Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Accident Analysis and Prevention 154 (2021) 106054

Contents lists available at ScienceDirect

Accident Analysis and Prevention


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

Factors associated with consecutive and non-consecutive crashes on


freeways: A two-level logistic modeling approach
Zhou Zichu a, Meng Fanyu b, c, *, Song Cancan a, Tay Richard d, Guo Zhongyin a, Yang Lili c,
Wang Weili a, e
a
Key Laboratory of Road and Traffic Engineering of the Ministry of Education, Tongji University, Shanghai, China
b
Academy for Advanced Interdisciplinary Studies, Southern University of Science and Technology, Shenzhen, China
c
Department of Statistics and Data Science, Southern University of Science and Technology, Shenzhen, China
d
School of Business IT and Logistics, RMIT University, Melbourne, Australia
e
Guizhou Transportation Planning Survey & Design Academy Co., Ltd, Guiyang, China

A R T I C L E I N F O A B S T R A C T

Keywords: A consecutive crash consists of a primary crash and one or more secondary crashes that occur subsequently in a
Consecutive crash short period of time within a certain distance. It often affects a relatively large area of road space and the traffic
Traffic hazard disruption created can be difficult for traffic managers to control and resolve. This study identifies the factors
Multi-level model
delineating a primary crash that results in secondary crashes within a minute from a regular crash that does not
Unobserved heterogeneity
Road safety
result in any secondary crashes. Random-effects, random-parameter and two-level binary logistic regression
models are applied to data collected on 8779 crashes on the freeway network of the Guizhou Province, China in
2018, of which 299 are consecutive crashes. According to the AIC values, the two-level logistic model out­
performs the other two models. Rear-end primary crashes have a significant random effect varying across road
segments on the occurrence of consecutive crashes. Various crash types (rear-end, roll-over and side-swipe),
tunnel crash and foggy weather are positively associated with the possibility to cause subsequent consecutive
crashes, whereas single-vehicle crash, truck involvement and the time periods with poorer natural lighting are
less likely to incur consecutive crashes. Recommendations are provided to minimize the possibility of the
occurrence of consecutive crashes on a freeway.

1. Introduction more burden to the traffic system such as congestion and subsequent
traffic crashes.
The increasing number of traffic collisions worldwide has been a Previous studies on secondary crashes have primarily focused on the
topic of interest to road safety researchers and practitioners for decades. occurrence and frequency of this type of crash. Various studies have
Factors contributing to crash risks under various circumstances have explored the occurrence of such crashes and its possible contributing
been identified for different types of crashes (Anastasopoulos et al., factors using logistic regressions (Karlaftis et al., 1999; Zhan et al., 2009;
2012; Wang and Feng, 2019; Yu and Abdel-Aty, 2013; Hong et al., 2019; Wang et al., 2016). Yang et al. (2014) and Xu et al. (2016) have gone
Hosseinpour et al., 2018; D. Lord and Mannering, 2010; Meng et al., beyond basic logistic regressions and incorporated its variants to explore
2017a, b; Rusli et al., 2018; Wang and Feng, 2019). rare events and unobserved heterogeneities, respectively. The occur­
Among the different types of crashes, the occurrence of secondary rence of secondary crashes is significantly dependent on the duration of
crashes is a special concern because of its impact on freeway operations, the primary crash, queue length, traffic condition and road geometric
cumulative social and economic costs, and potential preventability design (Yang et al., 2018). As for frequency of secondary crashes, mul­
(Yang et al., 2014; Vlahogianni et al., 2010). A secondary crash is tiple count data models on an aggregated level have been estimated and
defined as a crash occurring spatially and temporally within the impact compared, and factors including AADT, traffic composition, land use,
area of a primary crash (Owens, 2010). It occurs under the influence of number of lanes, and speed limit have been found to be positively
the hazards created by a primary crash downstream and may impose associated with secondary crash frequencies (Sarker et al., 2017).

* Corresponding author at: Academy for Advanced Interdisciplinary Studies, Southern University of Science and Technology, Shenzhen, China.
E-mail address: fymeng91@gmail.com (M. Fanyu).

https://doi.org/10.1016/j.aap.2021.106054
Received 6 June 2020; Received in revised form 7 October 2020; Accepted 19 February 2021
Available online 2 March 2021
0001-4575/© 2021 Elsevier Ltd. All rights reserved.
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

