using micro-simulated traffic conflicts as a …...transit signal priority schemes implemented in...
TRANSCRIPT
Using Micro-simulated Traffic Conflicts as a
Surrogate Safety Assessment Technique for
Evaluating Safety Performance of Transit Design
Alternatives at Signalized Intersections
by
Lu Li
A thesis submitted in conformity with the requirements
for the degree of Master of Applied Science
Graduate Department of Civil Engineering
University of Toronto
© Copyright by Lu Li 2015
ii
Using Micro-simulated Traffic Conflicts as a Surrogate Safety Assessment
Technique for Evaluating Safety Performance of Transit Design
Alternatives at Signalized Intersections
Lu Li
Master of Applied Science
Graduate Department of Civil Engineering
University of Toronto
2015
Abstract
This study focuses on crash prediction modelling at intersection-level using micro-
simulation to produce an effective surrogate safety assessment measure. The developed
crash prediction model followed generalized linear model with negative binomial error
structure to correlate the simulated traffic conflicts with the observed crash frequency in
Toronto, Ontario, Canada. Individual crash prediction models were developed for every
impact types and for transit-involved crash type. The resulting statistical performance
suggested adequate predictive ability. Based on the established correlation between the
simulated conflicts and observed crashes, scenarios were developed to investigate the
safety impacts of transit infrastructures by making hypothetical transit infrastructure
modifications in the micro-simulation networks. The findings implied that the existing
transit signal priority schemes implemented in Toronto had negative contributions on
iii
safety performance and that the existing near-sided stop positioning and streetcar transit
type were safer at their existing states than if they were replaced by their respective
counterparts.
iv
Acknowledgements
I would like to express my sincere appreciation to my supervisors Dr. Amer Shalaby and Dr.
Bhagwant Persaud. This thesis could not have been completed if not for their encouraging
academic guidance and supports. I would like to thank the City of Toronto for sharing the traffic
volume and signal timing data, the Toronto Transit Commission for sharing the transit service
data, and Taha Saleem for maintaining and sharing the crash data. I am grateful for the technical
and professional helps from Asmus Georgi and Mohamed Mahmoud.
I am highly indebted to my parents for their loving supports in my pursuing for this Master’s
degree. This academic endeavour would not have been motivated if not for their continuous and
unconditional care. Lastly, I am truly grateful to my friends for their inspiration and warmth in
supporting me throughout my academic pursuing and my life.
v
Table of Contents
List of Tables ............................................................................................................................... viii
List of Figures ................................................................................................................................ ix
List of Abbreviations ..................................................................................................................... xi
1 Introduction ............................................................................................................................. 1
1.1 Motivation ........................................................................................................................ 1
1.2 Scope and Objective ......................................................................................................... 3
1.3 Organization of Thesis ..................................................................................................... 5
2 Literature Review.................................................................................................................... 6
2.1 Conventional Safety Evaluation Measures ...................................................................... 6
2.2 Surrogate Safety Assessment ........................................................................................... 8
2.3 Micro-simulation Technique .......................................................................................... 12
3 Methodology ......................................................................................................................... 17
3.1 Stage 1 Pilot Study ......................................................................................................... 18
3.2 Stage 2 Micro-Simulation Model Construction ............................................................. 19
3.3 Stage 3a Model Calibration ............................................................................................ 20
3.3.1 Crash Prediction Model (CPM) .............................................................................. 21
3.3.2 Goodness-of-Fit ...................................................................................................... 23
3.3.3 Calibrated Parameters ............................................................................................. 25
vi
3.4 Stage 3b Model Validation ............................................................................................. 27
3.5 Stage 4 Scenario Tests.................................................................................................... 30
4 Pilot Study and Data Description .......................................................................................... 34
4.1 Pilot Study ...................................................................................................................... 34
4.2 Data ................................................................................................................................ 37
4.2.1 Study Area and Period ............................................................................................ 37
4.2.2 Crash Records ......................................................................................................... 38
4.2.3 Traffic Volume and Transit Service Information ................................................... 40
4.2.4 Signal Timing Plans and Intersection Geometry .................................................... 40
5 Model Construction, Calibration and Validation .................................................................. 42
5.1 Model Construction ........................................................................................................ 42
5.2 Model Calibration .......................................................................................................... 45
5.2.1 Driver Reaction Time (DRT) and Mean Headway Time (MHT) ........................... 45
5.2.2 Coverage Area of Analysis, Time to Collision, and Travelling Speed ................... 47
5.2.3 Peak Hour Ratio (PHR) .......................................................................................... 50
5.3 Model Validation............................................................................................................ 51
6 Scenario Tests ....................................................................................................................... 59
6.1 Effects of TSP ................................................................................................................ 59
6.2 Effect of Transit Stop Positioning .................................................................................. 65
vii
6.3 Effect of Transit Type .................................................................................................... 69
7 Conclusion and Future Work ................................................................................................ 73
References ..................................................................................................................................... 76
Appendix A List of Modelled Micro-simulation Networks ......................................................... 81
Appendix B Detailed Scenario Test Results ................................................................................. 89
viii
List of Tables
Table 3-1 Set-ups of Scenario Tests ............................................................................................. 31
Table 4-1 Descriptive Statistic for the Selected Sample ............................................................... 39
Table 5-1 Calibration Results of Parameters DRT and MHT....................................................... 46
Table 5-2 Calibration Results of Parameters Coverage Area, TTC, and Speed ........................... 48
Table 5-3 Calibration Results for the Inclusion of PHR ............................................................... 50
Table 5-4 Model Validation Results ............................................................................................. 52
Table 6-1 Predicted Results of TSP Treatments ........................................................................... 60
Table 6-2 Predicted Results of Transit Stop Positioning Treatments ........................................... 66
Table 6-3 Predicted Results of Transit Type Treatments ............................................................. 70
ix
List of Figures
Figure 2-1 Conflict Trajectory Diagram ....................................................................................... 12
Figure 3-1 Flow Chart of the Structure of the Study .................................................................... 17
Figure 3-2 Flow Chart of the Schematics of the Generation of Traffic Conflicts ........................ 21
Figure 4-1 Screenshot of the Pilot Study Network ....................................................................... 34
Figure 4-2 Observed Number of Crashes [2006 – 2010] at the Signalized Intersections of the
Selected Sample ............................................................................................................................ 38
Figure 4-3 Observed Crash Frequency [from 2006 to 2010] for the Selected Sample................. 39
Figure 5-1 Demonstration Network 1 - Lake Shore Blvd W and Marine Parade Dr ................... 43
Figure 5-2 Demonstration Network 2 - Finch Ave E and Tapscott Rd ........................................ 43
Figure 5-3 Demonstration Network 3 - Lake Shore Blvd E and Lower Jarvis St ........................ 44
Figure 5-4 CURE Plot - All Impact Type Crashes ....................................................................... 54
Figure 5-5 CURE Plot - Angle Crashes ........................................................................................ 55
Figure 5-6 CURE Plot - Rear-End Crashes .................................................................................. 55
Figure 5-7 CURE Plot - Side-Swipe Crashes ............................................................................... 56
Figure 5-8 CURE Plot - Transit-Involved Crashes ....................................................................... 56
Figure 6-1 Absolute Change in Predicted Crash frequency as a Result of Removing TSP [by
Transit Types] ............................................................................................................................... 63
x
Figure 6-2 Percentage Change in Predicted Crash Frequency as a Result of Removing TSP [by
Transit Types] ............................................................................................................................... 64
xi
List of Abbreviations
CPM - Crash Prediction Model
DeltaS - Speed Differential
DR - Deceleration Rate
DRT - Driver Reaction Time
GLM - Generalized Linear Model
MaxS - Maximum Speed
MHT - Mean Headway Time
NB - Negative Binomial
PET - Post-Encroachment Time
SPF - Safety Performance Function
SSAM - Surrogate Safety Assessment Model
TTC - Time to Collision
1
1 Introduction
Over the years, increasing efforts have been devoted into the field of traffic safety modelling. The
ability to evaluate the safety performance of an existing transportation infrastructure, or more
ideally a planned upcoming transportation infrastructure, could provide policy makers invaluable
understandings of how they could contribute towards safer roadway environment. Since
transportation infrastructures are generally irreversible in practice, the ability to predict the safety
impacts of planned transportation projects will stimulate a new facet of design considerations to
further ensure the engineering validity and robustness of transit designs.
Surface transit operation, albeit being an inseparable component of traffic flow, is often less
explored in comparison to general traffic. Aside from being dimensionally different from general
traffic, surface transit also exhibits many distinct operational characteristics, such as having transit
stops, inflexible lane usage, and occasionally priorities over other vehicles. As a result, this
academic endeavour was motivated to explore the safety impacts of transit operations, especially
those impacts not only to the transit vehicles themselves but also those imposed on general traffic.
Using micro-simulation as an effective surrogate safety assessment tool, this study seeks
quantitative answers to the questions of how, and to what extent do transit operations influence
overall intersection-level roadway safety performance.
1.1 Motivation
In the past, the designing stage of transit operations often emphasized on the more quantitatively
measurable norms such as capacity, delay, reliability, and transferability. Although some remarks
regarding safety impacts were occasionally provided, they were often of a qualitative and non-
deterministic nature. For example, in the long debated confrontation between near-sided versus
2
far-sided transit stops, both types have safety benefits and harms that are easily perceivable but
not deterministically quantifiable. Transit Signal Priority (TSP), an item that facilitates transit
movement at signalized intersections, also has had controversial safety impacts. TSP has often
been serviced in green extension individually, combined green extension and red truncation, or a
mixture of basic or more advanced priority algorithms; green extension and red truncation
respectively correspond to the practices of extending the green time and shortening the red time
for the transit-facilitating direction. The safety impacts of TSP have been especially difficult to
quantify given the stochastic trigger condition and operation of TSP, let alone being in the already
complex traffic operation environment. These challenges have motivated this academic endeavour
to investigate the safety impacts of transit infrastructures and furthermore to provide directions for
future transit planning.
Traditionally, the evaluation of safety performance has been done based on observed historic data.
This method involves using a set of observable characteristics as the explanatory variables to
establish a regression relationship with the response variable of interest, which has generally been
crash frequency. This method heavily relied on the representativeness of the observed data; as a
result, any inappropriate handling of population skewness may lead to undetected erroneous
conclusions. For the context of this study, since transit routes often service areas of high population
density, which arguably also have high traffic density, the effectiveness of using this method
diminishes. Crashes are generally rare events and thus require a long observation period or many
sites to establish an adequate sample size. Such long observation period of crash events, or
attaining a large sample of sites is impractical. In addition, the traditional regression analysis
method, while providing helpful understanding of the contribution of each explanatory variable,
has very limited sensitivity to detailed geometric, signal, and operational characteristics, which are
3
generally not easily quantifiable into explanatory variables. To overcome these difficulties, this
study has been motivated to adopt the micro-simulation technique as a surrogate assessment tool
for safety analysis. Micro-simulation emulates the microscopic interactions of individual roadway
users in the complex geometric, signal, and operational environment. Such interactions are an
integral component of the process through which crashes occur. A well-calibrated micro-
simulation model allows its users to observe the impacts of any experimental infrastructure
modification while maintaining all other explanatory variables status quo; this effectively allows
for a direct observation of the safety impact from, and only from the modifications made at a single
site.
1.2 Scope and Objective
Among the two commonly used measurements of safety performance, this study focuses on crash
frequency as the sole response variable. Crash frequency has generally been defined as the number
of crash occurrences within a pre-defined time period. Thus for the context of this study, the
quantitative term “crash frequency” alludes to the qualitative term “safety performance”, such that
an increasing crash frequency represents worsened safety performance and vice versa. It is worth
distinguishing crash frequency from the other safety performance measure, which is the crash
severity. Crash severity generally measures the probability of resulting in more severe injuries or
damages in the event of a crash. The evaluation of crash severity differs conceptually from that of
crash frequency and therefore does not fit within the scope of this thesis. Consequently, although
the overall safety performance is conceptually a mixed product of crash frequency and crash
severity, only the former one is used for assessing safety performance for the context of this study.
This study focuses on intersection-level safety performance. Despite of being only one component
of the large mass of traffic roadways, intersections are inherently concentrated areas of traffic
4
conflicts and crashes. It is not surprising that within the study region of the city of Toronto, more
than 75% of the recorded crashes are associated with intersection traffics. Therefore, all analyses
discussed within the context of this study, allude to intersection-level safety performance.
This thesis has two primary objectives. The first objective is to test the predictive capability of the
micro-simulation technique, when applied to assess the safety performance of transit
infrastructure. Using the simulated traffic conflicts as the explanatory variable and the observed
traffic crashes as the response variable, a set of Crash Prediction Models (CPMs) is developed to
describe the relationship between the two. The CPMs are calibrated using Generalized Linear
Modeling (GLM) with a Negative Binomial (NB) error structure. Ideally after careful calibration
and validation of the micro-simulation networks, if the CPM demonstrates strong statistical
significance and goodness-of-fit, the micro-simulation technique can be regarded as having
adequate predictive capability to emulate crash behaviours in reality.
The second objective is to examine the safety impacts of a number of hypothetical transit
infrastructure modifications. This would only be possible once the first objective has been
completed and the resulting CPMs have demonstrated adequate predictive capability. Through the
use of micro-simulation models, the safety performance of three main categories of transit
operation designs, namely the TSP, the positioning of transit stops, and the transit vehicle type are
investigated. For each category, modifications are made to the existing transit infrastructure in the
controlled environment of the micro-simulation networks; the resulting safety performance
following each infrastructure modification is monitored and compared with the existing condition.
Again, because other explanatory variables are held status quo, any change in safety performance
can be reasonably assumed to only be as a result of the modifications made. The goal of this
exercise is therefore to capture the independent responsiveness of the intersection-level safety
5
performance as a result of each individual modification. Such direct observations would otherwise
not have been possible in reality due to the complex and potentially interconnected characteristics
among the explanatory variables, let alone the need for a large sample of crash data.
1.3 Organization of Thesis
This thesis is organized in three sections, sequentially leading up to the objective of examining the
safety impacts of various transit infrastructure modifications. The first section discusses the
background that essentially forms the backbone of this thesis. In this section, past literature is
firstly reviewed, followed by a discussion of the methodologies adopted in this study. Then, this
section reviews the results of the pilot study, in which the viability of the available software and
data inputs were initially determined for this study. Lastly, this section discusses the collection of
these needed data, as well as other assumptions made for the construction of the micro-simulation
models.
The second section mainly focuses on the calibration and validation of the micro-simulation
technique adopted in this study. Following construction of the micro-simulation networks based
on the collected data, these networks are calibrated and validated. This section presents the
statistical significance and goodness-of-fit of the resulting CPMs, as well as a discussion of the
strength of the CPMs for prediction purposes.
The last section presents the scenario tests, in which hypothetical transit infrastructure treatments
are made to the existing micro-simulation networks. As mentioned, three main categories,
respectively TSP, transit stop positioning, and transit type, are investigated. Within each category
are several individual scenarios, such that each scenario examines one specific component of that
category. The resulting change in safety performance is discussed and compared in this section.
The practical implications of the change in safety performance are also discussed in this section.
6
2 Literature Review
The evaluation of safety performance has been an increasingly studied field as the ability to
understand contributing factors towards safer roadway designs is invaluable. An extensive number
of mathematical models has been developed around the world to test the statistical significance of
various explanatory variables that were observed in historical data. The feasibility and
extensiveness of most safety performance studies have been largely dependent on the availability
of data and the credibility of the data collection method. For most studies presented in this literature
review, collision data have been collected retrospectively. This means that each reported collision
had already been recorded individually at the time of the collision occurrence prior to the collision
data being used for safety analysis.
Although there have been extensive efforts spent on safety evaluation, the transit operation aspect
has been limitedly investigated. Surface transit operation has a crucial impacts on the flow of
traffic. Many elements of transit operation, such as its stop location, stop density, signal priority,
and even its acceleration/deceleration profile can differentiate the kinematic motion of a transit
vehicle from that of a general vehicle. However, only a small number of past researches has
focused on the safety impacts of transit operations on general traffics. More importantly, the safety
implications from these literatures have been mixed (Goh et al., 2013).
2.1 Conventional Safety Evaluation Measures
Crash record, when evaluate over a period of time, is a direct and commonly available piece of
data that represent the safety performance of a study region. Traditionally, crash frequency has
often been evaluated based on the study region’s observable factors. These factors often include
traffic volume, pedestrian volume, segment length, segment width, speed, as well as other binary
factors such as the presence of driveways, intersections, dedicated lanes, etc. In the past, crash data
7
from different areas in the world have been tested and researchers have reached a consensus on
that traffic volume is a key contributor towards crashes (El-Basyouny & Sayed, 2009; Hadayeghi,
Shalaby, & Persaud, 2007; Jonsson, Ivan, & Zhang, 2007; Mitra & Washington, 2007; Persaud,
Lord, & Palmisano, 2002). Traffic volume was often represented by vehicle-kilometer-travelled
for corridor-level study or by Annual Average Daily Traffic (AADT) volume for intersection-level
study. For intersection-level studies, AADT volume was often further categorized into volume in
the major direction and in the minor direction. In most of these literatures, it has been a generally
accepted practice to use GLM with NB error structure (or NB regression model for short) to model
crash frequency to account for the rare, discrete, and non-negative nature of crash occurrences.
However, other regression models were also tested. For example, El-Basyouny and Sayed (2009)
advocated the use of Multivariate Poisson-Lognormal Regression, in order to incorporate the
correlations among crashes of different severity levels. In this literature, it was demonstrated that
Multivariate Poisson-Lognormal Regression could have higher model precision and goodness-of-
fit, when compared with the conventional univariate NB regression models. In addition, due to the
ability to model crashes of different severity levels, this model formulation allowed for the
possibility to incorporate crash severity with crash frequency. Another CPM model structure, the
Negative Multinomial Regression Model was tested by Caliendo et al. (2007). In this literature it
was suggested that Negative Multinomial regression model with a carefully defined over-
dispersion parameter could have improved explanatory power than the conventional NB regression
model. Nonetheless, despite being relatively simpler to model when compared with the more
complex modelling structures, the conventional NB regression model is still a widely used GLM
structure for crash prediction.
8
In comparison with the larger pool of literatures in safety analysis, studies that focused on the
transit operation aspect of safety performance were less common. In many literatures, the transit
operation element has been omitted entirely except that transit volumes were anonymously
included in the total traffic volumes. Such omission inherently implied that a transit would be
indistinguishable from a general vehicle. In literatures that focused on the safety impacts of transit
operation, studies were generally at the corridor level that was in alignment with the investigated
transit route (Hedelin et al., 1996; Cheung et al. 2009; Goh et al. 2014). Many transit operation
elements were found to have statistically significant impacts to safety performance. For example,
in one of the literature, Cheung et al. (2008) used the NB regression model to investigate transit-
involved collisions at corridor-level in Toronto and uncovered that in addition to traffic volumes,
transit volume and near-sided stops also contributed to higher crash frequency. The corridor-level
analysis was followed by a zonal-level model, which included zonal transit operation elements
such as transit-kilometer-travelled, bus stop density, and percentage of near-sided stops; these
transit operation elements were also found to contribute to higher crash frequency. A very similar
result was found by Shahla et al. (2009), which also applied the NB regression model to study the
crash data in Toronto but focused on the intersection-level. In this literature it was also found that
near-sided stops and streetcars would contribute to high crash frequency and thus worsened safety
performance. Additionally, in this literature it was found that the presence of TSP was also
positively correlated with crash frequency.
2.2 Surrogate Safety Assessment
Aside from studying explanatory variables that were directly obtainable from field, researches
have also demonstrated that traffic conflict, a form of surrogate safety measure, is also applicable
for the evaluation of safety performance (Gettman et al., 2008). The idea of using traffic conflict
9
as a surrogate safety measure originated as early as 1967, when this technique was used to measure
the type and frequency of crashes at intersections (Perkins & Harris, 1967). The literature found
that drivers would react to traffic conflicts in an aggressively evasive maneuver in order to avoid
crashes. Thus, a traffic conflict was defined as the potential collision of any two vehicles had their
original moving trajectories be continued. In other words, this definition implied that every crashes
would originally be conflicts but only an unavoidable portion of the total conflicts would result in
crashes. Since then, the use of traffic conflict technique has gained more attention amongst
researchers and practitioners.
