Accident Analysis and Prevention 136 (2020) 105411

Contents lists available at ScienceDirect

Accident Analysis and Prevention

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

Risk factors of road accident severity and the development of a new system T
for prevention: New insights from China☆
Noureddine Benlagha*, Lanouar Charfeddine
Department of Finance and Economics, College of Business and Economics, Qatar University. P.O.X 2713, Doha, Qatar

ARTICLE INFO ABSTRACT

JEL classification: Road accident fatalities and accident severity costs have become top priorities and concerns for Chinese pol-
R41 icymakers. Understanding the principal factors that explain accident severity is considered to be the first step
G22 towards the adequate design of an accident prevention strategy. In this paper, we examine the contribution of
C13 various types of factors (vehicle, driver and others) in explaining accident severity in China. Unlike previous
studies, the analysis gives a particular focus on fatal accidents. Using a large sample of 405,177 observations for
Keywords:
4-wheeled vehicles in the year 2017 and various statistical and econometrics approaches (e.g., OLS, quantile
Accident severity
regression and extreme value theory), the results show that the factors explaining the severity of accidents differs
Fatal accidents
China significantly between normal and extreme severity accidents, e.g. across quantiles. Interestingly, we find that the
Quantile regression gender factor is only significant for fatal accidents. In particular, the analysis shows that male drivers have an
Extreme value theory increased likelihood of extreme risk taking. On the basis of these empirical findings, a new ratemaking approach
Prevention strategy that aims to improve road safety and prevention is discussed and proposed.

1. Introduction countries worldwide due to its direct and indirect effects on peoples’
well-being and government health expenditures (Gong et al., 2012; Loo,
Road traffic injuries are one of the top ten major causes of death in 2018).
the world, with more than 1.4 million people perishing on roads in According to current data,2 China contributed by 23 % of global
2016; 74 % of these were male fatalities. Road traffic injuries also re- road traffic deaths in 2016 (see for instance, Huang et al., 2018a,
present a very high cost, which has been evaluated to be approximately 2018b). For instance, according to the World Health Organisation more
3 % of the world’s GDP (Global status report on road safety by the than 700 people are killed in road accidents across China every day.
WHO, 20151). Despite several initiatives and actions by governments to Moreover, statistics show that during the last four years, the number of
improve road safety to reduce this global issue, the results remain un- fatalities witnessed an ascending trend reaching a maximum in 2017, of
satisfactory (GHE, 2016). In contrast, several empirical studies and 63,772 fatalities (WHO). These fatal accidents are generally associated
technical reports suggested that the global road safety situation has with serious accidents and generate very large material and morale
become worse over time (WHO, 2016); between 2000 and 2016, the losses to individuals and the community (WHO). The primary reasons
world ranking of traffic injuries in terms of causes of death progressed behind these road traffic injuries are the rapid urbanization and mo-
by three positions, i.e., from the tenth to the seventh position. Statistics torization that have been experienced in China because of its rapid
also show that 90 % of the global road traffic deaths are concentrated in economic growth during the last five decades. These rapid levels of
low- and middle-income countries (Global status report on road safety urbanization and motorization have resulted in increasing frequencies
by the WHO, 2015), where the rapid economic development of these of road traffic accidents, injuries and deaths. The appearance of these
countries has been accompanied by an exponential increase in the level excessively large losses has led to the need to understand the causes of
of motorization. Consequently, several researchers and policymakers these road traffic injuries as well as the contribution of all factors in
claim that combatting road traffic should be in the top priorities of all determining accident severity. This will allow Chinese policymakers

☆
We would like to thank the Editor, Professor Abdel-Aty, the Associate Editor, Professor Huang and the two anonymous referees for their helpful comments. Any
remaining errors are solely ours.
⁎
Corresponding author.
E-mail addresses: [email protected] (N. Benlagha), [email protected] (L. Charfeddine).
1
World Health Organization.
2
The Global Burden of Disease (GBD) (2016); http://ghdx.healthdata.org/gbd-results-tool.

https://doi.org/10.1016/j.aap.2019.105411
Received 5 July 2019; Received in revised form 11 October 2019; Accepted 21 December 2019
Available online 05 January 2020
0001-4575/ © 2019 Elsevier Ltd. All rights reserved.
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

and the Chinese government to adequately propose and establish new occupation have significant impacts on driving risk and insurance pri-
regulations and a prevention roadmap that can help improve traffic cing (Guo and Fang, 2013; Litman, 2005; Miyajima et al., 2007). On the
safety. other hand, several studies strongly associate the human and environ-
In the accident prevention literature, several studies have in- ment factors with weather conditions. Theofilatos (2017) investigates
vestigated the drivers of accidents severity such as Brian et al. (2011) accident likelihood and severity by incorporating real-time traffic and
for Canada, Hao et al. (2015) for the US, James et al. (2018) for Aus- weather data from urban arterials in a Bayesian logistic regression and
tralia, and Zhang et al. (2013) for China. The empirical findings are finite mixture models. The results indicate a generally mixed influence
very mixed and seem to depend on the empirical methodology used, the of traffic flow and speed variations on accident severity. Lee et al.
period examined, the type and the dataset sources employed (Chen (2018) use structural equation modelling to analyse the extent to which
et al., 2016; Fitzpatrick et al., 2017; Zeng et al., 2017; Shi et al., 2016; the rainfall intensity and water depth are responsible for traffic acci-
Kitali and Thobias Sando, 2017; Huang et al., 2018a, 2018b). In most dents using Seoul City, Korea as a case study. The model investigates
studies, two main sources of data have been used to investigate the four latent variables that represent the road; traffic, environmental, and
factors that determine accident severity. The first includes data ob- human factors as week as rain and water depth factors. The study re-
tained from government offices, administrations or ministries (Manner veals that traffic, environmental, and human factors; rain and water
and Wünsch-Ziegler, 2013; Feng et al., 2016; Lee et al., 2018 among depth factors; and road factors are mutually correlated with the acci-
others), which is also known as police-recorded data (Shinar et al., dent severity. In addition, in the existing literature, environmental and
1983; Austin, 1995; Conche and Tight, 2006; Loo, 2006). The second human factors are associated with the visibility conditions of the dri-
type of data is obtained from the insurance and reinsurance companies vers. Hassan and Abdel-Aty (2013) and Abdel-Aty et al. (2012) in-
(Krishnan and Carnahan, 1985; Mills and Hambly, 2011). Regarding vestigated mainly the risk factors under low visibility. Abdel-Aty et al.
the question of which of these two sources is more reliable, most studies (2012) used loop detectors data and automatic vehicle identification
suggest that insurance and reinsurance data sources are more relevant (AVI) sensors. They were able to predict more than 70 % of crashes
and can significantly enhance and improve our understanding of the under low visibility conditions. The results of the study reveal that both
factors that determine accident severity (Cohen, 2005; and González types of data can be used for low visibility accident prediction, helping
Dan et al., 2017). to improve road safety. Hassan and Abdel-Aty (2013) suggested that
The overall results from these studies can be grouped into two major significant traffic flow variables are different between visibility related
strands of literature. The first strand considers the driving behaviours crashes and crashes under clear visibility. More specifically, average
and habits as key risk factors causing the accident probability and the speed and average occupancy as well as average speed upstream in-
accident severity (e.g., Shaw and McMartin, 1977; Dahlen et al., 2012; creased accident risk under low visibility conditions. However, the
Sârbescu and Maricuţoiu, 2019). logarithm of the average occupancy downstream and the coefficient of
This first set of literature has emerged following the outstanding the variation of the speed upstream were the main risk factors in clear
study by Shaw and McMartin (1977), whose experimental design has weather.
become one of the most used methods to investigate the behavioural Overall, the previous studies reveal that the availability and the
factors of road safety (Farmer et al., 1997; Delen et al., 2006; robustness of the data are usually the most problematic to the in-
Deffenbacher et al., 2002; Sârbescu and Maricuţoiu, 2019; Papantoniou vestigation of risk factors affecting the accident severity. Moreover,
et al., 2019). Deffenbacher et al. (2002) employed an experiment in- although research has illuminated some important behavioural and
volving 290 participants to investigate the relationship between ag- environmental variables affecting the accident severity, to the best of
gressive driving behaviour and injury accidents. Their empirical find- our knowledge, no study to date has examined the risk factors affecting
ings show that the frequency and intensity of anger and the frequency both the normal and extreme severity in the case of China.
of aggressive behaviour are positively related to risky behaviour. Re- This paper falls under the second strand of literature exploiting
cently, Papantoniou et al. (2019) used driving simulation experiment several interactive factors to explore the determinants of traffic acci-
data to investigate several accident factors that may affect overall dent severity. This paper fills a gap in the literature on road safety by
driving error behaviour including driver distraction and driver char- providing a new methodological approach that allows accurate pre-
acteristics. The findings reveal that the impact of driver characteristics dictions of the factors simultaneously affecting the normal and extreme
and area type are the only statistically significant factors affecting the accident severities. Specifically, it uses large datasets from a Chinese
probability of driving errors. Sârbescu and Maricuţoiu (2019) employed insurance company that contain a significant amount of useful in-
a seven-week diary approach to investigate the intra-individual varia- formation that could help in determining the significant risk factors of
tion of dangerous driving behaviour. Their findings revealed that vio- the accident severity. In addition, although the number of accidents and
lations, verbal and physical aggression and vehicle aggression had si- the total fatalities in China are among the highest in the world, there
milar intra- to inter-individual variation ratios. In contrast, driving are currently few empirical and case studies investigating these im-
errors had a moderate variance at the intra-individual level. portant issues (see, for instance, Zhang et al., 2013; Jin et al., 2018;
Recently, several empirical studies have used the enormous Shao et al., 2019). Accordingly, our paper provides a novel and inter-
amounts of vehicle-related data produced by in-car sensors during esting case study on the accident severity and safety in China by ex-
every trip, commonly known as usage-based data, to explore the be- ploiting new insurance data.
haviour-based variables affecting road safety. A number of behavioural Our paper differs from previous studies of road safety and prevision
variables are examined such as hard deceleration, acceleration and in three ways. First, unlike previous studies, this paper focuses on fatal
swerve manoeuvres, and speeding (see, for instance, Staubach, 2009; accidents, which represent in our sample more than 60 % of the total
Pérez-Marín and Guillen, 2019). The results show that most accidents amount of losses.3 Second, this study employs a unique methodological
were influenced by multiple behavioural factors and that there is a approach that combines both the quantile regression (QR) approach
safety potential for advanced driver assistance systems (ADAS) that can and extreme value theory. The QR approach enables us to model the
support a driver in information assimilation and help to avoid distrac- heterogeneous relationship between the accident severity at various
tion and reduced activity. points of the conditional distribution with several risk factors. The re-
The second strand of literature considers the three sets of interactive sults of this first approach show strong evidence for the necessity to
factors: human, vehicle and environment (see, for instance, Haddon,
1968; Miyajima et al., 2007; Hassan and Abdel-Aty, 2013, Bergel-Hayat
et al., 2013; Theofilatos, 2017; Lee, et al. 2018). On the one hand, 3
The top 10% of accident severities in our sample represents 63.5% of the
studies showed that demographic variables such as age, gender and total amount of accident severities.

