is suitable for lar-
variables. The coding of each category is presented in Table 1, ger datasets without large memory demands, allowing the
which shows that the typical truck driver involved in a fatal crash researcher to include variables of different scales such as nominal,
can be characterized as male (97.1%) and middle-aged (82.0%), a continuous, and ordinal variables, in determining the number of
lower proportions of drivers have an invalid license (3.5%) and vio- clusters based on several statistical criteria, such as the Akaike’s
lation behaviors such as fatigued (0.9%), drinking (2.4%), and care- information criterion (AIC) and Bayesian information criterion
less (4.3%) driving, with relatively high proportions of history of (BIC) (Abenoza et al., 2017; Chang & Yeh, 2007; Feng et al., 2016;
crashes (14.5%) and conviction records such as license suspension Guo & Fang, 2013). In the present paper, all driver-related variables
(7.4%), speeding (23.6%), and other convictions (23.3%). mentioned in Table 1 are inputted into cluster analysis, and LCC is
To accurately describe the casualty level of the overall number fitted using Latent Gold software (v.4.5).
of deaths and injured persons involved in fatal crashes, the equiv- Table 2 shows indicators of fit for LCC with two through five
alent fatalities per accident is recommended to capture the equiv- cluster solutions. According to the results obtained for the degrees
alent impact of an injury relative to a fatality (Dupont, Martensen, of freedom and p-values, cluster solutions with two groups had
Papadimitriou, & Yannis, 2010; Evans, 1994; Feng, Li, Ci, & Zhang, significant test statistics (P < 0.001), suggesting that the latent
2016; Hu, Li, & Lee, 2010). For equivalent fatalities in the FARS classes in these models may not adequately account for hetero-
dataset, the equivalency rule of the KABCO scale is implemented geneity across individuals in the datasets. Smaller values of AIC,
as follows: 1 O-injury (no apparent injury), 1 C-injury (possible BIC, and L2 were obtained for cluster solutions with three groups.
injury), 1 B-injury (non-incapacitating injury), 1 A-injury (incapac- Considering the parsimony and stability of the results, the truck
itating injury), and 1 K-injury (killed) are equivalent to 0.0049, drivers were partitioned into three groups. In addition, the entropy
0.0148, 0.0310, 0.1107, and 1 fatalities, respectively (Blincoe, criterion (Celeux & Soromenho, 1996) and silhouette coefficient
Miller, Zaloshnja, & Lawrence, 2014). Thus, the equivalent fatalities (Zhou & Gao, 2014) were adopted to evaluate the clustering quality
of all persons involved in each accident can be calculated. The as in Eq. (2) and Eq. (3)–(5).
cumulative frequency distributions of equivalent fatalities are
Pn Pk
depicted in Fig. 1, which shows that a large difference occurs z¼1 pik lnðpik Þ
IðkÞ ¼ 1  i¼1
ð2Þ
among truck-involved fatal accidents. As shown in Fig. 1, the min- nlnð1=kÞ
imum, maximum, average, and standard deviation of the equiva-
lent fatalities are 1, 17.67, 1.22, and 0.96, respectively. And 85.9% where pik denotes the posterior probability that case i belongs to
of accidents have equivalent fatalities that are less than two per- cluster k and with the convention that pik lnðpik Þ = 0 if pik = 0.
sons, and only 5.5% of accidents have equivalent fatalities that
1 Xn bðiÞ  aðiÞ
are more than three persons. Referring to the method used in sim- SðkÞ ¼ i¼1 max½aðiÞ; bðiÞ
ð3Þ
n
ilar studies (Feng et al., 2016), we classified the crash severity into
three levels designated 1–3, with a higher severity level corre-
1 X
sponding to a more severe crash. Level 1 indicates less than two aðiÞ ¼ dði; jÞ ð4Þ
equivalent fatalities (85.9%), level 2 indicates greater than two nc  1 i;j2C ;i–j
z

but smaller than three equivalent fatalities (8.6%), and level 3 indi-
cates greater than three equivalent fatalities (5.5%). 1 X
bðiÞ ¼ minp;p–c ½ dði; jÞ ð5Þ
np i2C z ;j2C p
3.2. The classification of drivers
where aðiÞ is the average distance between case i and all other cases
A latent class clustering (LCC) was adopted to assign truck dri- in its cluster C z , and bðiÞ is the minimum average distance between
vers to several groups with different underlying probability distri- case i and all cases in any other cluster C p not containing case i.
156
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

Table 1
Statistical-descriptive analysis of truck-involved fatal crashes (N = 15,506).

Variable Categories Percentage Categories Percentage


Collision-related variables
Crash severity Level 1 85.9% Level 2 8.6%
Level 3 5.5%
Number of K-injuries 1 89.7% 2 8.1%
3 1.4% >3 0.8%
Number of A-injuries 0 82.1% 1 13.2%
2 3.0% >2 1.7%
Number of B injuries 0 76.6% 1 17.7%
2 4.1% >2 1.6%
Manner of collision Front-to-rear: 1, others: 0 19.9% Angle-front-to-side: 1, others: 0 16.1%
front-to-front: 1, others: 0 18.0%
Number of vehicles involved Multi vehicles: 1 84.7% Single vehicle: 0 15.3%
Environment-related variables
Atmospheric condition Clear: 0 84.2% Others (rain, cloudy, snow, etc.): 1 15.8%
Land use Rural: 1 66.6% Urban: 0 33.4%
Day of week Weekday: 0 84.5% Weekend: 1 15.5%
Time period Morning: 1, others: 0 36.0% Afternoon: 1, others: 0 41.0%
Evening: 1, others: 0 15.5% Night: 1, others: 0 7.5%
Season Spring: 1, others: 0 30.0% Summer: 1, others: 0 18.0%
Autumn: 1, others: 0 27.5% Winter: 1, others: 0 24.5%
Light condition Daytime: 1, others: 0 67.3% Dark-lighted: 1, others: 0 27.1%
Dark-not-lighted: 1, others: 0 4.8%
Road-related variables
Roadway alignment Straight: 0 83.4% Curve: 1 16.6%
Roadway grade Related variables: 0 72.9% Others (grade, uphill, etc.): 1 27.1%
Surface condition Dry: 0 86.6% Others (wet, ice, snow, etc.): 1 13.4%
Surface type Blacktop or asphalt: 0 86.6% Others (concrete, brick or stone): 1 13.4%
Speed limit 40 mile/h: 1, others: 0 11.7% 45–55 mile/h: 1, others: 0 48.8%
60–70 mile/h: 1, others: 0 36.6% >70 mile/h: 1, others: 0 2.9%
Number of lanes 2 lanes: 1 73.6% Others: 0 26.4%
Vehicle-related variables
Body type Truck-tractor: 1 64.4% Single unit truck: 0 35.6%
Truck weight 10,001–33,000 lbs: 0 26.5% 33,001 lbs. or more: 1 73.5%
Vehicle trailering No trailing units: 0 32.0% One or more trailing units: 1 68.0%
Vehicle age 5 years 43.7% >5 and  10 years 30.8%
>10 years 25.5%
Driver-related variables
Gender Male: 1 97.1% Female: 0 2.9%
Age 25 years: 1, others: 0 5.7% 25–65 years: 1, others: 0 82.0%
65 years: 1, others: 0 12.3%
License class compliance Valid: 1 96.5% Not valid: 0 3.5%
History of crashes Yes: 1 14.5% No: 0 85.7%
History of convictions Suspensions: 1, others: 0 7.4% Speeding: 1, others: 0 23.6%
DWI: 1, others: 0 0.6% Other convictions: 1, others: 0 23.3%
Driving behaviors Drowsy or fatigued: 1, others: 0 0.9% Under the influence of drugs:1 1.2%
Drinking driving yes: 1, no: 0 2.4% Careless: 1, others: 0 4.3%
Failure to keep in proper lane: 1, others: 0 6.2% Shoulder and lap belt used: 1, no:0 83.5%

