Accident Analysis and Prevention 48 (2012) 254–263

Contents lists available at SciVerse ScienceDirect

Accident Analysis and Prevention

journal homepage: [Link]/locate/aap

Estimation of rear-end vehicle crash frequencies in urban road tunnels

Qiang Meng a,∗ , Xiaobo Qu a,b
a
Department of Civil and Environmental Engineering, National University of Singapore, Singapore 117576
b
Griffith School of Engineering, Griffith University, Gold Coast, QLD 4222, Australia

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

Article history: According to The Handbook of Tunnel Fire Safety, over 90% (55 out of 61 cases) of fires in road tunnels
Received 17 July 2011 are caused by vehicle crashes (especially rear-end crashes). It is thus important to develop a proper
Received in revised form 1 January 2012 methodology that is able to estimate the rear-end vehicle crash frequency in road tunnels. In this paper,
Accepted 19 January 2012
we first analyze the time to collision (TTC) data collected from two road tunnels of Singapore and conclude
that Inverse Gaussian distribution is the best-fitted distribution to the TTC data. An Inverse Gaussian
Keywords:
regression model is hence used to establish the relationship between the TTC and its contributing factors.
Rear-end crash frequency
We then proceed to introduce a new concept of exposure to traffic conflicts as the mean sojourn time in
Time to collision
Inverse Gaussian regression model
a given time period that vehicles are exposed to dangerous scenarios, namely, the TTC is lower than a
Road tunnels predetermined threshold value. We further establish the relationship between the proposed exposure
to traffic conflicts and crash count by using negative binomial regression models. Based on the limited
data samples used in this study, the negative binomial regression models perform well although a further
study using more data is needed.
© 2012 Elsevier Ltd. All rights reserved.

1. Introduction 2008), and QRAFT model of Singapore (Meng et al., 2011a,b; Qu

et al., 2011). All these models acknowledge that the frequency of
Road tunnels are increasingly cost-effective infrastructures fire occurred in road tunnels is the most important contributing
which provide underground vehicular passageways for motorists factor for the risk assessment of road tunnels. The Handbook of Tun-
and commuters, especially in densely populated cities like nel Fire Safety (Beard and Carvel, 2005) points out that over 90% (55
Singapore. With the increasing traffic volume and urban devel- out of 61 fire cases) of tunnel fires are caused by vehicle crashes
opment as well as growing needs for land use in urban areas, (especially the rear-end crashes). In addition, according to crash
constructing road tunnels is becoming one alternative to enhance statistics in Singapore’s road tunnels, over 2/3 of crashes are cat-
the capacity and accessibility for road transport systems. However, egorized as rear-end crashes.1 Accordingly, on one hand, rear-end
fire disasters occurred in a road tunnel would result in catastrophic crashes are the major cause for fire in road tunnels; on the other
consequences due to the enclosed structure nature of tunnel sys- hand, rear-end crashes constitute around 70% out of all the crashes.
tems. For example, 39 people lost their lives in a fire disaster that It is, therefore, of great importance to develop a methodology that
happened in the Mont Blanc Tunnel from France to Italy in 1999; can estimate the rear-end vehicle crash frequency in road tunnels
and another disaster in Tauern Tunnel of Austria resulted in 12 (the “rear-end crash” henceforth are referred to as “R-E crash” for
fatalities (PIARC, 2008). These accidents have raised the aware- short).
ness among the public as well as the government on both the A number of studies have been conducted to predict/estimate
safety aspect of the tunnels and that of the road tunnel users. Thus, frequency of various types of crashes in highways using crash-
quantitative risk assessment (QRA) has been one of the require- frequency data. However, identification of the cause and effect
ments under the European Union (EU) Directive (2004/54/EC). In relationship is typical unavailable due to lack of microscopic traffic
Singapore, QRA for all major urban road tunnels longer than 240 m information (or the detailed driving data). Consequently, as pointed
is compulsory in accordance with the Project Safety Review (PSR) out by Lord and Mannering (2010), researchers have framed their
procedure manual for roads in the country (LTA, 2005). analytic approaches to study the factors that affect the number of
Several QRA models for road tunnels have been developed, crashes occurring in some geographical spaces over some speci-
including TuRisMo model of Austria, TUNPRIM model of the fied time periods by using various types of count-data regression
Netherlands, Italian risk analysis model, OECD/PIARC model (PIARC, models in accordance to some assumptions. These models include

∗ Corresponding author. Tel.: +65 6516 5494; fax: +65 6779 1635. 1
746 out of 1106 crashes (70%) are categorized as rear-end crashes in the CTE
E-mail addresses: ceemq@[Link] (Q. Meng), [Link]@griffi[Link] (X. Qu). road tunnel from 2006 to 2008.

0001-4575/$ – see front matter © 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/[Link].2012.01.025
 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263 255

Fig. 2. Traffic videos recorded from CTE road tunnel.

Fig. 1. General arrangement of KPE road tunnel.