By definition, secondary crashes are mainly caused by the queueing regression models are applied to data collected on 8779 crashes on the
effects generated from the primary crash, and hence the duration of the freeway network of Guizhou Province, China in 2018. To avoid biased
primary crash significantly affects the occurrence of secondary crashes results caused by unobserved spatial heterogeneity and hierarchical
(Yang et al., 2018). The spatiotemporal thresholds for identifying sec­ structure in the data, a random-effects (RE) binary logistic model, a
ondary crashes are relatively high, as cumulative effects are permitted random parameter (RP) binary logistic model and a two-level logistic
spatially and temporally in this case. Most previous studies have defined model are estimated, and their performances in predicting the likelihood
secondary crashes by providing fixed spatial-temporal thresholds, and of causing CCs are compared. Based on the results of the best fitting
the distance between the primary and secondary crashes is usually model, policy implications and recommendation are provided to prevent
longer than 2 miles, while the time span between the two crashes can be and mitigate the effects of CCs.
larger than 2 h (Latoski et al., 1999; Moore et al., 2004; Green et al.,
2012). Several studies adopted dynamic identification of secondary 2. Data
crashes based on either speed contour maps (Goodall, 2017; Park et al.,
2018) or shockwave models (Zheng et al., 2014; Mishra et al., 2016), in The crash database used in this study was provided by Guizhou
which queuing effects were still the fundamental cause of the secondary Traffic Information and Emergency Control Center, affiliated with the
crashes. Department of Transportation of the Guizhou Province. The data were
In previous research, the occurrence of an immediate secondary originally recorded by onsite traffic police officers for all crashes on
crash that ensues in a considerably short period of time within a certain national, provincial and local freeways in the province. Data on the
vicinity of the primary crash has not received enough attention. When speed limit and monthly traffic volume in 2018 were provided by the
the time span between the two crashes is short enough, the traffic queue Department of Transportation of the Guizhou Province.
will not develop into a significant problem, and the clearance time of the The definition of CC in this study refers to the definition of secondary
primary crash is less of an issue. Under this circumstance, congestion crash with fixed space and time thresholds, taking into consideration the
and the queuing effect will not play a major role in the occurrence of a differences in the mechanisms between these two types of crashes. As
secondary crash. queuing effect should not play a role in CCs, considerably shorter
The hazards produced by the primary crash mainly transmit up­ thresholds have been incorporated in defining CCs (Meng et al., 2020;
stream through vehicular behaviors and interactions, similar to the Yang et al., 2018). Hence, in this study, CCs were identified using two
mechanism of a chain-reaction crash (multiple-vehicle crash). This type conditions: (1) more than one crashes took place within 1 km in the
of immediate secondary crash is defined as a consecutive crash (CC) in same direction of the freeway within 1 min, and (2) primary crash
this study. A primary crash can induce multiple CCs within a short occurred on the downstream traffic of all subsequent crashes. In total,
period of time (named a CC series). This type of crash is distinctively there were 299 CCs on Guizhou freeways in 2018, with an average of
different from a chain-reaction crash, which is a single crash comprised 2.51 secondary crashes per primary crash. To identify factors associated
of a series of collisions happening over a short period time. Rather, a CC with the occurrences of CCs, data on 8480 other crashes that neither
series is a group of different crashes, each of which can involve one or caused nor were involved in any CCs were also extracted. Hence, there
multiple vehicles. were 8779 observations used in this study. For a CC, only the primary
Given the similarities between CCs and chain-reaction crashes, the crash was used in the analysis.
crash mechanisms of chain reaction crashes can be used as a starting Besides information on the crash characteristics and vehicles
block to analyze CCs. Sugiyama and Nagatani (2013) and Nagatani and involved, the crash database also included information on several
Yonekura (2014) posit that chain-reaction crashes can be caused by a environmental factors, such as the date, time, weather and location of
sudden slowdown and a lane change of vehicles in the lead. Simulation each crash. In this study the time of crash was recoded into six cate­
results from vehicle-following models showed that factors including gories: dawn (3 a.m. to 6.59 a.m.), morning (7 a.m. to 10.59 a.m.), noon
traffic conditions (headway and relative velocity) and the driver’s re­ (11 a.m. to 2.59 p.m.), afternoon (3 p.m. to 6.59 p.m.), evening (7 p.m.
action speed affected the occurrences of chain-reaction crashes (Naga­ to 10.59 p.m.) and night (11 p.m. to 2.59 a.m. of the next day). These
tani, 2015). categories were used in several previous studies (Meng et al., 2017a, b;
Besides the simple chain-reaction crashes, traffic hazards may also Pei et al., 2012).
spread via vehicular interactions in a discrete and irregular manner Crashes at special locations of the freeway (tunnel, ramp and open
based on traffic wave theory. Considering the differing driver alertness road) were assumed to create different risks for the occurrence of CCs
and swiftness in reaction, a hazard on the road may induce different compared to those happening on an open road because the driving en­
behaviors in upstream vehicles. Consequently, the hazard can still vironments at these locations might have different effects on drivers’
spread backwards to the following vehicles, but its influence on the hazard perception and reaction.
following vehicles’ behaviors may fluctuate when passing from one The severity of the crashes was extracted from the crash database and
vehicle to another. Thus, collisions may not happen to adjacent vehicles adopted as an indicator variable, as it includes comprehensive infor­
in a queue. Instead, a primary collision may trigger a secondary collision mation at the time the collision occurs such as the hitting mechanisms,
from several vehicles away. This effect may be magnified on roads with speed before collision and/or at-fault driver information (Alogaili and
high speed and busy traffic, such as freeways (Nagatani, 2015; Nagatani Mannering, 2020; Liu and Khattak, 2017; Duddu et al., 2018; Wang
and Yonekura, 2014), resulting in a higher risk of CCs. et al., 2019; Wen and Xue, 2020; Zhou and Chin, 2019). Crash severity is
CCs may hinder the traffic worse than simple chain-reaction crashes, defined as a dichotomous variable: 1 for crash causing at least one injury
as it consists of several crashes (each of which can involve one or mul­ and 0 for crash with property damage only (PDO).
tiple vehicles) and, unlike a simple multiple-vehicle crash, it often in­ In this study, vehicles were classified into two main categories: (1)
volves a larger area of the road and creates a greater disruption to the light vehicles which included two- and three-wheels vehicles, passenger
traffic system. Evacuation and emergency management of multiple and cars and light commercial vehicles, and (2) heavy vehicles which
separate crashes occurring in a limited road space are also relatively included trucks, buses and trailer trucks. Heavy vehicles deserve more
more cumbersome. Despite these challenges, little research has been attention in this study as these vehicles are large in sizes and tend to
conducted to identify the factors contributing to the occurrences of this induce more and propagate the hazardous driving conditions further to
type of crashes. vehicles following behind them. Note that a crash may involve more
This study aims to identify the factors delineating a primary crash than one vehicle and all vehicles involved are considered in this study.
that resulted in a CC from a regular crash that did not result in any CCs. Crash-related characteristics used in the models included crash type
Random-effects, random parameter and two-level binary logistic and number of vehicles. These factors would represent different types of

