category: literaturenote citekey: wongcontributoryfactorstrafficcrashes2007 title: Contributory factors to traffic crashes at signalized intersections in Hong Kong authors: "Wong, S.C.; Sze, N.N.; Li, Y.C." year: 2007 date: 2007-00-00 2007 doi: 10.1016/j.aap.2007.02.009 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2007.02.009" zotero_key: ZU67DZQQ zotero_storage: 79N7TJRX collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Wong, S.C."

status: converted

Accident Analysis and Prevention 39 (2007) 1107–1113

Contributory factors to traffic crashes at signalized intersections in Hong Kong

S.C. Wong ∗, N.N. Sze, Y.C. Li

Department of Civil Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong, China Received 4 November 2006; received in revised form 12 February 2007; accepted 17 February 2007

Abstract

Efficient geometric design and signal timing not only improve operational performance at signalized intersections by expanding capacity and reducing traffic delays, but also result in an appreciable reduction in traffic conflicts, and thus better road safety. Information on the incidence of crashes, traffic flow, geometric design, road environment, and traffic control at 262 signalized intersections in Hong Kong during 2002 and 2003 are incorporated into a crash prediction model. Poisson regression and negative binomial regression are used to quantify the influence of possible contributory factors on the incidence of killed and severe injury (KSI) crashes and slight injury crashes, respectively, while possible interventions by traffic flow are controlled. The results for the incidence of slight injury crashes reveal that the road environment, degree of curvature, and presence of tram stops are significant factors, and that traffic volume has a diminishing effect on the crash risk. The presence of tram stops, number of pedestrian streams, road environment, proportion of commercial vehicles, average lane width, and degree of curvature increase the risk of KSI crashes, but the effect of traffic volume is negligible.

Keywords: Signalized intersections; Crash incidence; Poisson regression; Negative binomial regression

1. Introduction

Conflicts between road users and traffic streams at intersections decrease capacity and increase the likelihood of crashes. However, efficient junction designs can substantially increase the traffic flow and enhance road safety, especially in congested urban road networks. Traffic signal control is a common control measure at junctions, especially in highly developed and crowded urban road networks that requires less land intake. Signalized junctions outperform other junction types because efficient signal phasing achieves the equalization of traffic delays at different approaches, resolves conflicts between different traffic streams, is flexible to variations in traffic demand, and allows crossing streams for pedestrians (Wong, 1996; Wong and Yang, 1997; Wong et al., 2006). In addition, traffic signal control can also reduce conflicts between cyclists, pedestrians, and vehicles traveling in different directions by separating their movements in time, thus maximizing traffic capacity and minimizing the crash risk.

Several models have been developed to assess the road safety performance of intersections for different junction types. Persaud et al. (2002) established a crash prediction model for injury and damage-only crashes at urban intersections in Canada that well described the relationship between the crash risk and traffic attributes, including traffic volume. The time-series crash data from the study explicitly revealed the temporal changes in safety conditions and enabled a comparison of the safety performance of junction types across different cities. Oh et al. (2003) proposed five rural intersection crash prediction models and considered the temporal and spatial transferability of the models; improved the model design, data collection, and standardization for stop-controlled and signalized intersections; and addressed concerns over the quality and availability of traffic flow data, site selection bias, and the adequacy of information on weather, environment, and driver characteristics.Bared et al. (2005) compared the safety performance of single point and tight diamond intersections using traffic flow information and crash data. Although they found no significant difference between the total numbers of crashes at the two intersections, the differences in the frequency of injury and fatality were significant, with the single point intersection being seemingly safer than the tight diamond intersection. Lyon et al. (2005) established a safety performance

^∗ Corresponding author. Tel.: +852 2859 1964; fax: +852 2559 5337. E-mail address: hhecwsc@hkucc.hku.hk (S.C. Wong).

function for a remedial design of urban signalized intersections that considered the possible influences of traffic flow, the number of approach lanes, and the number of pedestrian streams.

