A field investigation of red-light-running in Shanghai, China

Transportation Research Part F 37 (2016) 144–153

Contents lists available at ScienceDirect

Transportation Research Part F journal homepage: www.elsevier.com/locate/trf

A field investigation of red-light-running in Shanghai, China Xuesong Wang a,b,c, Rongjie Yu a,b,c,⇑, Chujun Zhong d a

Road and Traffic Key Laboratory, Ministry of Education, Shanghai 201804, China Jiangsu Province Collaborative Innovation Center of Modern Urban Traffic Technologies, SiPaiLou #2, Nanjing 210096, China College of Transportation Engineering, Tongji University, 4800 Cao’an Road, Shanghai 201804, China d Zachry of Civil Engineering, Texas A&M University, College Station 77840, United States b c

a r t i c l e

i n f o

Article history: Received 1 May 2014 Received in revised form 3 April 2015 Accepted 13 December 2015

Keywords: Signalized intersection Red-light-running Field investigation Comparison group Driver characteristics Random effects logistic regression model

a b s t r a c t Red-Light-Running (RLR) is the major cause of severe injury crashes at signalized intersections for both China and the US. As several studies have been conducted to identify the influencing factors of RLR behavior in the US, no similar studies exist in China. To fill this gap, this study was conducted to identify the key factors that affect RLR and compare the contributing factors between US and China. Data were collected through field observations and video recordings; four intersections in Shanghai were selected as the study sites. Both RLR drivers and comparison drivers, who had the opportunity to run the light but did not, were identified. Based on the collected data, preliminary analyses were firstly conducted to identify the features of the RLR and comparison groups. It was determined that: around 57% of RLR crossed the stop line during the 0–0.4 second time interval after red-light onset, and the numbers of red light violators decreased as the time increased; among the RLR vehicles, 38% turned left and 62% went straight; and at the onset of red, about 88% of RLR vehicles were in the middle of a vehicle platoon. Furthermore, in order to compare the RLR group and non-RLR group, two types of logistic regression models were developed. The ordinary logistic regression model was developed to identify the significant variables from the aspects of driver characteristics, driving conditions, and vehicle types. It was concluded that RLR drivers are more likely to be male, have local license plates, and are driving passenger vehicles but without passengers. Large traffic volume also increased the likelihood of RLR. However, the ordinary logistic regression model only considers influencing factors at the vehicle level: different intersection design and signal settings may also have impact on RLR behaviors. Therefore, in order to account for unobserved heterogeneity among different types of intersections, a random effects logistic regression model was adopted. Through the model comparisons, it has been identified that the model goodness-of-fit was substantially improved through considering the heterogeneity effects at intersections. Finally, benefits of this study and the analysis results were discussed. Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction Red-Light-Running (RLR) is a common traffic violation and the major cause of crashes at signalized intersections (Wang, Zhang, & Wang, 2011). In the United States, RLR is associated with about 260,000 crashes and 750 fatalities each year

⇑ Corresponding author at: College of Transportation Engineering, Tongji University, 4800 Cao’an Road, Shanghai 201804, China. Tel.: +86 21 69583946. E-mail address: [email protected] (R. Yu). http://dx.doi.org/10.1016/j.trf.2015.12.010 1369-8478/Ó 2015 Elsevier Ltd. All rights reserved.