2
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

decompose accident severity into two phases where during the first In the existing literature, several measures are used to assess the
phase (known as the normal phase) the determinants of accident se- performance of safety; these measures are based on defining a com-
verity are the same. However, during the second phase, which describes posite index including several safety factors and dimensions (Al Haji,
the extreme accident severities, the results show evidence of the ex- 2005; Ma et al., 2011; Chen et al., 2015; Wang et al., 2015; Tešić et al.,
istence of a switch in the determinants of accident severities. 2018). Regarding the China’s performance in road safety, few studies
Third, our paper differs also in the used methodology. Indeed, our have been conducted. Huang et al. (2016) have assessed the china’s
methodological approach is of particular interest for future road safety performance in road safety for the period 2002-2013. Their findings
studies. Essentially, we implemented interdisciplinary research area suggest that road traffic mortality in the period of study remains at a
that can be employed in the analysis of safety big data (SBD). Actually, high level,6 and safety is poor. Xiao et al. (2017) evaluated the effec-
we combined disciplines such as mathematics to identify the extreme tiveness of the law on road traffic mortality attributed to alcohol use
accidents, statistics to describe our safety data, and econometrics to and conclude that the used data are inadequate to determine the ef-
categorize the key safety factors to capture the most relevant accident fectiveness of the implementation of the more severe drunk-driving law
factors. Further, we used computer sciences to apply all the previous in China. Li et al. (2016) exploited several data sources to evaluate time
disciplines on the safety data. Actually, big data has several possible trends in road traffic fatalities (RTF) in China. Their findings indicate
applications in different fields (Shi and Abdel-Aty, 2015; Ouyang et al., that all data sources have limitations in coverage, representativeness or
2018). In particular, the use of big data and new tools of data analysis in overreliance on model specifications. As mentioned by Huang et al.
safety science becomes critical for organisations’ safety decision- (2018a, 2018b) the statistics and trends of the road traffic deaths pro-
making (Huang et al., 2018a, 2018b). Big safety data are gathered vided by several sources are commonly not coherent. However, despite
through various instruments, which are significantly evolving and in- the discrepancies in methodologies and estimates, all sources indicated
creasing (Stylianou et al., 2019). However, the lack of open access an increase in RTF before 2005 and a slight decrease after this year.
multiple-source data is one of the biggest problems facing the progress Overall, the divergent trends in road traffic fatalities and injuries
of studies on big road safety data Huang et al. (2017). across data sets affects the results of the assessment of the performance
Even if our data cannot be formally classified as big safety data,4 the of road safety in China. Accordingly, it is vital to reconcile data sources
exploited data in this paper are rich enough to provide useful results to further improve road safety. Indeed, if the used fatal data are in-
helping to develop policies in terms of road safety. Precisely, in addition complete and distorted, the prevention efforts will be misdirected
to the massive factors included in our models, to the best of our (Huang et al., 2018a, 2018b). Therefore, the elaboration of safety big
knowledge, this is the first study that attempts to explore the impact of data becomes a challenging task in developing road safety in China.
the unladen weight of the car on the accident severity.
The remainder of the paper is organized as follows. Section 2 pre- 3. Data and preliminary analysis
sents a description of the status of china’s road safety prevention.
Section 3 describes the data and the accident severity distribution. 3.1. Data description
Section 4 introduces the material and methods used to analyse the
factors affecting accident severity. Section 5 discusses the results. Sec- The dataset used in this study consists of a sample of 405,177 ob-
tion 6 develops a theoretical approach to improve road safety and servations for the year 2017 and includes only the four-wheeled ve-
prevention. Finally, Section 7 concludes the paper and presents direc- hicles and motorcycles. The data was obtained from an insurance
tions for future research. company in the Hubei Province headquartered in Wuhan7 (the in-
surance company covers the entire province of Hubei).8 The data con-
2. Current status of China’s road safety prevention tains policyholders who joined the insurer during the years 2012–2017.
All the reported policyholders are itemised in the insurance company at
This section provides an overview of the current situation of China the time of the data collection (2017). It includes all the information
road traffic safety based on new statistical reports and previous studies. that the insurer has about each of these policyholders. Especially, we
Because traffic accidents are responsible of large financial losses and have the information about the policy issued in the year the policy-
increasing number of injuries and fatalities, significant improvements holder joined the insurer and, for those who stayed for more than one
in the infrastructure and road safety legal system are required. On one year, considered their subsequent policies as well. The most relevant
hand, China, a country on the move, has built the world’s largest high- information used in this study are the following.
speed rail network, consisting of more than four million kilometres of
highways. This crucial project has achieved a massive rollout of elec- 3.1.1. Vehicle-related factors
tric, zero-tailpipe-emission, two wheeled vehicles (Fishman and Cherry, The variables that describe the vehicle’s characteristics used in this
2016; Jiang et al., 2017). On the other hand, since 2003, the Chinese study include the car price, number of seats, the unladen, and the car
Government has undertaken a series of strong initiatives to improve the use. Although all these variables are selected based on previous studies
road safety. The most important measures are: (i) the establishment of and the data availability, there is no clear direction in the existing lit-
the Inter-Ministerial Road Safety Forum (October 2003), (ii) the im- erature about which vehicle-related characteristics determine accident
plementation of the Road Traffic Safety Law (May 2004), (iii) the im-
plementation of more severe drunk-driving law in April 2011. 6
The pre-dicted maximum mortality (PMM) was estimated to be 0.174 from
The results of these measures was eminent with significant reduc-
health data in China.
tion in injuries and fatalities as well, both in absolute numbers and on a 7
Hubei Province is one of largest provinces in China. It is ranked as the 9th
population basis (Wang et al., 2019). Interestingly, the reported Chi- populated province with total population of 58,500,000 in 2015, ranked 13th in
nese police data indicate that the road traffic deaths decreased by 46 % term of area with 185,900 km2; ranked 8th in GDP. In addition, due to its
between 2002 and 2016.5 The fatality decrease are frequently attrib- central location in China, Hubei has clear regional transportation advantages.
uted to the significant expansions in transportation infrastructure in- The provincial government is promoting Wuhan, upgrading it to a modern city.
vestments (Jiang et al., 2017). Thus, Hubei province has many similarities with the largest provinces in China,
principally when considering the infrastructure development and the road ac-
cidents as well as other economic and social variables. These characteristics
4
Big Safety data is defined as: “the ensemble of all kinds of safety data, including motivates considering the Hubei Province as a representative case study of the
text, audio, video and picture…”(Quyang et al., 2018). China.
5 8
China Statistical yearbook, National Bureau of Statistics of China (2016) The name of the insurance company is confidential.