The remainder of this paper is organized as follows. In Section

2, the TTC is defined and data collected from Singapore’s road tun-
Poisson regression model (e.g. Miaou and Lum, 1993; Miaou,
nels are presented. In Section 3, the Inverse Gaussian regression
1994; Hauer, 2001), Negative binomial/Poisson-Gamma model
model is built to establish the relationship between TTC distri-
(e.g. Maycock and Hall, 1984; Malyshkina and Mannering, 2010a;
butions and corresponding traffic volumes. In Section 4, two R-E
Daniels et al., 2010), Zero-inflated Poisson and negative binomial
crash frequency estimation models are developed on the basis of
models (e.g. Miaou, 1994; Shankar et al., 1997; Malyshkina and
the derived TTC distributions. Several conjectures and recommen-
Mannering, 2010b; Lord et al., 2007), Conway–Maxwell–Poisson
dations for further studies are put forward in Section 5. Section 6
model (e.g. Lord et al., 2008, 2010), and others (e.g. Zhang and Xie,
concludes this study.
2007; Guo et al., 2010; Haque et al., 2010). The lack of the detailed
driving data on highways may make these statistical analysis mod-
els biased to reflect the fundamental cause and effect relationship. 2. TTC data collection
Lord and Mannering (2010) thus highlighted that the entirely new
direction of research could potentially open up if the anticipated Assume that there are two consecutive vehicles moving in the
availability of the detailed driving data and crash data are available. same direction on the same lane of a road tunnel. Let Lleader and
More detailed traffic data are obtainable in road tunnels com- Lfollower be the locations of the leading and following vehicles
pared to highways because most of road tunnels are equipped at a particular time, respectively. Correspondingly, let L̇leader and
with the closed circuit television (CCTV) cameras and/or an oper- L̇follower denote the speeds of the leading and following vehicles at
ation control center (OCC). For example, each of Singapore’s road the particular time. According to the TTC definition, namely, the
tunnels has been installed 2–4 CCTV cameras every 200 m and mon- time that remains until a collision between two vehicles would have
itored by a 24-h manned operation control center (OCC). These occurred if the collision course and speed difference are maintained,
CCTV cameras record real time and detailed traffic information. the TTC can be mathematically expressed by
In addition to hourly traffic volume and density, we can precisely 
Lleader − Lfollower − lleader
measure/estimate the time to collision (TTC) for two consecutive ,if L̇follower > L̇leader
TTC = L̇follower − L̇leader (1)
vehicles moving in the same lane of a road tunnel using traffic ∞ ,otherwise
videos. The TTC is defined as the time that remains until a colli-
sion between two vehicles would have occurred if the collision course where lleader is the length of the leading vehicle. Eq. (1) implies that
and speed difference are maintained (Hayward, 1972). The TTC has the TTC is measurable if we have real time traffic information.
been one of the well-recognized safety indicators for traffic con- To collect the TTC data in a road tunnel, the Kallang/Paya Lebar
flicts on highways (Farah et al., 2009; Svensson, 1998; Vogel, 2003). Expressway (KPE) and the Central Expressway (CTE) in Singapore
Minderhoud and Bovy (2001) further pointed out that the TTC is shown in Figs. 1 and 2 are selected. KPE and CTE are two vital infras-
inversely related to vehicle crash frequencies in road sections. It is tructures in Singapore’s road system. The first one has a total length
widely accepted as a safety indicator in highways. A TTC threshold of 12 km and 9 km of the expressway (Fig. 1) is built underground
value is usually chosen to distinguish relatively safe situation and as a road tunnel, serving the growing traffic demand of the north-
dangerous scenarios exposed to traffic conflicts (or critical encoun- eastern sector of Singapore. The second one, a 17-km expressway,
ters). It is acknowledged that the TTC threshold should be 2–4 s links the north and south of Singapore through the Central Business
(Minderhoud and Bovy, 2001; Vogel, 2003). District (CBD). 2.4 km of the expressway (Fig. 2) are laid under-
The objective of this study is to develop a novel R-E crash fre- ground and these portions of the CTE form the first road tunnel
quency estimation method on the basis of TTC distributions. The in Singapore. Both road tunnels are equipped with the 24-h OOC
TTC sample data are collected from the traffic videos of Singapore’s systems.
road tunnels. Based on the statistical analysis, we find that the We have requested 42-h tunnel traffic videos recorded by CCTV
Inverse Gaussian distribution is the best-fitted distribution model of these two tunnels from Land Transport Authority of Singapore,
for the collected TTC data. The Inverse Gaussian regression model including 14 locations for 3 typical time periods – morning peak
is thus employed to establish the relationship between TTC distri- hour: 8:00 am to 9:00 am, off-peak hour: 14:00 pm to 15:00 pm,
butions and the corresponding traffic volume. Having had the TTC evening peak hour: 19:00 pm to 20:00 pm – in March 2011. The
distributions, a R-E crash frequency estimation method is put up to TTC data are generated from these traffic videos in different time
reflect the relationship between the TTC distributions and the R-E periods with different traffic conditions. The procedure of mea-
crash frequencies. suring a TTC with respect to a particular car-following scenario
256 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263

Table 1

0.3746

0.3449
0.4948
0.3807
0.3600
TTC data.

K–S
Traffic volume (vehs/(h lane)) Number of data

Location 1 894 104

Location 2 963 65

Uniform (0, 29.82)

Uniform (0, 31.39)
Uniform (0, 36.84)
Uniform (0, 39.55)
Uniform (0, 29.57)
Location 3 1127 80
Location 4 1374 79

Distributions
Location 5 1672 93

Uniform
is summarized as follows. We first measure length of the lead-
ing vehicle (lleader ) in a car-following scenario. After that, the spot
speeds of the vehicles (L̇follower and L̇leader ) can be estimated by mea-

0.1814
0.1764

0.1934
0.2091
0.1408
suring the time taken by the vehicle to cover two lane-markers’

K–S
distance in the video. Then, the time headway (h) between the
leading and following vehicles is recorded. According to Vogel
(2003), the gap size (Lleader − Lfollower − lleader ) can be estimated by
(L̇follower × h − lleader ). Finally, the TTC for the car-following scenario

Expon (11.20)
Expon (12.30)
Distributions

Expon (9.26)
Expon (9.69)

Expon (7.26)
Exponential
could be calculated according to Eq. (1).
In the measurement, we display 30 frames per second and limit
the error of 0.03 s, which would yield data of comparable quality to
the radar-speed measurement method. 867 car following scenarios
occurred at various locations are examined, and 421 TTC data (TTC
with a finite value) with respect to different traffic volumes are

0.2385
0.2471
0.1756
0.1768
0.3199
obtained. Statistically, the number of TTC data with a finite value

K–S
should be equal to that of samples with infinite values. An infinite
TTC value indicates that the following vehicle will not be possible
catch up with the leading one, which is an absolutely safe situation.
We would focus on the probability distributions of TTC samples

Triang (0, 1.65, 29.88)

Triang (0, 2.41, 32.80)
Triang (0, 2.30, 31.50)

Triang (0, 2.10, 37.40)

Triang (0, 1.41, 40.60)
with finite values accordingly.