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

145

(Retting & Williams, 1996). In China, according to the statistics revealed by The Ministry of Public Security (2012), 4227 severe injury crashes and 789 fatalities between January and October 2012 were attributable to RLR. Given the similar traffic safety problems caused by RLR in China and the US, there are major gaps between the two countries in the aspects of traffic regulations, enforcement procedures, and signal settings. For example, traffic regulations in China do not allow vehicles to enter the intersection during the yellow phase, whereas it is legal in the US. Another difference is that red light cameras are more frequently used in China compared to the US. Regarding the signal settings, green signal countdown displays are commonly utilized in China to make it easier for drivers to anticipate the end of the green phase, to avoid entering the intersection during the yellow phase. Previous studies focused on RLR in the US have provided important findings regarding the characteristics of RLR behavior, which include the characteristics of drivers and corresponding driving conditions. However, RLR studies in China have only focused on the characteristics of RLR vehicles; no comparison studies that investigate factors that would separate RLR and non-RLR vehicles have been conducted. This study fills the gap through acquiring data for both RLR and non-RLR vehicles, and the development of models to identify the influencing factors for RLR events. Results from this study will be compared to studies in the US, which will further help us to understand the RLR events across different countries. Data of RLR drivers and comparison drivers (who did not run the red lights) were collected at four intersections in the urban area of Shanghai. Drivers’ genders, safety belt use, hand-held cell phone use, and presence of passengers were manually recorded by observers at each intersection and double checked through video recordings; drivers’ vehicle operations as they approached and traveled through the intersections were recorded by video cameras. The characteristics of RLR drivers and comparison drivers were then compared through preliminary analysis with a Chi-square test and systematic modeling analysis with an ordinary logistic regression model. However, the ordinary logistic regression models only have the capability of analyzing variables at the vehicle level; factors at the intersection level (such as position of traffic signals and lane markings) may also have substantial influence on red-light running behavior. Because these variables were not included in the ordinary logistic regression model analysis due to the small sample size, a random effects logistic regression model was utilized to capture the influence of unobserved heterogeneity across the intersections. 2. Background Previous RLR studies in the US have examined various aspects of RLR, which include RLR prevalence, frequency, antecedents (e.g., signal control, cycle length), and correlates (e.g., age, gender). For example, Retting and Williams (1996) conducted an on-site survey in Arlington County, Virginia, and observed 462 RLR drivers and 911 non-RLR drivers during 234 h of data collection. They found that 48% of RLR drivers entered the intersection 0.5–0.9 s after red onset; 34% at 1.0–1.4 s; 11% at 1.5–1.9 s; and 7% at least 2.0 s. It was also identified that longer red intervals were associated with higher RLR frequencies. Additionally, Yang and Najm (2007) analyzed 47,000 red light violations captured by enforcement cameras from 11 signalized intersections in the city of Sacramento, California. They found that the 8:00 PM to 5:00 AM off-peak period had fewer RLR drivers, but that the RLR drivers showed a higher probability of entering intersections two or more seconds after red onset. Regarding driver characteristics, RLR drivers have been found to be younger, less likely to be belted, have worse driving records, and drive smaller and older vehicles than the non-RLR drivers (Retting & Williams, 1996). Other researchers have also found RLR drivers to be unbelted (Porter & England, 2000); they are more likely to be male (Retting, Ulmer, & Williams, 1999) and without passengers (Porter & Berry, 2001). Signaling, traffic, and geometric variables have also been identified to be associated with RLR. When the yellow signal duration is under 3.5 s, drivers are more likely to run red lights compared to longer yellow signal timing (Brewer, Bonneson, & Zimmerman, 2002). However, longer yellow timing alone does not always eliminate the need for better enforcement (Retting, Ferguson, & Farmer, 2008), since it also has been observed that shorter duration light cycles are associated with higher rates of RLR (FHWA, 2009). Traffic environment variables have also been determined to have significant impact on red-light-running, including higher volumes, closer vehicle proximity to the intersection, and higher approaching speeds (Chang, Messer, & Santiago, 1985). Geometric variables associated with increased RLR rates include level and uphill approaches (FHWA, 2009), and wider approaches (Bonneson, Brewer, & Zimmerman, 2001). In addition, Elmitiny, Yan, and Radwan (2010) took advantage of three camera videos and observers to record driver behavior at an intersection in Orlando, and identified positive relationships between red-light-running and vehicle speed. The enforcement strategy of using red light cameras has been shown to reduce the frequency of RLR. An interrupted time series design study with a comparison group was conducted, and concluded that drivers are about 3.4 times more likely to run red lights when there is no camera present, as compared to intersections with cameras (Martinez & Porter, 2006). However, although the cameras reduce RLR behaviors, occurrence of rear-end crashes can increase (Shin & Washington, 2007). In China, studies related to RLR are limited. The primary method for investigating RLR has been through questionnaire analysis (Zhang, He, & Sun, 2009; Xian & Han, 2010; Bai, Qi, Zhao, 2007), sometimes combined with video recording (Bai et al., 2007). Wang (2006) examined the nature of the dilemma zone by looking at signal control and RLR. It was found that the wide use of red light running enforcement cameras has substantially reduced red light running behavior. Xian and Han (2010) studied responses of drivers during signal countdown using video observation, and found that about 40% of drivers crossed the intersection within the last seconds of the green interval. Based on a before–after survey, they concluded that the