2
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

hazards that a primary crash might induce, resulting in different



k
perception and reaction of drivers in vehicles following behind. The logit(πij ) = log(
πij
) = β0 + βk Xijk + εij (1)
base-line crash type (other crash types) consisted mainly of mild single- 1 − πij 1
vehicle crash types such as hitting freeway fixtures and slipping into
road-side ditches. where Xijk is the value of the k th independent variable for crash i on road
The descriptive statistics of the variables incorporated in the models segment j, β0 is the intercept of the model, βk is the estimated coefficient
are reported in Table 1. The mean and standard deviations (SDs) for for Xijk , and εij is the random error term following a logistic distribution.
continuous variable (traffic volume) are provided. Categorical variables
were coded into dummy variables, and the percentage of each category 3.2. Random-effects logistic model
were displayed.
Unobserved heterogeneity has been shown to widely exist in road
3. Methods crash modeling (Mannering and Bhat, 2014; Mannering et al., 2016;
Anastasopoulos and Mannering, 2009; Meng et al., 2019; Rusli et al.,
3.1. Binary logistic regression 2018). To account for this unobserved heterogeneity, a random effects
(RE) modeling approach is adopted by incorporating a random intercept
The aim of this study is to identify the factors delineating a primary term, νi , with a mean of 0 and a standard deviation of σ ν . The RE logistic
crash that results in secondary crashes within a minute from a regular function is described as follows:
crash that does not result in any secondary crashes. Since the outcome
πij ∑
k
variable is binary, the widely used the binary logistic regression is a logit(πij ) = log( ) = β0 + βk Xijk + νi + εij (2)
suitable method to analyze the data (Chin and Huang, 2009; Meng et al., 1 − πij 1

2017a,b; Nguyen-Phuoc et al., 2019; Theofilatos et al., 2019). A maximum likelihood estimation method is applied to estimate the
Let Yij denote the outcome of crash i on road segment j. Yij = 1 if coefficients, β0 and βk (McFadden, 1973). In addition to the variance, σ2ν ,
crash i is a primary crash and Yij = 0 if crash i is not a primary crash. The the proportion of variance explained by the panel effect is also
probability of Yij = 1 (denoted as πij ) is given by (McFadden, 1973): estimated:

Table 1
Descriptive statistics.
Variable name Description Mean/ Percentage SD Minimum Maximum

Dependent variable
Yes 3.4
Primary crash of consecutive crashes
No 96.6
Independent variables
Continuous variables:
Hourly traffic volume Hourly flow rate of the hour that the crash occurs, in thousand vehicles per hour 2.3 1.22 0.05 5.75
Categorical variables (%):
Crash severity Severe crash 1.2
Property damage only (base) 98.8
Speed limit 120 km/h 86.4
100 km/h 5.7
80 km/h (base) 7.9
Road geometry Curved line (radius>2000 m) 16.7
Curved line (500 m ≤ radius<2000 m) 35.1
Curved line (radius<500 m) 4.4
Straight line (base) 43.8
Location Tunnel 5.3
Ramp 4.4
Others (base) 90.3
Time period Dawn (3:00− 6:59) 6.0
Morning (7:00− 10:59) 16.8
Noon (11:00− 14:59, base) 24.5
Afternoon (15:00− 18:59) 25.4
Evening (19:00− 22:59) 18.1
Night (23:00− 2:59) 9.2
Truck involved in crash 26.6
Truck involvement
No truck involved 73.4
Trailer truck involved in crash 7.0
Trailer truck involvement
No trailer truck involved 93.0
Bus involvement Bus involved in crash 1.0
No bus involved 99.0
Crash type Rear-end 36.7
Rollover 5.5
Side-swipe 4.7
Others (base) 53.0
Weather Rainy 19.1
Foggy 62.5
frozen 1.5
Sunny (base) 16.9
Single-vehicle 58.1
Number of vehicles Double-vehicle 29.8
Multi-vehicle (base) 12.2

3
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

σ 2ν 3.5. Model comparison


ρ= (3)
σ ν + σ2ε
2

In this study, the Akaike Information Criterion (AIC) is used to select


where σ 2ε is the variance of εij . the best fitting model. AIC reflects the amount of information loss in
Note that the proportion of consecutive crashes is relatively small. each model (Akaike, 1973) and is estimated by:
This skewness can be addressed using an asymmetric distribution for the ̂
AIC = 2n − 2ln( L) (8)
error term (e.g., Burr-10 which gives rise to the skewed logistic model;
Tay, 2016; Jiang et al., 2021) or random intercept term. In study, a
where n is the number of estimated parameters, and ̂L is the maximum
random intercept with several different distributions, including asym­
value of the likelihood function for the model. The model with the
metric distribution such as the lognormal distribution, were tested and
lowest AIC value has the best performance in terms of using as few pa­
the model with the best fit was reported.
rameters as possible to maximize the likelihood function of the model.

3.3. Random parameter logistic model 4. Results

The RE model is sometimes considered as a special case of a random To identify the significant factors causing a series of CCs, three
parameter (RP) model where only the constant term is assumed to be models were estimated using the RE, RP and two-level binary logistic
random (Washington et al., 2020). In the RP model, the estimated co­ regressions. The same dataset was used in all models with the same sets
efficients for one or more independent variables are assumed to vary of dependent and independent variables. Before the regressions were
across the observations. The RP logistic function is given by: performed, Pearson correlation tests were conducted for all pairs of in­
dependent variables. All Pearson correlation values were smaller than
βik = βk + μik (4) 0.6, implying that there was no significant correlation among the in­
dependent variables. To check whether specification errors exist in the
where βik is the k th estimated coefficient for observation i, βk is the model, Ramsey’s Regression Equation Specification Error Test (i.e.,
estimated mean of the k th coefficient, and μik is a random term RESET test) have been conducted (Ramsey, 1969; Gujarati, 2009;
following a normal distribution with a mean of zero and a variance of σ2k . Ramalho and Ramlho, 2012; Tay, 2005a, b, 2003). The result of the
A simulated maximum likelihood estimation method with 200 Hal­ RESET test didn’t reject the null hypothesis, indicating that there was no
ton draws is applied to estimate the coefficients, βk and σk (McFadden, significant bias resulting from the omission of significant independent
1973). A z-test is applied to each estimated coefficient, and only co­ variables or other general specification bias in the model
efficients with a significant mean, βk , and a significant standard devia­ For the RE model, the S.D. of the random term, σ ν , and the proportion
tion (S.D.), σ k , are retained in the model as random coefficients. of variance, ρ, were both estimated. Variables with insignificant coeffi­
cient at the 90 % have been removed. As shown in Table 2, 10 estimated
parameters were significant at the 95 % confidence level, among which
3.4. Two-level logistic model 4 were positive and 6 were negative, with an additional one variable
being marginally significant at 90 % confidence level. The standard
To address the issue of spatial heterogeneity across road segments, a deviation of the random intercept was insignificant, and the ρ value was
two-level modeling scheme is adopted (Huang and Abdel-Aty, 2010). In
very low (2.55 × 10− 4 ), indicating that there was little or no
this study, level 1 is the crash and level 2 is the road-segment. The basic
heterogeneity.
binary logistic function can be rewritten in a two-level scheme as
To establish a RP logistic model, all parameters (including the con­
follows:
stant) were initially set to be random. Those with a significant mean and
( )
(
πij
) ∑K a significant S.D. at the 0.05 level were included in the next step as
logit πij = log = β0j + β1jk Xijk + εij (5) random parameters, and the rest of the parameters were set to be fixed.
1 − π ij
For the fixed parameters, only those whose estimated coefficient were
k=1