Traditionally without the aid of computers, traffic conflicts were collected by field observers
(Older & Spicer, 1976; Zegeer & Deen, 1978). However, it was soon acknowledged that human
observation of conflicts had an uncontrollable degree of subjectivity, which deteriorated the
validity of the data. Furthermore, the precision of whether an event should be classified as a
conflict largely depends on the judgement of the field observers, which in many cases had
inconsistent opinions. Nevertheless, the observed traffic conflicts were shown to have some
correlation with crash frequency. With the emerging computer-aided image recognition ability, the
use of video-camera to capture vehicle conflicts has also been continuously gaining attention. This
technique, if calibrated correctly, would greatly alleviate the labour pressure and remove human
errors. Another noteworthy advantage of using the traffic conflict technique was the avoided needs
for long observation periods to collect the rarely occurring and unwanted crash events. Traffic
conflicts by definition are more frequent events than traffic crashes. As a result, the traffic conflict
technique would be especially helpful for before-after studies, since the observation period would
have been much shorter for reaching an adequately large sample size of traffic conflicts than traffic
crashes. Autey et al. (2012) applied the video-captured traffic conflict technique to investigate the
10
safety impacts of decreased turning angle of channelized right turns. The data collection periods
were as short as four days for both before and after the treatment. Such data collection period
would likely have extended to several years, had crash frequency been used instead of traffic
conflicts. The findings from this literature suggested that the total hourly conflicts were reduced
by almost half due to the right turn treatment and furthermore demonstrated the capability of
applying the traffic conflict technique for before-after studies. One shortcoming of this literature
was the missing linkage between the observed traffic conflicts and the observed crash frequency.
Consequently, even though the results were undeniably promising, they were not yet relatable with
the more straightforward crash frequency.
Motivated by this missing linkage between traffic conflicts and crash frequency, El-Basyouny &
Sayed (2013) proposed one approach to model this linkage. In this literature, a two-phased
approach was proposed. In the first phase, a lognormal regression model was deployed to model
traffic conflicts based on observable explanatory variables such as traffic volume, area type, and
geometric characteristics. The modelled traffic conflicts were then taken as the explanatory
variable in the second phase, in which the crash frequency was modelled using the conventional
NB regression model. Since all observable explanatory variables were used in the first phase for
the modelling of traffic conflicts, in turn traffic conflicts were the sole explanatory variable in the
second phase. For both phases, their respective lognormal and NB regression model demonstrated
adequate goodness-of-fit. Also this literature confirmed the positive association between traffic
conflicts and traffic frequency with the use of the NB regression model. However, this study was
limited by its small sample size, which modelled only 51 signalized intersections.
Clearly, these literatures have strongly demonstrated that the traffic conflict technique is a viable
alternative in the field of safety evaluation. Regardless of the applied technology to capture the
11
vehicle trajectories, traffic conflicts could be computationally identified based on the observed
vehicle trajectories and other pre-defined kinematic parameters. Amongst the many thresholds are
the Time-to-Collision (TTC) and Post-Encroachment Time (PET). TTC is defined as the time from
when the two vehicles are on a collision course to when they would collide had their trajectories
not been changed. PET is the time between when the first vehicle initially occupied a spatial
location and when the second vehicle subsequently arrived at the same location (Gettman et al.,
2008). A figurative illustration of TTC and PET is shown by Figure 2-1 below. According to these
definitions, a conflict would be more dangerous and arguably more likely to result in a severe
crash, if it has lower values of TTC or PET. Although TTC and PET have been widely accepted
as criteria of identifying traffic conflicts, unfortunately no universal standards have been set on
their exact threshold values for which conflicting trajectories would be considered conflicts.
12
Figure 2-1 Conflict Trajectory Diagram
Again, TTC and PET are generally latent parameters that are complex to be observed directly from
the field, and are therefore obtained via post-data-processing based on the observed vehicle
trajectories that are comprised of detailed measurable variables such as spatial location and speed
profile. Currently, two common techniques of capturing vehicle trajectories are the video-captured
traffic conflict technique which was briefly discussed above, and the micro-simulation technique
which is discussed next.
2.3 Micro-simulation Technique
Micro-simulation has been a widely applied technique in many fields of transportation. A
cautiously constructed and well calibrated micro-simulation model grants its users the ability to
emulate real world traffic operations. However, the process of the micro-simulation technique can
13
be highly data demanding. The construction of micro-simulation networks often requires many
pieces of data, which sometimes may be from different sources, to achieve a desired level of
modelling precision. In addition, extensive calibration and validation efforts are generally needed
to validate that the micro-simulation model is representative of its emulated reality.
Once calibrated and validated, a micro-simulation model can be incredibly powerful in many
analyses. The most noteworthy value added of the micro-simulation technique is that it allows its
users to conduct scenario tests in a controlled environment in which hypothetical treatment can be
made. Such treatment would generally have been costly, time-consuming, and most importantly
irreversible for many transportation infrastructures in reality. In addition, the controlled
environment within the micro-simulation models allows for specific infrastructure treatment while
maintaining all other elements status quo. As a result, the observed changes in the response
variables could only be as a result of the treatments made. Such observations would have been
difficult and time-consuming in the uncontrollable reality, since other observable characteristics,
such as traffic volume and pedestrian volume, may be volatile from day to day. By applying the
micro-simulation technique, researchers can overcome these challenges and can often estimate the
effect of a proposed infrastructure treatment before actually undertaking the treatment. As a result,
the micro-simulation technique can be regarded as a powerful alternative to the traditional before-
after studies. In the past, most literatures have devoted efforts on developing micro-simulation
models to evaluate directly quantifiable norms such as flows, speed, queues, and other variables
of interests. It was not until more recently when researchers have been motivated to explore the
potentials of using the micro-simulation technique for evaluating traffic safety, which are more
difficult to quantify directly (Archer, 2004; Gettman & Head, 2003). In addition to the use of TTC
and PET thresholds as identifiers of traffic conflicts, other identifiers, including Maximum Speed
14
(MaxS), speed differential (DeltaS), and initial Deceleration Rate (DR), were also presented by
Gettman and Head (2003). MaxS measures the higher one of the two conflicting vehicles’
travelling speeds; DeltaS measures the difference in the travelling speeds of the two conflicting
vehicles. Thirdly, Initial DR measures the initial deceleration rate of the second vehicle in the
evasive action to avoid the potential crash. Intuitively, a conflict with a larger value of DR indicates
a higher probability of resulting in crash. On the other hands, the Maxs and DeltaS indentifiers
were argued to be more associated with the severity of the resulting crash, had the crash event
occurred.
The development of these kinematic identifiers of traffic conflicts lead to the development of the
software Surrogate Safety Assessment Model (SSAM) by the U.S. Federal Highway
Administration (Gettman et al., 2008). SSAM aimed to provide additional safety analyzing
capability to existing simulation softwares and currently supports Aimsun, Q-Paramics, TEXAS,
and VISSIM. By studying the vehicle trajectories outputted by the micro-simulation software,
SSAM is capable of recognizing traffic conflicts based on many identifiers such as TTC, PET, DR,
MaxS, DeltaS, etc. Gettman et al. (2008) also demonstrated that with the combined use of the
micro-simulation software VISSIM and SSAM, NB regression models could be developed to
correlate the simulated traffic conflict frequency with the observed crash frequency. A positive
association was found between the simulated traffic conflict frequency and the observed crash
frequency; such positive association was in alignment with that found from the video-captured
traffic conflict technique.
Since the development of SSAM, researchers have devoted more efforts in using micro-simulation
to study traffic safety and the results have been promising. Huang et al. (2013) used the micro-
simulation package VISSIM and found statistically significant correlation between the simulated
15
conflicts and observed conflicts obtained from using video-captured conflict techinique. In this
literature a two-staged model calibration procedure was applied, whereas the first stage calibrated
the model with the observed traffic volume, speed, and headways and the second stage calibrated
the model with the observed video-captured conflicts. A linear regression technique was applied
to correlate the simulated conflicts with the observed video-captured conflicts and strong positive
association was found. One limitation of this literature was the missing linkage between the
simulated conflict frequency and the observed crash frequency. In addition, this literature
somewhat reflected the data demanding nature of the micro-simulation technique, since during the
process many pieces of information were used in the two-staged model development.
Ariza et al., 2011 used the micro-simulation package Paramics and validated a large-scale network.
In this literature the simulated conflicts were also found to have a strong statistical relationship
with observed crashes. Two SPFs, respectively for corridor-level and intersection-level were
developed in this literature. In the comparison it was found that the corridor-level SPF performed
inadequately with poor goodness-of-fit. In contrast, the intersection-level SPF demonstrated a
better goodness-of-fit, thus suggesting more predictability at the intersection-level. The evaluation
of intersection-level safety was extended by Saleem et al. (2014), in which VISSIM was used to
confirm the strong statistical correlation between simulated conflicts and observed crashes. Also
in this literature, a hypothetical treament on modifying existing permissive left turns to protected-
permissive was also evaluated. The results suggested that such treatment would be beneficial and
would reduce the number of angle and turning crashes.
To the knowledge of the author, few literatures have applied the micro-simulation technique to
examine the effect of transit operation on road safety. Although there have been researches focused
on the safety of transit itself, to what extent does transit operation influence general traffic is yet
16
an under-explored field. The closest literature to the scope of this research is that by Goh et al.
(2014). In this literature, dedicated bus priority lanes were micro-simulated using Aimsun and the
effects of such infrastructure treatment on general traffic was examined. It was suggested that
dedicated bus lanes would improve the overall safety performance by reducing the number of
conflicts. However, the analyses in the literature were at the corridor-level along the investigated
transit route and consisted only a limited number of major arterial intersections. As a result, it
lacked the capability of investigating the safety impacts of many intersection-level transit
operation elements, such as transit stop positioning and TSP.
17
3 Methodology
For every micro-simulation model to effectively replicate its real world counterparts, significant
efforts in data acquisition, model calibration, and model validation must be done. Once the micro-
simulation model has been fully constructed, calibrated and validated, it can be considered
adequately representative of its real world counterparts. Then, hypothetical scenarios can be tested
in the controlled environment of the micro-simulation network. Figure 3-1 below presents a flow
chart illustrating the sequential steps undertaken in this study:
Figure 3-1 Flow Chart of the Structure of the Study
As depicted by the flow chart, items that are within the same stage were completed from the top
to bottom. An item did not need to be fully completed to proceed to the next item within the same
stage. However, every item(s) within the same stage must be fully completed in order to start the
first item in the subsequent stage. This chapter illustrates the methodologies undertaken for each
of the four stages of this study.
18
3.1 Stage 1 Pilot Study
Initially it was anticipated that the construction, calibration, and validation would be the most time
consuming items. Thus, a pilot study was first conducted to ensure these subsequent efforts can be
successfully and time-efficiently executed. Another importance of the pilot study was to test if the
available micro-simulation package, Paramics, is a suitable program for all the subsequent works.
More specifically, the pilot study was to investigate the following:
1) the capability of Paramics to incorporate the transit operation, pedestrian, and variable
signal operation elements into its simulation;
2) the capability of SSAM to interpret the trajectories of transit and pedestrian,
3) an acceptable study size such that the simulation time by Paramics and the analysis time
by SSAM are reasonable and not undesirably long; the study size should neither be too
small to avoid poor statistical significance; and,
4) data needs based on the capability of Paramics to implement them.
Despite being the least time-consuming item of all, the pilot study was crucial for all subsequent
efforts. In the pilot study an isolated intersection was created for a simulation duration of 90
minutes (including 30 minutes of warm-up period and 60 minutes of data collection). Vehicle
volume, transit service, and signal timing were introduced to the network. The trajectories for the
duration of the simulation were collected and analyzed in SSAM. Following this, pedestrian,
streetcar, and TSP were also introduced to examine the sensitivity of the trajectories in response
to these modifications.
19
3.2 Stage 2 Micro-Simulation Model Construction
Following the pilot study was the construction of the actual micro-simulation networks. At this
stage, the sample size of the number of isolated intersection networks to be studied was already
estimated from the pilot study. This sample size should be sufficiently large to avoid poor statistical
significance and at the same time not exceedingly large to be undesirably time-consuming.
Moreover, the sample size should be large enough such that any subgroup of interest within the
sample was also reasonably large. Here, a subgroup is a portion of the sample that exhibited a
common characteristic that would later be investigated in scenario tests, such as having streetcars
or TSP. Since random sampling technique was used, a large sample size of the subgroups would
allow for easier generalization to the population.
The micro-simulation models were constructed at the isolated intersection level. Every approach
upstream of the studied intersection would only extend upstream for 120 to 150 meters. The length
of the upstream link varied among intersections, but in general would not overextend into its
upstream intersections; it neither would be too short to avoid the study region becoming
unrepresentative.
The duration of the micro-simulation models was set to 90 minutes, which again included 30
minutes of warm-up period and 60 minutes of statistics collection period. The first 30 minutes of
warm-up period was to stimulate the initially empty networks with the specified vehicle volumes.
This would allow for a more accurate representation of the network once the warm-up period had
been completed. The 60 minutes duration of the simulation was injected with traffic volumes
observed from the PM peak hour so that the simulation would replicate the PM peak hour traffic
operation of the studied intersections. It was demonstrated in previous literature that, peak hour
20
operation, which ideally should represent the critical design performance, was both sufficient and
efficient for safety modelling (Gettman et al., 2008, Saleem et al., 2014).
3.3 Stage 3a Model Calibration
For every micro-simulation model, its calibration and validation must be carefully executed for it
to be representative of the reality. Calibration refers to the practice in which various parameters in
the model structure were adjusted for optimal statistical performance. The statistical performance
of the micro-microsimulation model often depends on a pre-defined set of criteria, which most
frequently are the goodness-of-fit of the simulated variables for fitting the observed variables. The
calibration was performed on a randomly selected subset from the entire sample size. Following
model calibration, the micro-simulation was validated on the entire sample size to ensure the
credibility of the model and further to demonstrate its ability to emulate real-world traffic
operation.
For the context of this study, since the micro-simulation networks were built at isolated intersection
level, the calibration work for such smaller-scale networks was considerably different from that of
the large-scale regional-level networks. Generally for large-scale networks, in which the traffic
volume in the simulation is obtained through traffic assignments, the simulation’s traffic volume
needs to be calibrated with the observed traffic volume. For this study, having the networks at
isolated intersection level avoided the need of a traffic assignment and thus allowed the observed
traffic volume to be used directly as the simulation’s traffic volume. Theoretically it would be ideal
to calibrate and validate the simulated crashes with the observed crashes. However, the micro-
simulation software does not directly simulate crashes but rather simulates vehicle trajectories,
which can be analyzed to produce conflicts. As a result, in this study the calibration and validation
21
of the model were done by assessing the goodness-of-fit of the simulated conflicts with the
observed crash frequency by applying the concept of SPF.
The generation of the traffic conflicts was a sequential step involving both Paramics and SSAM.
Firstly Paramics generated vehicle trajectories once the constructed networks were simulated.
Then taking these trajectories as an intermediary input, SSAM could analyze the trajectories and
output the corresponding traffic conflicts. A schematic of the generation of traffic conflicts is
illustrated by Figure 3-2 below:
Figure 3-2 Flow Chart of the Schematics of the Generation of Traffic Conflicts
3.3.1 Crash Prediction Model (CPM)
CPM belongs to the family of SPF that correlates a set of explanatory variables to the response
variable, which in this case is the observed crash frequency, leading to the terminology of “Crash
Prediction Model”. The simulated conflicts were used as the explanatory variable of the CPM. As
was presented in the literature review, it has been a generally accepted practice to use the GLM
with NB error structures, also known as the NB regression, for predicting frequency. This is
because the NB regression accounts for the discrete, non-negative, and rarely-occurring properties
of crash events.
Let 𝑌𝑖 denotes the crash frequency at intersection i, the expected mean and variation of the crash
frequency of NB regression follows that:
22
𝐸(𝑌𝑖) = �̂�𝑖 and
Equation 1
𝑉𝑎𝑟(𝑌𝑖) = �̂�𝑖 + 𝛼 ∗ �̂�𝑖2 Equation 2
where 𝑌�̂� represents the predicted crash frequency and α represents the dispersion parameter to be
estimated from the NB regression model. Note that the dispersion parameter α also uses one degree
of freedom during parameter estimation. α indicates the strength of the assumption of the negative
binomial error structure. If α is large, over-dispersion is present and the use of NB regression might
become inappropriate. In other words, a smaller value of α indicates better model fitting. It is worth
noting that some statistical software, such as R, uses the inverse of the dispersion parameter and
should not to be confused with the definition in this study (R Development Core Team, 2008). The
concept of over-dispersion will be discussed further in the Section 3.3.2. By applying the NB
regression, the CPM for the predicted crash frequency 𝑌�̂� resembles the following form:
�̂�𝑖 = 𝑒𝜆 ∗ 𝑋𝑖𝛽 or, ln �̂�𝑖 = 𝜆 ∗ 𝛽ln(𝑋𝑖)
Equation 3
where the explanatory variable 𝑋𝑖 represents the simulated traffic conflict frequency for each
intersection i, and 𝜆 and β are parameters of the CPM to be estimated. Since all simulations had a
duration of one hour, the explanatory variable 𝑋𝑖 is more conveniently referred to as “simulated
conflicts”. Similarly, since the recorded frequency for the investigated networks were collected
from the same study period, the response variable �̂� is more conveniently referred to as “predicted
crashes”. The first determination of the strength of the CPM is to examine whether the estimated
parameters were statistically significant. For each variable, the SAS software estimated the p-
value, for which smaller values indicate more evidence against the null hypothesis that the
variable’s parameter is zero. A p-value of less than 0.05 indicated 95% confidence in rejecting the
null hypothesis and the variable can be regarded as statistically significant. For this study, 95%
23
confidence was used in assessing the statistical significance of variables of interest. Since the
simulated conflicts were the sole explanatory variable being investigated, if the parameter for this
variable was determined to be non-significant, the model would have no predictive capability.
3.3.2 Goodness-of-Fit
The statistical software SAS was used for the estimation of parameters and the assessment of the
CPM’s goodness-of-fit. Since the NB regression belongs to the family of GLM, the PROC
GENMOD procedure of SAS software was used, which allowed for the maximum likelihood
estimation of the model parameters in the CPM (SAS Institute Inc., 2008). The goodness-of-fit
measures assessed the strength of the modelled relationship between the simulated conflicts and
the observed crashes. Three measures of goodness-of-fit were evaluated: the scaled deviance, the
Pearson χ2, and the Miaou’s R2.
The Scaled Deviance (SD) is the ratio that measures twice the difference between the maximum
log likelihood of the proposed CPM and the maximum achievable log likelihood of a “full” model
(or sometime referred to as the “saturated” model). A “full” model would have as many parameters
as the number of observations and thus would perfectly fits the observed crashes, providing a
benchmark for assessing the goodness-of-fit. McCullagh and Nelder (1989) demonstrated that the
SD for NB regression model follows the form:
𝑆𝐷 = 2∑[𝑦𝑖 ln (
𝑦𝑖
�̂�𝑖) − (𝑦𝑖 − 1/𝛼) ln (
𝑦𝑖 + 1/𝛼
�̂�𝑖 + 1/𝛼)]
𝑛
𝑖=1
Equation 4
where 𝑦𝑖 represents the observed crashes at the intersection i, and n is the total number of
intersections being studied.
24
The Pearson χ2 was the second goodness-of-fit measure, which is the summation of squared
residuals scaled by the variance of the modelled variable. The Pearson χ2 can be mathematically
illustrated as follow:
𝑃𝑒𝑎𝑟𝑠𝑜𝑛χ2 =∑
[𝑦𝑖 − �̂�𝑖]2
𝑉𝑎𝑟(𝑌𝑖)
𝑛
𝑖=1
, 𝑤ℎ𝑒𝑟𝑒𝑉𝑎𝑟(𝑌𝑖) = �̂�𝑖 + 𝛼 ∗ �̂�𝑖2
Equation 5
Both the SD and the Pearson χ2 are asymptotically distributed following χ2 distributions with n-p
degrees of freedom. Again, n is the sample size and the p is the number of parameters being
estimated. If the SD and the Pearson χ2 are smaller than the critical χ2 value at the modelled degree
of freedom, it can be concluded at 95% confidence that the NB regression model has provided an
adequate fit to the observed data. In addition, the effect of dispersion can be estimated by assessing
the ratios of either the SD or the Pearson χ2 divided by the n-p degree of freedom. More
specifically, both 𝑆𝐷/(𝑛 − 𝑝)and Pearsonχ2/(𝑛 − 𝑝) should be close to the value of 1.0 for
proper dispersion. If these indicators are larger than 1.0 then over-dispersion is present; conversely
if they are less than 1.0 then under-dispersion is present. For this study, values between 0.8 and
1.2 would be considered good fitting for the dispersion of the specified NB regression model.