Entropy varies between 0 and 1, and values over 0.90 denote a 2 and cluster 3 represent 32.8%, 35.6%, and 31.6% of the dataset,
clear cluster differentiation, as well as the interpretability of the respectively.
clusters. In addition, the value of silhouette coefficient ranges The truck drivers in cluster 1 were characterized by a high pro-
between 1 and 1, and the values are closer to 1 indicating better portion of male (99.8%), middle-aged (71%), and older (20.1%) dri-
clustering results. In the dataset, the silhouette coefficient and vers. High proportion of drivers with crash (26.8%), speeding
entropy criterion were 0.45 and 0.97, respectively, indicating a (41.9%), and conviction (41.6%) histories were also observed. How-
good separation among the clusters. Furthermore, discriminant ever, drivers in this cluster exhibited satisfactory driving behaviors,
analysis was also used to test the stability of the clustering results and no traffic violations, such as drunk driving, driving under the
(Abenoza et al., 2017), and the group membership results of cluster influence of drugs, or careless driving. In addition, the proportion
analysis indicated that 94.5% of the cases were correctly classified, of level 3 crash severity was lower than in cluster 2 (3.8%). Accord-
indicating stable clustering results within three groups. ing to these characteristics, the first cluster was called ‘‘middle-
After estimation of the parameters, the posterior probabilities aged and elderly drivers with low risk of driving violations and
of belonging to specific clusters to each case are calculated high historical crash records.”
according to the underlying statistical model in Latent Gold (de The truck drivers in cluster 2 exhibited satisfactory driving
Oña, López, Mujalli, & Calvo, 2013), and the case was assigned behaviors and no traffic violations, such as drunk driving, driving
to the cluster in which the probability was highest. As shown under the influence of drugs, or careless driving. In addition, dri-
in Table 3, each cluster corresponded to an underlying group of vers in this cluster were characterized by a higher proportion of
truck drivers characterized by a particular pattern of demograph- female (8%) and middle-aged (80%) drivers and a lower proportion
ics, conviction records, and violation behaviors. Cluster 1, cluster of having a valid license (90.3%) and using a seat belt (54%) than

157
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

Fig. 1. Cumulative frequency distributions of equivalent fatalities.

Table 2
Indicators of fit of the models with two to five cluster solutions.

Number of clusters L2 df P-value BIC AIC


2 8,548.58 13,971.00 <0.001 112,511.43 102,511.89
3 4,134.47 13,953.00 0.09 109,269.17 100,269.99
4 4,817.81 13,935.00 0.12 109,424.36 103,424.78
5 5,078.73 13,917.00 0.11 109,757.13 105,757.51

Table 3
Results of the driver classification (N = 15,506).

Variable Categories Overall Sample Cluster 1 Cluster 2 Cluster 3


(N = 15,506) N = 5,086 N = 5,516 N = 4,904
Gender Male 97.1% 99.8% 92.0% 100.0%
Female 2.9% 0.1% 8.0% 0.0%
Age Young driver 5.7% 8.9% 7.8% 0.0%
Middle-aged driver 85.6% 71.0% 80.0% 97.8%
Older driver 11.6% 20.1% 12.1% 2.2%
License Valid = Yes 96.5% 99.9% 90.3% 100.0%
Restraint system Belt used = Yes 83.5% 100.0% 54.0% 100.0%
History of crashes Crashes = Yes 14.5% 26.8% 16.0% 0.0%
History of convictions Suspensions = Yes 7.4% 0.0% 20.7% 0.0%
Speeding = Yes 23.6% 41.9% 27.7% 0.0%
Driving while intoxicated = Yes 0.6% 0.0% 1.8% 0.0%
Others = Yes 23.3% 41.6% 27.1% 0.0%
Driver drowsy or fatigued Drowsy = Yes 1.0% 0.3% 2.3% 0.0%
Influence of drugs Drug = Yes 1.2% 0.0% 3.4% 0.0%
Driver careless/inattentive Careless = Yes 4.3% 0.0% 12.1% 0.0%
Driver drinking Yes 2.4% 0.0% 6.7% 0.0%
Failure to yield right-of-way Yes 3.4% 0.0% 9.4% 0.0%
Failure to keep in proper lane Yes 6.2% 0.0% 17.3% 0.0%
Crash severity Level 1 86.0% 90.8% 83.0% 94.8%
Level 2 8.5% 5.6% 6.2% 4.5%
Level 3 5.5% 3.8% 10.8% 0.7%

other clusters. Moreover, a higher proportion of crashes (16%), cluster was named ‘‘drivers with high risk of driving violations
driving with a suspended license (20.7%), and drunk driving and high historical crash records.”
(1.8%) was apparent compared with the other groups. In addition, The truck drivers in cluster 3 were characterized by a high
a higher proportion of driving violation behaviors, such as drunk proportion of male (100%) and middle-aged (97.8%) drivers with
driving (6.7%), driving under the influence of drugs (3.4%), careless no history of traffic convictions or crash records. Moreover, these
driving (12%), and failure to drive in the proper lane (17.3%), were drivers were associated with no driving violations in crashes.
observed for this group. The proportion of level 3 crash severity The proportion of level 3 crash severity was the lowest among
was the highest among all groups (10.81%). Therefore, the second all groups (0.7%). Therefore, the third cluster was called

158
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

‘‘middle-aged drivers with no driving violations and conviction distribution, with the null hypothesis is that there is no parameter
records.” difference statistically between the full model and separate model.
The second log likelihood ratio test for transferability of the
4. Methodology coefficient for three separate models is individually compared as
follows:
The partial proportional odds (PPO) model is recommended as     
LRK 1 K 2 ¼ 2 LLK 1 K 2 bK 1 K 2  LLK 1 bK 1 ð10Þ
an effective alternative model for ordinal response data (Kaplan
 K1 K2