where β0j is the crash-level intercept and β1jk is the estimated coef­ significant at the 90 % level were kept. The final results of the RP model
ficient for Xijk . The road-segment-level model is specified as: were shown in Table 2. The parameter for “single-vehicle crash” were
the only random parameter in the RP model, whose mean and S.D. were

L
both significant at the 0.001 level, and the rest were all fixed parame­
β0j = γ00 + γ 0l Zjl + μ0j (6)
l=1 ters. 10 estimated fixed parameters were significant at the 95 % confi­
dence level while one additional variable was marginally significant at
β1jk = γ1k + μ1jk (7) the 90 % confidence level.
The estimation results of the two-level logistic model are shown in
where γ 00 and γ 1k are estimated intercepts on the road segment level; Zjl Table 2. The estimated random term of “rear-end” was significant at the
is l th independent variable representing environmental factors of road 95 % confidence level. All insignificant fixed coefficients have been
segment j, and γ 0l is the estimated coefficient for Zjl ; μ0j and μ1jk are the removed from the final model. Ten fixed coefficients were significant at
random effects varying across road segments for the crash-level inter­ the 95 % confidence level and one additional variable was marginally
cept and crash-level covariate k with means of zero and variances of σo 2 significant at the 90 % confidence level. For all categorical variables
and σ k 2 , respectively (Snijders and Bosker, 1999). Note that the random included in the model, at least one category was significant at the 95 %
effects, μ0j and μ1jk , are random across road segments and constant for all confidence level.
crashes on the same road segment, which enables unobservable spatial In terms of goodness-of-fit measures, the two-level model achieved a
effects to vary among the road segments (Kim et al., 2007). larger log-likelihood value (-800.22) compared with the RP model
A simulated maximum likelihood estimation method with 200 Hal­ (-1167.73) and the RE model (-1177.00). Moreover, the AIC value of the
ton draws is applied to estimate the coefficients (McFadden, 1973; two-level model (1632.4) were much smaller than those of the RP model
Train, 2009). A z-test was applied to each estimated coefficient to ac­ (2365.5) and the RE model (2378.0). Therefore, the two-level model
quire the statistical significance level. The road segments in this study is outperformed the RP model and the RE model by addressing the hier­
defined by the Transport Department for management purpose and vary archical nature of the data and accounting for the within-road-segment
between 40− 80 km in length. correlations. Hence, marginal effects were calculated for the two-level

4
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

Table 2
Estimation results for RE, RP and two-level regressions.
Variables RE model RP model Two-level model

Coefficient Std Err p-value Coefficient Std Err p-value Coefficient Std Err p-value

Fixed parameters
Road geometry:
- Curved line (radius>2000 m) − 0.694*** 0.200 0.001 − 0.080*** 0.231 0.001
Crash severity:
- Severe crash 1.573*** 0.783 0.045
Speed limit:
- 120 km/h 1.397*** 0.457 0.002
Location:
- Tunnel 0.902*** 0.183 0.000 1.102*** 0.211 0.000 2.148*** 0.461 0.000
Vehicle involvement:
- Truck involvement − 0.990*** 0.186 0.000 − 1.239*** 0.250 0.000 − 1.378*** 0.530 0.009
- Bus involvement 0.810* 0.416 0.052 1.008** 0.480 0.036 1.103* 0.651 0.090
Crash type:
- Rear-end 1.078*** 0.137 0.000 1.199*** 0.173 0.000 1.381*** 0.336 0.000
- Rollover 1.120*** 0.270 0.000 1.569*** 0.344 0.000 1.136** 0.488 0.020
Weather:
- Foggy 0.526*** 0.138 0.000 0.683*** 0.173 0.000 0.904*** 0.259 0.001
Number of vehicles:
- Single-vehicle crash − 0.622*** 0.128 0.000 − 5.897*** 1.914 0.002 − 0.659*** 0.229 0.004
Time period:
- Dawn − 0.940* 0.541 0.082
- Evening − 0.459*** 0.174 0.008 − 0.680*** 0.215 0.002 − 0.977*** 0.351 0.005
- Night − 1.239*** 0.374 0.001 − 1.666*** 0.533 0.002 − 1.643** 0.730 0.025
Constant − 3.696*** 0.169 0.000 − 3.872*** 0.207 0.000 − 5.688*** 0.555 0.000
Random parameters (S.D.)
Single-vehicle crash – – – 4.277*** 1.052 0.000 0.047 0.474 0.921
Truck – – – – – – 0.817 0.579 0.158
Rear-end – – – – – – 0.721** 0.353 0.041
Constant (σν ) 0.026 0.719 0.971 – – – 0.112 0.539 0.835
ρ 2.55E-4 2.05E-4 – – – – – – –
Number of observations 8779 8779 8779
Log-likelihood at convergence − 1177.00 − 1169.73 − 800.22
McFadden Pseudo R-squared 0.807 0.808 0.868
Chi-square 9816.29 9830.83 10569.83
AIC 2378.0 2365.5 1632.4

