Accident Prediction Models Title for Traffic Signals Presented by: Dr Shane Turner Beca Infrastructure Ltd, New Zealand Blair Turner ARRB Group, Victoria, Australia Professor Graham Wood Macquarie University, NSW, Australia Abstract A significant proportion of urban crashes occur at traffic signals. Many of the black-spots in both Australia and New Zealand cities occur at high volume and/or high speed traffic signals. Major intersections often have long cycle times and complex phasing arrangements, which can confuse and frustrate drivers and pedestrians. Right turning phases are often introduced at traffic signals to reduce right-turn-against crashes, but this in turn increases cycle lengths and overall delays. Some traffic signals have unusual layouts, such as highly staggered right turn bays, which may cause safety problems. Crash prediction models have been produced, in several separate research projects for traffic signals, examining the impact of traffic volumes, speed limit, signal phasing, number of pedestrians and cyclists and intersection layout on various crash types. These crash prediction models have enabled a better understanding of the impact of these factors on safety and also allow the safety impact to be quantified. This paper summarises the outcomes of each of the separate studies and key findings of the research. Introduction The majority of the higher volume intersections in our urban areas have signalised control. Traffic signals are often introduced at high volume priority controlled intersections and roundabouts to address capacity and safety problems. However many of the black-spots in our urban areas, particularly larger centres, are at traffic signals, because of the high exposure levels. A significant proportion of crashes occurred at traffic signals in NSW (22%), Victoria (17%) and Queensland (14%). This compares with 8% of crashes in New Zealand. Table 1 shows the number of injury crashes of each type in each State and within New Zealand, as well as the percent of fatal crashes at signals compared to all fatal crashes. Table 1: Average Crashes per year at Traffic Signals (over period 2001 to 2005) Country/State Fatal Serious Minor New South Wales* 35 (8%) 4390 Victoria 20 (6%) 800 2100 Queensland* 11 (3%) 679 1246 New Zealand 9 (2%) 117 659 *based on 2005 data only Considerable research has been undertaken internationally on the safety of traffic signals (Hall, 1986, Hauer et al., 1989, Taylor et al., 1996, Wang et al., 2003, Lyon et al., 2005 and Wong, 2007). The study by Hall (1986) was one of the first to utilise generalised linear modelling and a Poison error structure. The study draws on a sample size of 177 intersections throughout England, with around 30 in London. It considered eight different crash types, including right angle, single vehicle and pedestrian versus motorvehicle crashes. A number of variables were considered including traffic and pedestrian volumes, geometric features (approach width, number of lanes and sight distance), signal timings (inter-green time and cycle length), and the presence of antiskid surfacing. Hauer (1989) also used generalised linear models, but this time with a negative binomial error structure. Flow-only models were produced for the 15 major crash types for a sample of 145 intersections in Toronto. Wong et al. (2007) produced crash prediction models for the major crash types at 262 traffic signals in Hong Kong. This study also used generalised linear models with a negative binomial error structure. Key significant predictor variables included; traffic volume, degree of approach curvature, surround road environment, presence of tram stops, pedestrian volumes, proportion of commercial vehicles and average lane width. Wong’s paper also contains a literature review of recent research on crash prediction models at traffic signals. A number of crash prediction modelling studies have been undertaken of traffic signals in New Zealand, by Turner, and these are outlined in the following section. These studies have Page 1 The information provided in this publication is for information purposes only. Neither the author nor Beca assumes any duty of care to the reader or gives any representation or warranty in relation to the contents of this publication. You should not rely on this information as professional advice or as a substitute for professional advice. Specialist professional advice should always be sought for your particular circumstances. Accident Prediction Models for Traffic Signals become more detailed over time and the impact of a number of key variables, including, traffic flow, speed limit, number of pedestrian and cyclists, signal phasing and intersection layout, on the safety of traffic signals has been investigated. Of particular interest is the effect that speed limit and signal phasing have on safety. Several Australian states use traffic signals in high speed limit urban environments, including 70 and 80 km/h zones. It could be expected that given the higher speeds in such environments that the severity outcomes at such locations would be higher if a crash were to occur. Information from Victoria (again using information from 2001 to 2005) show that half of all fatal crashes at traffic signals occurred where the speed limits were 70 km/h or above (almost 10 fatal crashes per year), while over a quarter of serious injury crashes (264 per year) occur at such locations. The situation is similar in NSW, with almost 30% of fatal signal crashes in 2005 occurring at higher speed intersections (70 km/h or above). Although numbers are less in Queensland, 27% of fatal crashes (three of their 11 fatal signal crashes in 2005) occurred in 70 km/h areas. Traffic signals are not commonly used on high speed/rural roads in New Zealand. Where signals have been used the general experience has been that more severe crashes, be they fatal or serious crashes have been more prevalent. With a strong emphasis on reducing road trauma in New Zealand (and Australia) the use of traffic signals in high speed environments has to be questioned. But with traffic volumes growing and the increasing costs and effects (eg severance) of alternative grade separation options, particularly in urban areas, the use of traffic signals in such environments is often considered. The most desirable alternatives are standard or signalised roundabouts, but these are not always an option, due to construction constraints and their ability to handle high traffic volumes. They are also a major pinch point for cyclists (and to a lesser extent pedestrians), and large roundabouts have a poor safety record for cyclists. Crash models, like those contained in this paper, can provide useful future information for traffic engineers and decision makers that are attempting to trade-off costs of more expensive options and the safety consequences of installing traffic