Distributions
Triangular
3. Inverse Gaussian distribution for TTC

3.1. Statistical analysis for the TTC samples

A data analysis procedure is proposed in order to obtain the

best-fitted TTC distributions. Five commonly used distributions are
0.1198
0.1138
0.1024
0.1097
0.0781
examined in this study: Inverse Gaussian, Exponential, Normal,
K–S

Triangular, and Lognormal. The maximum likelihood estimation

(MLE) technique is employed to estimate the parameters involved
in a distribution. After obtaining the parameters for the five types
Lognorm (12.96, 13.86)

of distributions, the goodness-of-fit test is conducted to select the

Lognorm (11.53, 9.56)
Lognorm (9.27, 8.28)
Lognorm (9.71, 8.62)

Lognorm (7.24, 6.45)

best-fitted distribution among the given candidate distributions.

Kolmogorov–Smirnov (K–S) test, a nonparametric test method, has
Distributions

been widely applied to compare a sample with a reference probabil-

Lognormal

ity distribution in transportation studies (e.g. Ibeas et al., 2011; Páez

et al., 2011). In this study, the K–S test is also adopted to perform the
goodness-of-fit test. The K–S statistic quantifies a distance between
the empirical distribution function of the sample and the cumula-
tive distribution function of the reference distribution. In this study,
0.0968a

0.1017a
0.0813a
0.0651a
0.1003a

The K–S statistics of the best-fitted distributions.

a distribution with the lowest K–S test statistic is regarded as the

K–S

best-fitted distribution.
Following the above-mentioned data analysis procedure, we
analyze five sets of the TTC data collected at different locations with
Statistical analysis for the TTC samples.

respect to different traffic volumes, as shown in Table 1. Table 2

Inverse Gaussian

IG (11.20, 14.06)
IG (12.30, 11.01)
IG (9.26, 12.21)
IG (9.69, 12.88)

gives results of the best-fit analysis.

IG (7.26, 9.24)
Distributions

According to Tables 1 and 2, we can find that

(1) The Inverse Gaussian distribution is the best-fitted distribution

for all the five locations.2 Fig. 3(a) and (b) depicts the his-
tograms and empirical cumulative distribution function (CDF)
Location 1
Location 2
Location 3
Location 4
Location 5
Table 2

2
Inverse Gaussian distribution is a two-parameter family of continuous proba-
a

bility distributions with support on (0, ∞). Its probability density function is given
 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263 257

TTC Sample Mean - Traffic Volume Relationship

TTC Sample Mean Poly. (TTC Sample Mean)
18

16

TTC Sample Mean

14

12
y = -3E-05x 2 + 0.0655x - 28.747
10
R 2 = 0.9371
8

6
850 950 1050 1150 1250 1350 1450 1550 1650 1750
Traffic Volume (veh/hour lane)

Fig. 4. TTC sample mean–traffic volume relationship.

when traffic volume is high (>1600 vehs/(h lane)), the small dis-
tance headway would lead to low TTC values.
(5) The shape parameters () of the best-fitted Inverse Gaussian
distributions with respect to different traffic volumes are within
a relatively small range from 9.24 to 14.06.

3.2. Estimation of the parameters defining Inverse Gaussian

distribution

As discussed in Section 3.1, both Inverse Gaussian distribution

Fig. 3. The histograms and empirical CDF (traffic volume = 963 vehs/(h lane)). and Lognormal distribution fit the data very well. In the follow-
ing analysis, without loss of generality, we use an Inverse Gaussian
regression model to establish the relationship between TTC and
for data samples with the best-fitted distributions (traffic vol-
traffic volume. To formulate the inverse Gaussian regression model,
ume = 963 vehs/(h lane)).
let yi , i = 1, . . ., n, be n independent observations (TTC samples) dis-
(2) Lognormal distribution also performs very well for the five loca-
tributed as IG(i , ), in which the inverse of sample mean has a
tions (relatively small K–S values). In reality, other samples may
parabola relationship with traffic volume, namely:
suggest that Lognormal distributions are better. The two dis-
tributions have similar patterns. Let us take the Location 1 as 1
= ˇ0 + ˇ1 xi + ˇ2 xi 2 > 0 (2)
an example: P(IG(9.26, 12.21) ≤ 3) = 0.138 and P(Lognorm(9.27, i
8.28) ≤ 3) = 0.139.3 Indeed, the differences between the two dis-
where xi is traffic volume of TTC sample i. Whitmore (1983) derived
tributions are very marginal. In the following analysis, without
the pseudo maximum likelihood estimations of ˇ and  as
loss of generality, we will assume the samples follow Inverse
Gaussian distributions as suggested in Table 3. ˆ = (X  YX)−1 X  1
ˇ (3)
(3) The TTC data collected at different locations with respect to sim-
ilar traffic volume generally follows the same Inverse Gaussian ˆ = n
 (4)
distribution (e.g. Location 1 and Location 2). In other words, the 1T Y −1 1 − 1 X ˇ
ˆ
traffic volume could be considered as the contributing factor for
where Y is the diagonal matrix with ith diagonal elements being
TTC distributions. T
(4) The TTC sample mean and its inverse both have a parabola rela- yi , 1 is the n-vector of all ones and X = (1, xi , xi 2 ) . They are called
tionship with the traffic volume, as shown in Figs. 4 and 5. This the pseudo maximum likelihood estimations because the condition
is because the two contributing factors to TTC, distance head-
way and speed dispersion, are both dependent of the traffic Inverse of TTC Mean - Traffic Volume Relationship
volume. When traffic volume is low (<1000 vehs/(h lane)), the
TTC Mean - Traffic Volume Poly. (TTC Mean - Traffic Volume)
great speed dispersion could result in low TTC values. However,
Inverse of TTC sample mean