& Prato, 2012; Peterson & Harrell, 1990; Wang & Abdel-Aty, where LLK 1 K 2 b is the log-likelihood at convergence of a model
2008; Williams, 2006). By relaxing the proportional odds assump- using the converged parameters from the k2 model (using k2 data)
tion for some or all explanatory variables (Kaplan & Prato, 2012;  
while using the data for the k1 driver group; LLK 1 bK 1 is the log-
Mooradian, Ivan, Ravishanker, & Hu, 2013; Sasidharan & likelihood at convergence of the model using k1 driver group data.
Menéndez, 2014; Wang & Abdel-Aty, 2008), the PPO model can
obtain more correct and complete results than the ordered logit
5. Estimation results
(OL) model (Kaplan & Prato, 2012). For a more parsimonious and
easy-to-understand layout (Williams, 2006), a gamma parameter-
5.1. Model comparison
ization of the PPO model can be specified in terms of the probabil-
ity of the j-th injury severity for a given crash i as (Peterson &
In this study, PPO model was fitted using a user-written pro-
Harrell, 1990):
gram in Stata 15 (Williams, 2006). Three indicators including
0 0
exp½aj þ ðX i b þ T i cj Þ McFadden pseudo R-squared, the Akaike’s information criterion
PðY i > jÞ ¼ 0 0 where j (AIC), and the Bayesian information criterion (BIC) were employed
1 þ exp½aj þ ðX i b þ T i cj Þ
to measure the goodness-of-fit of the PPO and OL model. The
¼ 1; 2;    ; J  1 ð6Þ model comparison measures are shown in Table 4. The results
showed that higher pseudo R-squared and smaller AIC and BIC val-
where Yi denotes the discrete crash severity of a truck occupant in a
ues were obtained using the PPO model in each driver subgroup
crash i; Xi is a p1 vector of explanatory variables; b is a p  1 vec-
model, indicating that PPO had better performance than the OL
tor of estimated parameters, which is the same for all values of
model. As such, the PPO model was adopted and the estimation
crash severity; T i is a q1 vector of explanatory variables that vio-
results analyzed.
lates the proportional odds assumption (q  p); cj is a q  1 vector
of estimated parameters associated with a subset of variables T i ;
5.2. Model transferability
and aj is a cut-off point for the j-th cumulative logit. In this model,
each independent variable has one beta coefficient and J – 2 gamma Then, two different likelihood ratio tests were used to check the
coefficients. statistical justification of each driver subgroup model. First,
To obtain a tangible understanding of the strength and impor- LRall =218.62, which is greater than the critical value of the Chi-
tance of the effect of explanatory variable on crash severities, the square statistic with 50 degrees of freedom at the 99%
marginal effect was adopted to accurately and tangibly understand confidence-related variables (76.15), was utilized. Therefore, we
the probability change in crash severities when certain explanatory
achieved 99% confidence not supporting the null hypothesis, and
variables were increased or decreased by 1% (Train, 2003). Xie et al. the three separate driver subgroup models showed statistically dif-
(2012), Anarkooli and Hosseinlou (2015), and Uddin and Huynh
ferent model parameters than the overall model. Next, the values
(2017) recommended that the direct-marginal effect and cross- of LRK 1 K 2 for different driver group models were calculated, and
marginal effect of explanatory variables can be calculated by par-
the results of the second transferability are shown in Table 5. Each
tial derivatives of various explanatory variable in the crash severity
test statistics are greater than the critical value of the Chi-square
function, formulated as in Eqs. (7)–(8).
statistic, with corresponding degrees of freedom at the 99%
P
Direct EXijijk ¼ @Pij =@X ijk ¼ bk Pij ð1  Pij Þ ð7Þ confidence-related variables. Therefore, each driver group model
was statistically justified.
P
Cross EXim
ijk
¼ @Pim =@X ijk ¼ bk Pim P ij ð8Þ 5.3. Model results

where X ijk is the independent variable k associated with the j-th


The estimated results of the gamma parameterization of PPO
P
crash severity for crash i;Direct EXijijk represents the direct-marginal models for all three driver subgroups were obtained and are pre-
effect, which means the impact of changes in variable X ijk on Pij to sented in Table 6. From the results, significant factors contributing
be in j-th crash severity; and Cross EXim
P
represents the cross- to fatal crash severities for each driver subgroup were identified by
ijk
a statistical significance test. Then, the marginal effects (ME) of
marginal effect, which describes the impact of changes in the vari-
various explanatory variables were also evaluated and are shown
able X ijk of m-th crash severity (m – j) on the P im for crash i.
in Table 7, and the sign and magnitude of the marginal effects on
Two kinds of likelihood ratio tests are used to check the statis-
fatal crash severities were analyzed. As shown in Table 6, each
tical justification of the three developed models for different driver
explanatory variable had one beta coefficient and one gamma coef-
groups. Specifically, the first log likelihood ratio test for the trans-
ficient for the variable violating the parallel line assumption. Addi-
ferability of the coefficient from the full model to three separate
tionally, assumptions were being violated in variables such as
models is as follows:
atmospheric condition, surface type, and roadway alignment,
h XK  i
which had significant beta coefficients.
LRall ¼ 2 LLðball Þ  k¼1
LLk bk ð9Þ

where LLðball Þ is the log-likelihood at the convergence of the overall (1) Collision-related variables
 
model; LLk bk is the log-likelihood at convergence of the k-th dri- Front-to-front (coefficient = 1.01, 1.62, and 1.05; P-value <0.01,
ver group (k = 3); and LRall is a statistical value that follows thev2 0.01, and 0.01) and angled-front-to-side (0.23, 0.34, and 0.27;
159
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

Table 4
Fitness results of the ordered logit model and partial proportional odds model.

Fitness parameters Cluster 1 Cluster 2 Cluster 3


PPO OL PPO OL PPO OL
Log likelihood at zero 3,644.00 3,445.00 5,282.00 5,217.00 3,468.00 277.00
Log likelihood at convergence 3,319.00 3,254.00 3,963.00 4,107.00 3,208.00 2,634.00
Pseudo R-squared 0.15 0.10 0.25 0.20 0.12 0.08
AIC 6,716.00 6,873.00 8,043.00 8,291.00 6,376.00 6,315.00
BIC 6,967.00 6,979.00 8,422.00 8,538.00 6,438.00 6,462.00