signals in high speed areas. One would expect that signal phasing, high degrees of saturation and cycle length would impact on the safety of signalised intersections. However, there is limited research available internationally on the effect of these factors and traffic signal technicians/engineers ideally require such information to address safety issues. In the current environment the selection of safety improvement options is often by trial and error. While further research is required in this area, this paper does report on research into the factors effecting right-turn against crashes at traffic signals, including the provision of right turn phasing, in combination with intersection layout factors. This paper traces the development of crash models in New Zealand for traffic signals and the key findings that are likely to be of interest to traffic/transport engineers. The paper concludes with a discussion on key areas of future research. Signalised Intersection Crash Prediction Models The first generalised linear models in New Zealand for crashes at traffic signals were developed by Turner (1995) in the early 1990s, based on crash data from the mid to late 1980s. The modelling methods used were based on generalised linear modelling methods developed by Maycock and Hall (1984) and Hauer et al. (1989). Specialised macros were developed in Minitab to produce these models. The use of macros has enabled the modelling methods to be easily modified over the last 15 years so that other features, such as goodness-of-fit statistics, confidence interval routines and covariate analysis methods could be added. This compares to proprietary software products which are fairly inflexible. These initial models were flow-only, where the independent variable was the conflicting flow movements and the dependant variable was crashes. Following on from his PhD work Turner (2000) produced a series of flow-only crash prediction models for the most common urban and rural mid-block and intersection types, including traffic signals. Crash models were developed for three traffic signal types; cross-road signals with all legs two-way, cross-road signals where one road was two-way and the other was one-way, and signalised T-junctions (with all legs two-way). A strict selection criteria was developed to produce sample intersections that were as consistent as possible in terms of layout variables. Several of these models were incorporated into Appendix 6 of the New Zealand Project Evaluation Manual (Transfund, 1997) (now the Economic Evaluation Manual), when it was updated in 2001. Building on this previous research, Turner and Roozenburg (2004) added non-flow variables to the ‘right-turn-against’ crash prediction models at 4-arm traffic signals. The objective of this research was to consider a number of non-flow variables, and determine whether these are also key predictors for accident occurrence. The outcome was a more refined crash prediction model for this accident type. Of particular interest was the impact of right-turn phasing on accident occurrence. The key non-flow variables examined were: intersection geometry and layout (eg number of through lanes, right turn bay offset and intersection depth); right-turn signal phasing (eg filtered turns); and forward visibility to opposing traffic. Page 2 Accident Prediction Models for Traffic Signals Prior to 2004 no accident prediction models had been developed in New Zealand for the non-motorised modes of travel (walking or cycling). Internationally only a small number of studies have considered these ‘active modes’. Turner et al. (2005) developed models for pedestrian and cyclists accidents involving motor-vehicles at intersections and mid-blocks. These models investigated a small number of non-flow variables in addition to motor vehicle, pedestrian and cycle traffic volumes. This paper provides a summary of all the models that have been developed by Turner and colleagues for traffic signals. All the models presented are the preferred model forms for the given predictor variables in each study. The fit of these models to the data have been tested using the scaled deviance goodness-offit test and pass this test. Details on the goodness-of-fit testing procedures and of the results of the testing for each model can be found in the research papers and technical reports specified in the reference list. Table 2 shows that the major difference resulting from the increased speed limit is an increase in rear-end crashes and a large reduction in pedestrian crashes. The reduction in pedestrian crashes is likely a result of many low speed limit intersections being in commercial shopping areas, whereas high speed signals are typical on the urban-rural fringe or on multilane roads with little pedestrian activity. Cycle crashes are included within the four major crash types. Of particularly interest is the high number of cyclists involved in right-turn-against (28%) and crossing (28%) crashes at low speed traffic signals (Turner et al, 2005). The major crash types at signalised T-junctions are shown in Table 3. Table 3 – Major Crash Types at 3-arm Traffic Signals (source: New Zealand Crash database, CAS) Crash Type Low Speed Traffic Signal 2001 to 2005 High Speed Traffic Signals 2001 to 2005 Major Crash Types Right Turn-Against e.g. 34% 30% The major crash types at 4–arm traffic signals in New Zealand are shown in Table 2 for both low speed (50 to 70 km/hr) and high speed (80 km/h plus on at least one of the cross roads) traffic signals. Crossing e.g. 17% 10% Rear End e.g. 21% 30% Loss-of-Control e.g. 11% 12% Pedestrian 16% 0% Other 1% 18% Urban Traffic Signals - Flow-only Models Table 2 – Major Crash Types at 4-arm Traffic Signals (source: New Zealand Crash database, CAS) Crash Type Low Speed Traffic Signal 2001 to 2005 Signalised High Speed Traffic Signals Right Turn-Against e.g. 30% 33% Crossing e.g. 29% 26% Rear End e.g. 10% 22% Table 3 also shows that rear-end accidents are more prevalent at high speed T-junctions, but the effect is not as significant as at cross-roads. Again the percentage of pedestrian crashes is a lot higher at low speed signals. Crash Model Types Loss-of-Control e.g. 6% 5% Pedestrians 17% 2% Other 8% 12% There are three main flow-only model types that have been used internationally. In Type 1 models (Lau and May 1988, Hughes, 1991 and Transfund NZ, 1997) the total number of crashes is related to the sum of the traffic volumes entering the intersection. Logic however suggests that crashes cannot occur when there is zero traffic volume on at least one of the roads, and so this model form is no longer in common use. Page 3 Accident Prediction Models for Traffic Signals In Type 2 models the total number of crashes is related to the product of the traffic volumes on the approach roads. In Type 3 models (Maycock and Hall, 1984, Hauer et al., 1989, Turner, 