3
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

severity. The empirical findings are inconclusive and ambiguous (Al- extreme accident losses.
Ghamdi, 2002; Zhang et al., 2013; Benlagha and Karaa, 2017). Com- The results of the rejection of the normality hypothesis of the ac-
pared to previous studies, and to the best of our knowledge, this is the cident severity is confirmed by the Lilliefors test, which is an improved
first empirical study that makes use of the unladen weight of the vehicle version of the Kolomogorov-Smirnov (K-S) test statistic.
variable as a measure of risk factor. In fact, the unladen weight account Fig. 1 shows evidence for deviation of the empirical distribution
for all variables describing the real volume, size, and parts attached to from the standard normal distribution.
the vehicle. Consequently, it could be an important contributing factor To further comprehend the behaviour of the accident severity dis-
affecting the severity of accidents. Therefore, it is expected that vehicles tribution, we report in Table 2 the proportion of accidents according to
with higher values of unladen weight are expected to cause higher their degree of severity. It is remarkably shown that 1 % of accidents
accident severities. represents 24.9 % of the total accident severities and that the top 10 %
of the accident severities represent approximately 63.5 % of the total of
3.1.2. Driver-related factors losses in our studied dataset. This result can be interpreted as pre-
Similar to previous studies investigating road safety and accidents, liminary evidence for the existence of extreme values in the accident
we use the two variables, the age and gender of the drivers, considered severity distribution, which can induce a bias in the results if not ac-
as key risk factors that determine accident severity. We use the same counted for.
classification as adopted by the China traffic management bureau9, We complete the above descriptive analysis by analysing the box
where drivers are grouped into five age groups of < 25, 26–35, 36–45, and empirical Q-Q plots of accident severity. The box plot presented in
46–55, and > 56.10,11 The second driver-related variable is gender, Fig. 2a displays the variation in accident severity data without making
which is defined by a dummy variable given by: any assumptions concerning the underlying statistical distribution. In
this figure, the rectangle represents the cumulative distribution of ac-
1ifthedriverisaman cident severity arranged by value and the thin lines separate the ex-
Gen =
0ifthedriverisawoman treme observations from the standard observations (see, for instance,
Mcgill et al., 1978; Anselin, 2005). The existence of values outside the
In addition, unlike previous studies, we use the past involvement in thin lines indicates the existence of extremes in the accident severity
accidents (number of accidents) as a possible risk factor that may ex- data. The presence of extreme values in the accident severity data is
plain the accident severity. The number of past accidents can be con- confirmed by the analysis of the comparison between the theoretical
sidered as a proxy for the driver behavioural risk. It is established that normal and the empirical quantile-quantile as shown in Fig. 2b; it is
drivers with a high level of risk tend to be more involved in accidents clear that the QQ-plot does not lie on a straight line. The mismatching
than low-risk drivers. Accordingly, the number of accidents is assumed between the theoretical (normal) and empirical distributions is clear
to encapsulate all of the information on the risk behaviour of drivers. and more pronounced at the lower and upper quantiles.
In summary, the preliminary empirical findings highlight evidence
3.1.3. Other factors of the existence of extreme values in the accident severity distribution;
The other factors considered in this study include variables that are consequently, the traditional approaches may lead to biased results in
related to the insurance coverage of the vehicle. Explicitly, we consider the empirical analysis and to inaccurate predictions of risks (see, for
the premium paid by the policyholder (the driver), the maximum instance, Lane, 2000; Vernic, 2006). Bolance et al. (2008) used in-
coverage, the amount of refund and the paid amount as additional surance data to model the accident severity distribution; they restrict
explanatory variables. Particularly, the premium paid by the policy- their presentation to the normal, the skew-normal, and a kernel esti-
holders and the maximum coverage are expected to be positively as- mator. Eling (2011) extends the previous study by using skew-normal
sociated with the accident severity. The use of the premium paid by the and skew-student distributions. Kingphai and Chongcharoen (2016)
policyholder and the maximum coverage variables can allow us to used a set of probability distributions to fit to auto insurance claim data;
identify the possible existence of residual adverse selection or moral they find that the log-normal distribution performs well. However, their
hazard problems that may directly affect the accidents’ occurrences as study focused only on standard distributions, disregarding the im-
well as their severities (see, for instance, Chiappori and Salanié, 2000; portance of extreme value distributions.
Karaa and Benlagha, 2015; Fuller, 2014; Keane and Stavrunova, 2016).
4. Material and methods
3.2. The accident severity distribution
To investigate factors affecting the accident severity, we start the
A starting point before conducting the econometric analysis is to analysis with the standard OLS approach, then we make use of the
analyse the statistical distribution of our main variable of interest, e.g. quantile regression framework of Koenker and Bassett (1978), and fi-
the accident severity variable. Table 1 presents the descriptive statistics nally, we employ the extreme value theory where the peak over
of the variable in level and in its logarithmic forms. The table shows threshold approach will be employed to select the threshold that se-
that the severity distribution is right skewed indicating the presence of parates the extreme accident severities from the normal ones. Since the
an upper tail (i.e., positively skewed). Moreover, the kurtosis is greater OLS approach is very standard and well know, we introduce in the
than three in the Gaussian distribution. These statistics show that the following two subsections the Quantile regression and extreme values
accident severity distribution is non-normal and reveal the presence of approaches.

4.1. Quantile regression (QR) approach

9
According to the Traffic Management Bureau and China Road Traffic
Accidents Annual Statistical Reports. In applied statistical literature, it is well regonized that the standard
10
In our original data set, we have the variable namely the number of years of
linear regression technique is very restrictive and gives only a partial
driving experience. However, while this variable is considered as a main risk
view about the relationship between a set of explicative variables and
factor of accident severity, we find that it is highly correlated (problem of
multicolinearity) with the driver’s age. Consequently, this variables was the outcome variable (see for instance Koenker and Bassett, 1978). In
dropped from the analysis (A problem of multicollinearity can arise due to the many cases economists are more interested by comparing the entire
existence of high correlation between explicative variables). distribution rather than comparing only the means (see Bitler et al.,
11
We use also the robust command under Stata and the results is overall 2006; Djebbari and Smith, 2008; Jackson and Page, 2013). The quantile
similar than that when using the bootstrap technique. regression offer this capability of describing the relationship at different

4
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Table 1
Summary statistics for the accident severity.
Variable Mean Min Max SD Skewness Kurtosis Normality

Accident severity 5680.65 1.75 410000 17080.36 9.775 144.39 0.0060***

Accident severity (ln) 7.661 0.559 12.923 1.184 0.776 4.308 0.0060***

Note. This table displays the main summary statistics, including the mean, the minimum, the maximum, the skewness and the kurtosis of the accident severity. ***:
the normality of the data is rejected at the 1 % significance level. The amounts are in Yuan. (1 China’s Yuan equals = 1.15 $ at the time of the collection of the data).

Fig. 1. Normal versus empirical cumulative distributions.

Table 2 points in the conditional distribution of the dependent variable. In our

Accident severity distribution. specific context, the main advantage behind the use of the QR approach
Accident Percentile Total of accident Proportion of Number of
is to examine whether factors affecting accident severities differ sig-
severity severity accident observations nificantly across Quantiles.
severity Technically, the QR approach estimates the relationship between
the regressors and outcome using the conditional quantile function
>99% 84308.01 67877835.9 0.249 487
>95% 19299.35 141192770.4 0.519 2402
Q (y|x ) , where is the quantile. It formulates Q (y|x ) as a linear func-
>90% 9669.5 172893406.8 0.635 4792 tion of parameters,
>75% 3981 216686490.8 0.796 11980
>50% 1830.375 248564045.4 0.913 23958 Q (y|x ) = x ( )
Total 272182697.6 1 43132
The parameter vector ( ) is quantile dependent and can be esti-
Note. In this table, we considered only the non-zero severity accidents. The mated by solving the following minimization problem,
values are obtained using the cross classification of the accident severity for N
different percentiles. The amounts are in Yuan. (1 China’s Yuan equals = 1.15 $ min (yi xi )
at the time of the collection of the data). i=1

Fig. 2. a Box plot of the accident severity. b Q-Q plot of the accident severity.

5
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Where (u) is a check function for the quantile of the dependent If the parent distribution F were identified, the distribution of the
variable yi . This function is defined as (u) = u ( I (u < 0)) . The set threshold exceedances in (2) would also be known.
contains all quantile-dependent parameters ( 0, , 1, , …, k, ) where k In the peak over threshold method, the distribution of the ex-
refers to the number of regressors. ceedances is approximated by the generalized Pareto distribution.
Theoretically, it is proven that the estimator, ˆm, ,for m = 1, …, k , is The main results of the GPD are shown in the following theorem.
asymptotically normal with a variance-covariance matrix that can be
determined analytically. In this study, we use the Bootstrap technique Theorem 1. Let X1 , …, Xn be a sequence of independent random variables
to estimate the standard errors.11 For additional details about the sta- with common distribution function F , and let the following:
tistical proprieties of the quantile regression, we refer to Koenker and (3)Mn = max {X1 , …, Xn }
Bassett (1978), Koenker (2005), and Chen and Müller (2012). Denote an arbitrary term in the Xi sequence by X , and suppose that F
In our study, all the coefficients, m, , are relevant when designing satisfies a number of restricted conditions.
the prevention policies. Specifically, two questions are of particular
interest for policy makers and government. The first is whether the Pr {mn z} G (z )
estimated coefficients are significant across different quantiles?; and where
the second is whether the coefficients are statistically stable across
quantiles. While, the answer to the first question will allows us to de- z µ 1/
G (z ) = exp 1+
termine the factors that affect the accident severity outcome, the an- (4)
swer to the second question indicates whether the relationship differs
across quantiles. For the first test, we plot the estimated coefficient for for some µ, > 0 and . Then, for a sufficiently large u, the distribution
each factor and for several quantiles, function of (X µ) conditional on X > u is approximately the following:
= 0.02, 0.05, 0.10,0.15, …, 0.90, 0.95. Then, the null hypothesis of no 1/
y
effect will be rejected if the estimated coefficient does not fall in the H (y ) = 1 1+
confidence interval reported also in the graph. (5)
Regarding the coefficient stability across quantiles, it is important to Defined on {y : y > 0and (1 + y / ) > 0} , where = + (u µ) .
remember that, in this study, the most interesting equivalence occurs The family of distributions defined by Eq. (5) is called a generalized
across the upper quantiles, which correspond to those associated with Pareto family.
high accident severity. Consequently, we conduct a multiple hypothesis Theorem 1 implies that if block maxima have approximating dis-
testing to examine whether the distribution differ between the median tribution G , then threshold excesses have a corresponding approximate
and the upper tail. For this purpose, we propose to test the following distribution within the generalized Pareto family. Moreover, the para-
three hypothesis: meters of the generalized Pareto distribution of the threshold excesses
H0,1: = = are uniquely determined by those of the associated GEV distribution of
m,0.05 m,0.90 m,0.95
the block maxima. The shape parameter is dominant in determining
H0,2: = = the qualitative behaviour of the generalized Pareto distribution. If
m,0.50 m,0.95 m,95
< 0 , the distribution of excesses has an upper bound of u / , and if
H0,3: m,0.90 = m,0.95
< 0 , the distribution has no upper limit. Finally, if = 0 , the dis-
tribution is unbounded.
These hypotheses are tested for each regressor (for m = 1, …, k ) . In practice, various and rigorous steps are required to effectively
Note also that under the null hypothesis the three tests have the usual F- implement the peak over threshold approach to investigate the extreme
distribution as limiting distribution, where under the first and second accident severities. These steps are presented in Fig. 3 and additional
hypotheses the number of restrictions is two, and one restriction under theoretical and technical evidence is provided in the supplementary
the last hypothesis. materials.