146

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

green light countdown increased drivers’ competitive tendencies, resulting in more traffic crashes. Bai et al. (2007) observed an intersection in Harbin, a city in northeastern China, and found that RLR vehicle approach speeds were concentrated at about 45 km/h. Because previous RLR studies in China have only focused on the characteristics of RLR drivers, this study fills the gap through acquiring data for both the RLR and non-RLR vehicles and developing logistic regression models to identify the key factors that affect RLR probability. 3. Data collection 3.1. Intersection selection The following criteria were adopted to select intersections in this study: (1) (2) (3) (4)

Selected intersections must be located in different areas of the city. Approach volumes must be large enough to allow sufficient RLR events to be captured. Intersection signal controls must have left-turn phases. Intersections must have signal countdown indicators.

Based on the abovementioned criteria, four intersections in Shanghai were selected. The characteristics of four selected intersections are listed in Table 1. All the characteristics of the four intersections satisfy the criteria listed above. In order to acquaint the reader with the details of the four intersections, Fig. 1 additionally shows the direction of the lanes of the observed approaches. 3.2. Data collection procedures To ensure the reliability of collected data, two pre-data collection procedures were conducted to improve the collection plan before the formal data collection procedure. In addition, the manually recorded variables were double checked with high-quality video data. Data were collected for both the RLR group and the comparison (non-RLR) group in order to identify the key factors that would differentiate the RLR drivers and non-RLR drivers, which can be used to guide the development of countermeasures. Vehicles that went through the intersection after the traffic signal light turned red were labeled as RLR events. RLR vehicles were recorded on each successive cycle during the recording period; however, if there were more than one RLR vehicle within a cycle, only the last one was recorded in order to acquire a complete and accurate record of all the data elements on the manual record form. Drivers who did not go through the intersection when the signal light turned red on a cycle when there was an RLR vehicle, and who were within 1 car length of the RLR vehicle on an adjacent lane, were placed into the comparison group (see Fig. 2 for illustration). Vehicle and driver data were collected between 7:00 AM and 10:00 AM at four intersections using video recording devices and trained observers during April and May of 2012. The 7:00 AM and 10:00 AM time periods were selected to represent the morning peak hours. SONY HDR-CX180E HD video cameras were installed in front of the approach of each intersection (see Fig. 3). They were capable of recording the operation of all vehicles crossing the intersections and the signal light cycles for about 3 continuous hours. These high definition cameras (1440
1080 resolutions) allowed for a count of total traffic volume and the exact time each RLR entered the intersection after red onset. The videos recorded 471 RLRs during the investigation hours. However, the observers only recorded the last RLR of one light cycle, thus a total of 304 RLRs were left after video and manual records were matched and incomplete records were removed. A total of 317 comparison vehicles were recorded by observers. Fig. 3 below shows the intersection configurations and positions of the observers and cameras. O1 in Fig. 3 recorded the data of the RLR events and O2 recorded the characteristics of the control group, while C1 (in the same place as O1) was used to record intersection operations.

Table 1 Characteristics of the four observed intersections.