1995 and Maher and Summersgill, 1996) the conflicting flow crash types are related to the product of the conflicting traffic volumes. Type 3 models enable the prediction of each crash type and follow the logic that crashes involving two conflicting manoeuvres cannot occur when there is zero traffic volume for one or both conflicting movements. For these reasons most researchers prefer Type 3 models. Type 2 models are less popular than Type 3 models, but are seen as a practical alternative to the conflicting movement models when road safety engineers do not have access to conflicting volume data at intersections. The New Zealand Economic Evaluation Manual, EEM, (Land Transport NZ, 2006) contains both Type 2 and 3 models for this reason. The Type 2 models in the EEM do in many cases break the crashes down into the common crash types. A more detailed description of the model types is given in Turner and Nicholson (1998). The majority of the flow-only models presented in this paper, and developed by Turner since 1995 are conflicting movement models (Type 3 models), given their more direct link to causal factors compared with the other model types. Details on the latest modelling methods used by Turner are provided in Turner et al. (2006a). The four major all-vehicle crash types at cross-roads, and the corresponding models are shown in Table 4. Table 4: Signalised cross-road accident prediction models Accident Type NZ Accident Codes Equation (accidents per approach) Crossing (No Turns) HA A = 1.57 × 10 -4 ×q20.36 × q110.38 Right Turn Against LB A = 9.57 × 10 -5 ×q20.49 × q70.42 Rear-end FA to FE A = 1.66 × 10 -6 × Qe1.07 Loss-of – control C&D A = 2.96 × 10 -6 × Qe0.95 Others Rest of codes A = 1.26 ×10 - × Qe0.46 Crash Prediction Models for T-junctions The latest flow-only T-junctions models in New Zealand were also developed by Turner (2000), from a sample set of 30 intersections. The six conflicting flow movements used in the various models are shown in Figure 2. Figure 2: Conflicting and approach flow types (T-junction) Cross-road Crash Prediction Models The latest Type 3 crash prediction models in New Zealand for all major crashes types at cross-roads traffic signals were developed in the late 1990s (Turner, 2000), based on a sample set of 109 intersections (or 436 approaches). Figure 1 shows the conflicting movements that were used in the models for each approach ( Qe is the total traffic volume on each approach). For each of the four approaches the three turning movements on the approach of interest are q1 to q3. So there are for example four combinations of q2 and q11 at each intersection. Figure 1: Conflicting and approach flow types (Cross-roads) The four major all-vehicle crash types and models at T-junctions are shown in Table 5. Table 5: Signalised T-junction accident prediction models Accident Type NZ Accident Equation (accidents per approach) Codes Right Turn Against LB A = 1.08 × 10 -1 × q5-0.38 × q30.56 Rear-end FA to FD A = 7.66 × 10 -8 × Qe1.45 Crossing (Vehicle Turning) JA A = 2.67 × 10 -2 × q5-0.30 × q10.49 Loss-of – control C&D A = 1.91 × 10 -3 × Qe0.17 Others All other codes A = 1.69 × 10 -2 × Qe0.15 Page 4 Accident Prediction Models for Traffic Signals The large negative parameters in the ‘LB’ and ‘JA’ models indicate that intersections with higher flows have fewer crashes. This is an unexpected result. It is speculated that at high traffic flows the installation of right turn bays and exclusive right turn phases reduces crash occurrence. Further research is required to investigate this issue. Ideally the sample size should be increased from 30 up towards 100 or more intersections. major impact on crash occurrence at T-junction traffic signals. It is possible that signal phasing and layout are more important factors. Pedestrian & Cycle Crashes Turner et al. (2005) developed crash prediction models for pedestrians and cyclists. Prior to this research, crash prediction models in New Zealand had only been developed using motor vehicle volumes and for total crashes of each type. ie without consideration of the road user types involved in a crash. Pedestrian and cycle crashes are fairly common at most low speed urban traffic signals. It might be expected that the likelihood of such crashes would depend on the volumes of these road users travelling through the intersection, and so these flows are likely to be important for pedestrian and cycle crash prediction models. The cycle and pedestrian flows in the sample set vary from very low (close to zero) up to relatively high volumes. For example, the cycle volumes at traffic signals, per approach, vary from 20 to around 350 cyclists per day, with the majority of flows in the 100 to 250 flow range. Product of Link Models at Cross Roads and T-junctions Type 2 models were also produced for signalised cross-roads and T-junctions by Turner (2000). These models (Equation 1 and 2) are useful when conflicting flow data is not available at an intersection, as may be the case when only automatic tube counts are available for the approaches to an intersection. In both cases the model is used to calculate the total number of crashes per year. Equation 1 AT = 3.69 × 10 -3 × Qminor0.14 ×Qmajor0.46 Equation 2 Two model types were developed for cycle versus motor-vehicle crashes at signalised cross-roads. The first, a ‘same direction’ model predicts crashes on a single approach between cyclists either colliding with a stationary vehicle or moving motor-vehicle, travelling in the same direction. The second model is for rightturn-against crashes where a cyclist is travelling straight through the intersection and collides with a motor vehicle turning right. Table 6 and 7 present these two models and the proportion of cycle crashes that they represent at signalised cross-roads. Cycle movements are coded in a similar manner to motor-vehicle movements. Entering flows, for example Ce, are the sum of all cycle entering flows, for example c1 + c2 + c3. Figure 3 shows the movements graphically. AT = 1.73 × 10 -1 × Qstem0.12 × Qmajor0.04 Where, Qmajor is the highest of the two-way link volumes for crossroads or the main road flow for T-junctions Qminor is the lowest of the two-way link volumes for crossroads, and Qstem is the stem flow for T-junctions. It is interesting to note the relatively high coefficient at the front of Equation 2 and the low values of the two traffic flow exponents. This indicates that traffic flow does not have a Table 6: Signalised cross-road cycle crash prediction equations Crash Type Crash Codes Equation (crashes per approach) Proportion of Cycle Crashes Same Direction A, E, F, G A = b0 × Qe × Ce 35% Right Turn Against - Motor vehicle turning LB A = b0 × q7b1 ×c2b2 b1 b2 21% Table 7: Signalised cross-roads – cycle crash prediction model parameters Crash Type b0 b1 b2 Error Structure Same Direction 7.49 × 10 -4 0.29 0.09 Poisson