4.2. The peak over threshold approach 5. Results and discussion

In this study, we use the extreme value theory (EVT) to model and 5.1. OLS estimation results
predict the extreme accident severities. In the theoretical statistical
literature, two types of extreme value theory are used for this purpose: The results of the standard OLS estimation of the factors explaining
(1) the classical extreme value theory (EVT) which is based on the block the accident severity are reported in Table 3 below.
maxima, and (2) the peak over threshold (POT) method.
In this study, we will focus on the POT method, which is suitable for 5.1.1. Car-related variables
our purposes since only severity peaks greater than a selected threshold Table 3 shows that all the car related variables are highly significant
are considered. This method is based on the estimation of the gen- at the 1 % significance level. In particular, the car price is negatively
eralized Pareto distribution (GPD ), which is the unique non-degenerate related to the accident severity. Despite that fact that this empirical
distribution that approximates asymptotically the limiting distribution result seems to be unexpected, it can be explained by the fact that
of exceedances (Balkema and de Haan, 1974; Pickands, 1975; Benlagha drivers do not take a high risk if the car is expensive. The results also
et al., 2008; Grun-rehomme et al., 2007; Koenker and Bassett, 1978). show that the use of a car is negatively related to the accident severity.
Formally, X1 , …, Xn is a sequence of independent and identically In this modelling, the use of the car is a dummy variable that takes the
distributed random variables (e.g., losses or severities) having marginal value 1 if the vehicle is used for commuting and family use and 0
distribution function F . In the POT method, we consider as extreme otherwise. Accordingly, the negative sign of the variable “use of the
events those with severities exceeding a high threshold u (the threshold car” indicates that the severity decreases when the vehicle is used for
of accident severities). Designating an arbitrary term in the Xi sequence commuting. These results corroborate those obtained by Grun-re-
by X , the illustration of the stochastic behaviour of extreme events can homme et al. (2007) using French data and Wang et al. (2019) using
be given by the conditional probability, as follows: Chinese data. Empirically, the association between traffic accidents and
1 F (u + y ) the use of a car has been examined in different manners. For instance,
Pr {X > u + y |X > u}= ,y>0 Yang et al. (2009) and Chen and Jou (2019), among others, have dis-
1 F (u ) (2)
tinguished between two types of car use: private or public. Recently, Li-

6
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Fig. 3. The extreme value approaches.

Lu et al. (2019) differentiate between taxi and non-taxi driver related 5.1.3. Policy-related variables
risk factors. Table 3 also indicates that the impact of the premium paid by the
The results also show that the accident severity decreases with the insured on the accident severity is positive and highly significant. Ex-
number of seats. The number of seats allows identification of the type of plicitly, the accident severity increases with the insurance coverage.
car. If we consider compact and other cars as the two modalities for the Drivers with high vehicle damage insurance coverage are significantly
type of car (Lee et al., 2018), a positive correlation was revealed be- more likely to file accident claims than those with low damage in-
tween the vehicle type (compact) and the level of accident severity, surance coverage. More importantly, this can be explained by the pre-
which means that compact cars are vulnerable to the level of accident sence of moral hazard and adverse selection problems in the vehicle
severity. Conversely, in their study on the risk factors affecting fatal bus damage insurance market (Li et al., 2013).
accident severity, Feng et al. (2016) show that buses with more than 16 Given our large sample size, we propose examining whether the
seats are more likely to be involved in fatal accidents when compared factors affecting accident severities differ across quantiles. This com-
with buses with nine to 16 seats. parison is particularly interesting since it will help us identify the de-
Finally, the car’s characteristics variables, we find evidence for a terminant factors of accident severity in each percentile. This is parti-
positive and significant relationship between the accident severity and cularly interesting when designing individual policies or a general
the unladen weight of the vehicle. Since the unladen weight includes all strategy to improve road safety and accident prevention.
variables describing the real volume, size, and attached parts, it could While the OLS results show evidence of the existence of a significant
be an important contributing factor affecting the severity of accidents, relationship between the different factors considered in this study and
particularly in the case of accidents with large losses. the accident severity, the evidence for the existence of heterogeneity in
the accident severity questions the use of the standard OLS approach
(see Chen et al., 2016; Shannon et al., 2018). Consequently, the dif-
5.1.2. Driver-related variables ferent policies that are recommended to improve road safety could be
Concerning human factors, a driver’s age indicated a negative as- inefficient if the policymakers do not consider this heterogeneity in the
sociation with the accident severity. Explicitly, younger drivers have a accident severity distribution. As a first solution, the next section re-
significant effect on the level of accident severity. Actually, these ports the results of the quantile regression, which help in understanding
findings support risk-taking behaviours in young driver accidents. whether the factors affecting the accident severity differ between
Young drivers are more likely to be involved in road accidents than quantiles.
experienced drivers due to risk taking, such as excessive speeding and
the use of alcohol and drugs (de Boni et al., 2011; McDonald and 5.2. Quantile regression results
Sommers, 2016; Cordazzo et al., 2016). Moreover, the risk-taking be-
haviour of the young drivers is also reflected in traffic violations; see, The results of the estimation of the impact of the factors on the
for instance, Cordazzo et al. (2016) and Zhang et al. (2016). Although severity of accidents based on the quantile regression are reported in
the results of the existing literature have suggested that gender has an Fig. 4. As mentioned above, the most important feature of the quantile
effect on the likelihood of road accidents, our OLS estimates show that approach is its ability to identify, by quantile, the key factors that ex-
this variable is statistically non-significant. plain accident severity. Evidence of a significant difference regarding

Table 3
OLS estimation results.
Constant Car price Use N seats UW Gender Age N acc. Premium MSE

Coef. (p-value) 4.562 (.118) .216*** (.010) −.123*** (.034) −.026*** (.005) .047*** (.005) −.004 (.015) −.003*** (.001) .004 (.006) .116*** (.009) 257.49

Note. *** Significant at 1 % significance level.

7
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Fig. 4. The Quantile estimates of the risk factors.

factors explaining accident severity across quantiles suggests that pol- rejected for the second test where the coefficient of car price in the
icymakers and insurance companies should design group-oriented median regression is significantly different from the upper quantiles
strategies to improve road safety and increase accident prevention by (0.90 and 0.95).12
drivers. The results show that the impact of the use of the car on accident
severity is non-significant for all the percentiles except for the 90th one
where the estimated coefficient lies outside the confidence interval (see
5.2.1. Car-related variables Fig. 4). We find also that the sign of the parameter associated to the
Similar to the OLS results, we find that all the car’s related char- variable use is the same as that obtained by the OLS regression. The
acteristics are highly significant at the 1 % significance level. The re- implementation of the multiple test for comparison between the ex-
sults show that the impact of the car price on accident severity is sig- tremes percentiles 5th, 90th and 95th confirm the previous findings that
nificantly low along the first quartiles; between the 60th and the 85th the variable use is only significant at the upper quantiles. This latter
percentile the impact increased significantly. In the extreme quantiles, results indicate that the severity decreases when the vehicle is used for
the impact of this variable on the accident severity has diminished. The commuting.
results show also that there is strong evidence that the magnitude im- Regarding the number of seats variable, the quantile results high-
pact differs significantly across quantiles (see Table 4). For instance, the light that there is enough evidence to infer that the number of seats
three multiple tests (see Table 4) show that the null hypothesis is only negatively affects the accident severity, mainly, in the lower and upper
percentiles. Moreover, the multiple tests show that the magnitude of the
Table 4 impact of the number of seats varies among the quantiles. This is par-
Testing for equality between quantiles.
ticularly pointed when comparing between the lower and upper per-
q(0.05, .90, .95) q(0.5, .90, .95) q(.90, .95) centiles, 5th, 90th and 95th. This also confirmed when performing a
multiple test involving the median and the 90th and 95th.
Variable F-stat p-value F-stat p-value F-stat p-value
Finally, the results for the unladen weight variable show that this
Car characteristics variable explains the accident severity at all quantiles except in the 40th
Car Price 0.13 0.876 3.34 0.039 0.15 0.698 percentile where the associated coefficient is not significantly different
Use 3.10 0.045 0.76 0.468 0.52 0.472 from zero. The results show also that in the upper last quantile (the 90th
N. of seats 7.95 0.000 8.64 0.000 0.47 0.495
percentile), the impact of this variable on accident severity has de-
UW 60.16 0.000 10.33 0.000 2.32 0.127
Driver characteristics creased but remains highly significant. Regarding the question of
Gender 4.13 0.024 3.45 0.033 0.03 0.867
Age 8.91 0.000 0.16 0.853 0.24 0.625
Policy and other Characteristics 12
N. of Accident 0.11 0.898 0.11 0.894 0.08 0.776 A similar results (not reported here) show that the coefficient of the car
Premium 4.76 0.008 0.20 0.821 0.61 0.436 price obtained from the median regression is statistically different to that ob-
tained from lower quantiles regressions (0.05 and 0.10).