*

Intersection ID

I1

I2

I3

I4

Observed approaches Crossing approaches Area types Estimated volume Length of yellow duration in seconds Left-turn phase Countdown indicator

Lujiabang Rd. South Xizang Rd. Urban High* 3 Yes Yes

Zhaojiabang Rd. Dong’an Rd. Urban High 3 Yes Yes

Wuning Rd. Caoyang Rd. Suburban High 3 Yes Yes

Cao’an Rd North Jiasong Rd. Suburban High 3 Yes Yes

High refers to volume over 900 vehicles per 15 min during peak hours.

147

Lujiabang Rd.

Zhaojiabang Rd.

Caoyang Rd.

I4

Wuning Rd.

Cao'an Highway

Caoyang Rd.

Wuning Rd.

Cao'an Highway

North Jiasong Rd.

I3

Zhaojiabang Rd.

Dong'an Rd.

South Xizang Rd.

Lujiabang Rd.

Dong'an Rd.

I2

North Jiasong Rd.

I1

South Xizang Rd

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

Fig. 1. Diagrams of the four intersections.

Comparison group control group RLRred-light-running group stop line

Fig. 2. Comparison group.

3.3. Data elements and coding Traffic environment features, RLR behavior characteristics, and vehicle types for the RLR group and comparison group were acquired. The elements within each category and descriptions of how each variable were coded as shown in Table 2. Please note that the times of RLRs entering the intersection are coded into 10 behavior characteristics classes; the vehicle characteristics capacity of small cars is defined as fewer than five passengers, and large cars as over 10 passengers.

148

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

Fig. 3. Camera and observer locations.

Table 2 Data elements and coding. Category

Data element

Description & attributes

Traffic environment

Traffic volume Cycle length Approach lanes

Approach volume for through & left lanes of Westbound and Eastbound in seconds Numbers of through & left lanes

RLR behavior characteristics

Red-light-running Travel direction Red light running time (in seconds) Position in the platoon

Yes or no Through or left 0–0.4, 0.5–0.9, 1.0–1.4, 1.5–1.9, 2.0–2.4, 2.5–2.9, 3.0–3.4, 3.5–3.9, at least 4.0 s after red onset, or 3.0 s before green onset First or middle

Vehicle characteristics

License plate Car model Vehicle size

Local or nonlocal Passenger car or others Small, medium, or large

RLR and control driver characteristics

Gender Safety belt Hands held cell phone use Passenger presence

Male or female Yes or no Yes or no Yes or no

4. Preliminary analysis Preliminary analyses were conducted on the following aspects: hourly traffic volume, RLR frequency and distribution, RLR traffic behavior characteristics, vehicle characteristics, and RLR driver characteristics. Comparisons between the RLR group and the comparison group were conducted; the descriptive statistical and Chi-square test results are presented below. 4.1. Hourly traffic volume and RLR The hourly traffic volume and RLR frequency for each of the four studied intersections are shown in Table 3. The mean percentage of RLRs for the four intersections was 1.20% for through traffic and 4.08% for left-turn traffic, which indicates that left-turn traffic is more likely to include RLR vehicles. 4.2. Traffic behavior characteristics of RLR RLR behavior characteristics of travel direction, vehicle position in the traffic platoon, and time entering the intersection after red onset were analyzed as follows (Table 4). Approximately 40% of RLR vehicles turned left, so considering the volume of left-turn traffic as compared with through-traffic volume, left-turn RLRs are more frequently occurring than throughtraffic RLRs. About 88% of RLR vehicles were in the middle of a platoon, while only 12% were traveling in front of a platoon. Fig. 4 below shows the temporal distribution of the vehicles as they crossed the stop lines at the 4 intersections. Most RLR (57%) occurred within the first 0.4 s, with progressively fewer violations occurring as time into the red period continued. However, the probability of RLR again increased as the final 3 s before the green onset approached.