Table 5 severe crash severities than under dark-not-lighted condition (not


Results of the second transferability. significant (NS), 0.2%, and 0.7%).
k1 k2 Cluster 1 Cluster 2 Cluster 3
Cluster 1 0.00 1,900.00 (df = 39) 1,090.00 (df = 22)
(3) Road-related variables
Cluster 2 1,770.00 (df = 31) 0.00 1,240.00 (df = 22)
Cluster 3 1,298.00 (df = 31) 1,420.00 (df = 39) 0.00 The occurrence of a truck-involved in fatal crashes (level 3) in
Critical v2(P=0.01, v2(P=0.01, v2(P=0.01, curved (2.4% and 4.8%), graded (1.0% and NS), non-blacktop surface
value = 52.19 = 62.43 = 40.29
df=31) df=39) df=22)
(NS and 2.0%), or two-lane (0.2% and 1.8%) roadways showed an
increased likelihood of greater crash severities in driver groups 1
and 2; there were no significances in comparisons of group 3. This
P < 0.01, 0.01, and 0.01) collision manners were significantly posi- result may indicate that drivers in group 1 and group 2 cannot
tively associated with increased crash severities in all three driver effectively handle complicated roadway conditions. When the sur-
groups. Compared with other collision manners, front-to-front face condition was wet or icy compared with dry condition, the
could increase the likelihood mostly of level 2 (ME = 17.4%, likelihood of being involved in a level 3 crash severity increased
16.7%, and 9.2%) and level 3 (2.5%, 4.2%, and 1.9%) crash severity. significantly in group 1 (1.5%) and group 2 (2.2%), whereas drivers
Therefore, this collision manner could be considered the most dan- in group 3 (0.23, P < 0.05) were associated with decreased likeli-
gerous and merit greater attention. In addition, multi-vehicle hoods of severe crash severities.
involved crashes (2.89, 3.15, and 2.98; P < 0.05, 0.01, and 0.01) In addition, lower speed limits were associated with the mitiga-
were positively associated with severe crash severities (level 2 tion of fatal crash severities. Compared with a speed limit greater
and level 3) in all driver groups. Notably, compared with the dri- than 70 miles per hour, a speed limit less than 40 miles per hour
vers in group 1 and group 3 with fewer driving violation behavior, (13.8%, 14.7%, and 1.2%) and 45–55 miles per hour (2.3%,
drivers in group 2 showed increased likelihoods of level 2 (23.2% 5.7%, and 0.7%) could decrease the likelihood of a level 3 crash
vs. 18.8% and 20.1%) and level 3 (9.5% vs. 7.4% and 8.5%) crash severity significantly in all three driver groups. Furthermore, a
severities. higher speed limit of 60–70 miles per hour was associated with
an increased crash severity in groups 1 and 2.
(2) Environment-related variables
(4) Vehicle-related variables
Both adverse weather conditions such as rain and snow (0.28,
0.84, and 0.36; P < 0.05, 0.05, and 0.01) and rural area (0.17, 0.58, Truck tractors (0.6%, 3.0%, and 0.9%) with a heavy weight (2.2%,
and 0.31; P < 0.05, 0.01, and 0.01) were significantly associated 7.4%, and 2.5%) and trailing pull units (1.9%, 1.8%, and 1.7%) all
with increased fatal crash severities in all the three driver groups. increased the likelihood of being involved in a level 3 crash sever-
In comparison to the urban area, rural areas showed an increased ity in all three driver groups compared with other trucks. Notably,
likelihood of level 2 (5.1%, 4.6%, and 4.5%) and level 3 (3.1%, 3.2%, when the drivers in group 2 drove trailer trucks with heavier
and 2.4%) crash severities in all three driver groups. Notably, larger weights, the probability that they would be involved in severe
marginal effects of land use (3.2% vs. 3.1% and 2.4%) and adverse crash increases was even greater than in the other two driver
weather condition (3.4% vs. 3.2% and 2.8%) could be observed in groups.
driver group 2 compared with other two driver groups for level 3 Relative to a vehicle age of greater than 10 years, trucks with a
crash severity. vehicle age of less than 5 years were less likely to be involved in a
The season was also significantly associated with fatal crash severe crash in driver groups 1 and 3. When the vehicle age was
severities in all three driver groups. Compared with spring, sum- between 5 and 10 years, trucks (0.16, P < 0.05) were significantly
mer was associated with a slightly decreased likelihood of a level associated with decreased crash severities only in group 3.
3 crash severity (0.4%, 1.0%, and 0.7%) in all three driver
groups. In contrast, autumn (0.12, 0.27, and 0.10; P < 0.05, 0.05, (5) Driver-related variables
and 0.05) and winter (0.13, 0.31, and 0.16; P < 0.05, 0.05, and
0.05) were positively associated with crash severities. It should For driver-related variables, older drivers were associated with
be noted that winter (3.2% vs. 2.3%; 4.9% vs. 3.0%; and 3.9% vs. an increased likelihood of level 3 (1.5% and 7.7%) crash severity in
1.7%) was related to greater marginal effects of level 2 crash sever- driver groups 1 and 2. This finding may be caused by a poor health
ity than autumn. status and limited perception abilities among elderly drivers.
The day of week, light conditions, and time period were also sig- Moreover, compared with female truck drivers, male drivers
nificantly associated with fatal crash severities in all driver groups. (0.26, P < 0.01) had a significantly negative effect on crash sever-
With respect to crashes that occurred under day-light conditions, ity in driver group 2. It should be noted that female truck drivers
decreases in the likelihood of severe crash severities (level 2 and are few in number and that the role of gender in crashes requires
level 3) were observed under both dark-light and dark-not- further study.
lighted conditions. Particularly, driving under dark-lighted condi- A history of license suspension and other convictions in driver
tion (0.7%, 2.4%, and 2.4%) seemed to be more likely to reduce group 2 were significantly associated with decreased likelihoods

160
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

Table 6
Estimation results of gamma parameterization of PPO model in the three driver groups.

Variables Categories Cluster 1 Cluster 2 Cluster 3


Coefficient Coefficient Coefficient
Beta
Collision-related variables
Manner of collision Front-to-rear = yes: 1 0.07** 0.17*** 0.05**
Front-to-front = yes: 1 1.01*** 1.62*** 1.07***
Angle-front-to-side, right angle = yes: 1 0.23*** 0.34*** 0.27***
Number of vehicles involved Multi vehicle = 1 2.89** 3.15*** 2.98***
Environment-related variables
Atmospheric condition Rain, cloudy, snow etc. = 1 0.28** 0.84** 0.36***
Land use Rural = 1 0.17** 0.58*** 0.31***
Day of week Weekday = 1 0.31*** 0.38*** 0.31**
Light condition Dark-not-lighted = 1 – 0.12** 0.26**
Dark-lighted = 1 0.10** 0.24** 0.24**
Time period Afternoon = 1 – – –
Evening = 1 0.65** 1.22** 0.56**
Night = 1 0.54** 1.03** 0.46**
Season Summer = 1 0.05** 0.10** 0.06**
Autumn = 1 0.12** 0.27** 0.10**
Winter = 1 0.13** 0.31** 0.16**
Road-related variables
Roadway alignment Curve = 1 0.12** 0.36*** –
Roadway grade Grade, uphill, downhill etc. = 1 0.14** – –
Surface condition Wet, ice, snow etc. = 1 0.21** 0.17** 0.23**
Surface type Blacktop = 1 – 0.19** –
Speed limit 40 mile/h = 1 1.11** 1.49*** 0.96***
45–55 mile/h = 1 0.33** 0.46*** 0.66**
60–70 mile/h = 1 0.14** 0.19** –
Number of lanes per direction 2 lanes = 1 0.21*** 0.20*** –
Vehicle-related variables
Body type Truck-tractor = 1 0.08** 0.22*** 0.14**
Truck weight 33,001 lbs. or more = 1 0.19*** 0.11*** 0.17**
Vehicle trailering One or more trailing units = 1 0.26** 0.12** 0.25**
Vehicle age 5 years = 1 0.04** – 0.11**
>5 and  10 years = 1 – – 0.16**
Driver-related variables
Gender Male = 1 – 0.26** –
Age 25 years = 1 – – –
65 years = 1 0.20*** 0.40*** –
License class compliance Valid = yes: 1 – – –
Restraint system Shoulder and lap belt not used = 1 – 1.20*** –
Driver drinking Yes: 1 – 0.64*** –
History of crashes Yes: 1 0.07** 0.05** –
History of convictions Suspensions = yes = 1 – 0.21** –
Speeding = yes: 1 0.14** – –
Driving while intoxicated = yes: 1 – – –
Other convictions = yes: 1 0.09** 0.13** –
Driver behavior Drowsy or fatigued = yes: 1 – 0.85*** –
Under the influence of drugs = yes: 1 – 0.59*** –
Careless/inattentive = yes: 1 – 0.74*** –
Failure to keep in proper lane = yes: 1 – 1.03*** –

Gamma
Atmospheric condition Rain, cloudy, snow etc. = 1 0.72*** 0.18*** –
Number of vehicles involved Single vehicle = 1 2.25*** – 2.11***
Manner of collision Front-to-front = yes: 1 0.66 *** 0.49*** 0.45***
Angle - front-to-side, right angle = yes: 1 0.40 *** 0.51*** 0.44***
History of crashes Yes: 1 – 0.22*** –
Surface type Blacktop = 1 0.28*** – –
Roadway alignment Curve = 1 0.36*** – –

Alpha
Constant 1 0.50*** 0.28** 0.89**
Constant 2 2.35*** 2.37** 2.58**
** ***
Note: indicates values that are significant at the 0.05 level; indicates values that are significant at the 0.01 level; – indicates values that are not significant at the 0.05
level.

of level 3 (0.3% and 1.1%) crash severity. It should be noted that crashes (3.0%) in driver group 2 were associated with an increased
histories of speeding (3.8%) and crashes (0.5%) in driver group 1 likelihood of level 3 crash severity, which might be due to poor
were associated with a decreased likelihood of level 3 crash sever- driving habits.
ity, potentially because these drivers were cautious due to a fear of For driving violation behaviors in group 2, drunk driving (0.64,
repeat punishment. However, histories of speeding (0.7%) and P < 0.01), fatigue (0.85, P < 0.01), driving under the influence of

161
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

Table 7
The margin effects of fatal crash severities in the three driver groups.