8
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

whether the impact of the unladen weight variable on accident severity the influence of the premium paid by the insureds on the accident se-
differs across quantiles, the outcomes of multiple hypothesis tests re- verity for various subgroup sizes and different individuals in each
sults reported in Table 4 show that the answer is positive. For instance, group. The results show that an adverse selection problem is detected
we find that upper quantiles coefficients are significantly different from only for individuals having a certain high level of accident severity.
both the median and the lower quantiles confirming the previous re- Moreover, the adverse selection problem is also considered in the ex-
sults that the effect of the unladen variable on accident severity is treme quantiles.
quantile dependent. Overall, our results highlight the superiority of the quantile analysis
in revealing factors influencing the accident severity. It is worth men-
5.2.2. Driver-related variables tioning that the results show an oscillation in the impacts of several
Concerning the driver’s characteristics, the gender quantile results factors on the accident severity across quantiles (e.g., insurance pre-
provides new insights comparing to the OLS estimates and the existing mium, the use of the car, the unladen weight and gender, among
literature. Accurately, the results show that the gender does not affect others). This variability is largely observed in the extreme quantiles,
the accident severity in the mainstream lower percentiles. However, the i.e., mostly in the 80th, 85th, 90th and 95th quantiles. Interestingly, we
gender becomes a significant factor affecting the accident severity for find that the gender variable is a determinant factor in explaining the
the extreme upper percentiles (85th to 99th). This result is also con- accident severity at only the extreme quantiles. A conceivable explana-
firmed by the multiple hypothesis tests which show that the gender tion for this result is that the proportion of the extreme accident se-
estimated coefficients differ significantly across quantiles mainly be- verity in the data may be significantly high. Therefore, for a better
tween upper versus upper and lower versus median, e.g. (5th, 90th and assessment of the risk factors, it will be worthwhile to analyse the
95th) and (50th, 90th and 95th). Our results show that the accident se- pattern of these extreme severities for the full dataset as well as by
verity increases with male drivers13 in the upper quantiles. gender since gender is significant only in the highest quantiles.
When dealing with the age, the results show that this variable is
found to be statistically significant along the lower quartiles, e.g. the 5.3. EVT results for accident severity factors
estimated coefficients are outside the confidence interval until the 30th
percentile. The result show that the impact of the driver’s age on the In this section, we continue the analysis of factors explaining the
accident severity becomes again significant for the 70th to 95th, per- accident severity by focusing on the fatal accidents group for both
centiles (see Fig. 4). Furthermore, the results show, with strong evi- groups of the gender.14 In addition to the results discussed in Section
dence, that for only the lower quantiles the magnitude impact of the age 5.2., the analysis in this subsection is also motivated by the preliminary
differs significantly across quantiles (see Table 4). For the remaining results obtained in Section 2, which show that the scatter diagram of
percentiles, the null hypothesis of constancy of the age impact cannot accident severity distribution is fat tailed indicating evidence of the
be rejected. At this level, we suggest that for all types of accidents (i.e., existence of extreme values (i.e., fatal accident severities).
minor, moderate, serious, severe, critical, or extreme) young drivers are
more likely to be involved in road accidents than experienced drivers 5.3.1. Threshold selection
are. We start the extreme value investigation by using different ap-
proaches to select a suitable threshold that allows an accurate separa-
5.2.3. Policy-related variables tion between the normal and extreme severities (step 2 in Fig. 3). In this
Regarding the other variables, the results show that the number of study, we employed three different methods, namely, the mean residual
accidents is not a determinant variable for the accident severity what- life plot (MRL), L-moments plot (LM) and the threshold choice plot
ever the quantile used, e.g. none of the quantiles estimated coefficients function (TC).15 An appropriate analysis of these plots helps to identify
associated to this variable is significantly different from zero. A result the suitable threshold that separates the normal accident severity from
that is confirmed by the non-rejection of the null hypothesis of con- the extremes. Officially, using the mean residual life plot, a threshold
stancy of the quantiles parameters associated number of accident must be selected when the plot is nearly linear and the modified scale
variable. and shape estimates become constant.
However, the results show that the impact of the premium paid by Fig. 5 indicates that the mean residual life plot of the maximum
the driver or the owner of the car on the accident severity is significant severity thresholds is linear starting from a threshold of 8.5, where the
at the lower percentiles (from 2nd to 35th) (see Fig. 4). This latter result line becomes more stable until approximately 10. The results of the L-
is aligned with the outcomes of the first multiple hypothesis test, which moments plot and the threshold choice plot function (TC) confirm these
infers, with strong evidence that the impact of the premiums, differs last findings.16 In terms of percentiles, the suitable threshold is identi-
across quantiles when comparing the lower to upper quantiles, e.g. no fied to be between the 90th and 98th percentiles. For a robustness test,
effect in the upper quantiles and negative significant in the lower we applied threshold selection to subsets based on the gender of the
quantiles. Such findings have important implications for understanding drivers and we highlighted that a threshold of the 97th percentile is a
the adverse selection problem raised in the recent literature. Despite the reasonable choice for males and a threshold of the 95th percentile is a
extensive theoretical research on the adverse selection problem in in- reasonable choice for females. We must note that in practice, decision
surance markets, its empirical relevance, however, remains an issue of or threshold choices are not clear when using the threshold choice plot
considerable debate (see, for instance, Chiappori and Salanie, 2000; Li
et al., 2013; Benlagha and Karaa, 2017). Our findings offer a first 14
This analysis is motivated by the results obtained in subsection 4.2 that
compromise to the existing debate between the studies rejecting the
show that gender is a key factor in explaining accident severity at the upper
hypothesis of the presence of the adverse selection and those arguing quantiles.
the presence of the adverse selection problem in car insurance markets. 15
We notice that the extreme value estimation highly depends on the
Our findings indicate that this conflicting result in the literature threshold and sample size; specifically, the threshold selection involves a trade-
concerning the adverse selection is mainly due to the econometric off between bias and variance. To minimize this dilemma, it is recommended
model employed and the classification procedure implemented in the that several approaches be used to allow for an appropriate identification of a
previous studies. Our quantile on quantile approach helps to identify threshold.
16
For conciseness, we do not report the POT results here. All the tables and
figures related to the extreme value analysis results (i.e., threshold selection
13
The modal category for the gender, the most commonly occurring is male and fitting the GPD) can be obtained upon request from the corresponding
coded 1. author.

9
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Fig. 5. The mean residual life plot of the accident severity.

function. The findings related to the extreme value results have important
The next step of the peak over threshold approach (step 3 in Fig. 3) implications. Primarily, having a distribution that fits the accident se-
is to fit a generalized Pareto distribution to the real data. Table 5 pre- verities well will help in outlining suitable policies that aim to improve
sents the asymptotic maximum likelihood estimates of the GPD para- road safety, since it helps to predict extreme accident severities in terms
meters ( and ). The main statistical findings of the present POT of probability of occurrence and return levels.
modelling are that a suitable model for the accident severity could be a
distribution limited from the right. This result corroborates that ob-
5.3.2. The behaviour of risky drivers
tained by Pisarenko and Mikhail (2010). Furthermore, the different
Once the threshold value separating fatal from normal accident
thresholds employed for the scale and shape estimated parameters are
severity has been determined, the next step is to investigate the im-
statistically significant at the 5 % significance level. These statistical
portance of the different car, driver, and other factors in explaining the
findings provide some interesting facts.
fatal accident severity.
First, the shape parameter for the male is larger than that for the
Our first results reported in Fig. 4 in the supplementary documents
female (Table 5). This finding indicates that the proportion of male
indicate that the behaviour of the drivers depends on gender at the
drivers incurring extreme severities is significantly larger than that for
highest quantiles. A result that is confirmed by the extreme value
female drivers. This result is valid regardless of the selected threshold.
modelling shows that the extreme accident distribution varies by
The findings are consistent with the official statistics and empirical
gender (See Table 6). Thus, we believe that a further analysis in-
studies on Chinese road accidents and fatalities. As an illustration, the
vestigating the determinants of accident severities by gender will be
extreme accidents are highly correlated with the number of fatalities in
quite informative and will help in designing adequate prevention po-
road accidents; recent studies revealed that the number of fatal acci-
licies. To this aim, we estimated various econometric models to assess
dents is significantly higher for male than female drivers. In China, men
the factors affecting the male and female drivers at risk. In this part, we
were at a 2.9–3.4 times higher risk of age-adjusted road traffic mortality
used a sample containing the drivers selected using the results of the
compared with women between 2006 and 201617. In addition, this
peak over threshold approach.
result is observed in Europe18 with an approximately similar propor-
Our empirical finding shows that gender is only a key determinant
tion, where the road fatality rate of men for every year between 2006
of the accident severity costs for the extreme accident severity. This
and 2017 was more than three times the rate for women. In the US, for
result is in parallel with the findings of some previous studies such as
nearly every year between 1975 and 2017, the number of men who died
Williams (2003) and Curry et al. (2012). In particular, we found that
in crashes was more than twice the number of women who died in
male drivers increased the likelihood of extreme risk taking. The ana-
crashes.19
lysis of the extreme accident distribution shows that male drivers incur
Table 5 also shows that the scale parameter is statistically sig-
70.5 percent of the extreme accidents, while females represent only
nificant at the 5 % significance level for all the studied subsets of data.
29.5 percent when considering a threshold of 8.5 %; the values are 70.8
It is established that the scale parameter is commonly associated with
percent for male versus 29.2 percent for female drivers with a 5 %
the variability of the data. Thus, the variability of the accident severity
threshold.
increases with female drivers. This indicates that there is heterogeneity
Analysing the results by gender, we find that three common vari-
of accident severities among female drivers.
ables explain the accident severity, at least at the 5 % or 8.5 %
After the estimation of several models using different threshold le-
thresholds, which are the price of the car, the use of the car and the
vels, we used diverse information criteria to select the best model that
number of seats. The unladen weight, driver age, number of accidents
fits our data. Particularly, the deviance and AIC criteria reported in
and premium are significant only for male drivers.
Table 5 show that for all the studied samples the suitable model is a
It is important to note that the significance of the premium paid by
GPD with a threshold of 8.5.
the policyholders for men under both thresholds indicate strong evi-
dence for the existence of adverse selection among high-risk men.
17
Wang et al. (2019), see references. Conversely, the hypothesis of adverse selection is rejected for the fe-
18
CARE (EU road accidents database) or national publications. male drivers that have extreme accidents. This last finding is not
19
NCSA data resource website, Fatality Analysis System (FARS) completely in accordance with those obtained by the quantiles on
Encyclopedia. quantile regression using the entire dataset.

10
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

These last estimation results of the factors explaining the accident

UCL (scale)
severity are obtained after using the extreme value theory as tool for
discriminating the low-medium from the high-risk drivers. This ap-

8.951
7.072
7.088
proach provides more accurate results compared to the quantile re-
gression. Indeed the used extreme value approach (POT) based on
LCL (scale)

probability distribution and approximations as well, helps to identify a

suitable and non-arbitrary threshold (in contrast with the quantile ap-
8.951
7.061
7.088

proach) that could be employed as an accurate measure of separation

between different classes of risk.
)
−0.591 (2.1e−06)
)
−06

−06

6. Developing a new safety model

−0.750 (2.1e

−0.594 (2.0e

As mentioned above, reducing the occurrence and the severity of

Shape

extreme accidents is an ambitious target that can be achieved only

through an intensive effort that should involve all major stakeholders,
including national and local governments, civil society, and the private
7.067 (3.6e−03)
7.088 (2.1e−06)
8.951 (2.1e-06)

sector. The insurance industry should also be a fundamental part of this

coalition. Already playing an important, if somewhat hidden, role in the
u3=10.5

road safety agenda, the industry insures nearly one billion vehicles
Scale

globally, helping to reduce the costs of road accidents to society and the
economy.
Improvements in road safety benefit the society as well as the in-
UCL (scale)

surance industry. Broad-based insurance coverage ensures that health

2.274
2.275
2.273

and property costs for victims of road accidents are protected, but it
also benefits insurance companies by expanding their market. Similarly,
reducing the number or severity of crashes benefits all of us, while it
LCL (scale)

also reduces the volume of claims to insurance firms.