149

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153 Table 3 Hourly traffic volume and RLR frequency. Intersection ID

Drive direction

Volume (Veh)

RLR frequency

RLR percentage (%)

I1

Going through Turning-left Going through Turning-left Going through Turning-left Going through Turning-left

2237 565 3776 479 3029 268 1313 511

45 33 42 15 20 15 13 9

2.01 5.84 1.11 3.13 0.66 5.60 0.99 1.76

I2 I3 I4

Table 4 RLR travel direction and position in platoon. Item

Category

I1

I2

I4

RLR total

RLR proportion (%)

Drive direction

Going through Turning-left

111 82

83 29

I3 59 44

38 25

291 180

61.78 38.22

Platoon position

First Middle

18 175

16 96

2 101

21 42

57 414

12.10 87.90

300 267

RLR Frequency

250 200 150 88

100

50

50

25

12

12

2.0 to 2.4

2.5 to 2.9

2

0 0.0 to 0.4

300

0.5 to 0.9

1.0 to 1.4

1.5 to 1.9

0

4

3.0 to 3.4

3.5 to 4.0 to 3 3.9 seconds before green Time Distribution (seconds after red onset) onset

11 within 3 seconds before green onset

267 (57%)

RLR Frequency

250 200 150

(19%) 88

100

(10%) 50

50

(5%) 25

(3%) (2%) (0.3%) (3%) (0.0%) (0.6%) 12

12

2.0 to 2.4

2.5 to 2.9

2

0

4

11

0 0.0 to 0.4

0.5 to 0.9

1.0 to 1.4

1.5 to 1.9

3.0 to 3.4

3.5 to 4.0 to 3 3.9 seconds before green Time Distribution (seconds after red onset) onset

within 3 seconds before green onset

Fig. 4. Time distribution (in seconds) of RLRs entering the intersection (N = 471).

4.3. RLR vehicle characteristics The vehicle and driver characteristics for the RLR and comparison group were compared. Fig. 5 shows the vehicle registration (owner residence location), type (truck or passenger car), and size. From the figures it can be seen that: (i) local (in-state) vehicles were more likely to run the light than non-local (out-of-state) vehicles; (ii) passenger cars, mostly small in

150

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

size with quick acceleration, were, as expected, more likely to run red lights than trucks; and (iii) vehicle size seemed have no effect on RLR probability. 4.4. RLR driver characteristics Comparisons of the RLR and the control groups were made on driver gender, safety belt usage, passenger presence, and cell phone usage. From Fig. 6 it can be seen that: (i) female drivers ran red lights less frequently than male drivers; (ii) safety belt usage seems to have no influence on RLR; (iii) the presence of passengers would appear to reduce the RLR likelihood, as drivers without passengers are more likely to speed up to cross the intersection with the red onset; and (iv) there was no difference in cell phone usage between groups. 5. Modeling analysis In addition to the preliminary analysis, logistic regression models were developed for the purpose of identifying influencing factors in RLR events. The following sections illustrate the methodology of the logistic regression models and the modeling results. 5.1. Methodology There are two types of logistic regression models employed in this study: an ordinary logistic regression model and a random effects logistic regression model. Since the data were collected at four different intersections, the random effects logistic regression model was adopted to account for the unobserved heterogeneity among different locations. Bayesian inference technique was adopted here for the purpose of exploiting the MCMC (Markov Chain Monte Carlo) sampling to estimate the random effects models. The Bayesian random effects logistic regression models have been proven to be useful in accounting for the unobserved heterogeneity in different traffic safety studies (Huang, Chin, & Haque 2008; Yu & Abdel-Aty, 2013; Yu, Abdel-Aty, Ahmed, & Wang 2014). Suppose the RLR has the outcomes y = 1 or y = 0 with respective probability p and 1  p. The random effects logistic regression can be setup as follows:

y  Binomial ðpÞ  logit ðpÞ ¼ log

 p ¼ b0 þ Xb þ uj ðiÞ 1p

where b0 is the intercept, X is the vector of the explanatory variables, b is the vector of coefficients for the explanatory variables. uj is the random effects variable, which represents the intersection of specified random effects in this study. The intersection of random effects would account for the unobserved influence factors that affect RLR. The random effects are set to follow a normal distribution uj  Nð0; sÞ; j ¼ 1; 2; 3; 4 where s is the precision parameter and it was specified a gamma prior as s  Gamma (0.001, 0.001). For the explanatory variables, non-informative priors were set to follow normal distribution (Normal (0, 0.001)). Bayesian inference was employed in this study. For each model, three chains of 15,000 iterations were set up in WinBUGS (Lunn, Thomas, Best, & Spiegelhalter, 2000), and 5000 iterations were used in the burn-in step. The model convergences have