*, ** & *** denote statistically significant at the 90 %, 95 % & 99 % confidence level.

logistic model to reflect the quantitative impact of each significant probability to be negative in some segments. As law enforcement and the
variable on the propensity of CC occurrence. The marginal effects results driving environment may vary across road segments (40− 80 km),
are shown in Table 3. driving in different road segments may be associated with different
alertness levels and driving behaviors, leading to different probabilities
5. Discussion for a CC to occur.
As shown in Table 3, a rear end crash has 59.4 % higher probability
As the two-level binary logistic regression has a better performance to incur CCs in its behind than hitting objects. The positive effect of a
than the RE and the RP model, the following discussions are based on the rear end crash is consistent with the findings of Yang et al. (2013)’s
results of the two-level binary logistic regression. study on secondary crashes. A rear-end crash is often caused by the rapid
Rear-end crashes is the only variable that has a significant random deceleration of a lead vehicle in a queue, which has been found in
effect (β = 0.721; p = 0.041), together with a significant and positive simulation studies to cause chain-reaction crashes in the same lane
fixed effect (β = 1.381; p < 0.001). Note that although this effect varies (Sugiyama and Nagatani, 2013). Given that the traffic on an expressway
across road segments, it is mainly positive and has only a small is likely to consists of a mixture of vehicle types that are controlled by
drivers with different driving experiences and levels of skills, the hazard
perceptions and reactions of these drivers can be quite different. Hence,
Table 3 the hazards tend to spread non-linearly backwards and cause consecu­
Results of marginal effects of the two-level logistic regression. tive crashes.
Variable Marginal effect Compared to rear-end crashes, a rollover crash has a relatively
Severe crash 0.017***
weaker positive fixed effect (β = 1.136; p < 0.020), meaning that a
120 km/h 1.149*** rollover crash has a higher likelihood to trigger a CCs compared to other
Tunnel 0.097*** crash types (mainly hitting roadside objects crashes). A rollover crash is
Truck involvement − 0.193*** often the result of a driver losing control of his/her vehicle which is
Rear-end 0.594***
likely to cause considerable disruption on the surrounding vehicles and
Rollover 0.061***
Foggy 0.531*** an adverse effect on vehicles following behind.
Single-vehicle crash − 0.371*** Compared to crashes without trucks involvement, crashes with
Evening − 0.172*** trucks involved are less likely (β = -1.378; p = 0.009) to cause CCs in the
Night − 0.149*** traffic behind. A crash involving trucks is found to have 19.3 % lower
Constant − 0.081***
probability to cause CCs. Karlaftis et al. (1999) found that the involve­
***
denote the marginal effect was significant at the 99 % ment of trucks and semi-trucks increased the likelihood of a secondary
level.