0.16
y = 3E-07x 2- 0.0007x + 0.5007
0.14
R2 = 0.9532

by 0.12

 1/2  2
 0.1
 −(x − )
f (x; , ) = exp , 0 < x < ∞,
2x3 22 x 0.08

where  > 0 is the mean and  > 0 is the shape parameter. The distribution can be
0.06
viewed as the distribution of first passage time of a Wiener process with an absorbing
0.04
barrier, i.e., while the Gaussian describes a Brownian Motion’s level at a fixed time
850 950 1050 1150 1250 1350 1450 1550 1650 1750
(Wiener process), the inverse Gaussian describes the distribution of the time the
Brownian Motion takes to reach a fixed positive level. Traffic Volume (vehs/hour lane)
3
In this example, the TTC threshold is assumed to be as 3 s. Similar results are
obtainable if we assume the threshold is 2 s or 4 s. Fig. 5. Inverse of TTC mean–traffic volume relationship.
258 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263

Table 3
K–S tests.

Traffic volume (vehs/(h lane)) Number of samples (n) Distributions K–S values (Dn ) Critical value (K0.05 ) Test results

894 104 IG (9.02, 12.17) 0.0977 1.36 1.00 < 1.36

963 65 IG (10.26, 12.17) 0.1086 1.36 0.88 < 1.36
1127 80 IG (12.33, 12.17) 0.1496 1.36 1.34 < 1.36
1374 79 IG (12.83, 12.17) 0.0821 1.36 0.73 < 1.36
1672 93 IG (7.30, 12.17) 0.1026 1.36 0.99 < 1.36

ˆ0 + ˇ
ˇ ˆ 1 xi + ˇ
ˆ 2 xi 2 > 0 for all i is not guaranteed to be satisfied.4 IG(7.30, 12.17)
1.0
According to the collected 421 TTC data with different traffic vol-
umes, the estimated coefficients are
0.8
ˆ 0 = 5.606 × 10−1
ˇ (5)
ˆ 1 = −7.900 × 10−4
ˇ (6) 0.6

ˆ 2 = 3.21 × 10−7
ˇ (7)
0.4
ˆ = 12.17
 (8)

After obtaining the estimated coefficients, the TTC distributions 0.2

could be determined for different traffic conditions reflected by
their traffic volumes. In order to evaluate how well the Inverse 0.0
Gaussian regression model estimates the TTC distributions, we 0 5 10 15 20 25 30
compare the derived TTC distributions with the TTC samples (a) Empirical CDF & CDF generated by IG regression model
at different traffic volumes – 894 vehs/(h lane), 963 vehs/(h lane), IG(7.26, 9.24)
1127 vehs/(h lane), 1374 vehs/(h lane), and 1672 vehs/(h lane) – by 1.0
using the hypothesis test. The K–S test is applied to conduct the
hypothesis test. The null hypothesis is rejected at level ˛ if
0.8
√
nDn > K˛ (9)

where n is the number of samples, Dn is the K–S statistic, and K˛ is 0.6

the critical value (˛ = 0.05 in this study). The results of K–S tests are
reported in Table 3. 0.4
Table 3 shows that the regression model performs well. Fig. 6(a)
and (b) depicts the CDF of the best-fitted Inverse Gaussian distri-
0.2
bution and the CDF generated by Inverse Gaussian regression for
a TTC sample (traffic volume = 1672 vehs/(h lane)), respectively. As
can be seen in the figures, the samples could be well represented 0.0
0 5 10 15 20 25 30
by either of the two Inverse Gaussian distributions with different
(b) Empirical CDF & Best fitted IG CDF
parameters.
Fig. 6. Empirical CDF with Inverse Gaussian distributions (traffic vol-
4. Accident frequency estimation ume = 1672 vehs/(h lane)).

4.1. TTC threshold value and exposure to traffic conflicts distributions for road tunnel sections (Section 3), the exposure to
traffic conflicts in an hour can be quantified by.
As mentioned in Section 1, a TTC threshold value is usually cho-
sen to distinguish relatively safe situation and dangerous scenarios Nconflict () = (K × L − 1) × Pr(TTC(x) ≤ ) × 0.5 (10)
exposed to traffic conflicts (or critical encounters). Various opin- where K denote the traffic density; L is the length of a road tun-
ions can be found from the literature as to which value should be nel section; (K × L − 1) indicates number of gaps in the section;
used as the threshold value. Hirst and Graham (1997) reported that P(TTC(x) ≤ ) × 0.5 represents the probability of TTC less than the
a time-to-collision measure of 4 s could be used to discriminate threshold value  5 ; x is the traffic volume of the time period in
between cases where drivers unintentionally find themselves in a the road tunnel section. Note that only half of car following sce-
dangerous situation from cases where drivers remain in control. narios will result in finite TTCs and the other half is considered as
Hogema and Janssen (1996) presented a minimum TTC value of absolutely safe situations (infinite TTCs).
3.5 s for non-supported drivers and 2.6 s for supported drivers. It
is widely acknowledged that the TTC threshold should be 2 s to 4 s
4.2. Historical crash-damage database
(Minderhoud and Bovy, 2001; Vogel, 2003).
We define the exposure to traffic conflicts as the mean sojourn
Historical crash-damage (HCD) database of Singapore is used to
time in a given time period (e.g. an hour) that vehicles are exposed
examine the relationship between exposure to traffic conflicts and
to dangerous scenarios (or critical encounters), i.e. the TTCs are
lower than a predetermined threshold value . Having had the TTC
5
The infinite TTC samples are considered as absolutely safe situations. The coeffi-
cient 0.5 is adequate only in stable traffic conditions; the coefficient would be a little
4
The condition is guaranteed in this study since the traffic volume is with the greater than 0.5 for unstable traffic flows. For simplicity, the coefficient is assumed
range from 800 to 1700 vehs/(h lane). to be 0.5 in this study.
 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263 259

Table 4
Traffic volumes, density, length, and crash records for different time periods.