Record Frequency

300

255 (84%)

250

350 (71%) 225

300

(97%)

(94%)

(83%) 254

(84%) 266

250

250

200

300

298

294

200

200 150

150 92

100

150 100

100

49 50

50

0

19

10

0 RLR

Comparison

= 14.723, p = 0.001 Local

Nonlocal

(a) Vehicle registration

50

26 24

28 23

0 RLR

Comparison

= 2.549, p = 0.110 Passenger Car

Truck

(b) Car model

Fig. 5. Vehicle characteristics for RLR and controls.

RLR

Comparison

= 0.100, p = 0.951 Small

Medium

Large

(c) Vehicle size

151

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

Record Frequency

350 300

200 288 (95%)

281 (89%) 160

138

132

250 200

120

150

80

100 50

36

16

40 0

0

RLR

Comparison

RLR

= 7.510, p = 0.006 Male

350 302

Comparison 1 = 0.001, p = 0.978

Female

Yes

(a) Driver gender

Record Frequency

179(56%)

172(57%)

No

(b) Safety belt use 310

300

200 160

250 200

120

150

80

179 (48%)166 151

(41%) 125

100 50

(0.7%) 2

(2.2%) 7

0

40 0

RLR

Comparison = 2.611, p = 0.106 Yes No

(c) Hand cell phone use

RLR

Comparison = 2.618, p = 0.102 Yes No

(d) Passenger presence

Fig. 6. Driver characteristics comparisons: RLR and control.

been checked by monitoring the MCMC trace plots for the model parameters: if all values are within a zone without strong periodicities or tendencies, the model is considered convergent. Deviance Information Criterion (DIC) was selected as the evaluation measure. The DIC, recognized as a Bayesian generalization of AIC (Akaike information criterion), is a combination of the measure of model fitting and the effective number of parameters. A smaller DIC indicates a better model fitting. According to Spiegelhalter, Thomas, Best, and Lunn (2003), differences greater than 10 can rule out the reliability of a model with a higher DIC. Differences between 5 and 10 are considered substantial. 5.2. Modeling results The modeling results for the ordinary logistic regression model and the random effects logistic regression model are shown in Table 5. For Bayesian estimation results, significances of variables are determined by the credible interval. Given a 95% credible interval in Table 5, if the left bound (2.5%) and the right bound (97.5%) hold consistent estimates, this variable would be considered as significant at the 95% level. From the results it can be seen that the two models shared similar significant factors. A total of five variables are shown to significantly affect the RLR behavior; they include driving conditions, characteristics of drivers, and vehicle type. Among driving conditions, traffic volume is significant with a positive sign at the 95% level, which indicates that as the traffic volume increases, the likelihood of having a RLR would increase. Porter and England (2000) reached the same conclusion that traffic volume can be seen as an exposure factor for RLR. Travel with passengers is also significant with a negative sign at the 95% level: drivers traveling with passengers are less likely to run red lights. This can be understood as the presence of passengers increasing drivers’ attention to safety, and passengers also reminding drivers of the ending green light (Porter & Berry, 2001). The second significant driver characteristic is gender. Compared to female drivers, male drivers are more likely to cross the intersection at the onset of red (this variable is significant in the random effects logistic regression model at the 95% level, and is significant in the ordinary logistic regression model at the 90% level); this result is consistent with a previous study (Retting et al., 1999).