Right Turn Against - Motor vehicle turning 4.41 × 10 0.34 0.20 Negative Binomial -4 Page 5 Accident Prediction Models for Traffic Signals Figure 3: Cycle model variables Figure 4: Pedestrian model variables The small value of the exponent of cycle flows (b2) in Table 7 indicates a ‘safety in numbers’ effect where the crash rate per cyclist decreases substantially as the number of cyclists increases. Two models were developed for predicting pedestrian crashes at signalised cross-roads. The majority of all crashes involving pedestrians, not just those at signalised cross-roads, occur where vehicles are travelling straight along the road and the pedestrian is crossing. These crashes represent 50% of pedestrian crashes at signalised cross-roads. The second major type of pedestrian crashes at signalised cross-roads is where right turning vehicles collide with pedestrians crossing the side road. Table 8 and 9 present these two models. Pedestrian movements differ from motor-vehicle and cycle movements. For cross-roads, four movements are used, one for pedestrians crossing each approach. Figure 4 shows the movements used graphically. Table 8: Signalised cross-road pedestrian crash prediction equations Crash Type Crash Codes Equation (accidents per approach) Proportion of Ped. Crashes Crossing – vehicle intersecting NA, NB A = b0 × Qb1 × Pb2 50% Crossing – vehicle turning right ND, NF A = b0 × q4b1 × p1b2 36% Table 9: Signalised cross-roads – pedestrian crash prediction model parameters Crash Type b0 b1 b2 Error Structure Crossing – vehicle intersecting 7.28 × 10 -6 0.63 0.40 Negative Binomial Crossing – vehicle turning right 5.43 × 10 -5 0.43 0.51 Negative Binomial Unlike the cycle models, the ‘safety in numbers effect’ is not as pronounced, with exponents of flow being similar to those observed for motor vehicle flows. High Speed Traffic Signals Crash prediction models were developed from high speed traffic signals by Turner et al (2006b) using signalised intersections in Melbourne, Victoria and throughout New Zealand. There is a lot of variety in the layout and approach speed limits at high speed traffic signals. In this study only sites with at least two legs with a speed limit of 80 km/h or higher were selected. Sideroads typically had a speed limit of 50 km/h in New Zealand and 60 km/h in Melbourne. In some cases both intersecting roads had high speed limits. Both 3 and 4-arm intersections were included in the sample set. Table 10 shows the number of sites in New Zealand and Melbourne that met the selection criteria and where suitable data was available. As the number of signalised seagull intersections was small these were not analysed further in this study. Table 10 Numbers of High Speed Signalised Intersections in Study New Zealand Melbourne Total Crossroads 4 24 28 T-Junctions 8 7 15 Seagull 3 0 3 The typical mean-annual numbers of reported injury crashes for high-speed signalised T-junctions can be calculated using the crash prediction models in Table 11. Table 11 includes a column for the variable ϕVIC, which is a multiplication factor for sites in Melbourne, Victoria (Vic). This variable indicates that for most crash types the number of crashes at similar high-speed traffic signals in Victoria would be higher than in New Zealand. It is unclear why the crash risk in Victoria is higher than New Zealand. It could be due to differences in reporting rates (the models provided are for reported injury crashes) or due to different speed limits in the two countries, different road environments (most of the high speed traffic signals in Victoria are in the middle of the urban areas, whereas most New Zealand sites are on the urban-rural boundary). It may be due to different road layout standards or different driver behaviour. A larger sample set and more detail on the various contributing factors, eg speed limits would be needed to understand the differences. Page 6 Accident Prediction Models for Traffic Signals Table 11: High-speed Signalised T-junction crash prediction models Crash Type Equation (crashes per approach) Right-turn-against ARATT1 = 8.29 × 10 ×q3 -2 0.26 ×q5 0.29 Major -0.15 × ϕVIC Error Structure Sig. Model 2.85 Negative Binomial Yes 0.89 Poisson Yes Rear-end (Major Road) ARATT2 = 2.28 × 10 -2 × Q Crossing (Vehicle turning) ARATT3 = 3.18 × 10 -2 ×q10.12 × ϕVIC 1.67 Poisson Yes Loss-of-control (Major Road) ARATT4 = 5.77 × 10 × Q × ϕVIC 1.06 Poisson Yes Other (Major Road) ARATT5 = 1.82 × 10 -3 × Q × ϕVIC 2.81 Poisson No Other (Minor Road) ARATT6 = 1.40 × 10 -3 × Q × ϕVIC 5.04 Negative Binomial Yes -3 × ϕVIC ϕVIC 0.32 Major 0.37 Major 0.41 Minor The modelling shows that the sign on the straight through volume (q5) for the right-turn-against crash type is negative. This indicates that an intersection is safer with respect to this crash type when the through movement is high. It is unclear why this should be so. It is possible that it may be associated with the phasing of the signals, as these crashes occur when vehicles are ‘running the lights’, as the right turn phase is fully controlled (ie no filtering allowed). A very similar result was observed for this crash type at lower-speed urban signalised T-junctions. The modelling process was repeated for crashes at high-speed signalised crossroad intersections. Table 12 presents the models and multiplication factors for the signals located in Victoria, Australia. Again, the variables for intersections located in Victoria indicate that the numbers of crashes for similar traffic volumes are higher than in New Zealand. However, these models may not be particularly applicable to New Zealand, as unlike for signalised T-junctions, only a small number of New Zealand signalised crossroads were included in the sample set. Table 12: High-speed signalised crossroad crash prediction models ϕVIC Error Structure Sig. Model 3.95 Poisson Yes ARAXT2 = 2.17 × 10 -2 × q20.20 × ϕVIC 1.43 Negative Binomial Yes Rear-end ARAXT3 = 4.16 × 10 -7 × Qe1.18 × ϕVIC 9.91 Negative Binomial Yes Loss of Control ARAXT4 = 5.75 × 10 × Qe × ϕVIC 1.15 Negative Binomial Yes Others ARAXT5 = 1.04 × 10 -2 × Qe0.14 × ϕVIC 4.44 Negative Binomial Yes Crash Type Equation (crashes per approach) Crossing ARAXT1 = 6.82 × 10 × q2 Right-turn-against * -5 -5 0.31 0.70 ×q 0.35 11 × ϕVIC * the right running volume q7, was not found to be a key variable for this model. A model was produced that included q7, but the exponent on this variable was close to zero. This may be the result of a relatively small sample size. Page 7 Accident Prediction Models for Traffic Signals Non-Flow Variables – Signal Phasing & Layout The study (Turner 2004) on the affects of signal phasing has to date focused on the provision of right turning signal phasing at traffic signals as there is much debate within New Zealand on the safety merit of installing such phasing. Extra phasing adds to the cycle length