Given this strong position, there is an opportunity for the insurance
2.243
2.240
2.256

industry to assume a leadership role in furthering the cause of safer

roads. Certainly, there are various areas where insurance firms can
make a significant contribution.
−0.325 (2.0e−06)
−0.323 (0.001)
−0.322 (0.001)

First, insurance companies can promote new technologies that in-

form road users about driving behaviour through applications and other
tools. Linked with insurance policies and risk-based pricing, this can be
Shape

Note. The table displays the scale and the shape of the GPD with various thresholds (8.5, 9.5 and 10.5).

a game-changer. Second, by providing data-driven knowledge and re-

search, the insurance companies can actively contribute to a better
Estimation results for the best model with the POT approach (u1 = 8.5, u2 = 9.5 and u3 = 10.5).

understanding of the causes of road crashes and of safe road behaviour.

The LCL and UCL are the lower and upper limits for the interval estimate of the scale parameter.
2.265 (5.2e−03)
2.258 (0.009)
2.257 (0.010)

Moreover, insurers may actively invest in raising public awareness

of risk factors for road users (e.g., speeding, alcohol use, absence of
u2=9.5

seatbelts, and seasonality), with a focus on vulnerable road users such

Scale

as children, motorbikes and non-motorized vehicles. The insurance

companies can also be partners with public authorities and civil society
The values within the parentheses are the standard deviations of the estimates.
UCL (scale)

by sharing expertise and developing innovative methods to develop and

finance road safety policies and programmes.
1.402
1.397
1.442

Finally, risk-based pricing is a powerful tool in the promotion of

personal responsibility, as it links the causing of an accident to the
economics of paying for those consequences, which in turn gradually
LCL (scale)

leads to safer driving and behaviour changes. Actually, a number of

1.370
1.363
1.386

insurance companies employ a no-claim bonus rate mechanism to

promote personal responsibility. However, this pricing strategy con-
siders only the number of accidents caused by the drivers. It neglects
−0.178 (0.004)
−0.175 (0.004)
−0.188 (0.007)

the importance of the severity and the extreme losses. In this regard, we
propose a new pricing framework that considers the presence of ex-
treme accidents and aims to promote personal responsibility, affecting
Shape

the driver’s behaviour and leading to safer driving.

We start by considering a simple model with two types of accident
severities (for the insurer, standard claim and large claims). For each
1.386 (0.009)
1.380 (0.010)
1.414 (0.017)

type of accident, a driver requires insurance coverage and prevention

against risks of accidents. In this framework, we will focus on two types
u1=8.5

Scale

of accidents i = {1,2} (the normal and extreme accidents, respectively).

The risk manager (e.g. insurer) will explore two datasets j = {T , M } ,
where T corresponds to the total portfolio and M is the subpopulation
Entire dataset

containing the risky drivers (the male group in our study). For each
group of drivers, the insurer faces a random variable Ii, j that reveals the
Female
Table 5

occurrence of normal accidents (low risk) or extreme accidents (high

Male

risk).

11
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Table 6
Estimation results of the robust least square models.
Male Female

5% 8.5 % 5% 8.5 %

Coef. Std. Coef. Std. Coef. Std. Coef. Std.

Intercept 10.89*** 0.192 10.458*** 0.220 11.454*** 28.443 10.50*** 0.220

Car price (ln) −0.045*** 0.021 −0.033** 0.020 −0.038 −1.023 −0.163 0.592
Use −0.170*** 0.074 −0.094 0.066 −0.241*** −2.043 −0.198*** 0.066
N Seats 0.016 0.013 −0.029*** 0.011 −0.017 −0.817 −0.053*** 0.002
Unladen weight 0.032*** 0.009 0.0078 0.008 0.010 0.881 −0.003 0.789
Age −0.096*** 0.015 −0.084*** 0.001 −0.003 −1.479 −0.002 0.176
Accident N −0.006 0.015 0.012 0.013 −0.019 −0.640 0.024 0.023
Premium (ln) 0.057*** 0.018 0.048** 0.019 0.017 0.478 0.038 0.180
Adjusted Rw-S 0.138 0.059 0.158 0.0097
AIC 1692.33 3997.07 710.43 3917.535
Schwarz 1734.95 4049.33 751.78 3960.818

Note. *** Significant level at 1 % **Significant level at 5 %. * Significant level at.10 %

To enhance road safety by reducing the occurrence of accidents, the importance of managing the expected extreme accidents. In this
underwriters must propose different contracts that consider the pre- modelling, we proposed only two types of driver categories high and
vention efforts of the potential policyholder. In the meantime, the low risk by gender. However, this model can be easily extended to a
driver as a potential policyholder will select the appropriate or optimal general model by considering multiple groups or risk classes.
contract.
The probability of an accident will depend on the prevention efforts 7. Conclusion
of the driver. We denote by j the prevention effort that is not observed
by the insurer. In this respect, a standard accident occurs with a This paper investigates factors explaining accident severity by using
probability p1j ( j) , while a fatal (extreme) accident occurs with a very a large sample of 405,177 observations. Three types of accidents factors
low probability equal to p2j ( j) = 1 p1j ( j) ; however, the fatal accident were explored, and three methods of analysis were used. The three
causes large losses for the drivers and the insurers. In terms of road types of accident factors are the vehicle-related factors, driver-related
safety, the drivers may reduce the expected loss value through their factors, and other factors such as those related to the insurance cov-
prevention activities. By increasing his or her prevention activities, the erage of the vehicle. The three econometric approaches employed are
insured reduces the probability of occurrence of major disasters and the standard OLS, the quantile regression and the extreme value theory.
increases the probability of standard accidents. Facing these two types The first approach consists of estimating a multiple regression model
of accidents and considering the prevention activities of the drivers, we relating the accident severity to the abovementioned factors. This es-
present our new ratemaking approach that will be helpful in enhancing timation is performed for the whole sample assuming homogeneity in
road safety by reducing the probability of extreme accidents. terms of distribution in the sample. The second approach, the quantile
According to the economic expected equilibrium in the insurance regression, has the advantage of assessing whether factors explaining
activity, any portfolio composed of n policyholders should be con- the accident severity are the same across quantiles and whether the
sidered under the following constraint: impact differs across quantiles. Finally, the third approach is based on
2 n 2 n the observation and statistical evidence of the existence of fatal acci-
ALij = Pmij dents in our sample. This final approach allows us to assess the de-
i=1 j =1 i=1 j terminants of accident severity for only the group of fatal accidents. The
obtained results are used to design and propose a prevention strategy to
where
reduce the occurrence and severity of extreme accidents.
ALij represents the accident severity for type i and for the policy-
The standard OLS results provide evidence of the significance of
holder j , and
certain variables in affecting the accident severity such as car price, use
Pmij represents the pure premium paid by the policyholder j to
of car, number of seats, unladen weight, driver age, and premium. The
cover the risk i .
only two factors that do not explain the accident severity are gender
Considering the presence of extreme accidents in the pricing
and number of accidents. Since the multiple regression based OLS
strategy, the optimal aggregated pure premium collected by the insurer
method assumes that the power of explanation of each factor is
will be obtained by solving the following:
homogenous for the whole sample, quantile regression was used to
2 unveil the possible existence of significant differences across quantiles.
j = ij Pmij The results show evidence of significant differences between most of the
i=1 estimated coefficients across quantiles. In particular, the results reveal
2
that the two coefficients of gender and number of accidents are sig-
UC. =1
ij nificant in some special cases and levels of significance. While the re-
i=1
sults show that the number of accidents is significant in three quantiles
where but only at 10 %, the gender coefficient is significant in the upper tail
j represents the aggregated pure premium, and (extreme values). The third econometric approach used focuses the
ij represents the coefficients representing the weights of different analysis on the extreme accident severities. The EVT approach shows
accident losses. evidence that the gender plays an important role in explaining the fatal
This new pricing strategy is useful for any insurance company accident severity. In particular, we found that male drivers increased
aiming to actively participate in enhancing road safety by considering the likelihood of extreme risk taking.

12
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Due to the high proportion of extreme accident severity in terms of Declaration of Competing Interest
“money” cost for the insurance company and the government, and so-
cial cost for the population, a new ratemaking method has been dis- The authors declare that they have no known competing financial
cussed and proposed. interests or personal relationships that could have appeared to influ-
ence the work reported in this paper.

Appendix A. Additional descriptive analysis of fatal accidents

Fig. A1 clearly shows the presence of extreme values for the entire portfolio as well as the different studied groups of drivers. This indicates
clearly the difference in term of losses between men and women.
The box plot shows the presence of the extreme accident severities in the insurance data. However, it did not allow for modelling the distribution
of the detected extreme observations.

Fig. A1. The box plot of the accident severity data.

Appendix B. Model selection results

Table B1.

Table B1
Model selection results.
Threshold Data Men Women

Deviance 8.5 77653.78 65667.7 26647.06

AIC 77657.78 65671.7 26651.06
Rank 1/3 1/3 1/3
Deviance 9.5 135779.4 113940.3 46535.22
AIC 135783.4 114944.3 46539.22
Rank 2/3 2/3 2/3
Deviance 10.5 237173.5 206231.1 83427.78
AIC 237177.5 206235.1 83431.78
Rank 3/3 3/3 3/3

Appendix C. Model diagnostic for GDP

See Fig. C1.

After fitting the extreme value model to the data, the next step is to assess and interpret the fitted model based on the quantile and return levels
computed using the inverse of the distribution function. We calculate and graphically present the return levels on an annual scale so that the one-
year return level is the level that is expected to be exceeded annually.

13
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

Fig. C1. Diagnostic plots of the fitted generalized Pareto distribution, threshold = 8.5, for the portfolio.

Appendix D. Supplementary data

Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.aap.2019.105411.