5
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

crash due to queueing effect. However, the effect of truck involvement with greater interactions between vehicles and enormous energy dissi­
on secondary crashes is different when the secondary crashes are sup­ pation (Meng et al., 2020), and is thus more likely to influence the up­
posed to occur immediately after the primary crash when traffic queue stream traffic nearby and causes subsequent crashes in a short period of
has not been formed. Trucks on freeways have been shown to produce time.
lower risks for both single- and multiple-vehicle crashes as the average
driving speed drops with increasing proportion of trucks in the traffic 6. Conclusion
(Hiselius, 2004; Dominique Lord et al., 2005; Wang and Feng, 2019).
The massive size and weight of a truck serve as a natural warning to This paper investigates the effects of various factors related to a
other drivers. Drivers, especially the experienced ones, tend to keep a primary crash that are more likely to trigger a CC. 8779 crashes on the
longer headway or lower the speed while driving behind a truck on a freeway network of Guizhou Province, China were extracted, among
freeway to maintain an adequate sight distance and minimize the risk of which 299 were the primary crash of a consecutive crash. To account for
a crash. the possible unobserved heterogeneity and hierarchical nature existing
The estimated coefficient for single-vehicle crashes was statistically in the dataset, a RE, a RP and a two-level binary logistic regression
significant. The estimate for the mean effect is negative (β = -0.659; p = models were estimated. Based on the AIC values, the two-level binary
0.004), and the marginal effect implies that single-vehicle crashes are logistics model outperformed the RE and the RP model, indicating that
37.1 % less likely to trigger CCs on a freeway compared to a crash with 2 there was significant spatial heterogeneity that varies across the road
or more vehicles involved. A single-vehicle crash often creates less segment. Based on the results of the two-level model, rear-end crashes
disruption in the road as it is common for the vehicle to run off the road were found to have both fixed and random effects on the occurrences of
and hits a roadside object in this type of crash (Wang and Feng, 2019). consecutive crashes on a freeway. Other factors including foggy weather
The estimated coefficient for tunnel is positive and significant (β = and tunnel were found to be positively associated with the occurrences
2.148; p < 0.001), and its marginal effect implies a crash happening in a of CCs. Single-vehicle crash, truck involvement and nigh time (7 pm − 3
tunnel has a 9.7 % higher likelihood than a crash on an open freeway to am) were found to be negatively associated with CCs.
trigger secondary CCs. Previous research has discussed the negative ef­ Based on the results, several policies on traffic management are
fects of the constrained driving environment and reduced visibility in recommended to prevent the occurrence of CCs. Warning signs should
road tunnels on traffic safety and crash risks (Caliendo et al., 2019; Ma be placed where multiple-vehicle crashes tend to occur, such as ramps,
et al., 2016; Mehri et al., 2019; Yeung and Wong, 2013). These factors tunnels and toll stations, especially on roads that have relatively high
may also distract drivers from promptly reacting to hazards that a traffic volume and speed. Novice drivers should be educated about these
downstream crash produces. Moreover, given the constrained space high-risk areas and to keep a sufficiently long headway when following
inside a freeway tunnel, it is more likely for the hazards to propagate the front vehicle. Moreover, the police and other emergency services are
farther behind rather than diffuse quickly. advised to be more alert about the occurrences of rear-end and roll-over
The posted speed limit at 120 km/h has a significant positive coef­ crashes, and to develop targeted incident management plans for them,
ficient (β = 1.397; p = 0.002), implying that a crash on the freeway especially around noon time.
segment with a high speed limit (120 km/h) is associated with a higher The amount of information provided in the dataset for this study to
likelihood of a CC compared to the segment with a speed limit at 80 km/ investigate the contributing factors is quite limited and more research
h. The marginal effect (1.149) indicates that there is a 1.149 times should be conducted to explore other potential contributing factors.
higher probability for a CC to occur on a higher-speed freeway (120 km/ Future research can consider other characteristics of a CC, such as lateral
h) than on a lower-speed freeway (80 km/h). At a higher speed, drivers position of the primary CC, the number of secondary crashes and their
will have less time to react to any sudden incident in the downstream severity levels, and to model the relationships between these charac­
traffic (Sarker et al., 2017), which will result in a higher likelihood of a teristics and potential contributing factors. Vehicle-level or driver-level
CC. factors such as driver characteristics, driving errors and violations are
Compared to crashes occurring during “noon” (1100− 1459 h), also suggested to be considered as potential predictors that impacts the
crashes occurring during “evening” (1900− 2259 h; β = -0.977, p = occurrences of CCs. Traffic flow models can be developed to compare the
0.005) and “night” (2300− 0259 hrs; β = -1.643, p = 0.025) and are less theoretical mechanism of consecutive crashes with chain-reaction
likely to result in CCs. These results indicate that a crash taking place in a crashes.
poorer natural lighting condition has a lower probability to trigger
secondary crashes immediately. This result is consistent with Zhan et al. CRediT authorship contribution statement
(2009)’s finding that the daytime period (0600− 1900 hrs) has a rela­
tively higher propensity for secondary crashes to occur. Nagatani (2015) Zhou Zichu: Data curation, Formal analysis, Investigation, Project
suggests that drivers mostly follow the taillights of the vehicles in front administration. Meng Fanyu: Formal analysis, Funding acquisition,
when driving in poor lighting condition and may compensate for the Investigation, Methodology, Writing - original draft, Writing - review &
increased risk due to poorer sight distance by leaving a larger gap when editing. Song Cancan: Data curation, Project administration. Tay
following behind another vehicle. Richard: Methodology, Validation, Writing - original draft, Writing -
Relative to clear weather, foggy weather has a significant and posi­ review & editing. Guo Zhongyin: Supervision, Funding acquisition.
tive coefficient (β = 0.904, p = 0.001). The marginal effect implies that a Yang Lili: Supervision, Funding acquisition. Wang Weili: Funding
crash happening on a foggy day has a 53.1 % higher likelihood to trigger acquisition.
CCs. Since foggy weather reduces visibility and affects drivers’ cognitive
capacities (Deng et al., 2019; Outay et al., 2019), it may be more difficult Declaration of Competing Interest
for drivers to avoid the adverse consequences of a primary crash in the
front within a short period of time. The authors report no declarations of interest.
Crash severity has a significantly positive coefficient (β = 1.573, p =
0.045). Compared to PDO crashes, a severe crash has a 1.7 % higher Acknowledgements
likelihood to incur CCs in the upstream traffic. This effect is expected as
more severe crashes may distract other drivers more and cause more This study was supported by the National Natural Science Founda­
congestion and traffic disruption. Also, crash severity may represent tion of China under Grant 71771113, National Key R&D Program of
comprehensive information, such as hitting angle, lane positions and/or China under grant 2018YFC0807000, National Key R&D Program of
driving errors, etc. Therefore, a severe crash on a freeway is associated China under Grant 2019YFC0810705, and the Shanghai Sail Program