Time period R-E crash records Estimated Traffic Density Average Exposure to traffic conflicts
(2006–2008) volume (vehs/(km lane)) speed
(vehs/(h lane)) (km/h)

2s 3s 4s

7:00 am–8:00 am 11 1600 25 62 657 2024 3548

1:00 pm–2:00 pm 5 1200 16 73 263 829 1502
5:00 pm–6:00 pm 8 1400 20 70 364 1155 2070
8:00 pm–9:00 pm 20 1700 45 39 1566 4673 7998
9:00 pm–10:00 pm 17 1600 50 34 1341 4131 7243
11:00 pm–12:00 am 4 900 11 78 252 777 1374

Crash count - traffic conflicts relationship

TTC criterion = 2s TTC criterion = 3s TTC criterion = 4s

Linear (TTC criterion = 4s) Linear (TTC criterion = 3s) Linear (TTC criterion = 2s)
Number of crashes in 2006-

24
21 y = 0.0111x + 2.5929
y = 0.0037x + 2.4085
18 R2 = 0.9738
R2 = 0.9744
15
2008

12
y = 0.0022x + 2.2446
9 R2 = 0.9736
6
3
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
Exposure to traffic conflicts in the 3 years

Fig. 7. R-E crash count–traffic conflicts relationship with linear fit.

crash frequencies. According to the Motor Claims Framework (MCF) section, there is a 2.4 m wide shoulder and three 3.6 m traffic lanes
introduced by the General Insurance Association of Singapore (GIA), in each carriageway with a tunnel structural height of approxi-
in the event of a crash in expressways, everyone involved must mately 6 m high. Both the curvature and gradient are very gentle
inform the insurance company within 1 day using the GIA Motor in this section. In the current study, we just measure the TTC for
Accident Report form. In addition, according to Road Traffic Act vehicles in the mid lane to represent the traffic state. According to
in Singapore, another report must be made within 24 h of a crash the HCD database, only the longitudinal positions of crashes are
if an injury has occurred. The HCD database (2006–2008) has all obtainable and the latitude positions (shoulder lane, mid lane, or
the reported crash records, by means of either ways, occurred at median lane) are not reported. Therefore, it would be unlikely for
Singapore expressways from 2006 to 2008, which includes the time us to disaggregate the three lanes.
of crash, location of crash, crash type (e.g. rear-end, skidded, etc.), We assume that the traffic volumes in the road tunnel section
vehicle type (e.g. car–car, car–truck, etc.), number of slight injuries, in a specific time period would not have significant daily varia-
number of serious injuries, and number of fatalities.6 To sum up, tions. In accordance with the LTA policy, only the latest 2 years
there are 746 rear-end crashes in the CTE road tunnel from 2006 to traffic volume data are obtainable. Therefore, the traffic volumes
2008, causing 0 fatalities, 45 severe injuries, and 458 slight injuries. and densities from 2006 to 2008 are not obtainable for this study.
The average traffic volumes are estimated by LTA tunnel operators
4.3. Relationship between exposure to traffic conflict and crash on the basis of the 2010–2011 traffic volume and the summary traf-
frequency fic data in 2006–2008. Accordingly, due to the data unavailability,
we just use accurate R-E crash data and estimated average traffic
2 s, 3 s, and 4 s are considered as the TTC threshold values. From volume to illustrate the methodology in the case study. According
the HCD database we get the crash frequencies in a 1-km road tun- to Eq. (9), the exposure to traffic conflicts could be calculated. The
nel section in CTE road tunnel are 11, 5, 8, 20, 17, and 4 for he estimated traffic volumes, densities, lengths, and number of R-E
time period 7:00 am to 8:00 am, 1:00 pm to 2:00 pm, 5:00 pm to crashes are summarized in Table 4.
6:00 pm, 8:00 pm to 9:00 pm, 9:00 am to 10:00 am, and 11:00 pm
to 12:00 am from 2006 to 2008, respectively. In the 1-km tunnel 4.3.1. Preliminary analysis by using linear regression method
In this section, we analyze the relationship between exposure
to dangerous encounter and the crash frequency in a linear manner
6
According to the Cost of Road Traffic Accidents in Singapore, a serious injury is illustrated in Fig. 7. The cumulative residual (CURE) method is a
one who has suffer injuries such as fractures or a concussion and/or internal lesions, well-recognized method to examine the goodness-of-fit of models
crushed body parts or organs, severe cuts, or severe general shock requiring medical in transportation studies (AASHTO, 2010; Hauer, 2004; Hauer and
treatment or hospitalization that prevents the person from performing ordinary
tasks for at least 7 days; a slight injury refers to one who is transported to a hospital
Banfo, 1997). Fig. 8 depicts the cumulative residuals for the linear
from the scene in an ambulance or, otherwise, one who requires subsequent medical regression models. As can be seen from Fig. 8, the linear regression
treatment entailing hospitalization and medical leave of no less than 3 days. models perform well.
260 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263

Fig. 8. Plot of cumulative residuals against the proposed safety indicator.

The statistical results for the linear regression models are further conduct another linear regression model with 0 intercept,
reported in Table 5. Surprisingly, the P-values of the coefficients as shown in Fig. 9. This indicates that the crash rate may have a pro-
with respect to constant for the three linear regression models are portional linear relationship with the proposed exposure to traffic
all greater than 0.035. By contrast, the P-values of coefficients with conflicts. The corresponding proportional coefficient is defined as
respect to R-E crash frequency are all close to 0. That is to say, causation factor (P(t)) in a linear manner, which could be consid-
the coefficients with respect to intercept are very significant. We ered as the conditional probability that vehicle crashes could not
be avoided under dangerous encounters for 1 h.