The third goodness-of-fit measure was the Miaou’s 𝑅𝛼2, which is explicitly based on the dispersion
parameter α. This measure was proposed by Miaou (1996) and the formulation follows the form:
Miaou′s𝑅𝛼2 = 1 −
𝛼
𝛼𝑚𝑎𝑥
Equation 6
where 𝛼𝑚𝑎𝑥 is the maximum possible dispersion parameter estimated by using only a constant
term as the explanatory variable in the SPF. The Miaou’s 𝑅𝛼2 resembles the concepts of the R2,
which is the more traditional measure used to evaluate the goodness-of-fit of the ordinary least-
25
square regression. Miaou’s 𝑅𝛼2 ranges between the value 0 and 1, with higher 𝑅𝛼
2 indicating more
explanatory power of the NB regression model.
3.3.3 Calibrated Parameters
During the model calibration, key parameters were adjusted so that the resulting simulated
conflicts had a statistically defensible relationship with the observed crashes. Both Paramics’
simulation parameters and SSAM’s filtering parameters would dictate the simulated conflicts and
therefore would need to be calibrated. The process of parameter calibration was an iterative
approach of seeking the optimal set of parameters within both Paramics and SSAM such that the
resulting CPM would have the best performing goodness-of-fit.
In Paramics, Driver Reaction Time (DRT) and Mean Headway Time (MHT) are two of the core
parameters modifiable within network attribute. These parameters are closely connected with the
fundamental car-following, gap acceptance, and lane-changing models embedded in Paramics
(Duncan, 1997). DRT specifies the average delay in response time of the following vehicle in
response to a change in speed of the preceding vehicle. MHT specifies the average target headway
maintained between the preceding and the following vehicles. It is worth mentioning that both
parameters only represent the desired average value. The actual individual DRT and MHT
assigned to each driver vehicle unit respectively follow normal probability distributions with
means equal to the specified DRT and MHT. A reduction in either of the DRT or MHT would
result in more aggressiveness in average driving behaviour. Thus, both DRT and MHT must be
calibrated such that the simulation model had the most realistic driving behaviour for intersection-
level safety modelling.
Following adjusting the simulation parameters in Paramics, it was also essential to test the
sensitivity of the filtering parameters embedded in SSAM. These filtering parameters dictate the
26
proportion of raw conflicting trajectories to be pertained as conflicts. Two filtering criteria were
tested in this section, namely the coverage area of the analysis and the simultaneous adjustment of
TTC and travelling speed. The coverage area of the analysis specifies the lateral and longitudinal
range in which SSAM conducts its trajectory analysis. In other words, the coverage area dictates
the distance upstream of the intersection beyond which would be filtered out by SSAM. In reality,
if a crash did not occur within obvious proximity of the intersection, whether it would be classified
as midblock or intersection would largely depend on the recorder’s own judgement on an
acceptable upstream distance. Thus the goal of adjusting the coverage area was to match the
analysis area with the actual area in which crashes would be recorded as intersection crashes. Next,
TTC and travelling speed were also considered together as a filtering criteria to eliminate conflicts
from questionable or impractical driving behaviour. This filtering criteria was suggested by
Gettman et al. (2008). The criteria invovled setting a minimum TTC threshold of 0.05 seconds and
a minimum travelling speed of 16.1 km/hr. The minimum TTC threshold is to account for the rare
occasions in which two simulated vehicles occupy the same physical location, which is not
possible in reality. The minimum travelling speed threshold is included for the reason that a
conflict arguably would have been avoided or not reported in reality if both vehicles had been at
very low speed. Note that the maximum thresholds for TTC and PET were kept at their default
values, respectively 1.5 and 4.8 seconds. These two parameters, despite having an important
influence in dictating traffic conflicts computationally, are arguably more connected with crash
severity than with crash frequency. Since crash severity is a subject beyond the scope of this study,
the maximum thresholds for TTC and PET were not calibrated.
After the parameter calibration for both Paramics and SSAM, the CPM werealso tested to examine
the effect of introducing additional explanatory variables aside from the simulated conflicts.
27
Thereotically, every pieces of data that were initially used to develop the micro-simulation model,
should have already contributed to the generation of the simulated conflicts. Therefore the
introduction of new explanatory variables can only be those not used in the construction of the
micro-simulation model. In this study, the effect of Peak Hour Ratio (PHR) was investigated.
PHR was suggested by Saleem et al. (2014) to account for the representativeness of the peak hourly
traffic pattern to the daily traffic pattern. Since the simulation model captured only the PM peak
hour, its representativeness of the daily crash pattern should be investigated. In the study, the PHR
was assumed to be the same proportionality of the peak hourly traffic volume to the daily traffic
volume, which varied among intersections.
Ideally, the calibration of Paramics parameters, SSAM parameters, and the new explanatory
variable PHR should be simultaneous, such that all exhaustive combinative possibilities were
investigated. However, this would extend the time of the calibration process considerably. Thus
for this study model calibration was performed in sequential fashion, in which Paramics
parameters were calibrated first, SSAM parameters were calibrated second, and the effect of PHR
in the CPM was calibrated last. This sequence was in accordance with that in the generation of the
simulated conflicts. In addition, only the total simulated conflicts and total observed crashes were
correlated and compared in model calibration. In other words, the calibration of parameters was
based on the goodness-of-fit of modelling only the total observed crashes, which was not yet
dissectted by their impact types or by the involvement of transit until model validation.
3.4 Stage 3b Model Validation
Following model calibration on the randomly selected subset of samples, the optimal set of
parameters was applied to the entire sample size, thus validating the micro-simulation model. The
objective of the model validation was to ensure that the calibrated parameters in Paramics and
28
SSAM would also be adequate when applied to the entire sample size. This was tested by
examining the goodness-of-fit of the CPMs, when the entire sample had its simulated conflicts
tested with its observed crashes. A total of five CPMs were developed; the first one modelled the
relationship between all simulated conflicts and all observed crashes; the next three CPMs
modelled the relationship between angle, rear-end, and side-swipe impact type of simulated
conflicts and their respectively corresponding observed crashes; finally the last CPM modelled the
relationship between transit-involved simulated conflicts and transit-involved observed crashes.
Four goodness-of-fit measures were used in model validation. The first three measures were again
the SD, the Pearson χ2, and the Miaou’s R2, as were introduced in the Model Calibration section.
One additional graphical representation of the goodness-of-fit was used here, which was the
CUmulative REsidual (CURE) analysis. The CURE analysis is a powerful tool for assessing
whether the specified relationship between the response and explanatory variables is justifiable. In
this study, scaled residuals, rather than raw residuals, were used. Scaled residuals are the difference
between the estimated and the observed response variable, scaled by response variable’s standard
deviation. The scaled residuals are also sometimes referred to as the Pearson’s residuals, because
the summation of squared scaled residuals would lead to the Pearson’s χ2. However in the CURE
analysis, the cumulative scaled residuals are not squared and thus can be either positive or negative.
The observations are ranked based on the numerical values of the leading explanatory variable in
ascending order. The leading explanatory variable, which in this study would be the simulated
conflicts, is taken as the x-axis in the CURE analysis plot. The cumulative scaled residuals are
plotted as the y-axis. The formulation of the cumulative scaled residuals is as follow:
CumulativeScaledResiduals ateachobservationk
=∑𝑦𝑖 − �̂�𝑖
√𝑉𝑎𝑟(𝑌𝑖)
𝑘
𝑖=1
, 𝑓𝑜𝑟𝑘 ≤ 𝑛 Equation 7
29
where k is the k-th observation within the n number of total observations. For each observation,
the scaled residuals should be approximately unbiased and independent. Thus the pattern of plotted
cumulative scaled residuals should resemble that of the random walk phenomenon. Since the slope
of the cumulative scaled residuals plot is the scaled residuals, it can be used to identify biasness.
If the slope of the cumulative scaled residuals plot is consistently positive, the CPM underestimates
the response variable; conversely if the slope is consistently negative, the CPM over-predicts the
response variable. The summation of the scaled residuals (i.e. the cumulative scaled residuals when
k = n) should be approximately zero to indicate non-biasness; otherwise, a pattern in the direction
of cumulative scaled residuals suggests an unobserved systematic effect with increasing k. Such
systematic effect would also suggest that alternative model formulation, or even high-ordered non-
linear model structure might be possible.
In the CURE analysis, the upper and lower bands of the random-walk phenomenon were also
plotted to assist the assessment of the goodness-of-fit of the plotted cumulative scaled residuals.
The random-walk phenomenon, when sample size is large to invoke the central limit theorem,
would resemble normal distribution with a mean of zero and a variance of 𝜎∗2. The formulation
of 𝜎∗2 follows that (as was proven in Hauer & Bamfo (1997)):
𝜎∗2 = 𝜎2(𝑘)[1 −
𝜎2(𝑘)
𝜎2(𝑛)] Equation 8
where 𝜎2(𝑘) and 𝜎2(𝑛) are respectively the variance of the cumulative residuals when i=k-th
intersection and when i=n total number of intersections. The upper and lower bands are
respectively ±2𝜎∗ from the mean of zero, indicating the 95% confidence level of the range of the
random walk phenomenon. Ideally, the plots of the cumulative scaled residuals should fit within
the two bands for an appropriate fitting.
30
Once all goodness-of-fit measures were determined to be acceptable, the micro-simulation
networks and the CPM formulation would be considered representative of the reality and thus
validated. Since the sample was to be randomly selected from the population, this would imply the
established CPM should be representative of the entire population. Therefore, following successful
model validation, the effect on the response variable from any infrastructure change in the micro-
simulation networks could be monitored. This lead to the scenario tests, which is discussed next.
3.5 Stage 4 Scenario Tests
In this study, scenario tests refer to the practice of modifying an existing transit infrastructure in
the micro-simulation models and observing the resulting predicted crashes in response to the
modification. It is worth emphasized again that scenario tests are only possible once the parameters
within the micro-simulation models have been carefully calibrated and validated, such that the
models are a statistically defensible representation of the reality. Through the use of the micro-
simulation models, the effect of an infrastructure change can be observed in a closed and controlled
environment in which other factors would remain status quo. However, modification could only
be made to existing infrastructures for which data were initially available to describe their status
quo conditions. In other words, if an existing infrastructure setting in the micro-simulation model
was based on assumptions, such infrastructure would not be modified in the scenario tests. As was
outlined in the thesis objective, this study investigated the safety impacts of three main categories
of transit infrastructures, namely TSP, transit stop positions, and transit type.
For each scenario, only a number of qualified micro-simulation networks that exhibited the
infrastructure under investigation were included in the scenario. A total of nine scenarios were
designed and are described in more details in Table 3-1 below:
31
Table 3-1 Set-ups of Scenario Tests
Scenario
#
Investigated
Transit
Infrastructure
Investigated
Effects
Network
Selection Criteria
1
TSP
Removal of green extension Have green extension
2 Removal of red truncation Have red truncation
3 Removal of TSP Have TSP
4
Transit Stop
Positions
Near-sided to far-sided
Have near-sided stops along its
major direction
Do not have TSP
5 Near-sided to far-sided
Have near-sided stops along its
minor direction
Do not have TSP
6 Near-sided stop to no stop
Have near-sided stops along its
major direction
Do not have TSP
7 Near-sided to Far-sided
Have near-sided stops along its
TSP-servicing direction
Have TSP
8
Transit Types
Streetcar to Bus (1:1
replacement)
Have streetcar making through
movement
9
Streetcar to Buses (1:n
replacement for same
operating capacity)
Have streetcar making through
movement
As depicted by Table 3-1 above, a total of three categories were being investigated. The first
category investigated the safety impacts of TSP and consisted three scenarios. The three scenarios
respectively tested the removal of green extension, removal of red truncation, and removal of TSP
as a whole (i.e. both green extension and red truncation). Note that only the effect of removing a
TSP scheme was investigated. This is because it was practically more convenient to remove an
existing TSP scheme, than to design and implement a new TSP scheme, which would involve
making new assumptions for the TSP parameters including detection distance, decision point,
increment time, extension/truncation time, etc.
In the second category, four scenarios were designed to investigate how the positioning of transit
stops influences safety performance. The four scenarios respectively tested the effects of switching
32
near-sided transit stops to (1) far-sided stops along major directions at non-TSP intersections, (2)
far-sided stops along minor directions at non-TSP intersections, (3) no stops along major directions
at non-TSP intersections, and (4) far-sided stops along TSP-servicing directions at TSP
intersections. Note that in all four scenarios, the status quo condition was near-sided stops; this is
due to the fact that the majority of the transit stops in the study region was near-sided stops. When
the near-sided stops were switched to far-sided stops, the exact positioning of the far-sided stops
was largely dependent on the geometric environment of each individual network. In accordance
with the general practice of far-sided stops in the study area, the positioning of the far-sided stops
was generally 20 meters downstream of the intersection but may extend further downstream to
avoid obstructing driveways. In addition, the effect of near-sided stops versus far-sided stops was
studied separately for TSP and non-TSP serviced intersections, to avoid the biasness that TSP
generally favors far-sided transit stops with respect to delay performance. Lastly, one scenario was
designed to test the effect of removing transit stops along the major direction entirely to investigate
the safety impacts of having transit stops at intersection-level; the practical implication of this
scenario could be treated as if the transit stops were relocated to midblock.
In the third category, two scenarios were designed to investigate the safety impacts of streetcars
versus buses. In both scenarios, the networks being investigated had their through-moving
streetcars along their major directions being replaced with buses. The networks were only qualified
if their streetcars were making thru-movement; this criterion was to avoid (1) the under-
representation of networks in which streetcars, in very rare occasions, make left-turn/right-turns,
and (2) the biasness in mixing turning movements with thru movement, since the lane usage of a
streetcar’s turning movement is inherently different from a bus’s turning movement. Also, only
streetcars that operated under mixed right-of-way with general vehicles were being replaced in
33
these scenarios. The original median-lane streetcar stops were replaced with curb-lane bus stops.
Additionally, since originally the streetcar’s median-lane stops prohibited upstream vehicles to
pass stopped streetcars, this constraint was also removed when the median-lane streetcar stops
were replaced with curb-lane bus stops. In the first scenario, the streetcars were replaced with
buses according to one to one ratio; this practice was to maintain the same overall transit volume
and thus traffic volume. This scenario would investigate the safety impact of streetcars versus
buses with respect to the contribution of each transit vehicle unit. In the second scenario, each
streetcar was replaced by multiple buses such that the same operating capacities were met for the
transit routes. This scenario was more fair and practical from the transit operation perspective and
would provide insights regarding the overall safety impacts of each transit types.
Following the transit infrastructure modification in each of the nine scenarios, the micro-
simulation networks would be simulated again, using the same seed numbers as were used in the
simulations for the status quo condition. In Paramics, seed number specifies the random number
generation and the vehicle release sequence into the micro-simulation networks. Thus by using the
same seed numbers, the networks would service the same arrival pattern of vehicles released into
the networks for both before and after the infrastructure modification. This eliminated the
systematic randomness in the comparison process and therefore allowed for a more fair
comparison. Using the established CPMs relationship determined from model validation, the
predicted crashes were modelled for both the hypothetical scenarios and their corresponding status
quo condition. Then, the two predicted crashes were compared to examine the relative safety
performance.
34
4 Pilot Study and Data Description
4.1 Pilot Study
The intersection of King Street E and Church Street was used as a template for the micro-
simulation network of this pilot study. The micro-simulation network resembled the actual
physical geometry and all other characteristics, such as traffic volume, pedestrian volume, and
transit operation. Nevertheless, it is worth mentioning again that the objective of the pilot study
was to test the capability of the micro-simulation software, so that subsequent efforts in model
construction, calibration, and validation would be more efficient. A screenshot of this micro-
simulation network is illustrated by Figure 4-1 below:
Figure 4-1 Screenshot of the Pilot Study Network
35
From the pilot study, it was found that:
1) Paramics is capable of incorporating the transit operation element, including adjustable
stop location, dwell time, service headway, median-lane streetcars, and TSP. TSPs such as
green extension, red truncation, or transit pre-emption are not included in the basic set of
Paramics features and require additional user-written programs to override existing signal
operation. Also the re-allocation of unused green-time, which is a common signal operation
scheme in the city of Toronto, can also be simulated in Paramics using user-written
programs. Paramics also has the capability to model detailed pedestrian movement and
pedestrian-vehicle interaction with its Urban Analytics Framework (UAF).
2) SSAM, by default, has the ability to categorize conflicts by their impact type, including
rear-end, angle, or side-swipe. SSAM is also capable of identifying a conflict by its vehicle
type, thus allowing the categorization of transit-related conflicts. However the
identification of vehicle type is not an inherently available function and must be done
through manually filtering the detailed conflict output generated by SSAM. One critical
finding is that Paramics does not generate the trajectories of pedestrians. As a result, neither
pedestrian-pedestrian nor pedestrian-vehicle conflicts can be identified by SSAM. Despite
the inability to recognize pedestrian trajectory, the inclusion of pedestrian does influence
the trajectory of general vehicles due to yielding rules and additional pedestrian demand
on actuated green intervals. To the knowledge of the author, the only micro-simulation
program capable of generating pedestrian trajectory is VISSIM (U.S. Department of
Transportation, 2011). However, as SSAM was initially developed for the analysis of
vehicle-vehicle conflicts, the credibility this technique for analyzing pedestrian-vehicle and
pedestrian-pedestrian conflicts remains questionable.
36
3) Excluding the pedestrian element, Paramics takes approximately 5 minutes to simulate the
pilot study’s network 10 repeated times. Then, SSAM takes approximately another 5
minutes to analyze the trajectory files. Thus if 100 networks of isolated intersections, each
having 10 runs are to be simulated, it would take approximately 16.7 hours. If the
pedestrian element is to be incorporated into the simulation model, the simulation time
would be significantly longer by approximately 3 times. The analysis time by SSAM is
unaffected by the inclusion of pedestrians since pedestrian trajectory is not generated by
Paramics.
4) Traffic volume, vehicle composition, intersection geometry, and signal timing plan data,
are essential data for constructing the simulation networks. Additionally, transit schedules
are also needed as transit arrivals to an intersection affect upstream traffic and in some
cases override the existing signal operation by signal priority. In addition, data regarding
dwell times, stop location, and detailed override scheme of the TSP should also be acquired
for the most realistic representation of transit operation.
From the above findings, it was determined that 100 intersections should be constructed in total.
The chosen sample size of 100 was based on securing a sufficiently large sample size for
developing CPM and conducting scenario tests, while maintaining a reasonable simulation and
analysis time.
The pedestrian element was determined to be omitted from the micro-simulation model. The
inclusion of pedestrians would have incurred significantly longer time in both model construction
and especially in simulation. Also as was mentioned earlier, pedestrian-pedestrian and pedestrian-
vehicle conflicts could not be identified from Paramics simulations; so the value added of
introducing pedestrians would only be their presence to influence yielding rules and additional
37
pedestrian demands on green time at actuated phases. Most importantly, calibration of pedestrian
movements would have been very different compared to that of vehicle movements. It would also
require more data and significantly more efforts. Without careful calibration of pedestrian
behaviours, the inclusion of pedestrians might not necessarily make the micro-simulation model
more realistic. Therefore, even though pedestrians would have an influence on general traffic, they
were not implemented in the micro-simulation models.
4.2 Data
From the pilot study, all necessary pieces of data for the construction, calibration, and validation
of the micro-simulation models were identified. These include (1) Crash data, (2) Traffic volume
data, (3) Transit service data, (4) Signal Timing Plans, (5) TSP operation data if applicable, and
(6) Intersection Geometry data.
4.2.1 Study Area and Period
From the pilot study, 100 micro-simulation networks were determined to be built and studied. The
City of Toronto has a total of approximately 1970 signalized intersections, of which 647 have both
crossing streets being arterial roads. These 647 arterial intersections are this study’s population of
interest. So 100 signalized intersections were drawn randomly from this population. The random
selection process avoided biasness and ensured representativeness of the population. A detailed
list of the selected 100 intersections can be found in Appendix A. The time period of analysis was
chosen to be between 2006 and 2010, inclusively. A GIS map is provided below (Figure 4-2) to
visually depict the geographical location and the observed crash frequency for each of the 100
selected signalized intersections.