Specific assessments of both the operational and safety performance of signalized intersections have also been conducted. Li et al. (2004) applied the gray system method, which is a newly established multidisciplinary assessment system for subjects with inadequate information, to quantify the operational and safety performance of signalized intersections in urban areas using degree of saturation, average stopped delay, queue length, conflict ratio, and separation ratio as the five index parameters. Schattler and Datta (2004) attempted to identify the relationship between red-light violation and traffic crashes at signalized intersections and the possible intervention of signal phase changes on this relationship. A before-and-after study was also conducted to reveal the associations between signal phasing design, frequency of red-light violation, and the crash risk at signalized intersections. Kim et al. (2005) conducted a similar investigation to determine the influence of changes in the yellow time on the safety and operational performance of signalized intersections in Seoul. They found that the crash risk for left turning and straight moving traffic was reduced, although there was a concomitant decrease in junction capacity. Lum and Halim (2006) measured driver responses to a green signal countdown device at signalized intersections to determine whether drivers decided to stop or to cross during the critical phase-change period, as revealed by a change in the frequency of red-light violation with the influence of traffic volume controlled. However, possible interventions by geometric and traffic attributes may not have been adequately explored. Zhang and Prevedouros (2003) carried out a comprehensive study to evaluate the level of service, traffic delay, and safety index of signalized intersections, in which safety performance was measured in terms of the risk of potential conflicts, rather than observed crashes. A case study of two intersections illustrated the estimation of potential vehicle–vehicle and vehicle–pedestrian conflicts and accounted for the effects of geometric characteristics, traffic flow, capacity, signal timing, and perception–reaction time.

Research efforts have primarily been devoted to the production of specific crash prediction models, particularly for signalized intersections (Oh et al., 2004; Poch and Mannering, 1996), and count data models have been commonly applied to explore the relationship between traffic and geometric characteristics and the associated crash risk. Kumara and Chin (2005) applied Poisson regression to analyze the causal factors of crashes at signalized intersections in Singapore, in which the geometric, traffic flow, and control characteristics of 104 intersections were considered. Total flow, left-turning flow, number of signal phases, sight distance, right-turn channelization, surveillance cameras, median railings, and gradient were found to significantly affect the occurrence of crashes. Kumara et al. (2003) enhanced the effectiveness and efficiency of the regression technique for safety prediction by addressing concerns over violations of the statistical assumptions and regression deficiency.

Using the count data modeling approach, this study attempts to quantify the influence of various factors, including geometric design, road environment, traffic characteristics, and signal controls, on the different severity levels of crash incidents at signalized intersections in Hong Kong, controlling for the effect of exposure as measured by the aggregated traffic volume. Information is obtained from historical crash records, a traffic survey, and the signal design plans of 262 signalized intersections, and is incorporated into a Poisson regression model and a negative binomial regression model for killed and severe injury (KSI) and slight injury crashes, respectively. These specially developed crash prediction models are stratified by injury severity, and are able to adequately quantify the effect on the safety performance of the traffic characteristics, road environment, and geometric design of local urban road networks by applying the information from a comprehensive database of crash records, road infrastructure, and traffic controls. The models also attempt to determine the intervention effect of exposure on the association between the incidence of crashes and various contributory attributes by using a specific exposure variable – logarithmically transformed traffic flow – to reveal the non-linear relationship between traffic flow and the crash risk. The systematic specification with clusters of predictor variables and the exposure measure for the crash incidence models with different levels of injury severity at signalized intersections that is established in this study is rarely found in the field. Hopefully, the problematic attributes that contribute to the incidence of crashes of different severity levels at signalized intersections can be identified and appropriate remedial measures formulated.

2. Data

The crash information was obtained from the Traffic Accident Database System (TRADS) that is maintained by the Hong Kong Transport Department and the Hong Kong Police Force. In TRADS, crashes are divided into three categories according to injury severity level: fatal, serious, and slight. A fatal crash is a crash in which at least one person is killed immediately or is injured and subsequently dies of those injuries within 30 days of the accident; a serious injury crash is a crash in which one or more persons are injured and detained in hospital for more than 12 h; and a slight injury crash is a crash in which one or more persons are injured, but not to the extent that a hospital stay of more than 12 h is required. In addition to the crash risk, the effects of geometric design, traffic characteristics, and road environment on injury severity at signalized intersections are also considered. It is worth exploring the contributory factors that lead to more serious injuries in traffic crashes at signalized intersections. Hence, two separate crash prediction models are established: one for slight injury crashes and the other for killed and seriously injured (KSI) crashes.

We take advantage of the availability of a comprehensive set of traffic impact assessment (TIA) reports for the years 2002 and 2003 to obtain information about traffic volume, geometric design, road environment, and signal phasing. As the TIA studies were conducted for planning and design purposes and did not investigate the safety performance and crash record of the junctions, we envisage that there should be no marked biases in the sampling process. There are 262 signalized junctions with adequate traffic and geometrical information available for the analysis, which account for 15.7% (out of 1660) of the signalized intersections in Hong Kong.