Table 5
Statistical results of linear regression models. 4.3.2. Negative binomial regression models
Constant Crash frequency R-Sq
As suggested by Hauer et al. (1988), Lord and Mannering (2010),
and Miaou and Lum (1993), it is theoretically inappropriate to
Coefficient P-value Coefficient P-value model discrete and non-negative crash count data using the con-
2s 2.5929 0.035 0.0111 0.000 0.9738 ventional linear regression method. Generalized linear modelling
3s 2.4085 0.044 0.0037 0.000 0.9744 techniques (GLIM) have the advantages of overcoming the short-
4s 2.2446 0.058 0.0022 0.000 0.9736
comings associated with linear models. Therefore, the GLIM is
 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263 261

Crash count - traffic conflicts relationship

TTC criterion = 2s TTC criterion = 3s TTC criterion = 4s
Linear (TTC criterion = 4s) Linear (TTC criterion = 3s) Linear (TTC criterion = 2s)

Number of crashes in 2006-

24
21
y = 0.0135x
y = 0.0044x
R2 = 0.9097
18 R2 = 0.9207
15
2008

12 y = 0.0026x
9
R2 = 0.9282

6
3
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000
Exposure to traffic conflicts in the 3 years

Fig. 9. R-E crash count–traffic conflicts relationship with linear fit (0 intercept).

applied to fit the model using a negative binomial distributed error Table 6

Results of negative binomial regression models.
structure.
The Negative Binomial regression model considered in this ˛ ˇ P-value Log-likelihood AIC
study has the form as presented by Miaou (1994). Model 1 (2 s) 0.044 −4.114 0.000 −19.777 41.554

 (yi + (1/˛))
 1
1/˛  ˛ yi Model 2 (3 s) 0.036 −5.244 0.000 −19.760 41.519
i Model 3 (4 s) 0.031 −5.814 0.000 −19.767 41.493
p(Yi = yi ) = (11)
 (yi + 1) (1/˛) 1 + ˛i 1 + ˛i

yi = 0, 1, 2, . . . (12)

i = E(Yi ) = Nconflict,i exp(ˇ) (13) software. The three approaches deliver the same estimators for the
two parameters, presented in Table 6.
and the variance of Yi is
From Table 6, we can see the P-values are close to 0, indicat-
Var(Yi ) = i + ˛i 2 (14) ing that the three models perform well. The Log-likelihood value
and the Akaike Information Criterion (AIC) value for each model
where Yi is a random variable representing the number of crashes are also given in the table. Note that estimated models with high
in time period i; yi is the actual number of crash count in the Log-likelihood and low AIC values are preferred. Accordingly, the
time period; Nconflict,i is the exposure to traffic conflicts in the time performances of the three models with respect to different TTC
period; ˛ ≥ 0 and is referred to as dispersion parameter. According thresholds do not have significant differences. Table 7 depicts the
to the analysis in Section 4.3.1, it is reasonable to assume the mean estimated expected values of crash counts by the three models and
value of crash count i or E(Yi ) to be proportional to the exposure the actual crash count for the six data points in this study, which
to traffic conflicts. This model assumes an exponential rate function also indicate that the models perform very well. The proportional
exp(ˇ), which ensures that the crash rate is always non-negative. coefficients of the expected values of crash counts (with nega-
The parameters (˛ and ˇ) are estimated by the three approaches tive binomial assumption) over the exposure to traffic conflicts are
(hybrid, fisher, and Newton Raphson methods) using the SPSS e−4.114 = 0.0163, e−5.244 = 0.0053, and e−5.814 = 0.0030, respectively.

Crash count - traffic volume relationship

Crash count - traffic volume Linear (Crash count - traffic volume)

Linear (Crash count - traffic volume)

24
Number of crashes in 2006-

21
18
15
2008

y = 0.0082x
12
R2 = 0.5339
9
6 y = 0.0167x - 12.435
3 R2 = 0.7323
0
800 900 1000 1100 1200 1300 1400 1500 1600 1700 1800

Traffic volume

Fig. 10. Crash count–traffic volume relationship.