38
Figure 4-2 Observed Number of Crashes [2006 – 2010] at the Signalized Intersections of the
Selected Sample
4.2.2 Crash Records
The crash data for the 100 intersections were extracted from the crash records administered by the
City of Toronto. The data were by default recorded in person-based format, such that each
observation represented one victim involved in a crash. For each victim involved in a crash, the
dataset included a number of descriptive factors such as time, date, impact type, vehicle type,
severity, individual’s condition etc. Data handling was needed to convert the data into crash-based,
such as each observation represents one crash. For each crash, the descriptive factors of every
victims involved were also attached to that crash. This allowed a crash to be identified by its impact
types or whether it involved certain vehicle types. Among the sample size of 100 selected
intersections, the observed crash frequency ranged from 21 to 368 crashes within the 5 years of
39
analysis window. Figure 4-3 below illustrates the observed crash frequency within the sample,
along with sample average and population average. Table 4-1 presents the statistics of the crash
frequency and peak hour volume for the selected sample.
Figure 4-3 Observed Crash Frequency [from 2006 to 2010] for the Selected Sample
Table 4-1 Descriptive Statistic for the Selected Sample
Number of Intersections Peak Hour Volume Crash Frequency
Min Max Mean Min Max Mean Total
Full Sample [100] 1707 7991 3520 21 368 113 11263
With Bus Service [87] 1716 7991 3647 21 368 117 10219
With Streetcar Service [20] 1707 3875 2514 33 163 89 1778
With Both Bus and Streetcar Services [16]
1716 3875 2491 33 163 88 1319
With Near-Sided Stops [95] 1707 7991 3567 21 368 113 10593
With Far-Sided Stops [43] 1777 6438 3751 28 295 127 5476
With Green Extension [20] 1707 5202 3006 39 234 98 1961
With Green Extension and Red Truncation [10]
1707 3675 2382 39 133 75 754
0
50
100
150
200
250
300
350
400
1 11 21 31 41 51 61 71 81 91
Observed Crash Frequency [from 2006 to 2010] for the Selected Sample
Observed Crash Frequency at each intersection Sample Average Population Average
40
4.2.3 Traffic Volume and Transit Service Information
Traffic volume data for the selected 100 intersections were also obtained from the City of Toronto.
For each intersection, the date for when the traffic volume data were originally collected had to
range within 2006 and 2010 in order to match the study period. If multiple traffic counts had been
done within the study period for an intersection, the earlier one was used for consistency. As a
result, a large number of the selected intersections had their traffic volume data from 2006.
Transit service data were obtained from the Toronto Transit Commission. The transit service data
provides many information needed for the simulation, such as transit routes and service headway.
The 2005.10.15 copy of the transit service data was used as this represented the transit service
condition at the beginning of the study period. This was to correspond with the dates of the traffic
volumes, as mentioned earlier that a large number of the traffic volume counts was from 2006.
One piece of transit operation information that was not in the transit service data was the dwell
time at each transit stop. Dwell times are closely tied to the boarding and alighting profiles at each
particular transit stops. Due to the absence of this information, the dwell time at each transit stop
was assumed to be at the Paramics’ default value.
4.2.4 Signal Timing Plans and Intersection Geometry
Signal timing plans for the selected 100 intersections were obtained also from the City of Toronto.
Each traffic signal’s traffic controller system could be one of TranSuite, MTSS, or SCOOT. Thus
the format of the signal timing plans was not uniform and could vary, depending on their traffic
controller system. Each signal timing plan provided detailed signal operation schematics including
phases, movement priorities, green/amber/red times of each phase, gap extension, etc. Additional,
if an intersection had TSP, the TSP operation schematic was also provided in the signal timing
plan. The TSP operation schematic included all essential parameters needed to program the TSP
41
including the type of TSP (green extension or red truncation or both), maximum
extension/truncation time, incremental time, decision time, etc. If multiple signal timing plans
existed within the study period, the earlier version was used for consistency with other data.
Lastly the intersection geometry was obtained from the City of Toronto’s web tool “Interactive
Toronto Map” (City of Toronto, 2015). This tool allowed for an accurate distance measurement
from the map’s aerial image, which had very high image resolution and clarity. The high image
quality also allowed for the confirmation of the transit stops with the transit service data. The tool
had many versions of aerial image taken from years 2005, 2009, 2011 and 2012. The version of
the aerial image used for this study was from the year 2005, again for the purpose of consistency
with other data.
42
5 Model Construction, Calibration and Validation
5.1 Model Construction
For each of the 100 randomly selected intersections under investigation, an isolated micro-
simulation network was constructed. Within each micro-simulation network, the intersection was
modelled to the incorporate all infrastructure elements for which data were available. The
geometric layout for each intersection was first constructed in the micro-simulation models in
accordance with the measurements obtained from the 2005’s aerial images. Then, vehicle and
transit demands were created, respectively, in accordance with the vehicle volume data and the
transit service data. Also in this step, the proportion of heavy vehicles was adjusted to match the
observed proportion in the traffic volume data. The operating characteristics of transit units,
including the length, width, maximum speed, maximum acceleration, and maximum deceleration,
were also adjusted in the micro-simulation models in accordance with their actual performance.
Lastly, the traffic signal in each micro-simulation models was signalized in accordance with the
signal timing plans. In the cases where actuated signal operation, variable signal operation, or TSP
were present in the signal timing plans, additional control files were written manually to constitute
their corresponding signal operations in the micro-simulation. Figure 5-1, Figure 5-2, and Figure
5-3 below illustrate three demonstrations of the constructed micro-simulation networks.
43
Figure 5-1 Demonstration Network 1 - Lake Shore Blvd W and Marine Parade Dr
Figure 5-2 Demonstration Network 2 - Finch Ave E and Tapscott Rd
44
Figure 5-3 Demonstration Network 3 - Lake Shore Blvd E and Lower Jarvis St
Despite the efforts of using the available data exhaustively to construct the micro-simulation
models to the best replication of reality, two pieces of information were unavailable and required
assumptions to be made. The first one was the unavailability of transit dwell time at transit stops.
As a result, the default dwell time in Paramics was used, which assumed an arrival and alighting
rate of 12 passengers per hour and 2 seconds of boarding/alighting per passenger for every transit
stops. So for example if a transit service had a frequency of two transits per hour, each transit
would have 6 boarding passengers, 6 alighting passengers, and 12 seconds of dwell time (as
Paramics by default assumed simultaneous boarding and alighting). This assumption was arguably
conservative for locations with denser transit rider profile. The second assumption was the
omission of on-street parking. This omission was not due to the unavailability of parking regulation
but rather insufficient information regarding how on-street parking was used. Fortunately, on-
street parking was uncommon in the studied intersections; even if it was available, in most cases
it was prohibited during the studied PM peak hour. Nonetheless, in a few of the studied networks,
on-street parking was available and unrestricted, which lead to one source of inaccuracy.
45
5.2 Model Calibration
As was introduced in Chapter 3, the objective of the model calibration was to iteratively seek an
optimal set of parameters within Paramics and SSAM such that the resulting CPM has the best
goodness-of-fit (i.e. most representative of the reality). However, as both simulation by Paramics
and analysis by SSAM were very time-consuming processes, only 40 networks, randomly chosen
from the 100 networks, were simulated for this purpose. Given that the sampling method of 40
networks was by random selection, the resulting goodness-of-fit was expected to be representative
of the 100 networks of population. In every set of tested parameters, 10 simulation runs, each with
a distinct seed number from 100 to 109, were conducted. Only 10 simulation runs were chosen,
again because of the time-consuming nature of the simulation process. Theoretically, more
simulations runs would have provided more precision on the statistical average of the tested
parameters. This chapter discusses the resulting performance of each set of tested parameters and
the set of parameters selected for model validation.
5.2.1 Driver Reaction Time (DRT) and Mean Headway Time (MHT)
The default value of both DRT and MHT are 1.0 seconds in Paramics. However it was suggested
that MHT is set to 0.85 to 0.90 for urban areas (Department of Transportation Wisconsin, 2012).
10 different sets of DRT and MHT are tested. The first 9 sets were exhaustive combinations of
DRT values of 0.71, 0.55, and 0.40 and MHT values of 1.00, 0.86, and 0.50. The 10th set tested an
extremely aggressive scenario, in which the DRT and MHT were at low values of 0.40 and 0.30
respectively. The tested DRT and MHT values were below the default values of 1.0 for the reason
that vehicle generally behave more aggressively near intersections. The coverage area of analysis
was chosen to be 80 meters upstream of the intersection in SSAM and other parameters were kept
46
at their default values. The resulting goodness-of-fits for each set of tested parameters are
presented below in Table 5-1.
Table 5-1 Calibration Results of Parameters DRT and MHT
Testing Number 1 2 3 4 5
Tested
Paramics
Parameters
DRT 0.710 0.550 0.400 0.710 0.550
MHT 0.860 0.860 0.860 0.500 0.500
CPM
Parameter
Estimate
λ 1.3902 1.4438 1.4425 1.2928 1.3657
Significance of λ 0.0283 0.0198 0.0200 0.0352 0.0282
β 0.5450 0.5404 0.5447 0.5463 0.5400
Significance of β 0.0001 0.0001 0.0001 0.0001 0.0001
Dispersion
parameter 0.1498 0.1488 0.1487 0.1416 0.1455
CPM
Goodness-
of-fit
Scaled Deviance 41.0241 41.0201 41.0255 41.0127 41.0210
SD/(n-p) 1.0796 1.0795 1.0796 1.0793 1.0795
Pearson χ2 42.7675 43.5894 44.0964 42.9209 43.7166
Pearson χ2/(n-p) 1.1255 1.1471 1.1604 1.1295 1.1504
Miaou’s R2 0.4025 0.4065 0.4069 0.4352 0.4196
Testing Number 6 7 8 9 10
Tested
Paramics
Parameters
DRT 0.400 0.710 0.550 0.400 0.400
MHT 0.500 1.000 1.000 1.000 0.300
CPM
Parameter
Estimate
λ 1.3724 1.5159 1.4225 1.5525 1.4827
Significance of λ 0.0293 0.0154 0.0224 0.0107 0.0194
β 0.5441 0.5314 0.5508 0.5322 0.5198
Significance of β 0.0001 0.0001 0.0001 0.0001 0.0001
Dispersion
parameter 0.1468 0.1533 0.1491 0.1507 0.1512
CPM
Goodness-
of-fit
Scaled Deviance 41.0096 41.0282 40.9994 41.0307 41.0251
SD/(n-p) 1.0792 1.0797 1.0789 1.0798 1.0796
Pearson χ2 44.6518 42.8810 43.4974 44.4789 43.2664
Pearson χ2/(n-p) 1.1750 1.1284 1.1447 1.1705 1.1386
Miaou’s R2 0.4144 0.3885 0.4053 0.3989 0.3969
GLM Form: 𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]𝛽 ∗ 𝑒λ
47
In all 10 sets of tested parameters, the results were relatively similar. The estimate for the CPM
parameters β was significant at 95% confidence level for all 10 tests. In terms of goodness-of-fit,
the SD ranged from 40.9994 to 44.0964. Given that the degree of freedom for this CPM was 38,
the SD divided by the degree of freedom ranged from 1.0789 to 1.0798. The Pearson χ2 ranged
from 42.8810 to 44.6518. The Pearson χ2 divided by the degree of freedom ranged from 1.1255 to
1.1750. Both the SD/(n-p) and the Pearson χ2/(n-p) were close to 1.0 and thus the dispersion effect
was acceptable. The critical value of the χ2 distribution is 53.3835 for a degree of freedom of 38 at
the 95% confidence level. This critical value was not exceeded by any of the SD or Pearson χ2
from the 10 tested sets; therefore this result indicated that for every set of tested parameters, the
CPM was not rejected at 95% level of confidence.
The above calibration attempt suggested that all 10 sets of DRT and MHT were statistically
defensible for developing CPMs. In selecting the optimal set of DRT and MHT, the Miaou’s R2
was compared in addition to the SD and Pearson’s χ2. The baseline dispersion parameter, which
was obtained by having only one constant term as the CPM’s explanatory variable, was 0.2507.
The lowest dispersion parameter, and thus the highest Miaou’s R2 of 0.435, was observed in the
4th test. As a result, the 4th set of parameters with {DRT = 0.710 and MHT =0.500} was selected
as the most optimal amongst the 10 sets. Hence, this combination of DRT and MHT was used to
conduct subsequent calibration works for other model parameters. In addition, the parameters from
the first set {DRT = 0.710 and MHT =0.860} resulted in the lowest Pearson χ2, and were also kept
to conduct subsequent calibration works.
5.2.2 Coverage Area of Analysis, Time to Collision, and Travelling Speed
Since the size of each network being micro-simulated was approximately 120~150m upstream of
every approach, the absolute analysis area could only be as much upstream as this value.
48
Additionally approximately 20 meters at the start of every approaching link was set as the vehicle
release zone and could not be included in the coverage area of the conflict analysis. As a result,
the effective coverage area of each micro-simulation network could at most be 100m upstream of
every approach. The effects of using three different coverage areas in SSAM, respectively 65m,
80m, and 95m horizontally and vertically afar from the centre of the network, were investigated.
Each coverage area is denoted as { - L, L}, where L represents the distance from the centre of the
networks to each of the four directions.
Another tested parameter was the simultaneous use of the TTC and speed thresholds. As was
introduced in the Methodology section, these thresholds were applied simultaneously to filter out
the questionable or impractical conflicts. In SSAM, MaxS is a filtering parameter that defines the
greater of the two travelling speed of the two conflicting vehicles. Thus by setting a minimum
MaxS this ensured both conflicting vehicles would be travelling above this specified threshold.
This parameter was binary and so the minimum TTC and speed thresholds were either “used” or
“not used”. If the thresholds were “used”, any conflict that had a TTC of less than 0.05 seconds
and a MaxS of less than 16.1 km/hr would be filtered out by SSAM.
In tests 11 to 20, different combinations of the coverage area and the TTC and MaxS thresholds
were tested. The first five tests from 11 to 15 were based on the Paramics setting of DRT = 0.710
and MHT = 0.500 (previously test number 4). The next five tests from 16 to 20 were based on the
simulated result from the default Paramics setting of DRT = 0.710 and MHT = 0.860 (previously
test number 1). The resulting goodness-of-fit for these 10 new tests are shown in Table 5-2 below:
Table 5-2 Calibration Results of Parameters Coverage Area, TTC, and Speed
Testing Number 11 12 13 14 15
Tested
SSAM
Parameters
Coverage Area {-65,65} {-95,95} {-65,65} {-80,80} {-95,95}
TTC and MaxS
Thresholds Filter None None Used Used Used
49
CPM
Parameter
Estimate
λ 1.0600 1.4691 1.9908 2.0793 2.1660
Significance of λ 0.1238 0.0102 0.0005 0.0001 0.0001
β 0.5988 0.5099 0.5357 0.5034 0.4515
Significance of β 0.0001 0.0001 0.0001 0.0001 0.0001
Dispersion
parameter 0.1479 0.1392 0.1609 0.1560 0.1516
CPM
Goodness-
of-fit
Scaled Deviance 41.0301 41.0067 41.2265 41.2139 41.1967
SD/(n-p) 1.0797 1.0791 1.0849 1.0846 1.0841
Pearson χ2 43.7792 42.6048 42.0819 41.2839 40.8568
Pearson χ2/(n-p) 1.1521 1.1212 1.1074 1.0864 1.0752
Miaou’s R2 0.4101 0.4448 0.3582 0.3777 0.3953
Testing Number 16 17 18 19 20
Tested
SSAM
Parameters
Coverage Area {-65,65} {-95,95} {-65,65} {-80,80} {-95,95}
TTC and MaxS
Thresholds Filter None None Used Used Used
CPM
Parameter
Estimate
λ 1.0327 1.5678 2.0064 2.1202 2.1690
Significance of λ 0.1531 0.0074 0.0007 0.0001 0.0001
β 0.6218 0.5059 0.5530 0.5145 0.4958
Significance of β 0.0001 0.0001 0.0001 0.0001 0.0001
Dispersion
parameter 0.1537 0.1468 0.1656 0.1616 0.1580
CPM
Goodness-
of-fit
Scaled Deviance 41.0381 41.0113 41.2622 41.2335 41.2133
SD/(n-p) 1.0800 1.0792 1.0858 1.0851 1.0846
Pearson χ2 43.7692 42.4641 42.5535 41.3501 40.7824
Pearson χ2/(n-p) 1.1518 1.1175 1.1198 1.0882 1.0732
Miaou’s R2 0.3869 0.4144 0.3394 0.3554 0.3698
GLM Form: 𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]𝛽 ∗ 𝑒λ
Again, for all calibration tests, the estimates for the CPM parameters β were significant at 95%
confidence level, suggesting strong correlation between simulated conflicts and observed crashes.
The goodness-of-fit performance had more variation among the 10 tests. The scaled deviance
ranged from 40.0067 to 41.2622. The scaled deviance divided by the degree of freedom ranged
from 1.0791 to 1.0858. The Pearson χ2 ranged from 40.7824 to 43.7792. The Pearson χ2 divided
by the degree of freedom ranged from 1.0732 to 1.1521. Both the SD/(n-p) and the Pearson χ2/(n-
50
p) were reasonably close to 1.0 and thus the dispersion effect was acceptable. The critical χ2 value
of 53.3835 was again not exceeded by any of the Pearson χ2 from the 10 tested sets; therefore this
result indicated that for every set of tested parameters, the CPM was not rejected at 95% level of
confidence.
One interesting finding was that the introduction of the TTC and MaxS simultaneous filtering did
not seem to improve the overall goodness-of-fit of the CPM. In all tests in which this filtering
criterion was applied, the resulting CPMs had higher dispersion parameter. Thus this filter was not
applied for any subsequent calibration and validation works. Another interesting finding was that
the upstream distance of 95m seemed to result in the lowest dispersion parameter; this was intuitive
since this enclosed most of the study areas that were initially constructed. Ultimately, the
parameter set from test 12 was chosen as the most representative of the reality, as this set had the
lowest dispersion parameter and the highest Miaou’s R2. The use of this parameter set led to the
final calibration effort, which was to test the significance of the inclusion of PHR.
5.2.3 Peak Hour Ratio (PHR)
With the inclusion of the PHR, which was the proportion of the peak hourly volume to the daily
volume, the CPM form would consist two explanatory variables. This resulted in one additional
CPM model parameter and therefore a reduction in the degree of freedom by one. The parameter
estimates and the goodness-of-fit measures are presented in Table 5-3 below:
Table 5-3 Calibration Results for the Inclusion of PHR
Testing Number 21
CPM
Parameter
Estimates
λ 0.1531
Significance of λ 0.9172
β 1 0.5486
Significance of β 1 0.0001
β2 0.6034
Significance of β2 0.3539
51
Dispersion parameter 0.1448
CPM
Goodness-
of-fit
Measures
Scaled Deviance 41.0058
SD/(n-p) 1.1083 Pearson χ2 43.3159
Pearson χ2/(n-p) 1.1707
Miaou’s R2 0.422
GLM Form: 𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]𝛽1 ∗ [𝑃𝐻𝑅]𝛽2 ∗ 𝑒λ
Interestingly, with an estimated p-value of only 0.3539, the coefficient of the explanatory variable
PHR showed no statistical significance. The poor statistical strength of the PHR indicated that the
introduction of this explanatory variable did not improve the CPM, despite having a comparable
goodness-of-fit with the earlier CPMs. As a result, this estimated CPM form was inferior to the
CPM form from test 12. The final parameter set was therefore chosen as the ones used in test 12,
with DRT = 0.710, MHT = 0.500, and Coverage Area = {-95, 95}. All other parameters were kept
at their default settings in both Paramics and SSAM.