Several previous attempts have been made to determine the relationship between the incidence of crashes and traffic volume. Miaou (1994) suggested that traffic flow could be used as a proxy for the influence of traffic intensity or exposure on safety performance, rather than a direct predictor variable. Specific safety performance functions were established for the non-linear relationship between crash incidence and exposure, with the logarithm of the traffic volume incorporated into the prediction model (Fridstrom et al., 1995; Mountain et al., 1996; Qin et al., 2004). The results were consistent, and the crash risk was found to decrease as the traffic volume increased.

In the proposed prediction model in this study, the influences of geometric design, traffic characteristics, and road environment on the crash risk, which is defined as the crash count divided by the average annual daily traffic (AADT) controlling for the possible intervention of traffic intensity on the association measures, are effectively quantified using the logarithmically transformed AADT. It is anticipated that the proportionality of the relationship between the crash risk and traffic volume will be revealed.

Geometric characteristics are reasonably expected to influence the crash frequency at signalized intersections. The geometric attributes that are considered include the number of approaches, the number of approach lanes, the number of conflict points, the number of traffic streams, the average lane width, and the reciprocal of the average turning radius. The number of approach lanes is the sum of the number of lanes at every approach to an intersection. The turning radius refers to the centerline of vehicle maneuver, with an increase in the reciprocal average turning radius implying an increase in the curvature of the corresponding turning movement and zero denoting a straight-ahead movement. Together, these geometric attributes reflect the complexity of an intersection.

In terms of traffic characteristics, the presence of tram and LRT stops and the proportion of commercial vehicles are considered to address the influence of possible vehicle–pedestrian interactions on the crash risk and traffic composition, respectively. Given the impact of the road environment on safety performance, indicating variables for geographical locations (Hong Kong Island, Kowloon, and the New Territories) are also included.

Finally, as signal-phasing schemes significantly determine junction capacity and traffic delay, and thus operational performance, information on the number of signal stages and cycle time should reveal the association between signal operation and the crash risk. The number of pedestrian streams and the presence of right turning pockets are also considered to quantify the influence of the complexity of the signal phasing scheme at busy intersections on the incidence of crashes. The number of pedestrian streams refers to the number of temporally and spatially segregated signal phases that are provided for pedestrian crossings at a signalized intersection, with zero denoting an intersection with no at-grade pedestrian crossing (Table 1).

Table 1 Summary of the signalized intersections under investigation

	Range	Mean	S.D.
Dependent variables
Slight injury crash	Min: 0; Max: 23	1.28	2.81
Killed and seriously injured	Min: 0; Max: 5	0.32	0.72
(KSI) crash
Exposure
Log (AADT)	Min: 6.81; Max: 11.71	9.91	0.76
Numeral variables
Geometric design
Number of approaches	Min: 2; Max: 5	3.56	0.61
Number of approach lanes	Min: 2; Max: 18	7.27	3.49
Number of conflict points	Min: 0; Max: 30	8.46	9.13
Number of traffic streams	Min: 1; Max: 12	6.08	2.98
Average lane width	Min: 2.7; Max: 5.5	3.38	0.38
Reciprocal of the average turning radius	Min: 0; Max: 0.2	0.08	0.03
Traffic characteristics
Proportion of commercial vehicles	Min: 0.00; Max: 0.66	0.19	0.10
Signal phasing scheme
Number of signal stages	Min: 2; Max: 7	3.07	0.90
Cycle time	Min: 44; Max: 140	92.80	19.22
Number of pedestrian crossing streams	Min: 0; Max: 8	3.50	2.09
Indicator variables
Traffic characteristics
Presence of tram stops	1: Yes; 0: No	0.07
Presence of LRT stops	1: Yes; 0: No	0.04
Road environment
Hong Kong Island	1: Yes; 0: No	0.33
Kowloon	1: Yes; 0: No	0.42
Signal phasing scheme Presence of right turning pocket	1: Yes; 0: No	0.09

Number of observations = 262.

3. Method

The objective of this study is to estimate the influences of contributory factors such as geometric characteristics, traffic characteristics, road environment, and signal control on the incidence of crashes at signalized intersections in Hong Kong, controlling for the effect of exposure. To address the possible influences of these factors on the injury risk, two separate models are developed, one for crashes that result in slight injury and one for crashes that result in death or serious injury.