262 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263

Table 7 headways. For non-interrupt traffic flows with traffic volume from
Estimated expected values of crash counts.
900 vehs/(h lane) to 1700 vehs/(h lane), the TTC distributions may
R-E crash records Estimated by Estimated by Estimated by follow the Inverse Gaussian distributions (lognormal distributions
(2006–2008) model 1 (2 s) model 2 (3 s) model 3 (4 s) are also a good approximation) and traffic volume could be con-
11 10.74 10.68 10.59 sidered as the contributing factor to the distribution parameters. It
5 4.30 4.38 4.48 should be pointed out that these perspectives need to be validated
8 5.95 6.10 6.18 using more actual data from other expressways and/or urban road
20 25.59 24.67 23.87
tunnels.
17 21.91 21.81 21.62
4 4.12 4.10 4.10 The crash data from Singapore’s road tunnels shows that linear
or proportional relationship may not be good enough to reflect the
relations between crash count and traffic volume. Instead, the linear
These coefficients are the causation factors in regression models and proportional relationships perform very well between crash
with generalized linear manner. count and exposure to traffic conflicts. This may be because not
only traffic volume but also density is taken into consideration in
4.4. Remarks on the crash count–traffic flow relationship the proposed exposure to traffic conflicts.
Other than the TTC, the deceleration rate to avoid the crash
As for the crash count–traffic flow relationship, a series of stud- (DRAC) and the post encroachment time (PET) have also been con-
ies were carried out by several scholars on the basis of the actual sidered as good safety indicators to measure the safety level in
data. Jovanis and Chang (1986) proposed a method to model the roads (Meng and Weng, 2011; Cunto and Saccomanno, 2008). Fur-
relationship between miles travelled and crash count in a time- ther study may be conducted to establish the relationship between
space domain. Hauer et al. (1988) developed a model to estimate crash frequency and the above-mentioned two safety indicators.
the safety of a signalized intersection on the basis of the traffic The comparative analysis of these three safety indicators could also
flow and crash count. Miaou and Lum (1993) compared four regres- be studied accordingly. In addition, the model can also be applied to
sion models – two conventional linear regression models and two identify the hotspots in the urban road tunnels and/or expressways
Poisson regression models – in terms of their ability to model the (Cheng and Washington, 2005; Montella, 2010).
relationship among traffic flow, geometry design, and crash count
in highways. All the above-mentioned studies acknowledged that 6. Conclusions
it was not appropriate to apply the conventional linear regression
to model crash count and traffic flow. Fig. 10 depicts the crash In this study, a novel approach is proposed to estimate the R-
count–traffic volume relationship for the present study. As can be E crash frequency in road tunnels. We first conclude that Inverse
seen in the figure, neither of the linear regression (R2 = 0.7323) and Gaussian distribution is the best-fitted distribution to the TTC data
proportional regressions (R2 = 0.5339) performs well. Since traffic based on the best-fit analysis. Accordingly, an Inverse Gaussian
volume equals to the product of density and speed (Lieu et al., 1999) regression model is applied to establish the relationship between
the densities and speeds for two road sections with the same traffic the TTC and the corresponding traffic volume. A new concept of
volume may not be the same (e.g. the two points circled in Fig. 9). In exposure to traffic conflicts is defined as the mean sojourn time in
reality, the crash frequency is also closed related to the speed and a given time period that vehicles are exposed to dangerous sce-
density of a road section (Aarts and van Schagen, 2006). Hence, it narios, namely, the TTCs are lower than a predetermined threshold
is not appropriate to assume the linear relationship between crash value. A R-E crash frequency estimation model is then proposed
count and traffic volume. on the basis of the accident records provided by the HCD database
Comparatively, as can be seen from Figs. 7 and 9, both linear for Singapore’s road tunnels. We find that the expected value of
regression and proportional regression perform well in this study. crash frequency has a proportional linear relationship with the pro-
This is because not only traffic volume but also density is taken into posed exposure to traffic conflicts. Finally, several conjectures and
account in the proposed exposure to traffic conflicts. The results recommendations are proposed.
shows that, in Singapore’s road tunnels, the exposure to traffic con- It should be pointed out that the analysis with limited data sam-
flicts based method outperforms the traffic volume based approach. ples in this study may not be adequate to validate the relationship
between crash count and the proposed indicator. A further study
5. Discussions with more data is needed to further valid and calibrate the rela-
tionship.
Theoretically, linear regression models are not appropriate to
model discrete and non-negative crash count data. GLIM is proven
Acknowledgments
to be more effective to formulate the rare events such as crash
count. However, as illustrated in Section 4.3.1, the linear regres-
We are really grateful to the two anonymous referees and the
sion models also perform well according to the CURE method and
editors whose comments improved the presentation and the con-
correlation coefficients. Therefore, it is also acceptable to formu-
tent of the earlier version. Special thanks will also be expressed to
late the relationship between crash count and proposed index in
Ms Soh Ling Tim from Land Transport Authority of Singapore on
this study. The coefficients between crash counts (or expected val-
the data collection for this project. This study is supported by the
ues of crash counts) and the exposure to traffic conflicts are defined
innovation fund of Land Transport Authority of Singapore.
as the causation factor in this study. The proposed causation factor
P(t) reflects the conditional probability that vehicle crashes have
occurred when the vehicle are exposed to dangerous scenarios for References
1 h. The probability would be dependent on vehicle conditions,
Aarts, L., van Schagen, I., 2006. Driving speed and the risk of road crashes: a review.
drivers’ abilities, and the road geometries. We conjecture that this Accident Analysis and Prevention 38, 215–224.
factor could be a constant for a given road tunnel section in the AASHTO, 2010. Highway Safety Manual, 1st edition.
long run with a given TTC threshold value. The TTC values would Beard, A., Carvel, C., January 1, 2005. The Handbook of Tunnel Fire Safety. Thomas
Telford Publishing, London.
generally have a parabola relationship with traffic volume because Cheng, W., Washington, S., 2005. Experimental evaluation of hotspot identification
they will be affected by not only speed dispersion but also distance methods. Accident Analysis and Prevention 37, 870–881.
 Q. Meng, X. Qu / Accident Analysis and Prevention 48 (2012) 254–263 263