and lost time at most traffic signals, which in turn can increase delays for vehicles travelling straight through the intersection, which is often the predominant traffic movement. An increase in the cycle time also leads to more driver frustration, as drivers on average have to wait longer to proceed through the intersection. This in turn may lead to an increased safety risk as drivers may be more likely to run a red or amber signal when cycle times are longer. Given this uncertainty around the safety impacts of right turn phasing on ‘right-turn against’ crashes this crash type was investigated. The two movements of specific interest for this project were the through movement for each approach (of which there are four for most signalised cross-roads) (q2) and the opposing right turn (q7) as shown in Figure 5. Figure 5 – Conflicting flow movements for right-turn-against crashes Data Collection Data for this project was provided by three city councils in New Zealand; Christchurch, Hamilton and Palmerston North. Each of the councils provided traffic volumes, signal phasing and intersection geometry and layout information. A total of 455 approaches were selected for inclusion in the sample set for analysis and the development of crash prediction models. Table 13 shows the number of intersection approaches from each city in the sample set. Table 13 - Regional Distribution of Traffic Signal Approaches City Number of Approaches Christchurch 309 Palmerston North 71 Hamilton 75 Predictor Variables The data collected from each council is detailed below. a) Traffic volumes Traffic counts provided by councils were generally of one or two hours in duration and collected over two or three time periods, typically the morning and afternoon/evening peaks and an interpeak period. All traffic counts were disaggregated by approach and by movement ie left, through and right for each approach. Hourly factors were applied to the raw traffic counts to calculate an AADT for each straight and turning movement. The procedure to determine the AADT involved summing the raw traffic count of each movement and dividing that by the sum of the appropriate hourly factors (see Turner, 1995 for details). b) Signal phasing Signal phasing for right turn movements was separated into three categories; filtered, partially controlled and full controlled signal phasing. For many intersections there are two or more phase types for the four combinations of through and opposing right turn movements. ‘Filter Control’ (Phase 1) is no green signal phase (arrow) for traffic turning right. A red phase (arrow) may appear at certain times throughout the signal cycle for various reasons eg large opposing through movements, crossing pedestrians. ‘Partial Control’ (Phase 2) is a green phase (arrow) being displayed for right turners for part of the green phase allocated to each approach. Right turning traffic may filter turn at other times throughout the green phase, on the full green aspect. ‘Partial Control’ is associated with a lead or lag right turn phase (arrow). No differentiation is made between lead and lag right turn movements, as this information is not readily available for a number of intersections. ‘Partial Control’ also includes approaches where a green phase (arrow) for the right turn movement may only be displayed at certain times of the day or on demand. Generally such phases are applied when the right turning traffic volume is high. ‘Full Control’ (Phase 3) is a green phase (arrow) or red phase (arrow) being displayed for the right turning movement at all times of the day. Right turning traffic may not filter turn at other times throughout the signal cycle. A breakdown of approaches by signal phase type and number of opposing lanes is displayed in Table 14. Page 8 Accident Prediction Models for Traffic Signals Table 14: Phase Type and Number Opposing Lanes Non-Flow Variable Number of Approaches Phase 1 330 Phase 2 97 Phase 3 28 One Opposing Lane 292 Two or more Opposing Lanes 163 There are insufficient Phase 3 approaches in the sample set to produce results specifically for this type of signal phasing. The Phase 3 approaches were therefore combined with the Phase 2 approaches to assess the overall effect of right turn phases on accident occurrence. c) Intersection geometry and layout Intersection geometry and layout information was obtained from as-built drawings provided by each council. In Christchurch City the as-built drawings record the date the intersection was last upgraded, and hence those that were significantly upgraded in the period 1998 to 2002 could be removed from the sample. For the other two councils the date of upgrades was less clear, so each council was contacted to identify sites that were upgraded during this period. The following information was collected at each intersection by approach. Lane Information - The type (permitted movement), number and width of each lane. Of particular interest was the number of straight through lanes opposing each right turn. As the more lanes generally the greater the distance right turners have to travel to get to the safety of the side-road. Right Turn Lane Offset - The lateral distance between a prolongation of the right edge of the right turn lane and a prolongation of the left edge of the opposing right turn lane. d) Visibility The visibility of right turning drivers is often restricted by vehicles turning right from the opposing direction. In this study the visibility was measured from the limit line of the right turn lane to the centre of the rightmost opposing through lane (ie furthest from the kerb-line), as illustrated in Figure 6. The driver’s position in the right turn lane was assumed to be 1/3 of the width of the right turn lane from the right hand edge of that lane. In cases where the rightmost through lane was a shared through and right turn lane it was assumed that this lane is occupied by a right turning vehicle and through vehicles pass on a lane to the left, if that lane can be used for through movements. In the majority of cases the right turn lane was exclusive and hence it was not possible, due to the small sample size, to assess the safety of the shared lanes. Approaches with no opposing right turn lane (shared or exclusive) and those without a through lane when a vehicle is turning from a shared right and through lane, were removed from the sample set. Three methods were used to calculate visibility for right turning vehicles were: 1. Mathematical – where the visibility was calculated using trigonometry. This method could be used when opposing approaches are straight and close to parallel in alignment. A manual check from plans of 10% of all intersections was completed for quality control purposes. 2. Manual – where the visibility was measured directly from aerial photographs or other intersection geometry and layout plans. This method was used when opposing legs had a curved alignment or were not parallel. 