5.3 Model Validation
Upon calibrating the model and finding the parameter set that produced the optimal replication of
reality, model validation was executed. Here, all 100 networks were simulated in Paramics and
analyzed in SSAM using the parameter set determined from model calibration. The objective of
model validation was to confirm that the optimal parameter set determined from model calibration
remains suitable and statistically defensible when applied to the full study size. Again, each
network was simulated 10 times with seeds number from 100 to 109. In addition to the validation
of CPM of simulated conflicts with observed crashes of all impact types, CPMs were also
developed based on impacts type. Namely, CPMs for the impact types of angle, rear-end, and side-
swipe were also developed and validated. Additionally, simulated conflicts that involved transit-
vehicles were regressed against the observed crashes that involved transit vehicles. Again, the
52
filtering of transit-involved conflicts was not an inherent feature and was achieved manually by
studying the analyzed conflicts in Excel. The simulated angle, rear-end, and side-swipe, and
transit-involved conflicts were respectively validated against their crash counterparts, i.e. angle,
rear-end, side-swipe, and transit-involved crashes. The result of the validation is shown by Table
5-4 below:
Table 5-4 Model Validation Results
Model Parameters
All
Crash
Types
Angle
Crashes
Rear-
End
Crashes
Side-
Swipe
Crashes
Transit-
Involved
Crashes
Parameter
Estimate
λ 1.1789 1.0503 -0.5010 0.5765 -0.2476
Significance of λ 0.0031 0.0009 0.2310 0.1071 0.5075
β 0.5722 0.6104 0.7175 0.5647 0.5482
Significance of β 0.0001 0.0001 0.0001 0.0001 0.0001
Dispersion parameter 0.1931 0.1874 0.2660 0.3338 0.4950
Goodness-
of-Fit
Scaled Deviance 103.2580 102.3342 104.6372 103.6942 106.4388
SD/(n-p) 1.0537 1.0442 1.0677 1.0581 1.1323
Pearson χ2 109.2039 106.9563 122.5254 119.5977 103.4132
Pearson χ2 / (n-p) 1.1143 1.0914 1.2503 1.2204 1.1001
Miaou’s R2 0.4285 0.4144 0.5015 0.3014 0.2908
GLM Form: 𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]𝛽 ∗ 𝑒λ
As depicted by Table 5-4, the parameter estimates for each impact type differed among each other.
For the CPM for all crashes, both λ and β were statistically significant. It is worth noting that the
value of λ and β were similar to those estimated from model calibration. The SD and Pearson’s χ2
statistics were respectively 103.2580 and 109.2039, which did not exceed the critical χ2 value of
122.1077 for 98 degree of freedoms. The Miaou’s R2 was 0.4285, indicating a slight reduction
relative to the Miaou’s R2 computed from model calibration.
53
The CPMs for angle, rear-end, and side-swipe impact types differed slightly with the CPM for all
crashes. However, in all three impact-type-specific CPMs, their respective β were statistically
significant, indicating that simulated conflicts were strongly correlated with crashes. It is also
worth noting that the Pearson χ2 for rear-end and side-swipe models were relatively high and
possibly indicated over-dispersion. Nevertheless since their SDs were well within the acceptable
range, the CPMs were still used for subsequent analyses despite having potential over-dispersion.
When comparing the Miaou’s R2, rear-end GLM had the highest performance, indicating the
highest degree of variance explained within the model. In contrast, side-swipe GLM had the lowest
degree of variance explained within the model, which implied a less predictive nature of side-
swipe crashes using micro-simulation.
Also as shown by Table 5-4, the CPM for transit-involved crashes was adequately fitted. The
estimate for the β parameters was similar to that in the all-crash-types model. In the transit-
involved CPM, the SD and Pearson χ2 were also within the acceptable threshold of the critical χ2
value. When these two goodness-of-fit measures were scaled by the degree of freedom, the
resulting SD/(n-p) and Pearson χ2/(n-p) aligned closely to 1.0, indicating appropriate dispersion.
The Miaou’s R2 was however the lowest among all the CPMs presented in this section. This
indicated that a great proportion of the variance in the observed transit-involved crashes was yet
unexplained by the CPM, and furthermore implied that transit-involved crashes could be more
difficult to predict relative to general vehicle crashes. This difficulty in prediction may be due to
the significantly lower values in both the simulated transit-involved conflicts and observed transit-
involved crashes, when compared to their general vehicle counterparts. Nevertheless, given that
the coefficient estimate was strongly statistically significant and that both the SD and Pearson χ2
54
goodness-of-fit measures were acceptable, the transit-involved CPM was used for subsequent
analyses.
In addition to the three goodness-of-fit measures presented above, the fourth measure, the CURE
analysis was also conducted. In the CURE analysis, the cumulative scaled residuals for each of the
five CPMs presented earlier were plotted. The two standard deviation upper and lower bands of
each CPM’s corresponding random walk phenomenon, were also plotted to assist the graphical
illustration of the appropriateness of the modelling structure. The CURE analysis for the five
CPMs of all-impact-types, angle, rear-end, side-swipe, and transit-involved crashes are shown by
Figure 5-4 through Figure 5-8 below:
Figure 5-4 CURE Plot - All Impact Type Crashes
-12.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
8.0
10.0
12.0
0 200 400 600 800 1000 1200 1400 1600
Cu
mu
lati
ve S
cale
d R
esid
ual
s
Simulated Conflicts - All Impact Types
CURE Plot - All Impact Type Crashes
Cumulative Scaled Residuals
Upper 2σ Band
Lower 2σ Band
55
Figure 5-5 CURE Plot - Angle Crashes
Figure 5-6 CURE Plot - Rear-End Crashes
-12.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
8.0
10.0
12.0
0 20 40 60 80 100 120 140 160 180 200
Cu
mu
lati
ve S
cale
d R
esid
ual
s
Simulated Conflicts - Angle
CURE Plot - Angle Crashes
Cumulative Scaled Residuals
Upper 2σ Band
Lower 2σ Band
-12.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
8.0
10.0
12.0
0 200 400 600 800 1000 1200 1400
Cu
mu
lati
ve S
cale
d R
esid
ual
s
Simulated Conflicts - Rear-End
CURE Plot - Rear-End Crashes
Cumulative Scaled Residuals
Upper 2σ Band
Lower 2σ Band
56
Figure 5-7 CURE Plot - Side-Swipe Crashes
Figure 5-8 CURE Plot - Transit-Involved Crashes
The CURE analysis revealed that the cumulative scaled residuals fell within the acceptable two
standard deviations envelope of the random walk phenomenon for all five CPMs. For the all-crash-
-12.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
8.0
10.0
12.0
0 20 40 60 80 100 120 140 160 180
Cu
mu
lati
ve S
cale
d R
esid
ual
s
Simulated Conflicts - Side-Swipe
CURE Plot - Side-Swipe Crashes
Cumulative Scaled Residuals
Upper 2σ Band
Lower 2σ Band
-12.0
-10.0
-8.0
-6.0
-4.0
-2.0
0.0
2.0
4.0
6.0
8.0
10.0
12.0
0.0 10.0 20.0 30.0 40.0 50.0 60.0 70.0 80.0
Cu
mu
lati
ve S
cale
d R
esid
ual
s
Simulated Conflicts - Transit-Invovled
CURE Plot - Transit-Involved Crashes
Cumulative Scaled Residuals
Upper 2σ Band
Lower 2σ Band
57
types, angle, rear-end, and side-swipe CPMs, their corresponding cumulative scaled residuals plot
generally oscillated about the x-axis. Such oscillating pattern was ideal as it suggested no
systematic overestimation or underestimation of the model structure. For the transit-involved
CPM, the cumulative scaled residuals plot appeared to have continuously negative slope when the
explanatory variable was low, followed by continuously positive slope when the explanatory
variable was high. This upward concaving pattern suggested that the CPM may systematically
overestimate crashes when the simulated conflicts were low and conversely may underestimate
crashes when the simulated conflicts were high. This pattern also demonstrated that transit-
involved crashes could be more difficult to predict than general traffic crashes.
Based on the above model validation results, all five CPMs were considered acceptable to be used
for the purpose of crash prediction. This was based on the reasons that (1) the coefficients for the
simulated conflicts were highly statistically significant, and (2) the four goodness-of-fit measures,
respectively the SD, Pearson χ2, Miaou’s R2, and CURE analysis demonstrated appropriate fitting
for each of the CPMs. The equations for the five studied CPMs are therefore:
𝐴𝑙𝑙𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝐴𝑙𝑙𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]0.5722 ∗ 𝑒1.1789
Equation 9
𝐴𝑛𝑔𝑙𝑒𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝐴𝑛𝑔𝑙𝑒𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]0.6104 ∗ 𝑒1.0503
Equation 10
𝑅𝑒𝑎𝑟𝐸𝑛𝑑𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝑅𝑒𝑎𝑟𝐸𝑛𝑑𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]0.7175 ∗ 𝑒−0.5010
Equation 11
𝑆𝑖𝑑𝑒𝑆𝑤𝑖𝑝𝑒𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝑆𝑖𝑑𝑒𝑆𝑤𝑖𝑝𝑒𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]0.5647 ∗ 𝑒0.5765
Equation 12
𝑇𝑟𝑎𝑛𝑠𝑖𝑡 − 𝑖𝑛𝑣𝑜𝑙𝑣𝑒𝑑𝐶𝑟𝑎𝑠ℎ𝑒𝑠(𝑃𝑒𝑟5𝑌𝑒𝑎𝑟𝑠) = [𝑇𝑟𝑎𝑛𝑠𝑖𝑡 − 𝑖𝑛𝑣𝑜𝑙𝑣𝑒𝑑𝐶𝑜𝑛𝑓𝑙𝑖𝑐𝑡𝑠(𝐻𝑜𝑢𝑟𝑙𝑦)]0.5482 ∗ 𝑒−0.2476
Equation 13
58
As a result of the model validation efforts, Equations 9 to 13 were used as the basis for assessing
safety performance for the following scenario tests.
59
6 Scenario Tests
Following model validation, the established SPFs were applied to assess the safety impacts of
various hypothetical transit infrastructure modifications. As was introduced in the Methodology
section, nine scenarios were designed, with each scenario testing the effect of one individual
infrastructure change. This chapter discusses the results of the scenario tests and the implications
from these findings. For each scenario, the total predicted crashes in the status quo condition, the
absolute change in total predicted crashes relative to status quo, and the percentage change in total
predicted crashes relative to status quo, are presented. Since a reduction in predicted crashes is in
the more favourable direction, a reduction in predicted crashes is coloured in green. Conversely
an increase in predicted crashes is coloured in red.
Each network was simulated 20 times, respectively having seed numbers from 100 to 119. The
networks from the status quo case were also simulated again, using the same 20 seed numbers.
Note that even though the same seed numbers were used for both before and after the infrastructure
modification, this practice did not discourage the number of simulation runs. In fact, the number
of runs was increased from 10, which was used in model calibration and validation, to 20 runs.
This increase in simulation runs was to allow for a more exhaustive representation of how vehicles
arrive at the intersection in reality, which in turn would make the comparison more justifiable.
6.1 Effects of TSP
Out of the 100 constructed micro-simulation networks, 20 intersections were serviced with TSP.
All 20 of these TSP-serviced intersections had green extension while only 10 had red truncation.
Due to the relatively small subgroup size, all qualified intersections were selected for investigation.
The results of the three scenarios for investigating the effects of TSP are presented in the Table
6-1 below:
60
Table 6-1 Predicted Results of TSP Treatments
Scenarios 1 2 3
Transit Infrastructure Treatments
Removal
of Green
Extension
Removal
of Red
Truncation
Removal
of TSP
Number of Networks Studied 20 10 20
Total Predicted Crashes in Status Quo
All Impact
Types
2002.3 885.0 2002.3
Change in Total Predicted Crashes
relative to Status Quo -22.75 -14.11 -23.93
Percentage Change in Total Predicted
Crashes relative to Status Quo -1.14% -1.59% -1.20%
Total Predicted Crashes in Status Quo
Angle
Impact
Types
753.1 343.5 753.1
Change in Total Predicted Crashes
relative to Status Quo 3.62 -9.77 -1.85
Percentage Change in Total Predicted
Crashes relative to Status Quo 0.48% -2.85% -0.25%
Total Predicted Crashes in Status Quo
Rear-End
Impact
Types
722.0 309.5 722.0
Change in Total Predicted Crashes
relative to Status Quo -13.62 -5.39 -13.76
Percentage Change in Total Predicted
Crashes relative to Status Quo -1.89% -1.74% -1.91%
Total Predicted Crashes in Status Quo Side-
Swipe
Impact
Types
292.5 124.5 292.5
Change in Total Predicted Crashes
relative to Status Quo -5.75 -2.65 -2.46
Percentage Change in Total Predicted
Crashes relative to Status Quo -1.97% -2.13% -0.84%
Total Predicted Crashes in Status Quo All
Transit-
Involved
Crashes
93.8 41.2 93.8
Change in Total Predicted Crashes
relative to Status Quo 0.79 -0.04 1.51
Percentage Change in Total Predicted
Crashes relative to Status Quo 0.84% -0.09% 1.61%
The resulting predicted crashes suggested that the status quo case, which had TSP operation, was
more inferior in terms of safety performance. In scenario 1, in which green extensions were
removed amongst all the studied micros-simulation networks, the overall predicted crashes were
decreased by 1.14%. Except for the angle type crashes which were increased by 0.48%, rear-end
61
and side-swipe crashes were reduced by 1.89% and 1.97% respectively. The predicted crashes that
involved transit vehicles were increased by 0.84%, which is expected given that the green
extension no longer facilitated transit movement in this scenario. In scenario 2, in which red
truncations were removed, the predicted crashes for all impact types were reduced. Surprisingly,
the transit-involved crashes were predicted to have a very minor reduction of 0.09%, suggesting
marginal improvement for transit safety. Thus this suggested that red truncation could be inferior
to green extension since it did not improve safety performance in any of the studied aspects. In
scenario 3, the predicted crash were predicted to decrease by 0.25%, 1.91%, and 0.84%
respectively for angle, rear-end, and side-swipe crash types; the total crashes were predicted to
decrease by 1.20%. However, the transit-involved crashes were predicted to increase by 1.61%,
which was similar to scenario 1. Again, this increase in transit-involved crashes was intuitive since
TSP is no longer facilitating transit movements in scenario 3.
The overall negative association between TSP and safety performance was also observed by
Shahla et al. (2009). However in that literature, a conventional CPM was developed without the
use of conflicts from micro-simulation, but with traffic volumes and other variables, including a
TSP indicator variable that suggested that the implementation of TSP technology could increase
total crashes by as much as 28.4% at transit-serviced intersections.
However, it is worth mentioning that, although the overall change in predicted crashes was
negative, the individual changes for every networks greatly differed amongst themselves (see
Appendix B for individual changes). In fact, some particular studied intersections were
demonstrated to have increased crash frequency had the TSP schemes been removed. This was
likely due to the fact that not all TSP schemes were exactly identical amongst the studied
intersections, although some resemble similar TSP parameters. One of the key TSP parameters
62
that usually differed amongst intersections was the maximum extension/truncation time. In
addition, the direction and the transit type for which the TSP was servicing, were also influential
towards the individual safety performance. Thus, this motivated the endeavour to investigate
further into the individual characteristics of each studied intersection and if any particular pattern
could be observed. The individual characteristics of these studied intersections are also provided
in Appendix B. Upon investigation, it appeared that most of the characteristics did not demonstrate
an observable pattern in relation to the change in predicted crashes. However it was observed that,
the transit type for which the TSP is servicing, has a noteworthy influence. Taking scenario 3 as a
demonstration, the micro-simulated networks appear to respond more negatively when the transit
routes serviced by the TSP are streetcars. More specifically, when a TSP that was servicing
streetcars is removed, there was more reduction in predicted crashes relative to when a TSP that
was servicing buses was removed. Figure 6-1 below depicts the composition of the change in
predicted crashes in scenario 3:
63
Figure 6-1 Absolute Change in Predicted Crash frequency as a Result of Removing TSP
[by Transit Types]
Figure 6-1 suggests the TSPs that service streetcars and buses had different safety impacts. For
convenience, the networks that had their TSPs servicing streetcars are referred to as the “TSP-
streetcar” networks in this chapter and likewise the networks that had their TSPs servicing buses
are referred to as the “TSP-bus” networks. From Figure 6-2, it can be seen that the majority of the
changes in predicted crashes were from the TSP-streetcar networks. This phenomenon is also
observable when each impact type was investigated individually. In fact, for angle and side-swipe
impact types, the predicted crashes were even predicted to increase when the TSPs that service
buses are removed.
Out of the 20 investigated networks which had a total of 2002 predicted crashes, 12 networks had
their TSPs service streetcars; these 12 networks contributed to a total of 1021 predicted crashes,
which were approximately half of the total predicted crashes among the 20 networks. Using the
-30
-25
-20
-15
-10
-5
0
5
All Crashes Angle Rear-End Side-Swipe Transit-Involved
Change in Predicted Crash Frequency as a result of Removing TSP[by Transit Types]
Networks in which TSPs Service Streetcars Networks in which TSPs Service Buses
64
absolute changes shown in Figure 6-1 and the total predicted crashes of the respective two factions
of networks, the percentage changes are presented in Figure 6-2 below:
Figure 6-2 Percentage Change in Predicted Crash Frequency as a Result of Removing TSP
[by Transit Types]
Figure 6-2 confirms that the total predicted crashes were experiencing more reduction for the TSP-
streetcar networks relative to the TSP-bus networks. Again, the breakdown by impact types
suggests that the changes in predicted crash for the TSP-streetcar networks were one-sided in the
negative direction; however for the TSP-bus networks, the change in predicted crash was more
volatile. Again, the result suggests that if the TSPs in the TSP-bus networks were removed, the
angle and side-swipe crashes were predicted to increase. The transit-involved crashes for both
TSP-streetcar and TSP-bus networks were expected to increase as a result of the TSP removal,
which is intuitive since the transits no longer benefited from the signal priority.
The above investigation suggests that existing TSPs are less compatible with streetcars than with
buses. The removal of TSPs that service streetcars would result in less crashes and thus in the safer
-4.0%
-3.0%
-2.0%
-1.0%
0.0%
1.0%
2.0%
3.0%
All Crashes Angle Rear-End Side-Swipe Transit-Involved
Perchange Change in Predicted Crash Frequency as a result of Removing TSP[by Transit Types]
Networks in which TSPs Service Streetcars Networks in which TSPs Service Buses
65
direction. However the removal of TSPs that service buses would lead to more mixed results. The
exact reason of why TSPs were less compatible with streetcars than with buses remains unclear.
One hypothesis is that since general traffics are not allowed to pass streetcars when the streetcars
are stopped, the benefits from TSP to general traffics in the streetcars’ direction is inherently
diminished.
To summarize the findings from this subchapter in which the safety impacts of TSP were studied,
the above investigation implies that:
The existing TSP, whether it was in the form of green extension individually, red
truncation individually, or combined green extension and red truncation, did not contribute
to a safer environment for general traffics, albeit contributing to safer performance for
transit vehicles. In fact, the predicted crashes were expect to be reduced by approximately
2% had the TSP schemes be removed.
Solely from the safety perspective, the existing TSP was less compatible when it serviced
streetcars than when it serviced buses. When TSPs that serviced streetcars were removed,
a reduction in predicted crashes are expected for every impact types. In contrast, when
TSPs that serviced buses were removed, the resulting changes in predicted crashes were
more volatile and situational. In short, the existing TSPs that service streetcars were less
desirable for the safety of general traffics than those that service buses.