Cunto, F., Saccomanno, F.F., 2008. Calibration and validation of simulated vehicle Malyshkina, N., Mannering, F., 2010a. Empirical assessment of the impact of highway
safety performance at signalized intersections. Accident Analysis and Prevention design exceptions on the frequency and severity of vehicle accidents. Accident
40, 1171–1179. Analysis and Prevention 42, 131–139.
Daniels, S., Brijs, T., Nuyts, E., Wets, G., 2010. Explaining variation in safety perfor- Malyshkina, N., Mannering, F., 2010b. Zero-state Markov switching count-data mod-
mance of roundabouts. Accident Analysis and Prevention 42, 393–402. els: an empirical assessment. Accident Analysis and Prevention 25, 77–84.
Farah, H., Bekhor, S., Polus, A., 2009. Risk evaluation by modeling of passing behavior Maycock, G., Hall, R.D., 1984. Accidents at 4-Arm Roundabouts. TRRL Laboratory
on two-lane rural highways. Accident Analysis and Prevention 41, 887–894. Report 1120, Transportation and Road Research Laboratory, Crowthorne, UK.
Guo, F., Wang, X., Abdel-Aty, M., 2010. Modeling signalized intersection safety with Meng, Q., Weng, J., 2011. Evaluation of rear-end crash risk at work zone using work
corridor spatial correlations. Accident Analysis and Prevention 42, 84–92. zone traffic data. Accident Analysis and Prevention 43, 1291–1300.
Haque, M.M., Chin, H.C., Huang, H., 2010. Applying Bayesian hierarchical models to Meng, Q., Qu, X., Wang, X., Yuanita, V., Wong, S.C., 2011a. Quantitative risk assess-
examine motorcycle crashes at signalized intersections. Accident Analysis and ment modeling for non-homogeneous urban road tunnels. Risk Analysis 31,
Prevention 42, 203–212. 382–403.
Hauer, E., 2004. Statistical road safety modeling. Transportation Research Record Meng, Q., Qu, X., Yong, K.T., Wong, Y.K., 2011b. QRA model based risk impact analysis
1897, 81–87. of traffic flow in urban road tunnels. Risk Analysis 31 (12), 1872–1882.
Hauer, E., Banfo, J., 1997. Two tools for finding what function links the dependent Miaou, S.P., Lum, H., 1993. Modeling vehicle accidents and highway geometric design
variable to explanatory variables. In: Proceedings ICTCT 97 (International Coop- relationships. Accident Analysis and Prevention 25, 689–709.
eration on Theories and Concepts in Traffic Safety), Lund, Sweden, pp. 1–7. Miaou, S.P., 1994. The relationship between truck accidents and geometric design of
Hauer, E., Ng, J.C.N., Lovell, J., 1988. Estimation of safety at signalized intersections. road sections: Poisson versus negative binomial regressions. Accident analysis
Transportation Research Record 1185, 48–61. and Prevention 26, 471–482.
Hauer, E., 2001. Overdispersion in modelling accidents on road sections and in Minderhoud, M.M., Bovy, P.H.L., 2001. Extended time-to-collision measures for road
Empirical Bayes estimation. Accident Analysis and Prevention 33 (6), 799–808. traffic safety assessment. Accident Analysis and Prevention 33, 89–97.
Hayward, J.C., 1972. Near Miss Determination Through Use of a Scale of Danger Montella, A., 2010. A comparative analysis of hotspot identification methods. Acci-
(Traffic Records 384). Highway Research Board, Washington, DC. dent Analysis and Prevention 42, 571–581.
Hirst, S., Graham, R., 1997. The format and presentation of collision warnings. In: Páez, A., Trépanier, M., Morency, C., 2011. Geodemographic analysis and iden-
Noy, N.I. (Ed.), Ergonomics and Safety of Intelligent Driver Interfaces. tification of potential business partnerships enabled by transit smart cards.
Hogema, J.H., Janssen, W.H., 1996. Effects of Intelligent Cruise Control on Driv- Transportation Research Part A 45, 640–652.
ing Behavior. TNO Human Factors, Report TM-1996-C-12, Soesterberg, The PIARC Technical Committee C3.3 Road Tunnel Operation, 2008. Risk
Netherlands. Analysis for Road Tunnels, May 2008, [Link]
Ibeas, Á., Cordera, R., dell’Olio, L., Moura, J.L., 2011. Modelling demand in restricted ressources/publications files/4/2234,[Link], accessed 19
parking zones. Transportation Research Part A 45, 485–498. May 2011.
Jovanis, P.P., Chang, H.L., 1986. Modeling the relationship of accidents to miles trav- Qu, X., Meng, Q., Yuanita, V., Wong, Y.K., 2011. Design and implementation of a
eled. Transportation Research Record 1068, 42–51. quantitative risk assessment software tool for Singapore’s road tunnels. Expert
Land Transport Authority (LTA), 2005. Design Safety Submission for KPE [internal Systems with Applications 38 (11), 13827–13834.
report], Singapore. Shankar, V., Milton, J., Mannering, F.L., 1997. Modeling accident frequency as
Lieu, H., Gartner, N., Messer, C.J., Rathi, A.K., 1999. Traffic Flow Theory. U.S. Depart- zero-altered probability processes: an empirical inquiry. Accident Analysis and
ment of Transportation, Federal Highway Administration. Prevention 29, 829–837.
Lord, D., Geedipally, S.R., Guikema, S., 2010. Extension of the application of Svensson, A., 1998. A method for analyzing the traffic process in a safety perspective.
Conway–Maxwell–Poisson models: analyzing traffic crash data exhibiting under Doctoral Dissertation, University of Lund, Lund, Sweden.
dispersion. Risk Analysis 30, 1268–1276. The European Parliament and the Council of the European Union, 2004. Directive
Lord, D., Guikema, S., Geedipally, S.R., 2008. Application of the 2004/54/EC of the European parliament and of the council of 29 April 2004 on
Conway–Maxwell–Poisson generalized linear model for analyzing motor minimum safety requirements for tunnels in the Trans-European Road Network.
vehicle crashes. Accident Analysis and Prevention 40, 1123–1134. Official Journal of the European Union (April), L 167/39-91.
Lord, D., Mannering, F., 2010. The statistical analysis of crash-frequency data: a Vogel, K., 2003. A comparison of headway and time to collision as safety indicators.
review and assessment of methodological alternatives. Transportation Research Accident Analysis and Prevention 35, 427–433.
Part A 44, 291–305. Whitmore, G.A., 1983. A regression method for censored Inverse Gaussian data.
Lord, D., Washington, S.P., Ivan, J.N., 2007. Further notes on the application of Canadian Journal of Statistics 11, 305–315.
zero inflated models in highway safety. Accident Analysis and Prevention 39, Zhang, Y., Xie, Y., 2007. Forecasting of short-term freeway volume with v-support
53–57. vector machines. Transportation Research Record 2024, 92–99.