References Chiappori, P., Salanie, B., 2000. Testing for asymmetric information in insurance markets.
J. Political Economy 108 (1), 56–78. https://doi.org/10.1086/262111.
Cohen, A., 2005. Asymmetric information and learning: evidence from the automobile
Al Haji, G., 2005. Towards a road safety development index (RSDI). Development of an insurance market. Rev. Econ. Stat. 87 (2), 197–207. https://doi.org/10.1162/
international index to measure road safety performance. Linköping Studies in Science 0034653053970294.
and Technology, Licentiate Thesis, No. 1174. Development of Science and Conche, F., Tight, M., 2006. Use of CCTV to determine road accident factors in urban
Technology, Linköping University. areas. Accid. Anal. Prev. 38 (6), 1197–1207. https://doi.org/10.1016/j.aap.2006.05.
Abdel-Aty, M.A., Hassan, H.M., Ahmed, M., Al-Ghamdi, A.S., 2012. Real-time prediction 008.
of visibility related crashes. Transp. Res. Part C Emerg. Technol. 24, 288–298. Chiappori, P.A., Salanié, B., 2000. Testing for asymmetric information in insurance
Al-Ghamdi, A., 2002. Using logistic regression to estimate the influence of accident fac- markets. J. Polit. Econ. 108 (1), 56–78.
tors on accident severity. Accid. Anal. Prev. 34, 729–741. https://doi.org/10.1016/ Cordazzo, S., Scialfa, C., Jones, R., 2016. Modernization of the Driver Behaviour
S0001-4575(01)00073-2. Questionnaire. Accid. Anal. Prev. 87, 83–91. https://doi.org/10.1016/j.aap.2015.11.
Anselin, L., 2005. Exploring Spatial Data with GeoDa: A Workbook. Urbana. 016.
Austin, K., 1995. The identification of mistakes in road accident records: part 2, casualty Curry, A.E., Mirman, J.H., Kallan, M.J., et al., 2012. Peer passengers: how do they affect
variables. Accid. Anal. Prev. 27 (2), 277–282. https://doi.org/10.1016/0001- teen crashes? J. Adolesc. Health 50 (6), 588–594. https://doi.org/10.1016/j.
4575(94)00066-U. jadohealth.2011.10.016.
Balkema, A.A., de Haan, L., 1974. Residual life time at great age. Ann. Probab. 5, Dahlen, E., Edwards, B.D.E., Tubré, T., Zyphur, M.J., Warren, C.R., 2012. Taking a look
792–804. https://doi.org/10.1214/aop/1176996548. behind the wheel: an investigation into the personality predictors of aggressive
Benlagha, N., Grun-Réhomme, M., Vasechko, O., 2008. Les sinistres graves en assurance driving. Accid. Anal. Prev. 45, 1–9. https://doi.org/10.1016/j.aap.2011.11.012.
automobile: Une nouvelle approche par la théorie des valeurs extrêmes. Rev. Djebbari, H., Smith, J., 2008. Heterogeneous impacts in PROGRESA. J. Econometrics 145
MODULAD 39, 47–80. (1), 64–80. https://doi.org/10.1016/j.jeconom.2008.05.012.
Benlagha, N., Karaa, I., 2017. Evidence of adverse selection in automobile insurance Delen, D., Sharda, R., Bessonov, R., 2006. Identifying significant predictors of injury se-
market: a seemingly unrelated probit modelling. Cogent Econ. Financ. 5 (1), 1–15. verity in traffic accidents using a series of artificial neural networks. Accid. Anal.
https://doi.org/10.1080/23322039. Prev. 38 (3), 434–444.
Bergel-Hayat, R., Debbarh, M., Antoniou, C., Yannis, G., 2013. Explaining the road ac- de Boni, R., Bozzetti, M.C., Hilgert, J., et al., 2011. Factors associated with alcohol and
cident risk: weather effects. Accid. Anal. Prev. 60, 456–465. drug use among traffic crash victims in southern Brazil. Accid. Anal. Prev. 43 (4),
Bitler, M.P., Gelbach, J.B., Hoynes, H.W., 2006. What mean impacts miss: distributional 1408–1413. https://doi.org/10.1016/j.aap.2011.02.016.
effects of welfare reform experiments. Am. Econ. Rev. Insights 96 (4), 988–1012. Deffenbacher, J.L., Lynch, R., Oetting, E.R., Swaim, C., C.R, 2002. The driving Anger
https://doi.org/10.1257/aer.96.4.988. Expression Inventory: a measure of how people express their anger on the road.
Bolance, C., Guillen, M., Pelican, E., Vernic, R., 2008. Skewed bivariate models and Behav. Res. Ther. 40 (6), 717–737.
nonparametric estimation for CTE risk measure. Insurance: Math. Econ. 43, 386–393. Eling, M., 2011. Fitting insurance claims to skewed distributions: are the skew-normal
https://doi.org/10.1016/j.insmatheco.2008.07.005. and skew-student good models? Insurance Math. Econ. 51 (2), 239–248.
Brian, N.M., Andrey, J., Hambly, D., 2011. Analysis of precipitation-related motor vehicle Farmer, C.M., Braver, E.R., Mitter, E.L., 1997. Two-vehicle side impact crashes: the re-
collision and injury risk using insurance and police record information for Winnipeg, lationship of vehicle and crash characteristics to injuryseverity. Acc. Anal. Prev. 29
Canada. J. Saf. Res. 42 (5), 383–390. (3), 399–406.
Chen, T.Y., Jou, R.C., 2019. Using HLM to investigate the relationship between traffic Feng, S., Li, Z., Ci, Y., Zhang, G., 2016. Risk factors affecting fatal bus accident severity:
accident risk of private vehicles and public transportation. Transport. Res. Part A: their impact on different types of bus drivers. Accid. Anal. Prev. 86, 29–39.
Policy Pract. 119, 148–161. https://doi.org/10.1016/j.tra.2018.11.005. Fitzpatrick, C.D., Rakasi, S., Knodler, M.A., 2017. An investigation of the speeding-related
Chen, D., Müller, H.-G., 2012. Nonlinear manifold representations for functional data. crash designation through crash narrative reviews sampled via logistic regression.
Ann. Statist. 40 (1), 1–29. https://doi.org/10.1214/11-AOS936. Accid. Anal. Prev. 98, 57–63. https://doi.org/10.1016/j.aap.2016.09.017.
Chen, F., Wang, J., Yajuan Deng, Y., 2015. Road safety risk evaluation by means of im- Fishman, E., Cherry, C., 2016. E-bikes in the mainstream: reviewing a decade of research.
proved entropy TOPSIS–RSR. Saf. Sci. 79, 39–54. https://doi.org/10.1016/j.ssci. Transp. Rev. 36 (1), 72–91. https://doi.org/10.1080/01441647.2015.1069907.
2015.05.006. Fuller, D.L., 2014. Adverse selection and moral hazard: quantitative implications for
Chen, C., Zhang, G., Cathy Liu, X., Ci, Y., Huang, H., Ma, J., Chen, Y., Guan, H., 2016. unemployment insurance. J. Monet. Econ. 62–71, 108–122.
Driver injury severity outcome analysis in rural interstate highwaycrashes: a two- Gong, L., Li, H., Wang, D., 2012. Health investment, physical capital accumulation, and
level Bayesian logistic regression interpretation. Accid. Anal. Prev. 97, 69–78. economic growth. China Econ. Rev. 23 (4), 1104–1119. https://doi.org/10.1016/j.