To efficiently approximate the crash frequency in a discrete, non-negative, and random way, a count data model – either a Poisson regression or negative binomial regression model – is usually applied (Washington et al., 2003).

Under a Poisson distribution, the probability that an *i*th observation has yi crashes is given byP(yi) ⁼ exp(−λi)λyi i /yi!, where ^λⁱ is the expected number of crashes for the *ⁱ*th observation. The Poisson regression model is established by specifying the Poisson parameter with ^λˆi ⁼ exp(β Xi), where Xi denotes the vector for the independent variables and β the vector for the coefficient estimates.

| Results of the correlation analysis
Table 2 | | | | | | | | | | | | | |------------------------------------------------|------|-----------------------------------|-----------------------------|------------------------------|------------------------------|-----------------------|------------------------------------------------|-----------------------------------------|----------------------------|---------------|---------------------------------------------|--| | | | approaches
Log(AADT) Number of | approach lanes
Number of | conflict points
Number of | traffic streams
Number of | Average lane
width | turning radius
Reciprocal of
the average | Proportion of
commercial
vehicles | signal stages
Number of | Cycle
time | crossing streams
Number of
pedestrian | | | Log(AADT) | 1.00 | 0.24 | 0.43 | 0.17 | 0.23 | −0.25 | −0.05 | −0.01 | 0.19 | 0.04 | 0.11 | | | Number of approaches | | 1.00 | 0.44 | 0.45 | 0.51 | −0.06 | 0.27 | 0.11 | 0.29 | 0.19 | 0.39 | | | Number of approach lanes | | | 1.00 | 0.73 | 0.78 | −0.22 | −0.15 | 0.36 | 0.55 | 0.40 | 0.37 | | | Number of conflict points | | | | 1.00 | 0.95 | −0.10 | −0.17 | 0.30 | 0.62 | 0.37 | 0.34 | | | Number of traffic streams | | | | | 1.00 | −0.09 | −0.16 | 0.32 | 0.67 | 0.40 | 0.41 | | | Average lane width | | | | | | 1.00 | 0.10 | −0.05 | −0.09 | −0.10 | 0.04 | | | Reciprocal of the average | | | | | | | 1.00 | −0.04 | −0.15 | −0.06 | −0.01 | | | turning radius | | | | | | | | | | | | | | Proportion of commercial | | | | | | | | 1.00 | 0.24 | 0.26 | 0.18 | | | vehicles | | | | | | | | | | | | | | Number of signal stages | | | | | | | | | 1.00 | 0.35 | 0.30 | | | Cycle time | | | | | | | | | | 1.00 | 0.19 | | | Number of pedestrian | | | | | | | | | | | 1.00 | | | crossing streams | | | | | | | | | | | | | | | | | | | | | | | | | | |

The relationship between the crash frequency and the variables concerned can then be revealed with maximum likelihood methods by constructing the likelihood function with -L(β) ⁼ i exp(−eβ Xi)(eβ Xi) /yi!, for which the best estimate of β is obtained by maximizing the logarithm of the likelihood function, which is developed as LL(β) ⁼ i ⁻ exp(β Xi) ⁺ yiβ Xi ⁻ ln(yi!). To assess the goodness-of-fit of the proposed prediction model, a likelihood ratio statistic can be computed with LR = ⁻2(LL(βR) ⁻ LL(βU)), where LL(βR) is the restricted log likelihood function and LL(βU) the unrestricted log likelihood function.

Under the null hypothesis that there is no association between the dependent variable and the independent variables, the LRstatistic is ^χ2-distributed with a degree of freedom that is equal to the number of independent variables that are being considered. A significant LR-statistic indicates a good fit for the proposed model.

The misspecification of the Poisson regression model is generally attributed to the violation of the assumption of equality over the mean and variance of the count. When the data is subject to overdispersion (that is, the variance of the count is significantly greater than its mean), negative binomial regression should be adopted. Cameron and Trevedi (1990) suggested a test to illustrate the existence of overdispersion in the data in which the null and alternative hypotheses specified as follows.

$$\begin{cases} H_0 : \operatorname{Var}(y_i) = E(y_i) \ H_1 : \operatorname{Var}(y_i) = E(y_i) + \alpha g[E(y_i)] \end{cases},$$

where g[E(yi)] is determined by either g[E(yi)] = E(yi) or g[E(yi)] = *E*2(yi).