Variable Cluster 1 Cluster 2 Cluster 3


Level1 (%) Level2 (%) Level3 (%) Level1 (%) Level2 (%) Level3 (%) Level1 (%) Level2 (%) Level3 (%)
Collision-related variables
Collision manner: Front-to-rear = Yes:1 0.6 0.4 0.2 1.9 1.4 0.5 1.6 2.0 3.6
Collision manner: Front-to-front = Yes:1 19.8 17.4 2.5 20.9 16.7 4.2 11.1 9.2 1.9
Collision manner: Angle-front-to-side = Yes:1 4.5 2.9 1.6 6.2 8.3 2.1 5.3 3.5 1.8
Multi vehicle involved = Yes: 1 26.2 18.8 7.4 32.7 23.2 9.5 28.6 20.1 8.5
Environment-related variables
Adverse weather: rain, cloudy, snow, etc. = 1 8.6 5.4 3.2 8.0 4.6 3.4 4.5 1.7 2.8
Land use: rural = 1 8.2 5.1 3.1 7.8 4.6 3.2 6.9 4.5 2.4
Day of week: Weekday = 1 6.1 3.9 2.2 3.9 1.8 2.1 1.7 1.1 0.6
Light condition: Dark-not-lighted = 1 – – – 3.7 3.5 0.2 2.0 1.3 0.7
Light condition: Dark-lighted = 1 2.0 1.3 0.7 3.3 0.9 2.4 3.3 0.9 2.4
Time period: Evening = 1 7.4 5.5 1.9 9.8 5.6 4.2 6.5 4.7 1.8
Time period: Night = 1 6.2 4.3 1.9 7.2 4.8 2.4 6.0 4.0 2.0
Season: Summer = 1 1.2 0.8 0.4 2.7 1.7 1.0 1.6 0.9 0.7
Season: Autumn = 1 4.9 2.7 2.2 6.8 3.8 3.0 4.2 2.9 1.3
Season: Winter = 1 5.5 3.2 2.3 7.9 4.9 3.0 5.6 3.9 1.7
Road-related variables
Roadway alignment: curve = 1 3.4 1.0 2.4 6.6 1.8 4.8 – – –
Roadway: grade etc. = 1 2.7 1.7 1.0 – – – – – –
Surface condition: wet, ice, etc. = 1 4.2 2.7 1.5 3.1 0.9 2.2 7.3 5.1 2.2
Surface type: Blacktop = 1 – – – 4.7 2.7 2.0 – – –
Speed limit  40 mile/h = 1 21.7 7.9 13.8 26.1 11.4 14.7 3.6 2.4 1.2
Speed limit 45–55 mile/h = 1 6.4 4.1 2.3 7.9 2.2 5.7 8.8 9.5 0.7
Speed limit 60–70 mile/h = 1 2.8 4.9 2.1 3.1 0.9 2.2 – – –
Number of lanes per direction  2 lanes = 1 4.7 4.5 0.2 2.5 0.7 1.8 – – –
Vehicle-related variables
Body type: Truck-tractor = 1 1.6 1.0 0.6 4.2 1.2 3.0 2.7 1.8 0.9
Truck weight  33,001 lbs. = 1 4.6 2.4 2.2 10.3 2.9 7.4 7.4 4.9 2.5
One or more trailing units = 1 5.1 3.2 1.9 4.5 2.7 1.8 5.1 3.4 1.7
Vehicle age  5 years = 1 0.7 0.5 0.2 – – – 2.0 1.3 0.7
Vehicle age > 5 and  10 years = 1 – – – – – – 3.1 2.1 1.0
Driver-related variables
Gender Male = 1 – – – 5.2 6.5 1.3 – – –
Driver age  25 years = 1 3.9 0.8 3.1 1.8 0.5 1.3 – – –
Driver age  65 years = 1 4.0 2.5 1.5 7.6 0.1 7.7 – – –
License class compliance: Valid = 1 – – – 3.6 1.0 2.6 – – –
Restraint system: Shoulder belt not used = 1 – – – 12.8 10.1 2.7 – – –
Driver drinking: Yes = 1 – – – 11.6 8.4 3.2 – – –
History of crashes: Yes = 1 1.3 0.8 0.5 0.2 3.2 3.0 – – –
History of suspensions: Yes = 1 – – – 4.2 4.5 0.3 – – –
History of speeding: Yes = 1 2.7 1.1 3.8 2.3 3.0 0.7 – – –
History of driving while intoxicated: Yes = 1 – – – 3.3 0.9 2.4 – – –
History of other convictions: Yes = 1 2.0 1.3 0.7 1.5 0.4 1.1 – – –
Drowsy or fatigued: Yes = 1 – – – 9.2 7.1 2.1 – – –
Under the influence of drugs: Yes = 1 – – – 16.0 12.1 3.9 – – –
Careless/inattentive: Yes = 1 – – – 14.6 10.6 4.0 – – –
Failure to keep in proper lane: Yes = 1 – – – 19.4 16.8 2.6 – – –

Note: – indicates values that are not significant at the 0.05 level.

drugs (0.59, P < 0.01), careless driving (0.74, P < 0.01), and failure to Shinar, 2016). The clustering results classify truck drivers into
drive in the proper lane (1.03, P < 0.01) all were positively associ- three groups: (i) middle-aged and elderly drivers with low risk of
ated with severe crash severities (level 2 and level 3). Among them, driving violations and high historical crash records; (ii) drivers
drunk driving (8.4% and 3.2%), fatigue (7.1% and 2.1%), and careless with high risk of driving violations and high historical crash
driving (10.6% and 4.0%) all significantly increased the likelihood of records; and (iii) middle-aged drivers with no driving violations
experiencing level 2 and level 3 crash severities. In addition, failure and conviction records. The main differences among the three
to drive in the proper lane (16.8%) could mostly increase the like- groups were conviction records and driving violation behaviors.
lihood of being involved in a level 2 crash severity. This finding Then, the results of the PPO models for the three driver groups
may be related to the poor flexibility and heavy weights of trucks. were used to determine how risk factors make drivers more or less
Thus, it is necessary to focus on lane planning considering truck- likely to be involved in severe accidents.
related characteristics. The estimated results indicate that the risk factors, as well as
their impacts on different driver groups, are different. First, more
6. Discussion and conclusions driver-related risk factors, such as driving under the influence of
alcohol or drugs, fatigue, or carelessness, are significantly associ-
Latent class clustering was adopted based on drivers’ demo- ated with the high-risk group (group 2) only. Moreover, the prob-
graphic attributes, conviction records, and driving violation behav- ability of level 3 crash severity was highest in group 2. Thus,
iors, which were found in previous studies to influence the crash drivers in group 2 were not sufficiently responsible for themselves
severity (Li et al., 2005; Oppenheim, Oron-Gilad, Parmet, & and other occupants, and they were more likely to make mistakes