3. Field – where the visibility was measured in the field. This method was used where the visibility was restricted by vertical geometry. Intersection Depth - The distance between limit lines measured parallel with the road alignment. Figure 6: Illustration of visibility measurement Page 9 Accident Prediction Models for Traffic Signals Accident data b3 is a model parameter. It is applied as the power of C Accident data for this project was sourced from the Land Transport NZ’s CAS database, for the five-year period 1998 to 2002. All right turn against (Type LB) injury accidents in this period were assigned to the leg of the intersection on which the right turning vehicle originated. φ is a multiplicative model parameter for a discrete nonflow variable, which can take on one of two values. Crash Model Equations In previous research into crash prediction models in New Zealand the models for right-turn-against accidents only included the two conflicting traffic volumes, straight through and right turn. The form of this conflicting flow only model is: Equation 3 AT = b0 × q2b1 × q7b2 Where: b o is a model constant q2 is a the daily through traffic flow opposing right turning traffic b1 is a model parameter. It is applied as the power of q2 q7 is a the daily right turning traffic flow b2 is a model parameter. It is applied as the power of q7 The flow-only model form was modified to allow non-flow variables to be added. The form of the extended crash prediction model is as follows: In Equation 4 the φ enters in the format “exp (b x V)”, where V is the coded value (always chosen as “+1” and “-1”). The discrete model parameter then has the un-logged model value of “exp (b4 x code value)”. Three new variables were found to be important features; visibility, number of opposing lanes and signal phasing type. Other variables, such as opposing right turning flow and intersection depth were also investigated but were not significant. Visibility and intersection depth are continuous variables, and can be added as ‘c’ variables in Equation 4. Variables such as the number of opposing lanes or signal phasing are discrete variables and are added as multiplicative parameters (φ) in Equation 4. With the current model forms the continuous variable only applies over a particular range. An example of a continuous variable is the visibility in each direction at an intersection. This value can be a large number, but is generally not measured once it exceeds a particular threshold, around 400 to 500 m, as visibility above this level is not expected to improve safety. If a value of visibility of say 1 km is entered into the equation then the crash prediction will be very low, and not realistic. In this case the visibility measurement is outside the range over which visibility measurements are normally taken. In the event that actual visibility data was recorded and long visibility lengths were collected then a better model form would involve introducing this continuous variable as ‘exp (b3 × C1)’, rather than ‘c1b3‘. This study did not consider the correlation between the various variables. Correlation between variables can distort model parameters. Equation 4 AT = b0 × q2b1 × q7b2 × c1b3 ×φ Where: Crash Prediction Models a) Flow-only models AT is the predicted mean number of crashes over a given period, b0 is a model constant, q2 is a the daily through traffic flow opposing right turning traffic, b1 is a model parameter. It is applied as the power of q2, q7 is a the daily right turning traffic flow, b2 is a model parameter. It is applied as the power of q7, c is a continuous non-flow or non-conflicting flow variable – such as visibility (V) or intersection depth (I), both measured in metres To allow comparison with previous studies a flow-only model was produce containing only the two flows (daily through and right turn flows). The model has the following multiplicative form: AT = 8.09 × 10 -6 × q20.733 × q70.431 (crashes/year) Page 10 Accident Prediction Models for Traffic Signals Figure 7: Predicted mean number of accidents for different flow combinations Figure 7 shows how the predicted number of crashes is more dependent on the volume of conflicting through traffic ( q2 ) than the volume of right turning vehicles ( q7 ). This differs from the results in Table 4, which are models that were produced some years earlier. In Table 4 the exponent on both the through traffic (q2 ) and the right turning vehicles (q7 ) are similar, and close to that for q7 in the above model, while the constant value is a magnitude of 10 higher. It is unclear why this change has occurred, but may be a result of other factors, such as signal phasing options. This matter needs to be explored further. b) Visibility Models Initially the flow-only models were extended to include one additional non-flow variable. The first such variable was visibility. The model form is: AT = 1.29 × 10 -5 × q20.729 × q7439 × V -0.111 The resulting model was as follows. AT = 5.97 × 10 -5 × q20.664 × q70.446 × (V − VR + 100)-0.329 where V is the visibility in metres VR is the recommended visibility from AUSTROADS (in metres). The resulting model has a smaller log-likelihood than that of including visibility only (Table 15). Again it indicates that the more visibility the driver has the lower the accident rate. c) Opposing through lane models where V make this possible a value of 100 metres was added to the visibility. The value of 100 ensures that all the visibility values are positive. Higher values, such as 1,000 or 10,000, would result in the term being less sensitive to changes in visibility, which is not desirable. is the visibility in metres The negative exponent of visibility indicates that the number of accidents would decrease as the visibility increases. Although this model does predict a change in the number of accidents with varying visibility, an initial data analysis indicated that there was a stronger relationship between the number of crashes and the difference between the observed and recommended visibility (as specified in the Austroads guides to Traffic Engineering). To investigate this further a model was developed using the difference in visibility from the recommended visibility in Austroads. However because of the way generalised linear models are fitted, all the values have to be positive. To The second non-flow variable to be included in the crash prediction model is that of number of opposing through lanes. For this analysis a discrete variable was used for one opposing lane or more than one opposing lane. The resulting model is as follows: AT = 1.47 × 10 -4 × q20.435 × q70.389 × φL where φ L = 0.714 for a single opposing through lane = 1.401 for multiple opposing through lanes. Page 11 Accident Prediction Models for Traffic Signals The above factors for φL indicate that