To perform an overdispersion test, a simple linear regression model on Z is constructed as follows:

$$z = \beta w$$

where

$$z_i = \frac{(y_i - E(y_i))^2 - y_i}{\sqrt{2}E(y_i)}$$ and $w_i = \frac{g[E(y_i)]}{\sqrt{2}E(y_i)}$

A significant ^β indicates that ^α is significant, and that ^H⁰ should be rejected in favor of *H*1. This implies that the count is overdispersed and that negative binominal regression should be adopted.

For the negative binominal model, the predicted crash count is estimated by ^λⁱ = exp(β Xi ⁺ ^εi), where ^εⁱ is a gamma-distributed error with a mean and variance that are equal to 1 and α2, respectively, and α is the overdispersion parameter. Using a negative binominal distribution that is specified by:

$$P(y_i) = \frac{\Gamma[(1/\alpha) + y_i]}{\Gamma(1/\alpha)y_i!} \left(\frac{1/\alpha}{(1/\alpha) + \lambda_i}\right)^{1/\alpha} \left(\frac{\lambda_i}{(1/\alpha) + \lambda_i}\right)^{y_i},$$

the coefficient estimates can be obtained by using the maximum likelihood approach with the following function:

$$L(\lambda) = \prod_{i} \frac{\Gamma[(1/\alpha) + y_i]}{\Gamma(1/\alpha)y_i!} \left(\frac{1/\alpha}{(1/\alpha) + \lambda_i}\right)^{1/\alpha} \left(\frac{\lambda_i}{(1/\alpha) + \lambda_i}\right)^{y_i}.$$

Table 3 Results of the overdispersion test

	Slight injury crash	KSI crash
g[E(yi)] = E(yi)	3.04**	1.90
g[E(yi)] = [E(yi)]2	2.97**	1.85

^** Significant overdispersion at the 1% level.

4. Results

An important assumption in multivariate analysis is that all of the predictor variables are statistically independent. Multicollinearity refers to the violation of this assumption, and describes a situation in which the possible correlations between predictor variables are significant. To eliminate this problem, correlation analysis is conducted before the associations are measured. The number of approaches, approach lanes, conflict points, and signal stages are subject to remarkable correlations (as shown in Table 2), and thus problematic combinations of the offending variables are avoided in the analysis of the prediction models.

After the diagnosis of the individual variables and the systematic selection process, confined multivariable analyses are conducted with the statistically significant and unoffending variable sets. The predictor variables that are applied in the final count data models are Kowloon, the presence of tram stops, the proportion of commercial vehicles, the number of pedestrian streams, the average lane width, and the reciprocal of the average turning radius. To control for the possible intervention effects of exposure on the association measures, the logarithmically transformed average annual daily traffic is also incorporated.

As shown in Table 3, the slight injury crash count is subject to significant overdispersion at the 1% level. Therefore, negative binomial regression is applied to model the slight injury crashes, whereas the KSI crash count is satisfactorily dealt with using the parent Poisson model.

As has been stated, negative binomial regression is applied to model the likelihood of a slight injury crash in response to the considerable overdispersion. Table 4 shows that Kowloon (β = 1.08, ^p < 0.001), an increase in curvature (β = 11.15, ^p < 0.01), and the presence of tram stops (β = 1.20, ^p < 0.01) all significantly increase the incidence of a slight injury crash at the 1% level. More importantly, the traffic volume (β = 0.75, p < 0.001) significantly determines the incidence of slight injury crashes at the 1% level, and the corresponding coefficient of less than 1 implies that the crash incidence increases less proportionately than the traffic volume, which means that the marginal increase diminishes under a high traffic flow condition. Generally, the proposed negative binomial regression model fits well (χ² = 491.36) with the incidence of slight injury crashes at a 99% confidence level.

Poisson regression is applied to the model of the incidence of KSI crashes. Table 4 also shows that the presence of tram stops (β = 1.49, ^p < 0.001), the increase in the proportion of commercial vehicles (β = 4.09, ^p < 0.001), the increase in the number of pedestrian streams (β = 0.17, ^p < 0.01), and the decrease in the average lane width (β ⁼ ⁻1.20, ^p < 0.01) significantly increase the incidence of KSI crashes at the 1% level. Kowloon (β = 0.60, ^p = 0.02) and the increase in curvature (β = 9.07, ^p = 0.02) also determine the incidence of KSI crashes at the 5% significance level. However, the influence of traffic volume (β = 0.28, p = 0.10) on the incidence of KSI crashes is not statistically significant. The proposed Poisson regression model fits well (χ² = 64.94) with the observed number of KSI crashes at a 99% confidence level.