162
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

and cause more severe crashes. Therefore, it is recommended that and elderly drivers, which could be more helpful to identify risky
safety managers and planners pay more attention to historical traf- drivers.
fic violations and crash records of truck drivers, and carry out mea- Furthermore, risky driving behavior, such as driving under the
sures specific to different driver groups. Additionally, some safety influence of alcohol or drugs, fatigue, or carelessness, were signif-
managing recommendations are proposed as follows: (a) use of icantly associated with severe crash severities in group 2, which
an advanced driver information system (ADIS); (b) more severe is consistent with previous studies (Abegaz, Berhane, Worku,
penalties imposed on drivers who have poor traffic violation Assrat, & Assefa, 2014; Bédard et al., 2002; Behnood et al., 2014;
records; (c) use of an advanced driving assistance system (ADSS). Sullman et al., 2016). It is possible that drivers in group 2 had
Second, fewer roadway- and environment-related risk factors developed risky driving habits, and it was difficult to correct their
and minor marginal effects were identified in the low-risk groups driving behavior by punishment or warnings via traffic violations
(groups 1 and 3). Especially in group 3, several risk factors, such alone. Additionally, comprehensive countermeasures involving
as roadway alignment, number of lanes, driving violation behav- psychology, safety training, and advanced technological equipment
iors, and conviction records, were not significantly related to the are recommended for group 2.
drivers in this groups despite being significant for the other two Both older and female drivers showed an association with an
groups. This result may indicate that drivers in group 3 had devel- increased likelihood of driving violations and severe crash severi-
oped good driving habits and performed better in complex condi- ties in group 2. Oppenheim et al. (2016) also found that drivers’
tions than others. Therefore, it is recommended that safety specific violations could be traced to personal characteristics, such
managers provide some rewards to truck drivers in group 3, such as demographics and gender roles, and the health status and driv-
as awarding ‘‘excellent truck drivers” to affirm their good driving ing performance of drivers should be regularly checked. Further-
habits. Moreover, the driving habits and experience of drivers in more, license validity, license suspension, and other traffic
group 3 could be collected by interview and inquiry, which could violation records were related to the occurrence of crashes. Gene-
be used as an important reference for teaching and training novice viève et al. (2002) also found that previous crashes or convictions
truck drivers. were better predictors of subsequent crashes. Therefore, these
Third, the highest proportions of level 3 crash severities records could be used as an important basis for driving risk
(10.81%) and more risk factors and larger marginal effects were assessment.
found in group 2. In contrast, drivers in group 3 had no history of The results showed that risk factors such as adverse weather
crashes, no convictions, and no traffic violation behaviors, with conditions, winter, rural areas, and curved road segments were
the lowest proportion of level 3 crash severity. Additionally, fewer associated with an increased likelihood of severe crash severities
risk factors and lower marginal effects were identified in group 3. for all three driver groups. Additionally, Lemp et al. (2011), Dong
These results may indicate that drivers in group 2 are ‘‘the riskiest et al. (2015), Pahukula, Hernandez, and Unnikrishnan (2015),
ones,” drivers in group 3 are ‘‘the safest ones,” and drivers in group Wei et al. (2016), and Ouyang, Li, Xu, Wu, and Ke (2020) also found
1 are in the middle. Therefore, an evaluation report of driving per- that adverse weather conditions such as rain or snow events in
formance for each truck driver according to the historical traffic rural areas could increase casualties and property damage in
violation records and crash records of truck drivers should be truck-involved crashes. Hence, some policy recommendations can
established, which could be used to identify who is a risky or a safe be proposed based on these results that may be beneficial for mit-
driver. igating truck-involved fatal crash severities: (a) determine flexible
Both higher crash involvement rates and high conviction working timetables for truck drivers under different weather con-
records should be noted in groups 1 and 2, but almost no driving ditions and area types; (b) ascertain a rational design of alignments
violation behaviors, such as drinking or fatigue driving, were found and right-of-way for trucks; (c) use an advanced driving assistant
in group 1. The proportion of severe crash severities was lower in system, and assistance with driving in adverse conditions through
group 1 than group 2. It is speculated that punishment of crash timely GPS data and in-vehicle data records.
involvement or violation records has played a good educational The model results also indicated that truck tractors, trucks with
or warning role for the drivers in group 1. Similar to the results heavy weights, and using one or more trailing units were all
of Chandraratna, Stamatiadis, and Stromberg (2006), some previ- related to an increased likelihood of severe crash severities in the
ous records (history of license suspensions and history of driving three driver groups. Similarly, Zhu and Srinivasan (2011) and
while intoxicated) were associated with the alleviation of crash Islam et al. (2014) indicated that the overall injury severity in a
severity for high risk drivers (group 2), while history of crashes crash is significantly worse for heavy trucks and tractors pulling
and of speeding were associated with lowering the risk of severe van/box trailers. Notably, new vehicles were associated with a
crashes in group 1. Hence, it is suggested that an advanced driver decreased likelihood of severe crashes in group 3, potentially due
information system should be used to identify and record traffic to better braking performance. Similar results were obtained by
violations of truck drivers over time, correcting the driving behav- Lemp et al. (2011) and Islam et al. (2014). Therefore, some safety
iors of truck drivers by supervising or punishing drivers. designs such as truck body types and reasonable weights are rec-
In addition, the highest proportions of young (8.9%) and older ommended to truck manufacturers, with greater attention paid
(20.1%) drivers are noteworthy in group 1. Knight (2004) indicated to trucks with a vehicle age greater than 10 years.
that the likelihoods of fatal crash involvement are higher for young Among the collision-related variables, the results showed that
drivers with a history of crash and other traffic convictions. How- collision manners and the number of vehicles involved in a crash
ever, certain proportions of young and older drivers were capable were significant in all three driver groups. Notably, front-to-front
of learning from previous crashes or speed violation records. More- collisions were positively related to the crash severity for all driver
over, some elderly drivers with poor traffic violation and crash groups and could be considered the most dangerous collision man-
records were also found in group 2. However, they had a high risk ner. It is speculated that this finding may be linked to the heavy
of driving violations in the next crash, and they could be regarded weights of trucks and intense collisions in front-to-front crashes.
as dangerous drivers. The driving performance of elderly drivers in Additionally, multi-vehicle crashes were linked to an increased
these two groups was quite different in accidents. Therefore, likelihood of greater crash severities in all driver groups. Chen
besides demographics such as age and gender, managers should and Chen (2011), Xie et al. (2012), Islam et al. (2014) and
also focus on driving violation records and crash records of young Michalaki, Quddus, Pitfield, and Huetson (2015) also observed an