where the opposing through and right turning traffic volumes are held constant, then the predicted crash rate for an intersection with two or more opposing through lanes would be nearly twice that for the same flows as a single opposing lane. d) Signal phasing models Another non-flow variable that was investigated was that of the signal phasing with the model depending on whether the right turn was a filter turn or a fully or partially protected turn (green arrow). The resulting model is: AT = 6.68 × 10 -6 × q20.735 × q70.457 × φP where φP = 1.048 for a fully filtered turn = 0.954 for fully or partially protected turns. The two above φP factors indicate that the predicted number of crashes is higher, but only just for approaches with a fully filtered turn, than for approaches where there is at least a partially protected signal phase. The model indicates that a site with a fully filtered turn on average would have an accident rate 10% higher than one with a partially or fully protected turn. The closer the log-likelihood is to zero the more important the predictor variable. A large drop in the log-likelihood indicates that the variable explains a significant amount of the variability in the accident data. Table 15 indicates that the number of opposing through lanes has the greatest effect in reducing the loglikelihood and should be added to the model. The variable that appears to have the next best drop in log-likelihood is the visibility difference from recommended. Combining these two non-flow variables into a model produces the following model: AT = 3.94 × 10 -4 × q20.411 × q70.399 × (V − VR + 100)-0.190 × φL where V is the visibility in metres VR is the recommended visibility from AUSTROADS (in metres) φ L = 0.726 for a single opposing through lane = 1.378 for multiple opposing through lanes. f) Separate signal phasing models The model for approaches with filtered turns only is as follows: AT = 2.01 × 10 -5 × q20.516 × q70.586 × φL where e) Multiple non-flow variables The value of the log-likelihood can be use to assess how important each of the non-flow variables are to the model fit. Table 15 shows the respective log-likelihood for each of the models, when only one additional variable is added to the flowonly model. Table 15: Log-likelihood comparison – Models developed using entire dataset Model Log-likelihood Conflicting flow only model -384.75 Including: Visibility -384.58 Difference in visibility from recommended -383.61 Number of opposing through lanes -378.87 Signal phasing -384.63 Opposing right turning flow -383.86 Intersection depth -383.84 φ L = 0.718 for a single opposing through lane = 1.392 for multiple opposing through lanes. The resulting model indicates that crossing one lane of through traffic, without right turn signal phasing, is safer than crossing multiple lanes. The model where approaches are either fully or partially protected is as follows: AT = 2.88 × 10 -3 × q20.223 × q70.221 × φL where; φ L = 0.659 for a single opposing through lane = 1.518 for multiple opposing through lanes. The number of opposing lanes is more important for approaches with fully or partially protected right turn phases than for filtered turns. Fully or partially protected signal phases do not appear to be effective when there are multiple opposing lanes. Page 12 Accident Prediction Models for Traffic Signals Future Research Crash prediction modelling in New Zealand is getting to a stage at which the effects of a number of non-flow variables, such as visibility, signal phasing and intersection layout on crash rates can be quantified. Traffic engineers and planners can use such research to justify or reject safety improvement projects at traffic signals. There are a number of areas where further research would be beneficial to traffic engineers and planners. Key areas for future research include: 1. An analysis of signal phasing programs, lost time and cycle times at traffic signals on major crash types, including red light running and rear-end crashes, along with the right-turnagainst crashes. There is a lot of support for such research from traffic signal technicians and engineers. 2. The impact of green time allocation and level-of-service (or level of delay) for various approaches and movements on major crash types. 3. The impact of operating speed and speed limit, both on the main road and side-roads, on crash occurrence. The effectiveness of treatments, such as high friction surfaces on intersection approaches and extensions to all-red periods, in response to vehicles caught in the amber dilemma zone on intersection safety at high speed locations. 4. An analysis of gap acceptance and crash data to determine the traffic volumes levels at which priority controlled intersections and roundabouts need to be upgraded to traffic signals. 5. Research on model transferability methods. How can models developed in one country or state be transferred to another, eg between New Zealand and Victoria. Refer to Turner et al. 2007 for a discussion on this topic. Conclusions A number of crash prediction models have been developed for traffic signals over the last 15 years in New Zealand. Such models have a number of applications and allow traffic engineers, along with the results of ‘before and after’ studies to quantify and/or justify safety improvement works at traffic signals. A selection of the models produced so far are summarised in this paper. References are provided to the various reports and papers where a large number of other models are presented. Appendix A6 of the New Zealand Economic Evaluation Manual also has a number of models for traffic signals and other intersection and link types. The key findings to date from research on traffic signals are as follows: 1. Overall traffic volumes were not found to be strong predictor variables at signalised T-junction. The research indicates for some crash types that intersections with higher traffic volumes were actually safer than intersections with lower volumes. It is likely that other factors, such as the provision of right turn phasing and right turn lanes/bays are more important factors at such intersection as they reduced the crash risk at higher volume intersections. 2. The cycle-versus-vehicle crash models show a strong safetyin-number effect for cyclists at traffic signals. This finding indicates that strategies to concentrate cyclists on particular road corridors is safe than a more disperse approach to providing for cyclists. Ideally cyclists are concentrated on routes with low traffic volumes and speeds or where interaction with motorists is minimised. 3. The high speed crash models indicate that Victorian traffic signals are less safe than New Zealand traffic signals. This may be due to many of the Victorian intersections being in built up urban areas and typically higher speed limits on arterial and collector road in Australia. 