5. Discussion

With the comprehensive information that is obtained from the TIA studies, we attempt to quantify the influences of geometric design, traffic characteristics, road environment, and signal

Table 4 Results of the Poisson and negative binomial regressions

	Regression model
		Slight injury crash, negative binomial	KSI crash, Poisson
	β	(t)	β	(t)
Variables
Constant	−8.49	(−3.78)**	−2.76	(−1.15)
Log(AADT)	0.75	(4.88)**	0.28	(1.67)
Kowloon area (1: yes; 0: no)	1.08	(4.67)**	0.60	(2.39)*
Presence of tram stops (1: yes; 0: no)	1.20	(2.77)**	1.49	(4.00)**
Proportion of commercial vehicles	1.54	(1.36)	4.09	(3.86)**
Number of pedestrian streams	0.06	(1.19)	0.17	(3.13)**
Average lane width	−0.33	(−1.01)	−1.20	(−2.94)**
Reciprocal of the average turning radius	11.15	(2.80)**	9.07	(2.25)*
Goodness-of-fit assessment
Overdispersion parameter	1.47	(2.91)**	–	–
Restricted log likelihood function	−592.74		−204.13
Unrestricted log likelihood function	−347.06		−171.66
Likelihood ratio test	491.36**		64.94**

^* Statistically significant at the 5% level.

^** Statistically significant at the 1% level.

phasing design on the incidence of crashes at signalized intersections in Hong Kong while controlling for the effect of exposure.

The incidence of slight injury crashes is sensitive to the likelihood of vehicular conflicts because increasing numbers of vehicle–vehicle conflicts are likely to be induced as the traffic intensity increases, and thus the traffic volume should significantly determine the incidence of slight injury crashes at signalized intersections. However, the increase in the incidence of such crashes is less proportionate to the increase in traffic volume. The results show that the crash risk, which is defined as the crash count divided by AADT, decreases when the traffic volume increases, as revealed by the fact that the coefficient that is associated with log(AADT) is 0.75, which is less than unity. This may be attributable to the fact that traffic slows down at busy junctions, thus reducing the risk of losing control and other speed-related factors that are related to traffic crashes. This non-linear relationship between traffic volume and the incidence of slight injury crashes is consistent with previous findings, in which the coefficient that is associated with log(AADT) ranges from 0.25 to 0.56 in the studies of Oh et al. (2004) and Bared et al. (2005). Traffic flow data that are categorized by approach, traffic stream, or lane may be a more precise measure of the influence of traffic intensity on the incidence of crashes (Miaou, 1994; Wang et al., 2003), but these disaggregated data are not available for this study.

Incorporating the traffic volume with logarithmic transformation into the prediction model explicitly controls for the exposure effect on the crash incidence and the non-linear relationship between crash incidence and traffic flow, and also allows the effective measurement of the association between the crash rate, geometric design, road environment, and other attributes. The road environment noticeably determines the incidence of crashes. Signalized intersections in Kowloon have a higher risk of slight injury crashes, which may be attributable to difference in its land-use characteristics. The Kowloon urban area was developed early on in Hong Kong's history and is densely populated with numerous commercial outfits, and thus its high crash incidence may be associated with frequent off-road commercial activities. Aggressive driving behavior in this congested urban area may also induce a higher crash risk. The increase in curvature substantially hinders the maneuvers of the turning traffic stream and induces potential conflicts, especially for heavy trucks and buses, which also probably causes more crashes. Appropriate remedial measures, such as clearer road markings and warning signs, restrictions on off-road parking, and signal phase modifications, are thus essential to reduce crashes at signalized intersections.

The presence of tram stops also increases the number of conflicts among mixed traffic when vehicular speeds are more heterogeneous. At certain congested intersections in Hong Kong, vehicles are allowed to travel along the tram tracks, but this may further increase the crash likelihood due the lack of awareness among tram passengers of the passing of vehicular traffic on both sides of the tram stop. Protective railings at tram stops and the introduction of tram-only lanes could greatly reduce the incidence of crashes that are attributable to the presence of tram stops.