163
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

increase in serious driver injuries in multi-vehicle crashes in com- Chen, F., & Chen, S. (2011). Injury severities of truck drivers in single- and multi-
vehicle crashes on rural highways. Accident Analysis and Prevention, 43(5),
parison to other cases, potentially due to the involvement of more
1677–1688.
occupants in the crashes. Therefore, crashworthiness standards of Das, S., Sun, X., Wang, F., & Leboeuf, C. (2015). Estimating likelihood of future
trucks should be rationally ascertained considering different colli- crashes for crash-prone drivers. Journal of Traffic and Transportation Engineering
sion types and the number of vehicles involved. (English Edition), 2(3), 145–157.
de Oña, J., López, G., Mujalli, R., & Calvo, F. J. (2013). Analysis of traffic accidents on
The main contribution of this paper is the latent class clustering rural highways using Latent Class Clustering and Bayesian Networks. Accident
analysis of truck drivers based on their multiple attributes, and the Analysis & Prevention, 51, 1–10.
results can be used to reveal the heterogeneity among drivers with Depaire, B., Wets, G., & Vanhoof, K. (2008). Traffic accident segmentation by means
of latent class clustering. Accident Analysis & Prevention, 40(4), 1257–1266.
respect to the risk factors and crash severity level. In addition, var- Dong, C., Richards, S. H., Huang, B., & Jiang, X. (2015). Identifying the factors
ious risk factors associated with fatal crash severities for different contributing to the severity of truck-involved crashes. International Journal of
driver groups were identified, and these results are helpful to apply Injury Control and Safety Promotion, 22(2), 116–126.
Dong, C., Clarke, D. B., Richards, S. H., & Huang, B. (2014). Differences in passenger
various management strategies for specific driver groups. How- car and large truck involved crash frequencies at urban signalized intersections:
ever, explained variances of the fitting results for three driver An exploratory analysis. Accident Analysis & Prevention, 62, 87–94.
groups are relatively weak level. Possible explanations are missing Dupont, E., Martensen, H., Papadimitriou, E., & Yannis, G. (2010). Risk and protection
factors in fatal accidents. Accident Analysis & Prevention, 42(2), 645–653.
important variables or non-linear relationships between the vari- Evans, A. W. (1994). Evaluating public transport and road safety measures. Accident
ables. For future studies, other driver-related attributes, such as Analysis & Prevention, 26(4), 411–428.
mood and psychological characteristics, can be collected by face- Feng, S., Li, Z., Ci, Y., & Zhang, G. (2016). Risk factors affecting fatal bus accident
severity: Their impact on different types of bus drivers. Accident Analysis &
to-face or telephone interviews and be included in the analysis.
Prevention, 86, 29–39.
And multisource accident data fusion to obtain the latest and more Guo, F., & Fang, Y. (2013). Individual driver risk assessment using naturalistic
objective datasets can be used. Moreover, estimating non-linear driving data. Accident Analysis & Prevention, 61, 3–9.
relationships requires advanced methods such as neural network Guo, Y., Li, Z., Liu, P., & Wu, Y. (2019). Modeling correlation and heterogeneity in
crash rates by collision types using full bayesian random parameters
and other machine learning methods in future studies. Finally, multivariate Tobit model. Accident Analysis & Prevention, 128, 164–174.
heterogeneity across observations was not fully studied, and ran- Häkkänen, H., & Summala, H. (2001). Fatal traffic crashes among trailer truck
dom parameter models, such as a mixed logit model, could be drivers and accident causes as viewed by other truck drivers. Accident Analysis
and Prevention, 33(2), 187–196.
applied to provide new insights into understanding the related risk Hu, S.-R., Li, C.-S., & Lee, C.-K. (2010). Investigation of key factors for accident
factors in later studies. severity at railroad grade crossings by using a logit model. Safety Science, 48(2),
186–194.
Islam, S., Jones, S. L., & Dye, D. (2014). Comprehensive analysis of single- and multi-
Acknowledgments vehicle large truck at-fault crashes on rural and urban roadways in Alabama.
Accident Analysis & Prevention, 67, 148–158.
Kaplan, S., & Prato, C. G. (2012). Risk factors associated with bus accident severity in
The authors of this work are grateful to two anonymous review-
the United States: A generalized ordered logit model. Journal of Safety Research,
ers for their comments on this research paper. This research is 43(3), 171–180.
funded by the National Key R&D Program of China Khorashadi, A., Niemeier, D., Shankar, V., & Mannering, F. (2005). Differences in
(2018YFB1600900), National Natural Science Foundation of China rural and urban driver-injury severities in crashes involving large-trucks: An
exploratory analysis. Accident Analysis and Prevention, 37(5), 910–921.
(No. 51925801, and 71771049), and Postgraduate Research & Kim, J. K., Ulfarsson, G. F., Kim, S., & Shankar, V. N. (2012). Driver-injury severity in
Practice Innovation Program of Jiangsu Province (No. single-vehicle crashes in California: A mixed logit analysis of heterogeneity due
KYCX19_0105). to age and gender. Accident analysis and prevention, 50(1), 1073–1081.
Knight, S. (2004). The fast and the fatal: Street racing fatal crashes in the United
States. Injury Prevention, 10(1), 53–55.
References Lemp, J. D., Kockelman, K. M., & Unnikrishnan, A. (2011). Analysis of large truck
crash severity using heteroskedastic ordered probit models. Accident Analysis &
Prevention, 43(1), 370–380.
Abegaz, T., Berhane, Y., Worku, A., Assrat, A., & Assefa, A. (2014). Effects of excessive
Li, G., Baker, S. P., Qiang, Y., Grabowski, J. G., & McCarthy, M. L. (2005). Driving-
speeding and falling asleep while driving on crash injury severity in Ethiopia: A
while-intoxicated history as a risk marker for general aviation pilots. Accident
partial proportional odds model analysis. Accident Analysis and Prevention, 71,
Analysis & Prevention, 37(1), 179–184.
15–21.
Lombardi, D. A., Horrey, W. J., & Courtney, T. K. (2017). Age-related differences in
Abenoza, R. F., Cats, O., & Susilo, Y. O. (2017). Travel satisfaction with public
fatal intersection crashes in the United States. Accident Analysis & Prevention, 99,
transport: Determinants, user classes, regional disparities and their evolution.
20–29.
Transportation Research Part A: Policy and Practice, 95, 64–84.
Massie, D. L. (1995). Traffic accident involvement rates by driver age and gender.
Alver, Y., Demirel, M. C., & Mutlu, M. M. (2014). Interaction between socio-
Accident Analysis & Prevention, 27(1), 73–87.
demographic characteristics: Traffic rule violations and traffic crash history for
Mercer, G. W. (1987). Influences on passenger vehicle casualty accident frequency
young drivers. Accident Analysis & Prevention, 72, 95–104.
and severity: Unemployment, driver gender, driver age, drinking driving and
Bédard, M., Guyatt, G. H., Stones, M. J., & Hirdes, J. P. (2002). The independent
restraint device use. Accident Analysis & Prevention, 19(3), 231–236.
contribution of driver, crash, and vehicle characteristics to driver fatalities.
Michalaki, P., Quddus, M. A., Pitfield, D., & Huetson, A. (2015). Exploring the factors
Accident Analysis & Prevention, 34(6), 717–727.
affecting motorway accident severity in England using the generalized ordered
Behnood, A., Roshandeh, A. M., & Mannering, F. L. (2014). Latent class analysis of the
logistic regression model. Journal of Safety Research, 55, 89–97.
effects of age, gender, and alcohol consumption on driver-injury severities.
Mooradian, J., Ivan, J. N., Ravishanker, N., & Hu, S. (2013). Analysis of driver and
Analytic Methods in Accident Research, 3-4, 56–91.
passenger crash injury severity using partial proportional odds models. Accident
Blincoe, L., Miller, T. R., Zaloshnja, E., & Lawrence, B. A. (2014). The economic and
Analysis & Prevention, 58, 53–58.
societal impact of motor vehicle crashes, 2010 (Report # DOT HS-812-013).
Morgan, A., & Mannering, F. L. (2011). The effects of road-surface conditions, age,
Washington, DC: National Highway Traffic Safety Administration.
and gender on driver-injury severities. Accident Analysis & Prevention, 43(5),
Brodie, L., Lyndal, B., & Elias, I. J. (2009). Heavy vehicle driver fatalities: Learning’s
1852–1863.
from fatal road crash investigations in Victoria. Accident Analysis & Prevention,
Moustaki, I., & Papageorgiou, I. (2005). Latent class models for mixed variables with
41(3), 557–564.
applications in Archaeometry. Computational Statistics & Data Analysis, 48(3),
Celeux, G., & Soromenho, G. (1996). An entropy criterion for assessing the number
659–675.
of clusters in a mixture model. Journal of Classification, 13(2), 195–212.
Oppenheim, I., Oron-Gilad, T., Parmet, Y., & Shinar, D. (2016). Can traffic violations be
Chang, L., & Mannering, F. (1999). Analysis of injury severity and vehicle occupancy
traced to gender-role, sensation seeking, demographics and driving exposure?
in truck- and non-truck-involved crashes. Accident Analysis and Prevention, 31
Transportation Research Part F: Traffic Psychology and Behaviour, 43, 387–395.
(5), 579–592.
Osman, M., Paleti, R., Mishra, S., & Golias, M. M. (2016). Analysis of injury severity of
Chang, L. Y., & Chien, J. T. (2013). Analysis of driver injury severity in truck-involved
large truck crashes in work zones. Accident Analysis & Prevention, 97, 261–273.
crashes using a non-parametric classification tree model. Safety Science, 51(1),
Ouyang, P., Li, X., Xu, C., Wu, J., & Ke, Z. (2020). Safety effects of road pavement
17–22.
resurfacing: a case study of city-wide scale projects in China. Journal of
Chandraratna, S., Stamatiadis, N., & Stromberg, A. (2006). Crash involvement of
Transportation Safety & Security, 1–20.
drivers with multiple crashes. Accident Analysis & Prevention, 38(3), 532–541.
Pahukula, J., Hernandez, S., & Unnikrishnan, A. (2015). A time of day analysis of
Chang, H.-L., & Yeh, T.-H. (2007). Motorcyclist accident involvement by age, gender,
crashes involving large trucks in urban areas. Accident Analysis & Prevention, 75,
and risky behaviors in Taipei, Taiwan. Transportation Research Part F: Traffic
155–163.
Psychology and Behaviour, 10(2), 109–122.