6.2 Effect of Transit Stop Positioning
Most of the micro-simulation networks had at least one pair of transit stops. This was not surprising
since the studied population was arterial roadway intersections. As a result, a great majority of the
micro-simulation networks qualified for scenarios 4 and 6, which investigated transit stop
66
positioning along the major direction at intersections without TSP service. 80 out of the 100 micro-
simulation networks did not have TSP and 72 out of these 80 intersections had near-sided stops
along their major direction; then 35 intersections were randomly selected from these 72 qualified
intersections to be investigated for scenario 4 and 6. For the network selection in scenario 5, 39
out of the 80 non-TSP serviced intersections had near-sided stops along their minor direction and
25 intersections out of these 39 were randomly selected to be investigated. For these three
scenarios, the number of studied intersections was smaller than the number of actually qualified
intersections in order to sustain a reasonable simulation and analysis time. For scenario 7, only 15
out of the 20 TSP-servicing intersections had near-sided stops along its TSP direction; thus all 15
networks were investigated in this scenario. The results of the four scenarios that study the effect
of transit stop positioning were presented by Table 6-2 below:
Table 6-2 Predicted Results of Transit Stop Positioning Treatments
Scenarios 4 5 6 7
Transit Infrastructure Treatments
Near-sided
to Far-
sided
[Major
Direction
Non-TSP]
Near-sided
to Far-
sided
[Minor
Direction
Non-TSP]
Near-sided
to No Stop
[Major
Direction
Non-TSP]
Near-
sided to
Far-sided
[TSP
servicing
Direction]
Number of Networks Studied 35 25 35 15
Total Predicted Crashes in
Status Quo
All
Impact
Types
4664.7 3201.7 4664.7 2002.3
Change in Total Predicted
Crashes relative to Status
Quo
97.89 77.39 -92.11 0.32
Percentage Change in Total
Predicted Crashes relative to
Status Quo
2.10% 2.42 % -1.97% 0.02%
Total Predicted Crashes in
Status Quo Angle
Impact
Types
1524.1 1048.0 1524.1 753.1
Change in Total Predicted
Crashes relative to Status
Quo
71.72 54.37 -25.03 13.24
67
Percentage Change in Total
Predicted Crashes relative to
Status Quo
4.71% 5.19% -1.64% 2.59%
Total Predicted Crashes in
Status Quo
Rear-
End
Impact
Types
1883.7 1286.3 1883.7 722.0
Change in Total Predicted
Crashes relative to Status
Quo
35.77 32.16 -46.99 -4.82
Percentage Change in Total
Predicted Crashes relative to
Status Quo
1.90% 2.50% -2.49% -1.05%
Total Predicted Crashes in
Status Quo
Side-
Swipe
Impact
Types
656.4 447.0 656.4 292.5
Change in Total Predicted
Crashes relative to Status
Quo
18.81 10.59 -13.64 2.76
Percentage Change in Total
Predicted Crashes relative to
Status Quo
2.87% 2.37% -2.08% 1.43%
Total Predicted Crashes in
Status Quo
All
Transit-
Involved
Crashes
170.2 119.3 170.2 93.8
Change in Total Predicted
Crashes relative to Status
Quo
25.66 9.59 -30.92 -1.96
Percentage Change in Total
Predicted Crashes relative to
Status Quo
15.08% 8.03% -18.17% -3.07%
The resulting predicted crashes suggest that near-sided stops were inferior to far-sided stops in
terms of safety performance. In both scenario 4 and 5, the predicted crashes were higher than status
quo, suggesting that far-sided stops would be less safe regardless of whether the transit was along
the direction of heavier or lighter traffic. It can be seen that the results from scenario 4 and 5 are
similar, both suggesting approximately 2% increase in predicted crashes had the near-side stops
be relocated to far-sided. In addition, not only was the total predicted crashes expected to increase,
it was noticed that the individual changes in predicted crashes among the studied networks were
also one-sided in the increasing direction. 31 out of 35 studied networks in scenario 4 and 20 out
of 25 studied networks in scenario 5 suggested increase in predicted crashes. This trend was also
68
observable in each of the angle, rear-end, and side-swipe impact type studies. Furthermore, the
transit-involved crashes were also expected to increase, suggesting that far-sided stops were
neither safer for general traffics nor for transits, when compared to near-sided stops. This
worsening in safety performance was likely due to the possibility that when traffic queue forms
upstream of a stopped transit at a far-sided stop, the queue may overextend into the perimeter of
the intersection itself, which is inherently a more volatile region of traffic conflicts; however at
near-sided stops, when traffic queue forms upstream of a stopped transit, they would not risk
queuing into the intersection.
In scenario 6 in which transit stops were removed entirely along the major direction, the predicted
crashes unsurprisingly decreased by approximately 2%. Again, this trend is also observable when
dissected by impact-types. The transit-involved crashes experienced a greater reduction of 18.2%,
which was intuitive since transits flows would be steadier without the stops. It is worth mentioning
that this scenario only suggests that the total predicted crashes would decrease at intersection-level,
which may not represent the global effect. It is likely that midblock crashes could have increased
as a result of the relocation of transit stops from intersection to midblock. However, the analysis
of midblock crashes would be beyond the scope of this study.
The result from scenario 7 was interesting and differentiated itself with the earlier scenarios. It can
be seen that in this scenario, the change in predicted crashes was only an increase of 0.02%,
suggesting neither large improvement nor deterioration. This differed from the 2.10% and 2.42%
that were observed respectively from scenario 4 and 5. In addition, despite having only 0.02% total
increase in this scenario, the changes in predicted crashes at each individual studied networks were
actually more volatile. Some networks were shown to have decreased predicted crashes as a result
of the relocation of transit stops while others were shown to have increased predicted crashes.
69
After dissecting the predicted crashes by their impact types, the rear-end crashes were predicted to
decrease, while the angle and side-swipe crashes were predicted to increase. The transit-involved
crashes were interestingly displaying a decrease, implying that far-sided stops were slightly safer
for transits in the presence of TSP.
As a summary, the above investigation suggests that:
When near-sided stops were relocated to far-sided at intersections without TSP, the
resulting safety performance deteriorated, regardless of whether the transit line was in the
direction of heavier or lighter volume
When near-sided stops were relocated to mid-block, the resulting safety performance
improved at the intersection-level; however the global effect was unexplored
When near-sided stops were relocated to far-sided along the TSP-servicing direction, the
resulting safety performance was uncertain and volatile for each individual intersection;
however the overall safety performance was expected to be neither an improvement nor a
deterioration
6.3 Effect of Transit Type
19 of the 100 micro-simulated intersections had at least one transit route serviced by streetcars.
Then, 15 out of these 20 had their streetcars moving in the through direction and in mixed right-
of-way. Therefore all 15 networks were investigated in scenario 8 and 9, in which these streetcars
were replaced with buses. In scenario 8, the replacement ratio was 1:1 such that streetcars were
replaced by buses following the same service schedule. In scenario 9, the original service headway
of the streetcar was shortened by a factor of m and then used as the new service headway of the
bus, to ensure equal transit capacity relative to status quo. The factor m was obtained from the
70
simple inversion of the ratio of the streetcar’s peak hour passenger counts to the bus’s peak hour
passenger counts, which are published by the Toronto Transit Commissions (Toronto Transit
Commission, 2015). This factor m turned out to be 51/74, or 0.69. Thus if the original service
headway for the streetcar was for example 10 minutes, the new service headway for the
hypothetical buses would be 6 minutes 54 seconds. The results for this exercise is shown by Table
6-3 below:
Table 6-3 Predicted Results of Transit Type Treatments
Scenarios 8 9
Transit Infrastructure Treatment Streetcars to Buses
(1:1 replacement)
Streetcars to Buses
(1:m replacement for
same operation
capacity)
Number of Networks Studied 15 15
Total Predicted Crashes in
Status Quo
All Impact
Types
1135.5 1135.5
Change in Total Predicted
Crashes relative to Status
Quo
19.04 36.55
Percentage Change in Total
Predicted Crashes relative to
Status Quo
1.68% 3.22%
Total Predicted Crashes in
Status Quo
Angle
Impact
Types
432.7 432.7
Change in Total Predicted
Crashes relative to Status
Quo
10.97 15.66
Percentage Change in Total
Predicted Crashes relative to
Status Quo
2.54% 3.62%
Total Predicted Crashes in
Status Quo
Rear-End
Impact
Types
358.0 358.0
Change in Total Predicted
Crashes relative to Status
Quo
6.20 15.27
Percentage Change in Total
Predicted Crashes relative to
Status Quo
1.73% 4.26%
71
Total Predicted Crashes in
Status Quo
Side-
Swipe
Impact
Types
190.3 190.3
Change in Total Predicted
Crashes relative to Status
Quo
3.77 5.34
Percentage Change in Total
Predicted Crashes relative to
Status Quo
1.98% 2.80%
Total Predicted Crashes in
Status Quo
All
Transit-
Involved
Crashes
60.0 60.0
Change in Total Predicted
Crashes relative to Status
Quo
14.75 22.45
Percentage Change in Total
Predicted Crashes relative to
Status Quo
24.57% 37.39%
The above results illustrate that by replacing the studied streetcars with buses, the predicted crashes
actually increased. This trend is observable for each of the three impact types and is more obvious
for transit-involved crashes. The overall increase was approximately 1.7% in scenario 8 and 3.2%
in scenario 9. The transit-involved crashes were predicted to increase by as much as 24.6% and
37.4% respectively in scenario 8 and 9. This implies that along the existing streetcar routes,
through-moving streetcars would not be made more safe had them been replaced by through-
moving buses. This might be because the existing infrastructures along the streetcar routes, such
as left-turn restrictions or the absence of left-turn signal, had already made the intersections more
favourable for streetcar operation. Thus the replacement of existing streetcars to buses would not
bring forth safety benefits.
It is worth emphasizing that the above results are unidirectional and do not provide any conclusion
for the other direction of replacement, i.e. if buses were being replaced by streetcars. Buses can
operate in a more flexible environment where the intersections have left-turn lanes and left-turn
signals, which would first have to be removed to even allow for a reasonable streetcar operation.
72
In addition, the above investigation was only comparing through-moving streetcars with through-
moving buses. Although rare, in reality streetcars occasionally make left or right turns along their
routes of operation, whether they are turning at intersection or turning into their terminal. Thus the
safety aspect of the turning movements of streetcars and buses remains unexplored. As a result, if
a streetcar make many turns along its servicing route, the strength of the above finding would
indisputably be hindered.
73
7 Conclusion and Future Work
Through the combinative use of micro-simulation and CPM, the safety performance at arterial
intersections have been investigated in this study. The methodology presented in this study
suggests a flexible approach of modelling not only the intersection-level safety performance, but
also the transit operation aspect and its influence on general traffics. Within the presented
methodology, the data acquisition, model calibration, and model validation were the key steps
undertaken to ensure the technical validity of the following analyses. From the scenario tests, the
safety impacts of three facets of transit operations, namely the TSP, positioning of transit stop, and
transit type were investigated. The results clearly suggest that transit operations have a certain
level of influence towards the overall safety performance at intersection-level. More specifically,
the results imply that:
1. The existing TSPs do not improve the safety performance, especially when the TSPs are
servicing streetcar routes. By removing the TSPs, crash frequency is expected to be
reduced slightly.
2. Along a direction that is not serviced by TSP, the existing near-sided stops have better
safety performance than their corresponding hypothetically placed far-sided stops.
However along a direction that is serviced by TSP, the overall safety performance is
expected to be similar for near-sided and far-sides stops, while the individual performance
is more volatile.
3. The existing streetcars have better safety performance than if they were to be replaced
with buses, possibly due to the fact that existing lane usages are already favouring streetcar
operation.
74
Despite the practical implications of the above findings, there were limitations within the applied
methodology that may somewhat challenge the strength of the findings. The first limitation was
due to the absence of two pieces of data, which were the transit dwell time and the on-street parking
usage. This absence of these data led to the assumption of transit dwell time and the omission of
on-street parking during the construction of the micro-simulation model. Had these data been
available, the micro-simulation models could be constructed to incorporate these elements and
arguably be more realistic. Nonetheless, careful considerations must be given to ensure these data
would fit the study period and the duration of the micro-simulation.
The second limitation was the exclusion of the pedestrian element in the micro-simulation, due to
the inability to identify pedestrian-vehicle conflicts. If pedestrians had been introduced, the
yielding priorities for turning movement would be significantly more complex due to pedestrian
crossing. As a result, pedestrian movements will generally act as moving barricades to the flow of
traffic, which at the moment have unclear impacts in safety performance. The inclusion of the
pedestrian elements, or even more preferably the integration of pedestrians as transit riders, are
definitely a worthwhile direction for future researches.
The last limitation is the separation of the investigated crash frequency and the less discussed crash
severity. Again, in reality, the overall safety performance is conceptually an end product of a
mixture of crash frequency and crash severity. However in this study, only the facet of crash
frequency has been discussed, which inherently assumed that a fatal crash is indifferent from a
minor-injury crash. If crash frequency and crash severity can be simultaneously modelled, for
example in the form of crash frequency weighted by severity, the resulting safety implications
would indisputably have more practical values. Thus clearly this is also a worthwhile direction for
future researchers to pursue.
75
With more advanced micro-simulation packages and perhaps improved SPF modelling structures
to be introduced in the future, the field of traffic safety modelling will definitely attract more
researchers. It is the author’s hope that, this thesis will not only provide planners directions for
safer transit infrastructure designs, but also strengthen the foundation of safety modelling for future
researchers.
76
References
Archer, J. (2004). Methods for the Assessment and Prediction of Traffic Safety at Urban
Intersections and their Application in Micro-simulation Modelling. Royal Institute of
Technology.
Ariza, A. (2011). Validation of Road Safety Surrogate Measures as a Predictor of Crash
Frequency Rates on a Large-Scale Microsimulation Network. Master Thesis, University
of Toronto.
Autey, J., Sayed, T., & H.Zaki, M. (2012). Safety Evaluation of RIght-Turn Smart Chennels
Using Automated Traffic Conflict Analysis. Accident Analysis and Prevention, 45, 120-
130.
Caliendo, C., & Guida, M. (n.d.). Microsimulation Approach for Predicting Crashes at
Unsignalized Intersections Using Traffic Conflicts.
Caliendo, C., Guida, M., & Parisi, A. (2007). A Crash-Prediction Model for Multilane Roads.
Accident Analysis and Prevention, 39(4), 657-670.
City of Toronto. (2015). Toronto Maps. Retrieved from City of Toronto:
http://www1.toronto.ca/wps/portal/contentonly?vgnextoid=15dede0230460410VgnVCM
10000071d60f89RCRD
Department of Transportation Wisconsin. (2012, October). Suggested paramics Settings.
Retrieved from Micro-simulation Guidelines, Department of Transportation, Wisconsin:
http://www.wisdot.info/microsimulation/index.php?title=Suggested_Paramics_Settings
77
Duncan, G. (1997). Paramics Technical Report:Car-Following, Lane-Changing and Junction
Modelling. Quadstone Paramics.
El-Basyouny, K., & Sayed, T. (2009). Collision Prediction Models using Multivariate Poisson-
lognormal Regression. Accident Analysis and Prevention, 41(4), 820-828.
El-Basyouny, K., & Sayed, T. (2013). Safety Performance Functions Using Traffic Conlicts.
Safety Science, 51(1), 160-164.
Gettman, D., & Head, L. (2003). Surrogate Safety Measures from Traffic Simulation Models.
Transportation Research Record: Journal of the Transportation Research Board,
1840(1), 104-115.
Gettman, D., Pu, L., Sayed, T., & Shelby, S. (2008). Surrogate Safety Assessment Model and
Validation: Final Report. The Federal Highway Administration.
Goh, K., Currie, G., Sarvi, M., & Logan, D. (2013). Road Safety Benefits from Bus Priority.
Transportation Research Record: Journal of the Transportation Research Board,
2352(1), 41-49.
Goh, K., Currie, G., Sarvi, M., & Logan, D. (2014). Experimental Microsimulation Modeling of
Safety Impacts of Bus Priority. Transportation Research Record: Journal of the
Transportation Research Board, 2402(1), 9-18.
Hadayeghi, A., Shalaby, A. S., & Persaud, B. N. (2007). Safety Prediction Models: Proactive
Tool for Safety Evaluation in Urban Transportation Planning Applications.
Transportation Research Record: Journal of the Transportation Research Board
2019(1), 225-236.
78
Hauer, E., & Bamfo, J. (1997). Two Tool for Finding what Function Links the Dependent
Variable to the Explanatory Variables. Proceedings of the ICTCT 1997 Conference.
Lund, Sweden.
Hedelin, A., Björnstig, U., & Brismar, B. (1996). Trams - A Risk Factor for Pedestrians.
Accident Analysis and Prevention, 28(6), 733 - 738.
Huang, F., Liu, P., Yu, H., & Wang, W. (2013). Identifying if VISSIM Simulation Model and
SSAM Provide Reasonable Estimates for Field Measured Traffic Conflicts at Signalized
Intersections. Accident Analysis and Prevention, 50, 1014-1024.
Jonsson, T., Ivan, J. N., & Zhang, C. (2007). Crash Prediction Models for Intersections on Rural
Multilane Highways: Differences by Collision Type. Transportation Research Record:
Journal of the Transportation Research Board, 2019(1), 91-98.
McCullagh, P., & Nelder, J. A. (1989). Generalized Linear Models (Vol.2). London: Chapman
and Hall.
Miaou, S.-P. (1996). Measuring the Goodness-of-Fit of Accident Prediction Models. (No.
FHWA-RD-96-040).
Mitra, S., & Washington, S. (2007). On the Nature of Over-Dispersion in Motor Vehicle Crash
Prediction Models. Accident Analysis and Prevention, 39(3), 459-468.
Older, S. J., & Spicer, B. R. (1976). Traffic Conflicts - A Development in Accident Research.
Human Factors: The Journal of the Human Factors and Ergonomics Society, 18(4), 335-
350.
79
Perkins, S. R., & Harris, J. I. (1967). Criteria for Traffic Conflict Characteristics: Signalized
Intersections. General Motors Corporation. Electro-Mechanical Department. Research
Publication GMR-632.
Persaud, B., Lord, D., & Palmisano, J. (2002). Calibration and Transferability of Accident
Prediction Models for Urban Intersections. Transportation Research Record: Journal of
the Transportation Research Board, 1784(1), 57-64.
R Development Core Team. (2008). R: A Language and Environement for Statistical
Computing. Vienna, Austria. Retrieved from http://www.R-project.org
Saleem, T., Persaud, B., Shalaby, A., & Ariza, A. (2014). Can Microsimulation be Used to
Estimate Intersection Safety? Transportation Research Record: Journal of the
Transportation Research Board, 2432(1), 142-148.
SAS Institute Inc. (2008). SAS/STAT 9.2 User's Guide, PROC GENMOD. Cary, NC, USA: SAS
Institute Inc.
Shahla, F., Shalaby, A. S., Persaud, B. N., & Hadayeghi, A. (2009). Analysis of Transit Safety at
Signalized Intersections in Toronto, Ontario, Canada. Transportation Research Record:
Journal of the Transprotation Research Board, 2102(1), 108-114.
Toronto Transit Commission. (2015). General Information. Retrieved from Schedules and Maps:
https://www.ttc.ca/Routes/General_Information/General_Information.jsp
U.S. Department of Transportation. (2011, April 4). SSAM 2.1.6 Release Notes. Retrieved from
Federal Highway Administration:
http://www.fhwa.dot.gov/downloads/research/safety/ssam/ssam2_1_6_release_notes.cfm
80
Zegeer, C. V., & Deen, R. C. (1978). Traffic Conflicts as a Diagnostic Tool in HIghway Safety.
Transportation Research Record 667, 48-55.