4. As the number of opposing through lanes increases the number of right turn against crashes increases. Partially or fully controlled right turning phases are effective in reducing right turn against crashes. Right turn signal phasing is more effective for single opposing lanes than two or more opposing lanes. This may be due to the safety disbenefit that results from increased cycle time at larger multi-lane intersections. 5. That the inter-visibility between right turners and opposing through traffic, when a vehicle is in the opposing right turn lane, is an important factor in right turn against crash occurrence. The crash prediction models presented in this paper (and the more detailed technical reports that are available on each study) can be used to compare the safety performance of traffic signals in different speed environments, with different mixes of road users (motor vehicles, cyclists and pedestrians), for different crash types and for different signal phasing and layouts. There are a number of areas were further research would be beneficial to traffic engineers, as outlined above. The findings to date indicate the usefulness of such research going forward. Page 13 Accident Prediction Models for Traffic Signals References Hall, R. D 1986, “Accidents at 4-arm single carriageway urban traffic signals”, Transport and Road Research Laboratory Contractor Report CR65, UK. Hauer, E, Ng J.C & Lovell J 1989, “Estimation of safety of signalised intersections”, Transportation Research Record 1185, p. 48-61, USA. Hughes, B. P 1991, “Accident Predictions at Traffic Signals”, Main Roads Department, Perth, Australia. Land Transport NZ 2006, “New Zealand Economic Evaluation Manual”, Land Transport NZ, NZ. Lau. M.Y.K and May A.D 1988, “Injury accident prediction models for signalised intersections” Transportation Research Record 1172, pp 58-67. Lyon, C, Haq, A, Persaud, B and Kadama, S.T 2005, “Safety performance functions for signalised intersections in large urban areas – development and application to evaluate left-turn priority treatment, Transportation Research Record, 1908, p 165 to 171, USA . Turner, S.A, Wood, G.R & Roozenburg, A 2006a, “Accident Prediction Models for Roundabouts”, 22nd ARRB Conference, Canberra, Australia. Turner, S.A, Wood, G.R and Roozenburg, A.P 2006b, “Accident Prediction Models for High Speed Intersections (Both Rural and Urban)”, 22nd ARRB Conference, Canberra, Australia. Turner, S.A, Persaud, B, Lyon, C, Roozenburg, A and Chou, M 2007, “International Crash Experience Comparisons Using Prediction Models”, 86Th Annual Transportation Research Board Meeting, Washington DC, United States of America. Wang, Y, Ieda, H and Mannering, F 2003, “Estimating rear-end accident probabilities at signalised intersections: occurrencemechanism approach”, Journal of Transportation Engineering, ASCE, 129,4 p 377-384, USA . Wong, S.C, Sze, N.N and Li, Y.C 2007, “Contributing Factors to Traffic Crashes at Signalized Intersections in Hong Kong”, 86th Annual Transportation Research Board Meeting, Washington DC, United States of America. Maher, M.J & Summersgill. I 1996, “A comprehensive methodology for the fitting of predictive accident models”, AAP 28 Vol 3, pp281-296. Maycock, G & Hall, R.D 1984, “Accidents at four-arm roundabouts”, Transport and Road Research Laboratory Report LR1120, UK. Taylor, M.C, Hall, R.D & Chatterjeek, K 1996, “Accidents at 3-arm traffic signals on urban single-carriageway roads”, TRL Report 135, Transport Research Laboratory, Wokingham RG 40 3GA, UK. Transfund NZ 1997, “New Zealand Project Evaluation Manual (has been superseded)” Land Transport NZ, (Appendix A6 updated in 2001). Turner, S.A 1995, “Estimating Accidents in a Road Network”, PhD Thesis, Dept of Civil Engineering, University of Canterbury. Turner, S.A & Nicholson, A.J 1998, “Intersection Accident Estimation: The role of intersection location and non-collision flows”, AAP Vol 30 No 4. pp 505-517. Turner, S.A 2000. “Accident Prediction Models”. Transfund New Zealand Research Report No. 192, Transfund NZ, Wellington, New Zealand. Turner, S.A & Roozenburg, A.P 2004, “Accident Prediction Models at Signalised Intersections: Right-Turn-Against Accidents, Influence of Non-Flow Variables, Unpublished Road Safety Trust Research Report. Turner, S.A, Roozenburg, A.P & Francis, T 2005, “Predicting Accident Rates for Cyclists and Pedestrians”, Land Transport NZ Research Report 289, Wellington, NZ. Page 14 Accident Prediction Models for Traffic Signals Shane Turner Shane is a Technical Director in the consulting firm Beca Infrastructure Ltd. He is based in the Christchurch office where he leads a team of ten transport professionals. He is also the national transport research manager for Beca. Shane completed his BE (Hons) from the University of Auckland in 1990 and his PhD, specialising in the development of accident prediction models, at the University of Canterbury in 1995. He was appointed an Adjunct Senior Fellow at the University of Canterbury in 2006, where he teaches in the Master of Transport course. Shane is also a member of the TRB Committee on Safety Data, Analysis and Evaluation (2008 to 2011). Shane has extensive experience, through leading many road safety research projects, in safety related research, particularly the development of crash prediction models. He is the author of over 20 road safety publications. Graham Wood Graham’s current role is Professor of Statistics in the Department of Statistics, Macquarie University, Sydney. Graham has worked at the University of Canterbury, Central Queensland University and Massey University, prior to moving to Sydney. He is the author of 80 internationally refereed papers in mathematics and statistics, about 40 other miscellaneous publications and has co-authored two books. Graham has been involved with the fitting and development of accident prediction models in New Zealand for the past fifteen years, leading recently to the publication of three papers in the area. Graham’s current research interests are in mathematical and statistical modeling, particularly in optimisation, traffic accident prediction modeling and bioinformatics. Blair Turner Blair joined ARRB Group Ltd at the end of 2004, and has a number of years experience in road safety, both in Australasia and Europe. He initially worked for the New Zealand Government (LTSA) before moving to the UK to continue his career. He has been involved in a wide range of road safety research projects, road safety audits and investigation of crash locations, and production of road safety reviews (including a review of the UK Road Safety Strategy). Before moving to Australia, Blair also spent time at the UK Home Office where he further developed his research skills. He is currently a Senior Research Scientist at ARRB Group, and responsible for research on road safety engineering risk assessment. Page 15