14
N. Benlagha and L. Charfeddine Accident Analysis and Prevention 136 (2020) 105411

chieco.2012.07.002. Adolescent drivers’ perceptions about inattention and peer passengers. J. Pediatr.
González Dan, J.R., Arnaldos, J., Darbra, R.M., 2017. Introduction of the human factor in Nurs. 31, e375–e382.
the estimation of accident frequencies through fuzzy logic. Saf. Sci. 97, 134–143. Mcgill, R., Tukey, J.W., Larsen, W.L., 1978. Variations of box plots. Am. Statistician 32
https://doi.org/10.1016/j.ssci.2015.08.012. (1), 12–16. https://doi.org/10.1080/00031305.1978.10479236.
Grun-rehomme, M., Benlagha, N., Vasyechko, O., 2007. Une approche locale de la gestion Mills, B.N., Hambly, D., 2011. Analysis of precipitation-related motor vehicle collision
des sinistres graves en assurance automobile. Insurance Risk Manage. Montreal 75 and injury risk using insurance and police record information for Winnipeg, Canada.
(3), 409–430. J. Safety Res. 42 (5), 383–390. https://doi.org/10.1016/j.jsr.2011.08.004.
Guo, F., Fang, Y., 2013. Individual driver risk assessment using naturalistic driving data. Miyajima, C., Nishiwaki, Y., Ozawa, K., Wakita, T., 2007. Driver modeling based on
Accid. Anal. Prev. 61, 3–9 10.1016/j.aap.2012.06.014. driving behavior and its evaluation in driver identification. Proc. IEEE 95, 427–437.
Haddon Jr., W., 1968. The changing approach to the epidemiology, prevention, and Papantoniou, P., Yannis, G., Christofa, E., 2019. Which factors lead to driving errors? A
amelioration of trauma: the transition to approaches etiologically rather than de- structural equation model analysis through a driving simulator experiment. IATSS
scriptively based. Am. J. Public Health 58, 1431–1438. Res. 43 (1), 44–50.
Hassan, H.M., Abdel-Aty, M.A., 2013. Predicting reduced visibility related crashes on Pérez-Marín, A.M., Guillen, M., 2019. Semi-autonomous vehicles: usage-based data evi-
freeways using real-time traffic flow data. J. Safety Res. 45, 29–36. dences of what could be expected from eliminating speed limit violations. Accid.
Hao, W., Kamga, C., Janice Daniel, J., 2015. The effect of age and gender on motor Anal. Prev. 123, 99–106.
vehicle driver injury severity at highway-rail grade crossings in the United States. J. Ouyang, Q., Chao Wu, C., Huang, L., 2018. Methodologies, principles and prospects of
Safety Res. 55, 105–113. applying big data in safety science research. Saf. Sci. 101, 60–71. https://doi.org/10.
Huang, H., Yin, Q., Schwebel, D.C., Li, L., Hu, G., 2016. Examining road traffic mortality 1016/j.ssci.2017.08.012.
status in China: a simulation study. PLoS One 11 (4), e0153251. https://doi.org/10. Shannon, D., Murphy, F., Mullins, M., Eggert, J., 2018. Applying crash data to injury
1371/journal.pone.0153251. claims - an investigation of determinant factors in severe motor vehicle accidents.
Huang, H., Yin, Q., Schwebel, D.C., Ning, P., Hu, G., 2017. Availability and consistency of Accid. Anal. Prev. 113, 244–256. https://doi.org/10.1016/j.aap.2018.01.037.
health and non-health data for road traffic fatality: analysis of data from 195 coun- Shao, B., Zhigen Hu, Z., Liu, Q., Chen, S., He, W., 2019. Fatal accident patterns of building
tries, 1985–2013. Accid. Anal. Prev. 108, 220–226. https://doi.org/10.1016/j.aap. construction activities in China. Saf. Sci. 111, 253–263.
2017.08.033. Shaw, J.I., McMartin, J.A., 1977. Personal and situational determinants of attribution of
Huang, H., Chang, F., Schwebel, D.C., Ning, P., Cheng, P., Hu, G., 2018a. Improve traffic responsibility for an accident. Hum. Relat. 30 (1), 95–107. https://doi.org/10.1177/
death statistics in China. Science. Insights, Letters. 362 (6415), 650. 001872677703000106.
Huang, L., Wu, C., Wang, B., Ouyang, Q., 2018b. Big-data-driven safety decision-making: Shi, Q., Abdel-Aty, M., 2015. Big data applications in real-time traffic operation and
a conceptual framework and its influencing factors. Saf. Sci. 109, 46–56. https://doi. safety monitoring and improvement on urban expressways. Transp. Res. Part C
org/10.1016/j.ssci.2018.05.012. Emerg. Technol. 58 (Part B), 380–394. https://doi.org/10.1016/j.trc.2015.02.022.
Jiang, B., Liang, S., Peng, Z., Cong, H., Levy, M., Cheng, Q., Wang, T., Remai, J., 2017. Staubach, M., 2009. Factors correlated with traffic accidents as a basis for evaluating
Transport and public health in China: the road to a healthy future. Lancet. 390, Advanced Driver Assistance Systems. Accid. Anal. Prev. 41 (5), 1025–1033.
1781–1791. Pickands, J., 1975. Statistical inference using extreme order statistics. Ann. Statist. 3 (1),
James, P.T., Baldock, M.R.J., Dutschke, J.K., 2018. Trends in the crash involvement of 119–131. www.jstor.org/stable/2958083.
older drivers in Australia. Accid. Anal. Prev. 117, 262–269. Pisarenko, V., Mikhail, R., 2010. Heavy-Tailed Distributions in Disaster Analysis.
Jackson, E., Page, M.E., 2013. Estimating the distributional effects of education reforms: a Springerhttps://doi.org/10.1007/978-90-481-9171-0.
look at Project STAR. Econ. Educ. Rev. 32, 92–103. https://doi.org/10.1016/j. Sârbescu, P., Maricuţoiu, L., 2019. Are you a “bad driver” all the time? Insights from a
econedurev.2012.07.017. weekly diary study on personality and dangerous driving behavior. J. Res. Pers. 80,
Jin, W., Deng, Y., Jiang, H., Xie, Q., Shen, W., Weijian, H., 2018. Latent class analysis of 30–37.
accident risks in usage-based insurance: evidence from Beijing. Accid. Anal. Prev. Shinar, D., Treat, J.R., McDonald, S.T., 1983. The validity of police reported accident
115, 79–88. data. Accid. Anal. Prev. 15 (3), 175–191.
Karaa, I., Benlagha, N., 2015. Testing for Asymmetric Information in Tunisian Automobile Shi, J., Xiao, Y., Atchley, P., 2016. Analysis of factors affecting drivers’ choice to engage
Insurance Market. Mediterr. J. Soc. Sci. 6 (3), 455–464. https://doi.org/10.5901/ with a mobile phone while driving in Beijing. Transp. Res. Part F Traffic Psychol.
mjss.2015.v6n3s1p455. Behav. 37, 1–9. https://doi.org/10.1016/j.trf.2015.12.003.
Keane, M., Stavrunova, O., 2016. Adverse selection, moral hazard and the demand for Stylianou, K., Dimitriou, L., Abdel-Aty, L., 2019. Big data and Road safety: a compre-
Medigap insurance. J. Econom. 190 (1), 62–78. hensive review. In: Antoniou, C., Dimitriou, L., Pereira, F. (Eds.), Mobility Patterns,
Kingphai, K., Chongcharoen, S., 2016. Estimating the claim severity distribution using Big Data and Transport Analytics, https://doi.org/10.1016/B978-0-12-812970-8.
variable neighborhood search. Am. J. Appl. Sci. 13, 1400–1406. https://doi.org/10. 00012-9.
3844/ajassp.2016.1400.1406. Theofilatos, A., 2017. Incorporating real-time traffic and weather data to explore road
Kitali, A.E., Thobias Sando, P.E., 2017. A full Bayesian approach to appraise the safety accident likelihood and severity in urban arterials. J. Safety Res. 61, 9–21.
effects of pedestrian countdown signals to drivers. Accid. Anal. Prev. 106, 327–335. Tešić, M., Hermans, E., Lipovac, K., Pešić, D., 2018. Identifying the most significant in-
https://doi.org/10.1016/j.aap.2017.07.004. dicators of the total road safety performance index. Accid. Anal. Prev. 113, 263–278.
Koenker, R., 2005. Quantile Regression, Econometric Society Monographs, n. 38. https://doi.org/10.1016/j.aap.2018.02.003.
Cambridge University Press, New York. Vernic, R., 2006. Multivariate skew-normal distributions with applications in insurance.
Koenker, R., Bassett, G., 1978. Regression quantiles. Econometrica 46 (1), 33–50. https:// Insurance: Math. Econ. 38 (2), 413–426. https://doi.org/10.1016/j.insmatheco.
doi.org/10.2307/1913643. 2005.11.001.
Krishnan, K.S., Carnahan, J.V., 1985. Analysis of the effect of car size on accident injury Wang, L., Ning, P., Yin, P., 2019. Road traffic mortality in China: analysis of national
probability using automobile insurance data. Accid. Anal. Prev. 17 (2), 171–177. surveillance data from 2006 to 2016. Lancet Public Health 4, e245–e255. https://doi.
Lane, M.N., 2000. Pricing risk transfer transactions. ASTIN Bull. 30 (2). org/10.1016/S2468-2667(19)30057-X.
Lee, O., Chae, J., Yoon, T., Hojin Yang, H., 2018. Traffic accident severity analysis with Wang, L., Shi, Q., Abdel-Aty, M., 2015. Predicting crashes on expressway ramps with real-
rain-related factors using structural equation modeling – a case study of Seoul City. time traffic and weather data. Transp. Res. Rec. J. Transp. Res. Board. 32–38. https://
Accid. Anal. Prev. 112, 1–10. doi.org/10.3141/2514-04.
Li, Q., He, H., Liang, H., Bishai, D.M., Hyder, A.A., 2016. One outcome, many trends: Williams, A., 2003. Teenage drivers: patterns of risk. J. Saf. Res. 34, 5–15. https://doi.
understanding national data sources for road traffic fatalities in China. Am. J. Public org/10.1016/S0022-4375(02)00075-0.
Health 106 (10), 1793–1795. https://doi.org/10.2105/ajph.2016.303287. Xiao, W., Ning, P., Schwebel, D.C., D.C, H., G, 2017. Evaluating the effectiveness of im-
Li, C.-S., Liu, C.-C., Peng, S.-C., 2013. Bundled automobile insurance coverage and acci- plementing a more severe drunk-driving law in China: findings from two open access
dents. Accid. Anal. Prev. 50, 64–72. https://doi.org/10.1016/j.aap.2012.03.026. data sources. Int. J. Environ. Res. Public Health 14 (8), 832. https://doi.org/10.
Li-Lu, S., Dan, L., Tian, C., Meng-Ting, H., 2019. Analysis on the accident casualties in- 3390/ijerph14080832.
fluenced by several economic factors based on the traffic-related data in China from Zhang, Z., He, Q., Tong, H., Gou, J., Li, X., 2016. Spatial-temporal traffic flow pattern
2004 to 2016. Chin. J. Traumatology 22 (2), 75–79. https://doi.org/10.1016/j.cjtee. identification and anomaly detection with dictionary-based compression theory in a
2019.02.002. large-scale urban network. Transport. Res. Part C: Emerging Technol. 71, 284–302.
Litman, T., 2005. Pay-As-You-Drive pricing and insurance regulatory objectives, Journal Zhang, G., Yaub, K.W., Chen, G., 2013. Risk factors associated with traffic violations and
of Insurance Regulation. National Association of Insurance Commissioners. 23, 3. accident severity in China. Accid. Anal. Prev. 59, 18–25.
Loo, B.P.Y., 2018. Unsustainable Transport and Transition in China, 1st edition. Yang, J., Corinne, P.A., Gang, C., Erin, H., Scott, F., Marizen, R., 2009. Incidence and
Routledgehttps://doi.org/10.4324/9781315677941. characteristics of school bus crashes and injuries. Accid. Anal. Prev. 41, 336–341.
Ma, Z., Shao, C., Ma, S., Ye, Z., 2011. Constructing road safety performance indicators https://doi.org/10.1016/j.aap.2008.12.012.
using fuzzy delphi method and Grey Delphi Method. Expert Syst. Appl. 38 (3), Zeng, K.-H.C., Chan, F.-H., Niebles, J.C., Sun, M., 2017. Agent-centric risk assessment:
1509–1514. https://doi.org/10.1016/j.eswa.2010.07.062. accident anticipa-tion and risky region localization. IEEE Conference on Computer
Manner, H., Wünsch-Ziegler, L., 2013. Analyzing the severity of accidents on the German Vision and Pattern Recognition (CVPR) 2–7.
Autobahn. Accid. Anal. Prev. 57, 40–48. WHO, 2015. World Health organization report.
McDonald, C.C., Sommers, M.S., 2016. “Good passengers and not good passengers:”

15