81
Appendix A List of Modelled Micro-simulation Networks
Scenarios Scenario 1 Scenario 2 Scenario 3 Scenario 4 Scenario 5 Scenario 6 Scenario 7 Scenario 8 Scenario 9
# Intersection
Name
Removal of Green Extension
Removal of Red
Truncation
Removal of TSP
Near-Sided to Far-sided
stops [Major
Directions without
TSP]
Near-Sided to Far-sided
stops [Minor
Directions without
TSP]
Near-Sided to No stops [Minor
Directions without
TSP]
Near-Sided to Far-sided
stops [TSP Directions]
Streetcar to Bus (1:1
replace-ment)
Streetcar to Bus (1:m
Replace-ment for
Same Capacity)
1 Bloor St E and Jarvis St
2 Church St and Front St E
3 Church St and King St E
4 Church St and Gerrard St E
5 Yonge St and King St E
6 Yonge St and Wellesley St E
7 Yonge St and Eglinton Ave E
8 Bay St and Front St W
9 Bay St and Bloor St W
10 University Ave and King St W
82
11 Eglinton Ave W and Bathurst St
12 Lawrence Ave W and Avenue Rd
13 Avenue Rd and Wilson Ave
14 Kingston Rd and Dixon Ave
15 Church St and Park Rd
16 Dundas St W and Ossington Ave
17 Dundas St W and Dovercourt Rd
18 Eglinton Ave E and Mccowan Rd
19 Lake Shore Blvd E and Lower Jarvis St
20 Lake Shore Blvd W and Marine Parade Dr
21
Lake Shore Blvd W and Colonel Samuel Smith Park Dr
22 Parliament St and Gerrard St E
23 Sherbourne St and King St E
24 Sherbourne St and Dundas St E
83
25 King St W and Spadina Ave
26 Danforth Ave and Broadview Ave
27 Bathurst St and Harbord St
28 Bathurst St and Dupont St
29 Bathurst St and Glencairn Ave
30 Bloor St W and Dundas St W
31 Bloor St W and Keele St
32 Danforth Ave and Greenwood Ave
33 Danforth Ave and Main St
34 Gerrard St E and Carlaw Ave
35 Danforth Rd and Midland Ave
36 Mccowan Rd and Lawrence Ave E
37 Danforth Rd and St Clair Ave E
38 Lawrence Ave E and The Donway E
84
39 Lawrence Ave E and Pharmacy Ave
40 Lawrence Ave E and Midland Ave
41 Lawrence Ave E and Markham Rd
42 Lawrence Ave W and Marlee Ave
43 Dufferin St and Lawrence Ave W
44 Lawrence Ave W and Caledonia Rd
45 Weston Rd and Lawrence Ave W
46 Lawrence Ave W and Scarlett Rd
47 Dixon Rd and Royal York Rd
48 O Connor Dr and Pape Ave
49 O Connor Dr and Coxwell Ave
50 O Connor Dr and St Clair Ave E
51 Eglinton Ave W and Keele St
52 St Clair Ave E and Pharmacy Ave
85
53 St Clair Ave W and Old Weston Rd
54 St Clair Ave W and Runnymede Rd
55 Keele St and Gulliver Rd
56 Keele St and Wilson Ave
57 Dufferin St and Glencairn Ave
58 Bayview Ave and Finch Ave E
59 Coxwell Ave and Mortimer Ave
60 Midland Ave and Progress Ave
61 Kennedy Rd and Glamorgan Ave
62 Ellesmere Rd and Pharmacy Ave
63 Mccowan Rd and Ellesmere Rd
64 Ellesmere Rd and Morningside Ave
65 Albion Rd and Kipling Ave
66 Sheppard Ave E and Willowdale Ave
86
67 Kennedy Rd and Sheppard Ave E
68 Browns Line and Horner Ave
69 Martin Grove Rd and The Westway
70 Finch Ave W and Arrow Rd
71 College St and Dovercourt Rd
72 College St and Lansdowne Ave
73 Dupont St and Spadina Rd
74 Dupont St and Dovercourt Rd
75 Martin Grove Rd and Rathburn Rd
76 Ellesmere Rd and Helicon Gt
77 Finch Ave W and Islington Ave
78 Finch Ave W and Martin Grove Rd
79 Finch Ave W and Alness St
80 Finch Ave E and Victoria Park Ave
81 Finch Ave E and Kennedy Rd
82 Pharmacy Ave and Huntingwood Dr
87
83 Kingston Rd and Scarborough Golf Club Rd
84 Ellesmere Rd and Orton Park Rd
85 Sheppard Ave E and Morningside Ave
86 Kennedy Rd and Huntingwood Dr
87 Morningside Ave and Military Trl
88 Steeles Ave E and Pharmacy Ave
89 Steeles Ave E and Warden Ave
90 Steeles Ave E and Kennedy Rd
91 Steeles Ave W and Islington Ave
92 Steeles Ave W and Signet Dr
93 Victoria Park Ave and Gordon Baker Rd
94 Finch Ave E and Tapscott Rd
95 Warden Ave and Mcnicoll Ave
96 Carlingview Dr and
88
International Blvd
97 Brimley Rd and Mcnicoll Ave
98 Tapscott Rd and Mclevin Ave
99 Morningside Ave and Milner Ave
100 Morningside Ave and Sewells Rd
Total Selected Networks
20 10 20 35 25 35 15 15 15
89
Appendix B Detailed Scenario Test Results
This Appendix details the resulting predicted crashes rates in each of the scenario test. The notation
in each of the following tables is “∆ (B)”, where ∆ is the change in predicted crash frequency
relative to status-quo case and B is the predicted crash frequency for status-quo case. A positive ∆
indicates an increase in predicted crash frequency and vice versa. The summations amongst the
studied intersections are also provided at the bottom of each table. Since a reduction in predicted
crash is the more desirable direction, a negative total changes in predicted crashes is coloured in
green. Conversely, a positive total changes in predicted crashes is coloured in red.
Scenario 1 With Green Extension Without Green Extension
Intersections (20 in total) All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Church St and King St E 0.5 (78.2) 1.0 (35.3) -0.2 (23.8) 0.4 (12.4) -0.2 (6.4)
Dundas St W and Ossington Ave 1.4 (62.2) 0.5 (33.1) 0.4 (15.5) 0.5 (11.5) 0.1 (4.6)
Dundas St W and Dovercourt Rd -1.1 (43.4) -0.5 (15.9) -0.3 (11.2) -0.2 (9.0) -0.2 (1.4)
Parliament St and Gerrard St E -0.4 (90.0) 0.7 (32.5) -0.3 (31.1) -0.3 (13.4) 0.1 (6.1)
Sherbourne St and King St E 1.5 (59.1) 2.4 (23.9) 0.2 (18.0) -0.1 (7.7) -0.1 (3.7)
Sherbourne St and Dundas St E 0.1 (66.6) 0.2 (35.4) 0.1 (18.6) -0.3 (8.5) 0.5 (4.6)
Danforth Ave and Broadview Ave -6.1 (161.3) -1.5 (53.6) -3.0 (68.3) -1.6 (20.1) 0.0 (5.9)
Bathurst St and Glencairn Ave 0.8 (106.9) -1.4 (38.6) 0.6 (39.2) 0.5 (14.8) 0.1 (4.3)
Bloor St W and Dundas St W -6.2 (133.4) -0.5 (62.0) -2.9 (45.5) -2.0 (23.1) -0.1 (6.8)
Danforth Ave and Main St -2.8 (91.9) 0.4 (37.1) -1.6 (32.0) -0.5 (11.2) 0.2 (5.2)
Gerrard St E and Carlaw Ave -1.5 (69.7) -1.0 (36.9) -0.4 (18.9) -0.4 (11.0) -0.2 (3.5)
Dufferin St and Glencairn Ave 0.2 (82.0) 0.6 (31.5) -0.1 (27.5) 0.2 (11.5) 0.0 (3.8)
Mccowan Rd and Ellesmere Rd 1.1 (156.6) 1.1 (44.9) 0.3 (65.9) 0.5 (22.9) 0.2 (6.7)
Sheppard Ave E and Willowdale Ave
0.0 (93.5) 0.0 (27.2) 0.0 (35.3) 0.0 (11.0) 0.0 (2.3)
Finch Ave W and Arrow Rd 1.7 (128.3) 0.7 (53.5) 0.9 (48.9) 0.0 (14.3) -0.1 (4.7)
College St and Dovercourt Rd -0.4 (65.8) 1.0 (25.4) -0.3 (21.3) -0.4 (7.6) 0.2 (1.9)
College St and Lansdowne Ave -0.2 (99.8) 0.3 (26.5) -0.2 (32.9) 0.0 (23.6) 0.2 (4.9)
Finch Ave W and Islington Ave 0.7 (199.7) 0.0 (70.6) 1.0 (85.1) -0.8 (31.2) 0.0 (8.9)
Finch Ave W and Alness St -14.3 (146.4) -2.7 (43.8) -8.1 (61.4) -1.6 (19.2) -0.3 (4.3)
90
Finch Ave E and Tapscott Rd 2.2 (67.4) 2.2 (25.5) 0.5 (21.8) 0.5 (8.6) 0.4 (3.6)
Summation of Changes in Predicted Crashes
-22.75 3.62 -13.63 -5.75 0.79
Summation of Predicted Crashes in Status Quo Case 2002.3 753.1 722.0 292.5 93.8
Percentage Difference relative to Status Quo Case -1.14% 0.48% -1.89% -1.97% 0.84%
Scenario 2 (10 in total) With Red Truncation Without Red Truncation
Intersections All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Church St and King St E -1.0 (78.2) -1.2 (35.3) -0.3 (23.8) 0.0 (12.4) -0.4 (6.4)
Dundas St W and Ossington Ave 0.7 (62.2) -0.9 (33.1) 0.5 (15.5) 0.4 (11.5) 0.4 (4.6)
Dundas St W and Dovercourt Rd 0.1 (43.4) -0.6 (15.9) 0.1 (11.2) 0.2 (9.0) 0.2 (1.4)
Sherbourne St and King St E -1.7 (59.1) -1.5 (23.9) -0.4 (18.0) -0.6 (7.7) -0.1 (3.7)
Danforth Ave and Broadview Ave -1.8 (161.3) -0.3 (53.6) -0.7 (68.3) -0.9 (20.1) -0.2 (5.9)
Bathurst St and Glencairn Ave -5.2 (106.9) -2.4 (38.6) -1.9 (39.2) -1.7 (14.8) 0.1 (4.3)
Danforth Ave and Main St -1.0 (91.9) -1.7 (37.1) -0.2 (32.0) 0.2 (11.2) -0.1 (5.2)
Gerrard St E and Carlaw Ave 0.1 (69.7) 0.0 (36.9) 0.2 (18.9) -0.3 (11.0) 0.0 (3.5)
College St and Dovercourt Rd 0.4 (65.8) 0.9 (25.4) 0.0 (21.3) -0.1 (7.6) 0.1 (1.9)
Finch Ave W and Alness St -4.7 (146.4) -2.0 (43.8) -2.8 (61.4) 0.2 (19.2) -0.3 (4.3)
Summation of Changes in Predicted Crashes -14.11 -9.77 -5.39 -2.65 -0.04
Summation of Predicted Crashes in Status Quo Case 885.0 343.5 309.5 124.5 41.2
Percentage Difference relative to Status Quo Case -1.59% -2.85% -1.74% -2.13% -0.09%
Scenario 3 (20 in total) With TSP (both Green Extension and truncation)
Without TSP
Intersections All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Church St and King St E -0.8 (78.2) -0.3 (35.3) -0.3 (23.8) -0.1 (12.4) -0.3 (6.4)
Dundas St W and Ossington Ave 0.3 (62.2) 0.6 (33.1) -0.4 (15.5) 0.7 (11.5) 0.3 (4.6)
Dundas St W and Dovercourt Rd 0.1 (43.4) -0.6 (15.9) 0.4 (11.2) -0.4 (9.0) 0.4 (1.4)
Parliament St and Gerrard St E -0.4 (90.0) 0.7 (32.5) -0.3 (31.1) -0.3 (13.4) 0.1 (6.1)
Sherbourne St and King St E -3 (59.1) -1.3 (23.9) -1.2 (18.0) -0.4 (7.7) -0.1 (3.7)
91
Sherbourne St and Dundas St E 0.1 (66.6) 0.2 (35.4) 0.1 (18.6) -0.3 (8.5) 0.5 (4.6)
Danforth Ave and Broadview Ave -6.6 (161.3) -3.1 (53.6) -3.2 (68.3) -1.0 (20.1) 0.0 (5.9)
Bathurst St and Glencairn Ave 4.3 (106.9) 0.7 (38.6) 2.2 (39.2) 0.6 (14.8) 0.2 (4.3)
Bloor St W and Dundas St W -6.2 (133.4) -0.5 (62.0) -2.9 (45.5) -2.0 (23.1) -0.1 (6.8)
Danforth Ave and Main St -3.7 (91.9) -0.2 (37.1) -1.9 (32.0) -0.7 (11.2) 0.3 (5.2)
Gerrard St E and Carlaw Ave -0.7 (69.7) -0.1 (36.9) -0.3 (18.9) -0.3 (11.0) -0.1 (3.5)
Dufferin St and Glencairn Ave 0.2 (82.0) 0.6 (31.5) -0.1 (27.5) 0.2 (11.5) 0.0 (3.8)
Mccowan Rd and Ellesmere Rd 1.1 (156.6) 1.1 (44.9) 0.3 (65.9) 0.5 (22.9) 0.2 (6.7)
Sheppard Ave E and Willowdale Ave 0.0 (93.5) 0.0 (27.2) 0.0 (35.3) 0.0 (11.0) 0.0 (2.3)
Finch Ave W and Arrow Rd 1.7 (128.3) 0.7 (53.5) 0.9 (48.9) 0.0 (14.3) -0.1 (4.7)
College St and Dovercourt Rd -1.5 (65.8) 0.1 (25.4) -1.0 (21.3) 0.4 (7.6) 0.2 (1.9)
College St and Lansdowne Ave -0.2 (99.8) 0.3 (26.5) -0.2 (32.9) 0.0 (23.6) 0.2 (4.9)
Finch Ave W and Islington Ave 0.7 (199.7) 0.0 (70.6) 1.0 (85.1) -0.8 (31.2) 0.0 (8.9)
Finch Ave W and Alness St -11.6 (146.4) -3.2 (43.8) -7.4 (61.4) 1.0 (19.2) -0.6 (4.3)
Finch Ave E and Tapscott Rd 2.2 (67.4) 2.3 (25.5) 0.4 (21.8) 0.5 (8.6) 0.4 (3.6)
Summation of Changes in Predicted Crashes
-23.93 -1.85 -13.76 -2.46 1.51
Summation of Predicted Crashes in Status Quo Case
2002.3 753.1 722.0 292.5 93.8
Percentage Difference relative to Status Quo Case
-1.20% -0.25% -1.91% -0.84% 1.61%
Scenario 4 (35 in total) Near-Sided Stops (Along Major Direction at non-TSP
intersections) Far-sided Stops
Intersections All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Yonge St and King St E 1.2 (51.3) 1.3 (13.6) 0.2 (15.1) 0.4 (10.4) 0.4 (3.0)
Bay St and Front St W 2.5 (72.0) 1.4 (39.2) 0.7 (19.4) 0.7 (11.2) 0.7 (3.8)
Bay St and Bloor St W 2.5 (95.2) 2.9 (37.5) 0.9 (33.3) -0.5 (12.4) -0.2 (3.4)
Avenue Rd and Wilson Ave -0.2 (190.8) 1.8 (70.9) -1.0 (81.5) 0.7 (25.8) 0.9 (5.6)
Eglinton Ave E and Mccowan Rd 1.7 (80.5) 3.7 (28.1) -0.1 (27.0) 0.4 (12.5) 1.0 (6.0)
Bathurst St and Dupont St 7.6 (109.5) 10.1 (50.2) 1.3 (37.3) 0.9 (15.5) 1.4 (3.8)
Mccowan Rd and Lawrence Ave E 5.4 (232.5) 3.4 (61.3) 2.1 (111.2) 2.3 (31.2) 0.9 (7.3)
Danforth Rd and St Clair Ave E -0.8 (139.1) 0.8 (30.9) -1.0 (60.4) 0.7 (16.2) 0.1 (4.4)
Lawrence Ave E and Pharmacy Ave 2.2 (175.0) 1.8 (54.5) 1.0 (77.3) 0.3 (20.1) 1.2 (4.5)
92
Lawrence Ave E and Markham Rd 4.3 (114.5) -1.9 (45.3) 2.7 (41.2) 1.1 (16.9) 0.5 (4.7)
Lawrence Ave W and Marlee Ave 1.3 (167.2) 2.8 (43.7) 0.3 (74.4) -0.1 (20.3) 1.0 (4.9)
Dufferin St and Lawrence Ave W 5.1 (176.5) 3.5 (49.3) 2.5 (77.2) 0.4 (25.4) 1.4 (6.8)
Weston Rd and Lawrence Ave W 1.3 (91.0) 2.3 (31.2) 0.3 (28.8) -0.1 (19.6) 0.4 (6.4)
Dixon Rd and Royal York Rd 9.4 (165.8) 7.7 (64.6) 4.0 (68.6) 0.8 (19.7) 0.1 (4.2)
O Connor Dr and Pape Ave 1.4 (64.5) 3.9 (30.2) -0.3 (18.3) -0.2 (10.1) 0.9 (6.3)
O Connor Dr and St Clair Ave E 1.3 (134.2) 0.8 (13.2) 0.9 (59.4) -0.5 (17.6) 0.0 (4.6)
Bayview Ave and Finch Ave E 5.2 (220.3) 2.1 (66.9) 3.0 (102.3) 0.6 (28.7) 1.6 (4.4)
Coxwell Ave and Mortimer Ave 1.2 (79.3) 0.8 (29.7) 0.4 (26.2) 0.3 (11.9) 0.8 (3.0)
Midland Ave and Progress Ave 1.5 (111.2) 2.3 (36.2) 0.1 (42.5) 0.5 (14.6) 1.3 (3.8)
Kennedy Rd and Glamorgan Ave -0.6 (110.5) -0.6 (38.5) 0.0 (41.1) -0.4 (15.6) 0.7 (3.3)
Ellesmere Rd and Morningside Ave 5.9 (137.5) 1.6 (46.7) 2.7 (56.3) 1.8 (15.2) 0.8 (5.8)
Albion Rd and Kipling Ave 2.6 (110.9) 0.8 (34.7) 0.9 (41.1) 1.0 (18.0) 0.9 (4.8)
Kennedy Rd and Sheppard Ave E 2.9 (183.9) 3.0 (53.9) 0.8 (80.1) 0.8 (27.7) 0.7 (5.7)
Browns Line and Horner Ave 2.4 (105.6) 1.2 (32.1) 0.9 (40.7) 0.7 (13.2) 1.1 (2.9)
Finch Ave W and Martin Grove Rd 1.8 (93.9) 0.5 (45.3) 0.6 (30.5) 1.0 (11.9) 0.2 (6.1)
Finch Ave E and Victoria Park Ave 5.2 (103.9) 1.5 (43.7) 2.5 (35.3) 0.6 (16.4) 1.8 (6.1)
Finch Ave E and Kennedy Rd 0.5 (176.4) -0.7 (49.0) -0.2 (74.8) 1.1 (29.9) 0.2 (5.6)
Ellesmere Rd and Orton Park Rd 8.4 (96.6) 3.2 (29.5) 3.9 (35.7) 0.9 (13.4) 1.6 (5.7)
Sheppard Ave E and Morningside Ave 4 (85.6) 0.9 (32.5) 1.7 (28.3) 1.0 (13.7) -0.6 (4.2)
Kennedy Rd and Huntingwood Dr 2.9 (149.3) 1.4 (43.4) 1.4 (64.5) 0.5 (16.0) 0.2 (2.9)
Steeles Ave E and Pharmacy Ave 2.5 (171.7) 4.2 (65.7) 0.1 (69.1) 0.8 (26.4) 1.1 (5.5)
Steeles Ave E and Warden Ave -0.5 (213.4) 2.5 (64.1) -0.9 (96.9) -0.1 (30.8) 0.7 (6.4)
Steeles Ave E and Kennedy Rd 1.3 (204.2) -1.0 (60.4) 1.4 (89.7) -0.3 (33.4) 0.1 (7.0)
Steeles Ave W and Islington Ave 2.9 (162.9) 1.1 (58.2) 1.3 (66.6) 0.5 (23.8) 1.5 (3.7)