6
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

under Grant 19YF1451800. We would also like to acknowledge the Meng, F., Xu, P., Song, C., Gao, K., Zhou, Z., Yang, L., 2020. Influential factors associated
with consecutive crash severity: a two-level logistic modeling approach. Int. J.
Guizhou Transport Information and Emergency Control Center and
Environ. Res. Public Health 17 (15), 5623.
Department of Transportation of Guizhou Province for providing the Nagatani, T., 2015. Chain-reaction crash in traffic flow controlled by taillights. Phys. A
crash database and the traffic volume data. Stat. Mech. Appl. 419, 1–6. https://doi.org/10.1016/j.physa.2014.10.055.
Nagatani, T., Yonekura, S., 2014. Multiple-vehicle collision induced by lane changing in
traffic flow. Phys. A Stat. Mech. Appl. 404, 171–179. https://doi.org/10.1016/j.
References physa.2014.02.028.
Nguyen-Phuoc, D.Q., Tran, A.T.P., De Gruyter, C., Kim, I., Su, D.N., 2019. Turn signal use
Akaike, H., 1973. Maximum likelihood identification of Gaussian autoregressive moving among car drivers and motorcyclists at intersections: a case study of Da Nang,
average models. Biometrika 60 (2), 255–265. Vietnam. Accid. Anal. Prev. 128, 25–31. https://doi.org/10.1016/j.
Alogaili, A., Mannering, F., 2020. Unobserved heterogeneity and the effects of driver aap.2019.03.012.
nationality on crash injury severities in Saudi Arabia. Accid. Anal. Prev. 144, Outay, F., Ahmar, A.U.H., Kamoun, F., Yasar, A.U.H., Sommer, C., Jabeur, N., El-
105618. Amine, S., 2019. Investigation of the impact of a wireless Fog Warning System with
Anastasopoulos, P.C., Mannering, F., 2009. A note on modeling vehicle accident respect to road traffic on a highway. Pers. Ubiquitous Comput. 23 (5–6), 893–899.
frequencies with random-parameters count models. Accid. Anal. Prev. 41 (1), https://doi.org/10.1007/s00779-018-1151-4.
153–159. https://doi.org/10.1016/j.aap.2008.10.005. Owens, N., 2010. 2010 Traffic Incident Management Handbook Update. Rep. No.
Anastasopoulos, P.C., Mannering, F., Shankar, V.N., Haddock, J.E., 2012. A study of FHWAHOP-10-013. Federal Highway Administration, Office of Transportation
factors affecting highway accident rates using the random-parameters tobit model. Operations, Washington, DC.
Accid. Anal. Prev. 45, 628–633. https://doi.org/10.1016/j.aap.2011.09.015. Pei, X., Wong, S.C., Sze, N.N., 2012. The roles of exposure and speed in road safety
Caliendo, C., De Guglielmo, M.L., Russo, I., 2019. Analysis of crash frequency in analysis. Accid. Anal. Prev. 48, 464–471. https://doi.org/10.1016/j.
motorway tunnels based on a correlated random-parameters approach. Tunn. aap.2012.03.005.
Undergr. Space Technol. 85, 243–251. https://doi.org/10.1016/j.tust.2018.12.012. Ramalho, E., Ramlho, J., 2012. Alternative versions of the RESET test for binary response
Chin, H.C., Huang, H.L.L., 2009. Safety assessment of taxi drivers in Singapore. Transp. index models: a comparative study. Oxf. Bull. Econ. Stat. 74, 107–130.
Res. Rec. (2114), 47–56. https://doi.org/10.3141/2114-06. Ramsey, J.B., 1969. Tests for specification errors in classical linear least-squares
Deng, C., Wu, C.Z., Cao, S., Lyu, N.C., 2019. Modeling the effect of limited sight distance regression analysis. J. R. Stat. Soc. Ser. B 31 (2), 350–371.
through fog on car-following performance using QN-ACTR cognitive architecture. Rusli, R., Haque, M.M., Afghari, A.P., King, M., 2018. Applying a random parameters
Transp. Res. Part F-Traffic Psychol. Behav. 65, 643–654. https://doi.org/10.1016/j. Negative Binomial Lindley model to examine multi-vehicle crashes along rural
trf.2017.12.017. mountainous highways in Malaysia. Accid. Anal. Prev. 119, 80–90. https://doi.org/
Duddu, V.R., Penmetsa, P., Pulugurtha, S.S., 2018. Modeling and comparing injury 10.1016/j.aap.2018.07.006.
severity of at-fault and not at-fault drivers in crashes. Accid. Anal. Prev. 120, 55–63. Sarker, A.A., Paleti, R., Mishra, S., Golias, M.M., Freeze, P.B., 2017. Prediction of
Gujarati, D.N., 2009. Basic Econometrics. Tata McGraw-Hill Education. secondary crash frequency on highway networks. Accid. Anal. Prev. 98, 108–117.
Hiselius, L.W., 2004. Estimating the relationship between accident frequency and Snijders, T.A., Bosker, R.J., 1999. An Introduction to Basic and Advanced Multilevel
homogeneous and inhomogeneous traffic flows. Accid. Anal. Prev. 36 (6), 985–992. Modeling. Sage, London.
Hong, J., Tamakloe, R., Park, D., 2019. A comprehensive analysis of multi-vehicle Sugiyama, N., Nagatani, T., 2013. Multiple-vehicle collision in traffic flow by a sudden
crashes on expressways: a double hurdle approach. Sustainability 11 (10). https:// slowdown. Phys. A Stat. Mech. Appl. 392 (8), 1848–1857. https://doi.org/10.1016/
doi.org/10.3390/su11102782. j.physa.2013.01.009.
Hosseinpour, M., Sahebi, S., Zamzuri, Z.H., Yahaya, A.S., Ismail, N., 2018. Predicting Tay, R., 2003. Marginal effects of changing the vehicle mix on fatal crashes. J. Transp.
crash frequency for multi-vehicle collision types using multivariate Poisson- Econ. Policy 37 (3), 439–450.
lognormal spatial model: a comparative analysis. Accid. Anal. Prev. 118, 277–288. Tay, R., 2005a. Drink driving enforcement and publicity campaigns: are the policy
https://doi.org/10.1016/j.aap.2018.05.003. recommendations sensitive to model specification? Accid. Anal. Prev. 37 (2),
Huang, H., Abdel-Aty, M., 2010. Multilevel data and Bayesian analysis in traffic safety. 259–266.
Accid. Anal. Prev. 42 (6), 1556–1565. Tay, R., 2005b. General and specific deterrent effects of traffic enforcement: Do we have
Jiang, C., Tay, R., Lu, L., 2021. A Skewed Logistic Model of Two-unit Bicycle-vehicle Hit- to catch offenders to reduce crashes? J. Transp. Econ. Policy 39 (2), 209–223.