152

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

Table 5 Modeling results of the logistic regression models. Variables

s

# of observations DIC *

Random effects logistic regression model

Estimates

Estimates

3.25** 0.48** 0.59* 0.89** 0.41** 0.011**  621 772.82

Intercept In-state (vs. out-of-state) Male (vs. female) Passenger vehicle (vs. other vehicle types, e.g., trucks) Travel with passengers (vs. drive alone) Traffic volume

**

Ordinary logistic regression model 95% credible interval 2.5%

97.5%

5.01 0.009 0.11 0.001 0.82 0.005 –

1.96 0.018 1.34 2.30 0.008 0.018 –

2.84** 0.26** 0.47** 0.59** 0.52** 0.015** 2.12 621 749.62

95% credible interval 2.5%

97.5%

5.11 0.002 0.02 0.006 0.96 0.002 0.16

0.58 0.74 1.22 0.74 0.09 0.029 7.44

Significant at 90% level. Significant at 95% level.

4.0 [3]

2.0

[4]

[1] [2]

0.0

-2.0

-4.0

Intersection 1

Intersection 2 Intersection 3 Intersection 4

Intersections Fig. 7. Boxplot of intersection random effects.

Two vehicle information variables were found to be significant: (i) in-state vehicles are more likely to run red lights compared to out-of-state vehicles; (ii) the greater incidence of passenger vehicle RLRs over that of other vehicle types is significant in both models at the 95% level. Based on the DIC values, it can be seen that the random effects logistic regression model outperformed the ordinary logistic regression model with a substantially lower DIC (749.62 vs. 772.82). This may be attributed to the random effects logistic regression model’s accounting for unobserved heterogeneity among the four different intersections. For example, by comparing the intersection based random effects (uj ), from Fig. 7 it can be seen that intersection 2 has the lowest RLR likelihood among the four intersections while RLRs are more likely to be detected at intersection 4. 6. Conclusion and discussion In this study, four intersections in Shanghai were selected for the collection of field data, through both manual observing and video recording, on red-light running. Instead of only focusing on characteristics of RLR vehicles, this study compared RLR and non-RLR groups to identify through logistic regression models the key influencing factors from the aspects of driver characteristics, vehicle types, and driving conditions. Unobserved heterogeneity across different intersections was considered through additional random effects. Preliminary analysis utilized histograms and Chi-square testing to investigate RLR behavior: (i) left-turning traffic was seen to have a higher probability of RLR, which may be due to insufficient left-turn green time; (ii) the percentage of RLR for the leading vehicles in a platoon was about 12%, while more RLRs (about 88%) were caused by drivers in the middle of a platoon; (iii) about 57% of the violations occurred within the 0.4 s following red onset, which contradicts Retting and Williams’s study (1996) where 48% of RLR drivers entered the intersection 0.5–0.9 s after red onset; (iv) out-of-state vehicles were less likely to run red lights than in-state vehicles; (v) male drivers were more likely to run red lights, consistent with previous analyses; (vi) drivers without passengers are more likely to run red lights than drivers with one or more passengers; (vii) safety belt use and hand-held cell phone use showed no significant effects on RLR.