Steeles Ave W and Signet Dr 1.5 (87.9) 0.3 (29.8) 0.7 (31.6) 0.3 (10.9) 0.1 (3.7)
Summation of Changes in Predicted Crashes 97.89 71.72 35.77 18.81 25.66
Summation of Predicted Crashes in Status Quo Case 4664.7 1524.1 1883.7 656.4 170.2
Percentage Difference relative to Status Quo Case 2.10% 4.71% 1.90% 2.87% 15.08%
Scenario 5 (20 in total) Near-Sided Stops (Along Minor Direction at non-TSP
intersections) Far-sided Stops
93
Intersections All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Lawrence Ave W and Marlee Ave -1.7 (104.4) 1.1 (20.3) -1.2 (41.9) -0.1 (14.0) 0.5 (1.4)
Dufferin St and Lawrence Ave W 1.7 (51.3) 1.0 (13.6) 0.6 (15.1) 0.1 (10.4) 0.5 (3.0)
Weston Rd and Lawrence Ave W 2.1 (72.4) 0.4 (26.0) 0.7 (22.6) 0.8 (12.9) 0.6 (3.8)
Dixon Rd and Royal York Rd 1.4 (109.5) 1.8 (50.2) 0.3 (37.3) 0.1 (15.5) 0.4 (3.8)
O Connor Dr and Pape Ave 1.9 (232.5) 0.3 (61.3) 1.2 (111.2) 0.4 (31.2) -1.3 (7.3)
O Connor Dr and St Clair Ave E -3.1 (175.0) 1.6 (54.5) -2.3 (77.3) -0.4 (20.1) 0.2 (4.5)
Bayview Ave and Finch Ave E 4.2 (114.5) 1.0 (45.3) 1.9 (41.2) 1.1 (16.9) 0.8 (4.7)
Coxwell Ave and Mortimer Ave 4.4 (176.5) 1.3 (49.3) 2.5 (77.2) 0.4 (25.4) 0.7 (6.8)
Midland Ave and Progress Ave 4.1 (168.9) 4.1 (49.8) 1.8 (75.0) 0.0 (18.3) 0.0 (5.8)
Kennedy Rd and Glamorgan Ave 1.7 (91.0) 2.9 (31.2) 0.3 (28.8) 0.0 (19.6) 0.1 (6.4)
Ellesmere Rd and Morningside Ave 33.6 (165.8) 26.1 (64.6) 14.5 (68.6) 2.7 (19.7) 0.7 (4.2)
Albion Rd and Kipling Ave 2.0 (64.5) 0.3 (30.2) 0.9 (18.3) 0.3 (10.1) 0.5 (6.3)
Kennedy Rd and Sheppard Ave E 0.4 (93.8) -0.1 (31.3) 0.1 (31.6) 0.4 (17.7) 0.2 (5.4)
Browns Line and Horner Ave -1.8 (220.3) -0.1 (66.9) -1.1 (102.3) -0.5 (28.7) 0.7 (4.4)
Finch Ave W and Martin Grove Rd 1.0 (79.3) 1.9 (29.7) 0.2 (26.2) -0.3 (11.9) 0.2 (3.0)
Finch Ave E and Victoria Park Ave -1.1 (111.2) -1.6 (36.2) -0.3 (42.5) 0.0 (14.6) 0.4 (3.8)
Finch Ave E and Kennedy Rd 5.2 (137.5) 0.9 (46.7) 2.7 (56.3) 1.2 (15.2) 0.6 (5.8)
Ellesmere Rd and Orton Park Rd 0.9 (110.9) 0.0 (34.7) 0.7 (41.1) -0.4 (18.0) 0.9 (4.8)
Sheppard Ave E and Morningside Ave 6.1 (183.9) 2.9 (53.9) 3.0 (80.1) 1.1 (27.7) 1.8 (5.7)
Kennedy Rd and Huntingwood Dr 0.7 (90.1) 0.7 (26.1) 0.0 (32.8) 0.6 (13.0) 0.4 (3.5)
Steeles Ave E and Pharmacy Ave 1.8 (93.9) 0.5 (45.3) 0.6 (30.5) 1.0 (11.9) 0.2 (6.1)
Steeles Ave E and Warden Ave 7.9 (103.9) 0.3 (43.7) 3.7 (35.3) 2.4 (16.4) 0.4 (6.1)
Steeles Ave E and Kennedy Rd -0.1 (149.3) 4.7 (43.4) -1.1 (64.5) 0.0 (16.0) 0.0 (2.9)
Steeles Ave W and Islington Ave 3.1 (213.4) 1.3 (64.1) 2.2 (96.9) -0.6 (30.8) -0.4 (6.4)
Steeles Ave W and Signet Dr 0.9 (87.9) 1.0 (29.8) 0.2 (31.6) 0.2 (10.9) 0.4 (3.7)
Summation of Changes in Predicted Crashes 77.39 54.37 32.16 10.59 9.59
Summation of Predicted Crashes in Status Quo Case 3201.7 1048.0 1286.3 447.0 119.3
Percentage Difference relative to Status Quo Case 2.42% 5.19% 2.50% 2.37% 8.03%
94
Scenario 6 (35 in total) Near-Sided Stops (Along Major Direction at non-TSP
intersections) No Transit Stops
Intersections All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Yonge St and King St E -0.3 (51.3) -0.6 (13.6) -0.1 (15.1) 0.1 (10.4) -0.3 (3.0)
Bay St and Front St W -5.0 (72.0) -1.5 (39.2) -2.2 (19.4) -0.8 (11.2) -2.1 (3.8)
Bay St and Bloor St W -3.1 (95.2) 0.7 (37.5) -1.7 (33.3) -0.9 (12.4) -1.9 (3.4)
Avenue Rd and Wilson Ave -8.0 (190.8) -2.2 (70.9) -5.0 (81.5) 0.0 (25.8) -1.2 (5.6)
Eglinton Ave E and Mccowan Rd -9.1 (80.5) -2.8 (28.1) -3.8 (27.0) -1.7 (12.5) -3.1 (6.0)
Bathurst St and Dupont St -3.7 (109.5) -0.7 (50.2) -2.0 (37.3) -0.6 (15.5) -0.3 (3.8)
Mccowan Rd and Lawrence Ave E -3.2 (232.5) -2.0 (61.3) -1.5 (111.2) -0.7 (31.2) -0.8 (7.3)
Danforth Rd and St Clair Ave E -4.2 (139.1) -2.1 (30.9) -1.9 (60.4) -0.9 (16.2) -0.5 (4.4)
Lawrence Ave E and Pharmacy Ave -2.7 (175.0) -1.8 (54.5) -1.3 (77.3) -0.2 (20.1) -0.7 (4.5)
Lawrence Ave E and Markham Rd -2 (114.5) -2.6 (45.3) -0.7 (41.2) 0.3 (16.9) -0.7 (4.7)
Lawrence Ave W and Marlee Ave -5.1 (167.2) -1.0 (43.7) -2.7 (74.4) -1.2 (20.3) -2.0 (4.9)
Dufferin St and Lawrence Ave W -0.3 (176.5) -0.6 (49.3) 0.3 (77.2) -0.9 (25.4) 0.6 (6.8)
Weston Rd and Lawrence Ave W -3.5 (91.0) -0.8 (31.2) -1.8 (28.8) -0.3 (19.6) -0.3 (6.4)
Dixon Rd and Royal York Rd 7.3 (165.8) 4.5 (64.6) 3.3 (68.6) 1.0 (19.7) -1.3 (4.2)
O Connor Dr and Pape Ave -1.7 (64.5) 0.7 (30.2) -1.5 (18.3) 0.7 (10.1) -0.3 (6.3)
O Connor Dr and St Clair Ave E 0.1 (134.2) 0.2 (13.2) 0.3 (59.4) -0.6 (17.6) -0.3 (4.6)
Bayview Ave and Finch Ave E -3.6 (220.3) -0.3 (66.9) -2.1 (102.3) -0.9 (28.7) 0.4 (4.4)
Coxwell Ave and Mortimer Ave -0.8 (79.3) -0.1 (29.7) -0.3 (26.2) -0.1 (11.9) -0.4 (3.0)
Midland Ave and Progress Ave -2.1 (111.2) -2.0 (36.2) -0.8 (42.5) 0.0 (14.6) -0.9 (3.8)
Kennedy Rd and Glamorgan Ave -2.3 (110.5) -0.3 (38.5) -1.3 (41.1) -0.3 (15.6) -0.9 (3.3)
Ellesmere Rd and Morningside Ave -4 (137.5) -0.8 (46.7) -2.2 (56.3) -0.6 (15.2) -1.0 (5.8)
Albion Rd and Kipling Ave 0.6 (110.9) 1.1 (34.7) 0.1 (41.1) 0.0 (18.0) 0.1 (4.8)
Kennedy Rd and Sheppard Ave E -0.8 (183.9) 0.1 (53.9) -0.5 (80.1) -0.3 (27.7) -0.4 (5.7)
Browns Line and Horner Ave -1.9 (105.6) -1.6 (32.1) -0.9 (40.7) 0.2 (13.2) -0.6 (2.9)
Finch Ave W and Martin Grove Rd 0.5 (93.9) 0.5 (45.3) -0.1 (30.5) 0.5 (11.9) -0.5 (6.1)
Finch Ave E and Victoria Park Ave -4.3 (103.9) -1.6 (43.7) -2.0 (35.3) -0.5 (16.4) -2.2 (6.1)
Finch Ave E and Kennedy Rd -2.4 (176.4) -1.2 (49.0) -1.4 (74.8) -0.1 (29.9) -0.2 (5.6)
Ellesmere Rd and Orton Park Rd -8.4 (96.6) -1.7 (29.5) -3.7 (35.7) -2.2 (13.4) -2.4 (5.7)
Sheppard Ave E and Morningside Ave -0.3 (85.6) 0.4 (32.5) -0.5 (28.3) 0.6 (13.7) -0.7 (4.2)
95
Kennedy Rd and Huntingwood Dr -1.9 (149.3) 0.2 (43.4) -1.1 (64.5) -0.4 (16.0) -1.0 (2.9)
Steeles Ave E and Pharmacy Ave -2.5 (171.7) 0.9 (65.7) -2.0 (69.1) 0.0 (26.4) -2.1 (5.5)
Steeles Ave E and Warden Ave -3.6 (213.4) -1.6 (64.1) -1.7 (96.9) -1.1 (30.8) -0.8 (6.4)
Steeles Ave E and Kennedy Rd -5.5 (204.2) -1.7 (60.4) -2.6 (89.7) -1.7 (33.4) -0.5 (7.0)
Steeles Ave W and Islington Ave 0.1 (162.9) -0.2 (58.2) -0.2 (66.6) 0.2 (23.8) 0.5 (3.7)
Steeles Ave W and Signet Dr -4.1 (87.9) -2.5 (29.8) -1.5 (31.6) -0.6 (10.9) -2.3 (3.7)
Summation of Changes in Predicted Crashes -92.11 -25.03 -46.99 -13.64 -30.92
Summation of Predicted Crashes in Status Quo Case 4664.7 1524.1 1883.7 656.4 170.2
Percentage Difference relative to Status Quo Case -1.97% -1.64% -2.49% -2.08% -18.17%
Scenario 7 Near-Sided Stops (Along TSP-servicing Direction at TSP
intersections) Near-Sided Transit Stops
Intersections (15 in total) All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Church St and King St E 3.9 (78.2) 4.1 (35.3) 1.2 (23.8) -0.1 (12.4) 0.0 (6.4)
Dundas St W and Ossington Ave 0.1 (62.2) -1.0 (33.1) 0.3 (15.5) 0.2 (11.5) -0.2 (4.6)
Dundas St W and Dovercourt Rd 0.8 (43.4) 0.8 (15.9) 0.2 (11.2) 0.0 (9.0) -0.1 (1.4)
Parliament St and Gerrard St E 2.2 (90.0) 2.5 (32.5) 0.2 (31.1) 0.9 (13.4) 0.1 (6.1)
Sherbourne St and King St E -2.2 (59.1) 1.1 (23.9) -1.5 (18.0) 0.0 (7.7) -1.3 (3.7)
Sherbourne St and Dundas St E 1.5 (66.6) 1.6 (35.4) 0.4 (18.6) 0.0 (8.5) -0.2 (4.6)
Danforth Ave and Broadview Ave -4.4 (161.3) -0.6 (53.6) -2.4 (68.3) -0.8 (20.1) -0.2 (5.9)
Bathurst St and Glencairn Ave 2.9 (106.9) 2.6 (38.6) 0.8 (39.2) 0.5 (14.8) 0.1 (4.3)
Gerrard St E and Carlaw Ave -0.8 (69.7) -1.1 (36.9) -0.2 (18.9) 0.1 (11.0) -0.2 (3.5)
Dufferin St and Glencairn Ave -0.1 (82.0) 0.6 (31.5) -0.6 (27.5) 0.7 (11.5) -0.7 (3.8)
Finch Ave W and Arrow Rd -2.3 (128.3) 0.2 (53.5) -1.6 (48.9) 0.0 (14.3) 0.1 (4.7)
College St and Dovercourt Rd -3.9 (65.8) 0.6 (25.4) -2.2 (21.3) -0.3 (7.6) -0.4 (1.9)
College St and Lansdowne Ave -2.6 (99.8) -0.3 (26.5) -1.2 (32.9) -0.6 (23.6) 0.1 (4.9)
Finch Ave W and Alness St 1.0 (146.4) 0.3 (43.8) 0.4 (61.4) 0.3 (19.2) 0.2 (4.3)
Finch Ave E and Tapscott Rd 4.3 (67.4) 1.7 (25.5) 1.2 (21.8) 1.9 (8.6) 0.6 (3.6)
Summation of Changes in Predicted Crashes 0.32 13.24 -4.82 2.76 -1.96
Summation of Predicted Crashes in Status Quo Case 1327.2 511.3 458.2 193.2 63.8
Percentage Difference relative to Status Quo Case 0.02% 2.59% -1.05% 1.43% -3.07%
96
Scenario 8 Thru-moving Streetcars Thru-moving Buses
(1:1 replacement)
Intersections (15 in total) All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Church St and King St E 1.0 (78.2) 1.2 (35.3) 0.3 (23.8) -0.1 (12.4) 0.6 (6.4)
Yonge St and King St E 2.6 (51.3) -0.2 (13.6) 1.1 (15.1) 0.6 (10.4) 1.7 (3.0)
University Ave and King St W 3.6 (62.4) 1.6 (43.6) 1.7 (12.1) 0.1 (10.3) 1.8 (5.2)
Dundas St W and Ossington Ave 1.7 (62.2) 0.0 (33.1) 0.6 (15.4) 0.8 (11.5) 0.7 (4.6)
Dundas St W and Dovercourt Rd 2.2 (43.4) 0.8 (15.9) 0.7 (11.2) 0.5 (9.0) 1.3 (1.4)
Lake Shore Blvd W and Marine Parade Dr -3.5 (122.5) -0.9 (23.4) -1.6 (50.1) -0.8 (18.9) -0.3 (4.5)
Lake Shore Blvd W and Colonel Samuel Smith Park Dr 0.0 (72.4) 0.1 (26.0) -0.2 (22.6) 0.5 (12.9) -0.1 (3.8)
Sherbourne St and King St E 1.3 (59.1) 3.4 (23.9) -0.3 (18.0) 0.2 (7.7) 1.5 (3.7)
Sherbourne St and Dundas St E 0.9 (66.6) -5.6 (35.4) 2.0 (18.6) 0.6 (8.5) -0.7 (4.6)
King St W and Spadina Ave -0.5 (97.6) 1.4 (30.5) -0.8 (33.2) 0.2 (19.3) 0.4 (5.5)
Bathurst St and Harbord St 2.2 (94.4) 1.7 (31.9) 0.8 (35.3) 0.3 (9.5) 1.6 (2.3)
Gerrard St E and Carlaw Ave 1.8 (69.6) 0.0 (36.9) 1.0 (18.9) 0.2 (11.0) 1.2 (3.5)
St Clair Ave W and Old Weston Rd 4.0 (93.8) 5.0 (31.3) 1.0 (31.6) 0.5 (17.7) 2.5 (5.4)
College St and Dovercourt Rd 2.7 (62.1) 1.8 (25.4) 0.9 (19.3) 0.4 (7.6) 1.9 (1.2)
College St and Lansdowne Ave -1.1 (99.8) 0.6 (26.5) -0.8 (32.9) 0.0 (23.6) 0.7 (4.9)
Summation of Changes in Predicted Crashes 19.04 10.97 6.20 3.77 14.75
Summation of Predicted Crashes in Status Quo Case 1135.5 432.7 358.0 190.3 60.0
Percentage Difference relative to Status Quo Case 1.68% 2.54% 1.73% 1.98% 24.57%
Scenario 9 Thru-moving Streetcars Thru-moving Buses (1:m replacement for same operating capacity)
Intersections (15 in total) All Impact
Types Angle Rear-End
Side-Swipe
Transit-Involved
Church St and King St E 2.5 (78.2) 0.2 (35.3) 1.3 (23.8) 0.4 (12.4) 1.4 (6.4)
Yonge St and King St E 4.7 (51.3) 1 (13.6) 2.1 (15.1) 0.4 (10.4) 2.4 (3)
University Ave and King St W 4.6 (62.4) -2.1 (43.6) 3.3 (12.1) 1 (10.3) 2.4 (5.2)
Dundas St W and Ossington Ave 3.3 (62.2) 1.1 (33.1) 0.9 (15.4) 1.2 (11.5) 1 (4.6)
Dundas St W and Dovercourt Rd 2.7 (43.4) 0.9 (15.9) 1 (11.2) 0.4 (9) 1.6 (1.4)
Lake Shore Blvd W and Marine Parade Dr -1.4 (122.5) 0.7 (23.4) -0.6 (50.1) -0.7 (18.9) 0.8 (4.5)
97
Lake Shore Blvd W and Colonel Samuel Smith Park Dr 2.7 (72.4) 1.1 (26) 1.2 (22.6) 0.2 (12.9) 0.6 (3.8)
Sherbourne St and King St E 5.6 (59.1) 7.8 (23.9) 0.9 (18) 0 (7.7) 2.8 (3.7)
Sherbourne St and Dundas St E 1.4 (66.6) -5.8 (35.4) 2.4 (18.6) 0.5 (8.5) -0.2 (4.6)
King St W and Spadina Ave 1 (97.6) 3 (30.5) 0.1 (33.2) -0.3 (19.3) 1.2 (5.5)
Bathurst St and Harbord St 2.2 (94.4) 2.8 (31.9) 0.3 (35.3) 1 (9.5) 2.3 (2.3)
Gerrard St E and Carlaw Ave 1.3 (69.6) -1.8 (36.9) 1.1 (18.9) 0.4 (11) 0.9 (3.5)
St Clair Ave W and Old Weston Rd 3.3 (93.8) 3.3 (31.3) 0.8 (31.6) 0.6 (17.7) 2.6 (5.4)
College St and Dovercourt Rd 2.1 (62.1) 1.7 (25.4) 0.6 (19.3) 0.1 (7.6) 1.8 (1.2)
College St and Lansdowne Ave 0.7 (99.8) 1.6 (26.5) 0 (32.9) 0.2 (23.6) 1 (4.9)
Summation of Changes in Predicted Crashes 36.55 15.66 15.27 5.34 22.45
Summation of Predicted Crashes in Status Quo Case 1135.5 432.7 358.0 190.3 60.0
Percentage Difference relative to Status Quo Case 3.22% 3.62% 4.26% 2.80% 37.39%
98
Detailed Operational Characteristics for the 20 TSP-featured intersections in Scenario 3
Intersections
Maximum
Green
Extension
(seconds)
Maximum
Red
Truncation
(seconds)
Peak
Hourly
Volume
Percentage of
Peak Hourly
Volume in the
TSP-servicing
direction
Cycle
Time
(seconds)
Transit
Type
receiving
TSP
Joint
Headway in
the direction
receiving
TSP
(minutes)
Church St and King St E 30 9 2,504 51.3% 70 Streetcar 1.85
Dundas St W and Ossington Ave 30 8 2,033 50.4% 70 Streetcar 5.25
Dundas St W and Dovercourt Rd 30 6 1,692 67.4% 70 Streetcar 5.25
Parliament St and Gerrard St E 16 2,344 100% (All direction) 70 Streetcar/Bus
4.33 (E-W) 9.50 (N-S)
Sherbourne St and King St E 30 12 1,633 48.1% 70 Streetcar 1.71
Sherbourne St and Dundas St E 16 2,072 100% (All direction) 70 Streetcar
5.25 (E-W) 7.00 (N-S)
Danforth Ave and Broadview Ave 30 23 3,675 29.4% 98 Streetcar 2.27
Bathurst St and Glencairn Ave 30 5 2,983 79.2% 80 Bus 6.00
Bloor St W and Dundas St W 30 3,875 38.0% 80 Streetcar 2.27
Danforth Ave and Main St 16 15 2,862 33.3% 79 Streetcar/Bus 2.65
Gerrard St E and Carlaw Ave 30 2 2,263 54.9% 70 Streetcar 4.33
Dufferin St and Glencairn Ave 30 2,979 80.3% 80 Bus 6.25
Mccowan Rd and Ellesmere Rd 9 5,138 32.3% 118 Bus 5.13 Sheppard Ave E and Willowdale
Ave 16 3,342 52.7% 110 Bus 30.00
Finch Ave W and Arrow Rd 16 4,899 66.6% 110 Bus 5.00
College St and Dovercourt Rd 30 7 1,935 58.4% 70 Streetcar 4.33
College St and Lansdowne Ave 16 1,718 54.0% 80 Streetcar 6.00
Finch Ave W and Islington Ave 16 7,204 61.0% 108 Bus 6.50
Finch Ave W and Alness St 14 21 4,200 67.5% 108 Bus 6.50
Finch Ave E and Tapscott Rd 16 2,477 61.1% 98 Bus 2.80