and-run Crashes (accepted). Traffic Injury Prevention. Tay, R., 2016. Comparison of the binary logistic and skewed logistic (Scobit) models of
Karlaftis, M.G., Latoski, S.P., Richards, N.J., Sinha, K.C., 1999. ITS impacts on safety and injury severity in motor vehicle collisions. Accid. Anal. Prev. 88, 52–55.
traffic management: an investigation of secondary crash causes. J. Intell. Transp. Theofilatos, A., Yannis, G., Kopelias, P., Papadimitriou, F., 2019. Impact of real-time
Syst. 5 (1), 39–52. traffic characteristics on crash occurrence: preliminary results of the case of rare
Kim, D.G., Lee, Y., Washington, S., Choi, K., 2007. Modeling crash outcome probabilities events. Accid. Anal. Prev. 130, 151–159. https://doi.org/10.1016/j.
at rural intersections: application of hierarchical binomial logistic models. Accid. aap.2017.12.018.
Anal. Prev. 39 (1), 125–134. Train, K.E., 2009. Discrete Choice Methods With Simulation. Cambridge university press.
Liu, J., Khattak, A.J., 2017. Gate-violation behavior at highway-rail grade crossings and Vlahogianni, E.I., Karlaftis, M.G., Golias, J.C., Halkias, B.M., 2010. Freeway operations,
the consequences: using geo-spatial modeling integrated with path analysis. Accid. spatiotemporal-incident characteristics, and secondary-crash occurrence. Transp.
Anal. Prev. 109, 99–112. Res. Rec. 2178 (1), 1–9.
Lord, D., Mannering, F., 2010. The statistical analysis of crash-frequency data: a review Wang, X., Feng, M., 2019. Freeway single and multi-vehicle crash safety analysis:
and assessment of methodological alternatives. Transp. Res. Part A-Policy Pract. 44 influencing factors and hotspots. Accid. Anal. Prev. 132, 105–268. https://doi.org/
(5), 291–305. https://doi.org/10.1016/j.tra.2010.02.001. 10.1016/j.aap.2019.105268.
Lord, D., Middleton, D., Whitacre, J., 2005. Does separating trucks from other traffic Wang, J., Liu, B., Zhang, L., Ragland, D.R., 2016. Modeling secondary accidents
improve overall safety? Transp. Res. Rec. 1922 (1), 156–166. identified by traffic shock waves. Accid. Anal. Prev. 87, 141–147.
Ma, Z., Chien, S.I.J., Dong, C., Hu, D., Xu, T., 2016. Exploring factors affecting injury Wang, K., Bhowmik, T., Yasmin, S., Zhao, S., Eluru, N., Jackson, E., 2019. Multivariate
severity of crashes in freeway tunnels. Tunn. Undergr. Space Technol. 59, 100–104. copula temporal modeling of intersection crash consequence metrics: a joint
https://doi.org/10.1016/j.tust.2016.06.013. estimation of injury severity, crash type, vehicle damage and driver error. Accid.
Mannering, F., Bhat, C.R., 2014. Analytic methods in accident research: methodological Anal. Prev. 125, 188–197.
frontier and future directions. Anal. Methods Accid. Res. 1, 1–22. https://doi.org/ Washington, S., Karlaftis, M.G., Mannering, F., Anastasopoulos, P., 2020. Statistical and
10.1016/j.amar.2013.09.001. Econometric Methods for Transportation Data Analysis. CRC press.
Mannering, F.L., Shankar, V., Bhat, C.R., 2016. Unobserved heterogeneity and the Wen, H., Xue, G., 2020. Injury severity analysis of familiar drivers and unfamiliar drivers
statistical analysis of highway accident data. Anal. Methods Accid. Res. 11, 1–16. in single-vehicle crashes on the mountainous highways. Accid. Anal. Prev. 144,
https://doi.org/10.1016/j.amar.2016.04.001. 105667.
McFadden, D., 1973. Conditional Logit Analysis of Qualitative Choice Behavior. Xu, C., Liu, P., Yang, B., Wang, W., 2016. Real-time estimation of secondary crash
Mehri, A., Sajedifar, J., Abbasi, M., Naimabadi, A., Mohammadi, A.A., Teimori, G.H., likelihood on freeways using high-resolution loop detector data. Transp. Res. Part C
Zakerian, S.A., 2019. Safety evaluation of lighting at very long tunnels on the basis of Emerg. Technol. 71, 406–418.
visual adaptation. Saf. Sci. 116, 196–207. https://doi.org/10.1016/j. Yang, H., Bartin, B., Ozbay, K., 2013. Investigating the characteristics of secondary
ssci.2019.03.018. crashes on freeways. In: Washington, DC92nd Annual Meeting of the Transportation
Meng, F., Wong, W., Wong, S.C., Pei, X., Li, Y.C., Huang, H., 2017a. Gas dynamic Research Board, Vol. 2.
analogous exposure approach to interaction intensity in multiple-vehicle crash Yang, H., Ozbay, K., Xie, K., 2014. Assessing the risk of secondary crashes on highways.
analysis: case study of crashes involving taxis. Anal. Methods Accid. Res. 16, 90–103. J. Saf. Res. 49, 143.
https://doi.org/10.1016/j.amar.2017.09.003. Yang, H., Wang, Z., Xie, K., Ozbay, K., Imprialou, M., 2018. Methodological evolution
Meng, F., Xu, P., Wong, S.C., Huang, H., Li, Y.C., 2017b. Occupant-level injury severity and frontiers of identifying, modeling and preventing secondary crashes on
analyses for taxis in Hong Kong: a Bayesian space-time logistic model. Accid. Anal. highways. Accid. Anal. Prev. 117, 40–54.
Prev. 108, 297–307. https://doi.org/10.1016/j.aap.2017.08.010. Yeung, J.S., Wong, Y.D., 2013. Road traffic accidents in Singapore expressway tunnels.
Meng, F., Wong, S.C., Yan, W., Li, Y.C., Yang, L., 2019. Temporal patterns of driving Tunn. Undergr. Space Technol. 38, 534–541. https://doi.org/10.1016/j.
fatigue and driving performance among male taxi drivers in Hong Kong: A driving tust.2013.09.002.
simulator approach. Accid. Anal. Prev. 125, 7–13. https://doi.org/10.1016/j.
aap.2019.01.020.

7
Z. Zichu et al. Accident Analysis and Prevention 154 (2021) 106054

Yu, R., Abdel-Aty, M., 2013. Multi-level Bayesian analyses for single- and multi-vehicle Zhan, C., Gan, A., Hadi, M., 2009. Identifying secondary crashes and their contributing
freeway crashes. Accid. Anal. Prev. 58, 97–105. https://doi.org/10.1016/j. factors. Transp. Res. Rec. 2102 (1), 68–75.
aap.2013.04.025. Zhou, M., Chin, H.C., 2019. Factors affecting the injury severity of out-of-control single-
vehicle crashes in Singapore. Accid. Anal. Prev. 124, 104–112.

You might also like