As KSI crashes are rare events, the incidence of such crashes is less sensitive to exposure. An increase in the KSI crash likelihood is largely attributable to the crash mechanism and energy dissipated upon collision, and thus no significant association can be found between the occurrence of KSI crashes and traffic volume. These rare events are not subject to the dispersion effect that is observed for slight injury crashes, and thus the Poisson regression works well in predicting KSI crashes.

The number of pedestrian crossing streams increases with the incidence of KSI crashes at signalized intersections. This increase in injury risk may be attributable to the strong collision force that is exerted on pedestrians in vehicle–pedestrian crashes. Crashes that involve pedestrians may be associated with jaywalking behavior or other pedestrian–vehicle interactions on the crossing streams. The presence of tram stops also leads to more pedestrian–vehicle interactions because of the pick-up, drop-off, and pedestrian crossing activities in these areas. It is therefore recommended that grade-separated pedestrian crossing facilities be provided to avoid potential vehicle–pedestrian conflicts and thus enhance road safety.

The traffic composition also determines the incidence of KSI crashes at signalized intersections, with the proportion of commercial vehicles – mainly heavy trucks and buses – increasing the injury risk. This is attributable to the stronger force of impact and higher energy dissipation upon collision in crashes that involve these types of vehicles. The aggressive driving behavior of occupational drivers, such as driving at higher speeds and engaging in dangerous maneuvers, also increases the injury risk for other road users. Restrictions on the flow of commercial vehicles may not be practical, but educational driver improvement schemes would obviously enhance road safety.

Finally, a decrease in the average lane width increases the likelihood of KSI crashes at signalized intersections. This increase in the injury risk may be attributable to the stronger force of impact that results from the reduced time for driver reaction and corresponding defensive vehicle maneuvers that is due to the reduced space. The application of appropriate remedial measures to the road infrastructure and attempts to widen the roads could allow more time for defensive behavior and thus minimize the injury risk. However, where road widening is impractical, efficient traffic channelization through the modification of the signal stage, both temporally and spatially, should be sufficient to reduce crashes between conflicting turning traffic streams.

We note that the casual contribution to crashes at intersections may not be fully inferred from the regression analysis. In the forgoing discussion, we merely attempt to hypothesize the possible associations between the dependent and independent variables to highlight areas that may be worthy of further exploration.

6. Conclusions

The study that is reported herein attempts to evaluate the associations between the incidence of crashes, geometric design, traffic characteristics, road environment, and traffic control at signalized intersections in Hong Kong, controlling for the influence of exposure. Information from historical accident records, a traffic survey, and the signal phase design of 262 signal intersections, which account for one-sixth of the total signalized intersections in Hong Kong, is incorporated into the proposed count data models in a comprehensive approach to the determination of the safety performance of signalized intersections.

Poisson regression and negative binomial regression are applied to estimate the incidence of KSI crashes and slight injury crashes, respectively. The road environment, curvature, and presence of tram stops significantly determine the incidence of slight injury crashes, which is attributable to the increased likelihood of vehicle–vehicle and vehicle–pedestrian conflicts that result from the higher traffic intensity and more frequent pedestrian activities that are associated with these factors. However, the presence of tram stops, number of pedestrian streams, road environment, traffic composition, lane width, and curvature also noticeably determined the incidence of KSI crashes. KSI crashes are rare incidents that are unlikely to be sensitive to traffic volume, and thus these associations may be attributable to the force of impact upon collision. Grade-separated pedestrian crossing facilities and efficient traffic channelization could enhance the efficiency, capacity, and safety of signalized intersections.

In an extended study, the crash prediction model for signalized intersections could be improved by incorporating information on the volume of conflicting traffic streams, pedestrian flow, and junction types. A comprehensive model could be established with disaggregated crash data on traffic volume, approaches, and lanes obtained from an extensive traffic survey and stratified by collision type and injury severity.

Acknowledgements

We would like to thank Mr. T.F. Leung and his colleagues at the Road Safety and Standards Division of the Transport Department of the Hong Kong Special Administrative Region (HKSAR) Government for helping to obtain the crash data for the research work that is reported in this paper. The work that is described here was jointly supported by grants from the University Research Committee of the University of Hong Kong (10206125, 10207394), the Research Grants Council of the Hong Kong Special Administrative Region of China (Project No.: HKU7187/05E), the National Natural Science Foundation of China (70629001), and the National Basic Research Program of China (2006CB705503).