164
Y. Yuan, M. Yang, Y. Guo et al. Journal of Safety Research 76 (2021) 154–165

Pai, C.-W., & Saleh, W. (2008). Modelling motorcyclist injury severity by various Zhou, H. B., & Gao, J. T. (2014). Automatic method for determining cluster number
crash types at T-junctions in the UK. Safety Science, 46(8), 1234–1247. based on silhouette coefficient. Advanced Materials Research, 951, 227–230.
Peterson, B., & Harrell, J. (1990). Partial proportional odds models for ordinal Zhu, X., & Srinivasan, S. (2011). A comprehensive analysis of factors influencing the
response variables. Journal of the Royal Statistical Society. Series C (Applied injury severity of large-truck crashes. Accident Analysis & Prevention, 43(1),
Statistics), 39(2), 205–217. 49–57. https://doi.org/10.1016/j.aap.2010.07.007.
Sasidharan, L., & Menéndez, M. (2014). Partial proportional odds model—An
alternate choice for analyzing pedestrian crash injury severities. Accident Yalong Yuan is a PhD candidate in the Jiangsu Key Laboratory of Urban ITS at
Analysis & Prevention, 72, 330–340. Southeast University, Nanjing, China. He graduated with a M.S. and B.S. in trans-
Slovic, P., Lichtenstein, S., & Fischhoff, B. (1984). Modeling the societal impact of portation engineering from Harbin Institute of Technology (HIT). His research
fatal crashes. Management Science, 30(4), 464–474. interests cover accident analysis, safety management, and driving behavior.
Sullman, M. J. M., Stephens, A. N., & Pajo, K. (2016). Transport company safety
climate - the impact on truck driver behaviour and crash involvement. Traffic Min Yang is a Full Professor in the Department of the transportation engineering at
Injury Prevention, 18(3), 306–311. Southeast University, China. And his research interests mainly include public transit
Train, K. (2003). Discrete choice methods with simulation. New York: Cambridge planning and management, intelligent transportation system, passenger percep-
University Press. tion, and hub optimization.
Uddin, M., & Huynh, N. (2017). Truck-involved crashes injury severity analysis for
different lighting conditions on rural and urban roadways. Accident Analysis & Yanyong Guo is an Assistant Professor in the Department of the transportation
Prevention, 108, 44–55. engineering at Southeast University, China. And his research interests mainly
Wang, X., & Abdel-Aty, M. (2008). Analysis of left-turn crash injury severity by include pedestrians, speed management, intersection safety, wrong-way driving,
conflicting pattern using partial proportional odds models. Accident Analysis & and safety management.
Prevention, 40(5), 1674–1682.
Weaver, B., & Bédard, M. (2013). The influence of stimulants on truck driver crash Soora Rasouli is a Full Professor in the Department of the Built Environment at
responsibility in fatal crashes. Forensic Science International, 228(1–3), 15–20. Eindhoven University of Technology (TU/e). She leads the Urban Planning and
Wei, Z., Wang, X., & Zhang, D. (2016). Truck crash severity in New York city: An Transportation group, where the focus is on developing models on citizens’ travel
investigation of the spatial and the time of day effects. Accident analysis and behavior with special attention paid to the recent advances in transport technology
prevention, 99, 249–261. including EV, autonomous driving and Mobility as a Service (MaaS).
Williams, R. (2006). Generalized Ordered logit/partial proportional odds models for
ordinal dependent variables. The Stata Journal, 6, 58–82. Zuoxian Gan is an Assistant Professor in College of Transportation Engineering at
Xie, Y., Zhao, K., & Huynh, N. (2012). Analysis of driver injury severity in rural single- Dalian Maritime University, China. He received the M.A. degree in transportation
vehicle crashes. Accident Analysis & Prevention, 47, 36–44. engineering from Chang’an University in 2015, and the Ph.D. degree in trans-
Yang, B., Guo, Y., & Xu, C. (2019). Analysis of freeway secondary crashes with a two- portation engineering from Southeast University in 2019. His current research
step method by loop detector data. IEEE Access, 7, 22884–22890. interests include travel behavior analysis, transport geography, and traffic data
Yuan, Y., Yang, M., Wu, J., Rasouli, S., & Lei, D. (2019a). Assessing bus transit service mining.
from the perspective of elderly passengers in Harbin, China. International
Journal of Sustainable Transportation, 13(10), 761–776. Yifeng Ren received her bachelor degree of transportation engineering from
Yuan, Y., Yang, M., Gan, Z., Wu, J., Xu, C., & Lei, D. (2019b). Analysis of the risk factors Southeast University, Nanjing, China, in 2019. She is currently pursing master
affecting the size of fatal accidents involving trucks based on the structural degree at Southeast University. And her research interest includes accident analysis,
equation model. Transportation Research Record, 2673(8), 112–124. machine learning, public transit planning, as well as data mining and modeling.
Zaloshnja, E., & Miller, T. R. (2004). Costs of large truck-involved crashes in the
United States. Accident Analysis & Prevention, 36(5), 801–808.

165

You might also like