X. Wang et al. / Transportation Research Part F 37 (2016) 144–153

153

Modeling results concurred with the preliminary analysis: in-state drivers, male drivers, and passenger vehicles were more likely to run red lights. Modeling results also showed that an increase in traffic volume increased the likelihood of RLR. A comparison of these results with similar studies in the US shows consistency, which indicates that the influencing factors of RLR are similar, despite the different jurisdictions and driving behaviors. One exception is that previous US studies concluded that seat-belt use would affect RLR probability, but that variable is not significant in this study. This may be due to the low proportion of seat-belt use in China, which therefore could not be used to distinguish unsafe driving behavior. Given the unobserved heterogeneity among different observation locations, the random effects logistic regression model development was successful methodology. Compared to the ordinary logistic regression model, this model provided better model fitting; this study demonstrated that the random effects logistic regression model is capable of analyzing RLR behavior while considering the heterogeneity between intersections. Acknowledgements This study was jointly sponsored by the National Key Technology Support Program (2014BAG01B03) and Chinese National Science Foundation (51522810). References Bai, Z., Qi, X., & Zhao, Y. (2007). Video-based detection of vehicles running red lights violation research. Channel Science. Bonneson, J., Brewer, M., & Zimmerman, K. (2001). Review and evaluation of factors that affect the frequency of red-light-running. Report No. FHWA/TX-02/04027-1. Texas: Texas Transportation Institute. Brewer, M., Bonneson, J., & Zimmerman, K. (2002). Engineering countermeasures to red-light-running. In Proceeding of the ITE 2002 spring conference and exhibit (CD-ROM). Washington, DC: Institute of Transportation Engineers. Chang, M., Messer, C., & Santiago, J. (1985). Timing traffic signal change intervals based on driver behavior. Transportation Research Record, 1027, 20–30. Elmitiny, N., Yan, X., & Radwan, E. (2010). Classification analysis of driver’s stop/go decision and red-light running violation. Accident Analysis and Prevention, 42, 101–111. FHWA, 2009. Engineering countermeasures to reduce red-light running. Publication FHWA-SA-10-005.Washington, DC. Huang, H., Chin, H., & Haque, M. (2008). Severity of driver injury and vehicle damage in traffic crashes at intersections: A Bayesian hierarchical analysis. Accident Analysis and Prevention, 40, 45–54. Lunn, D., Thomas, A., Best, N., & Spiegelhalter, D. (2000). Winbugs-a Bayesian modelling framework: Concepts, structure, and extensibility. Statistics and Computing, 10, 325–337. Martinez, K., & Porter, B. (2006). Characterizing red light runners following implementation of a photo enforcement program. Accident Analysis and Prevention, 38, 862–870. Porter, B., & Berry, T. (2001). A nationwide survey of self-reported red light running: Measuring prevalence, predictors, and perceived consequences. Accident Analysis and Prevention, 33, 735–741. Porter, B., & England, K. (2000). Predicting red-light running behavior: A traffic safety study in three urban settings. Journal of Safety Research, 31, 1–8. Retting, R., Ferguson, S., & Farmer, C. (2008). Reducing red light running through longer yellow signal timing and red light camera enforcement: Results of a field investigation. Accident Analysis and Prevention, 40, 327–333. Retting, R., Ulmer, R., & Williams, A. (1999). Prevalence and characteristics of red light running crashes in the United States. Accident Analysis and Prevention, 31, 687–694. Retting, R., & Williams, A. (1996). Characteristics of red light violators: Results of a field investigation. Journal of Safety Research, 27, 9–15. Shin, K., & Washington, S. (2007). The impact of red light cameras on safety in Arizona. Accident Analysis and Prevention, 39(6), 1212–1221. Spiegelhalter, D., Thomas, A., Best, N., & Lunn, D. (2003). Winbugs user manual. Cambridge: MRC Biostatistics Unit. The Ministry of Public Security, People’s Republic of China, 2012. . Accessed on April 28, 2014. Wang, J. (2006). Red-light-running analysis at urban intersection. Central South Highway Engineering, 4, 123–130. Wang, L., Zhang, C., & Wang, X. (2011). Characteristic parameters of vehicle trajectories predicted for urban intersection through a red light. Highway and Transportation Research, 28, 108–112. Xian, H., & Han, H. (2010). The impact of green light countdown on intersection traffic safety. China Safety Science Journal, 20, 9–13. Yang, C., & Najm, W. (2007). Examining driver behavior using data gathered from red light photo enforcement cameras. Journal of Safety Research, 38(3), 311–321. Yu, R., & Abdel-Aty, M. (2013). Multi-level Bayesian analyses for single- and multi-vehicle freeway crashes. Accident Analysis and Prevention, 58, 97–105. Yu, R., Abdel-Aty, M., Ahmed, M., & Wang, X. (2014). Utilizing microscopic traffic and weather data to analyze real-time crash patterns in the context of active traffic management. IEEE Transactions on Intelligent Transportation System, 15, 205–213. Zhang, J., He, Y., & Sun, X. (2009). Urban intersection countdown show on the driving behavior. Traffic Information and Safety, 27, 2009.