References

Bared, J., Powell, A., Kaisar, E., Jagannathan, R., 2005. Crash comparison of single point and tight diamond interchanges. J. Transport. Eng. ASCE 131 (5), 379–381.

Cameron, A., Trevedi, P., 1990. Regression based tests for overdispersion in the Poisson model. J. Econometrics 46, 347–364.
Fridstrom, L., Ifver, J., Ingebrigtsen, S., Kulmala, R., Thomsen, L.K., 1995. Measuring the contribution of randomness, exposure, weather, and daylight to the variation in road accident counts. Accident Anal. Prev. 27 (1), 1–20.
Kim, H.J., Son, B., Lee, S., Park, J., 2005. Improving yellow time method of leftturning traffic flow at signalized intersection. In: Computational Science and ITS applications—ICCSA 2005, PT 2 Lecture Notes in Computer Science 3480, pp. 787–797.
Kumara, S.S.P., Chin, H.C., 2005. Application of Poisson underreporting model to examine crash frequencies at signalized three-legged intersections. Transport. Res. Record 1908, 46–50.
Kumara, S.S.P., Chin, H.C., Weerakoon, W.M.S.B., 2003. Identification of accident causal factors and prediction of hazardousness. Transport. Res. Record 1840, 116–122.
Li, J., Yue, Z.Q., Wong, S.C., 2004. Performance evaluation of signalized urban intersections under mixed traffic conditions by gray system theory. J. Transport. Eng. ASCE 130, 113–121.
Lum, K.M., Halim, H., 2006. A before-and-after study on green signal countdown device installation. Transport. Res. F 9, 29–41.
Lyon, C., Haq, A., Persaud, B., Kodama, S.T., 2005. Safety performance functions for signalized intersections in large urban areas—development and application to evaluation of left-turn priority treatment. Transport. Res. Record 1908, 165–171.
Miaou, S., 1994. The relationship between truck accidents and geometric design of road sections: Poisson versus negative binomial regressions. Accident Anal. Prev. 26 (4), 471–482.
Mountain, L., Fawaz, B., Jarrett, D., 1996. Accident prediction models for roads with minor junctions. Accident Anal. Prev. 28 (6), 695–707.
Oh, J., Lyon, C., Washington, S., Persaud, B., 2003. Validation of FHWH crash models for rural intersections: lessons learned. Transport. Res. Record 1840, 41–49.
Oh, J., Washington, S., Choi, K., 2004. Development of accident prediction models for rural highway intersections. Transport. Res. Record 1897, 18–27.
Persaud, B., Lord, D., Palmisano, J., 2002. Calibration and transferability of accident prediction models for urban intersections. Transport. Res. Record 1784, 57–64.
Poch, M., Mannering, F., 1996. Negative binomial analysis of intersection accident frequencies. J. Transport. Eng. ASCE 122, 105–113.
Qin, X., Ivan, J.N., Ravishanker, N., 2004. Selecting exposure measures in crash rate prediction for two-lane highway segments. Accident Anal. Prev. 36, 183–191.
Schattler, K.L., Datta, J.K., 2004. Driver behavior characteristics at urban signalized intersections. Transport. Res. Record 1862, 17–23.
Wang, Y., Ieda, H., Mannering, F., 2003. Estimating rear-end accident probabilities at signalized intersections: occurrence-mechanism approach. J. Transport. Eng. ASCE 129 (4), 377–384.
Washington, S.P., Karlaftis, M.G., Mannering, F.L., 2003. Statistical and Econometric Methods for Transportation Data Analysis. CRC Press, New York, United States.
Wong, S.C., 1996. Group-based optimisation of signal timings using the TRAN-SYT traffic model. Transport. Res. B 30, 217–244.
Wong, S.C., Yang, H., 1997. Reserve capacity of a signal-controlled road network. Transport. Res. B 31, 397–402.
Wong, C.K., Wong, S.C., Tong, C.O., 2006. A lane-based optimization method for the multi-period analysis of isolated signal-controlled junctions. Transportmetrica 2, 53–85.
Zhang, L., Prevedouros, P.D., 2003. Signalized intersection level of service incorporating safety risk. Transport. Res. Record 1852, 77–86.

wongcontributoryfactorstrafficcrashes2007.md 49 KB 文件歷史 原始文件