RELATIONSHIP BETWEEN TRAFFIC OPERATIONS AND ROAD SAFETY By GUSTAVO RIENTE DE ANDRADE A DISSERTATION PROPOSAL PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2018
2018 Gustavo Riente de Andrade
To my family and to those whose ultimate sacrifice paved the way for safer roads
4 ACKNOWLEDGMENTS I would like to express my gratitude to the many people who supported me and made this dissertation possible. My special thanks to my advisors, Dr. Lily Elefteriadou and Dr. Coca Ferraz, for believing in this project from the beginning and welcoming me to their research groups. I would also like to extend these acknowledgements to all other professors who I had the opportunity to collabo rate with and contributed in my growth as a researcher and professional. Special thanks to Dr. Jos Setti, Dr. Cira Pitombo and Dr. Andr Cunha at USP, and to Dr. Siva Srinivasan, Dr. Yafeng Yin Dr. Washburn and Dr. Ruth Steiner at UF. My further apprecia tion goes to the Mc Trans Center, prof. Bill Sampson and Vishal Khanapure, for the financial support thought these years and for being a valuable lab where research and practice joined together. To the USDOT UTC Program under ST RIDE Center, that funded part of this research. Special thanks to Ines Aviles Spadoni. I am also thankful to t he So Paulo state transport agency ARTESP, toll highway association ABCR SP and affiliated highway concessionaries, which kindly provided the data for this research. Than ks to t h e financial support from the CNPq, Brazil I would a lso like to thank Tectran Group, which gave me the opportunity to experience real world problems and formulate the research questions that motived me to pursue a path to academia. My acknowledgeme nts go to everybody whose support was crucial the life as a graduate student. Special thanks to Heloisa Belo and Beth Ortega at USP and, at UF, to the wonderful staff at Weil 365 and UFTI Mrs. Debra, Blair, Doretha, Keri and
5 Jennifer. My thanks to suppor tive people from the International Center, my international adviser Mrs. Ethel Porras. Special thanks to Mrs. Debra Anderson, who kindly helped me navigate through the life in this new country. A section of this acknowledgements is rightfully dedicated to all of those who have worked hard to build the bridge between University of So Paulo and University of Florida, in the form of a novel PhD Dual Degree Agreement. Special t hanks to Dr. Ana Paul a Larocca and Cesar Derisso at USP, and to Nancy McIlrath at UF who demonstrated the good will and determination to go much beyond th e i r daily duties to make this possible. T hanks t o the many f riends I made who not only helped but made this journey a fun ride: Magaly Romo, Fernando Hirose, Fernando Piva, Gaurav Sult ania, Kiarash Fariborzi, Fabio and Renata Sasahara, Ash Omidvar, Marilo Martin Gasulla, Yinan Zheng and Zhibin Chen. My profound gratitude to my parents Silvestre and Regina my sister Carolina and family members, who have always incentiv iz ed me and inspir ed to go higher and higher. Finally, my loving thanks to my wife Karla and my son Rafael, whose tender souls and sheer determinations overcame every distance to support me in this endeavor.
6 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 8 LIST OF FIGURES ................................ ................................ ................................ .......... 9 LIST OF ABBREVIATIONS ................................ ................................ ........................... 10 ABSTRACT ................................ ................................ ................................ ................... 12 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 14 2 LITERAT URE REVIEW ................................ ................................ .......................... 16 Crash Prediction at Signalized Intersections ................................ ........................... 16 Crash Prediction on Expressways ................................ ................................ .......... 19 Safety Operations Relationship ................................ ................................ ........ 19 Safety Operations Modelling ................................ ................................ ............ 20 Summary ................................ ................................ ................................ .......... 21 Crash Severity Distribution on Expressways ................................ .......................... 25 Studies on M odeling C rash S everity U sing D iscrete C hoice M odels ................ 25 Summary ................................ ................................ ................................ .......... 27 3 EFFECTS OF MOBILIT Y, SAFETY AND EMISSIONS ON SIGNAL TIMING OPTIMIZATION ................................ ................................ ................................ ...... 31 Methodology ................................ ................................ ................................ ........... 32 Signal Timing Optimization ................................ ................................ ............... 32 Em issions Model ................................ ................................ .............................. 34 Test Scenarios ................................ ................................ ................................ ........ 35 Base Scenario ................................ ................................ ................................ .. 35 Alternative Sc enarios ................................ ................................ ....................... 37 Sensitivity Analysis ................................ ................................ ................................ 38 Application ................................ ................................ ................................ ........ 41 Results ................................ ................................ ................................ ............. 43 Discussion on the Trade off between Measures ................................ .............. 45 Ranking of Influencing Variables by Measure ................................ .................. 46 Impact of Weight Schemes on Signal Timing ................................ ................... 48 Final Remarks ................................ ................................ ................................ ......... 50 4 CRASH AND TRAFFIC DATABASE FOR EXPRESSWAYS SEGMENTS ............. 52
7 5 RELATIONSHIP BETWEEN CRASH RATE AND OPERATIONS ON EXPRESSWAYS ................................ ................................ ................................ .... 60 Relationship between Density and C rash Rate ................................ ....................... 61 Modelling the Relationship Between Density and Crash Rate ................................ 65 Influencing Factors ................................ ................................ ........................... 65 Proposed Model ................................ ................................ ............................... 69 Comparison Between the Proposed Model and the HSM ................................ 73 Discussion ................................ ................................ ................................ ........ 74 Final Remarks ................................ ................................ ................................ ......... 75 6 ESTIMATION OF CRASH SEVERITY ON EXPRESSWAYS ................................ 77 Methodology Overview ................................ ................................ ........................... 78 Results ................................ ................................ ................................ .................... 79 Marginal Effects ................................ ................................ ................................ ...... 81 Crash Type ................................ ................................ ................................ ....... 83 Time and Environment ................................ ................................ ..................... 84 Geometry and Design ................................ ................................ ...................... 84 Traffic Characteristics ................................ ................................ ....................... 85 Alternative Model ................................ ................................ ................................ .... 87 Final Remarks ................................ ................................ ................................ ......... 89 7 CONCLUSIONS ................................ ................................ ................................ ..... 91 LIST OF REFERENCES ................................ ................................ ............................... 95 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 101
8 LIST OF TABLES Table page 2 1 Selected Safety Model Details ................................ ................................ ............ 18 2 2 Database size and characteristics of previous crash prediction work. ................ 24 2 3 Database size and characteristics of previous crash severity work. ................... 30 3 1 Variables included in the sensitivity analysis and variation range. ...................... 38 3 2 Random trajectory generator. ................................ ................................ ............. 42 3 3 Kruskal Wallis test for all variables and performance measures. ....................... 47 3 4 Results for different optimization schemes, per intersection. .............................. 49 4 1 Frequency of main road characteristics, number of crashes and exposure. ....... 54 4 2 Discrete variables included in the database. ................................ ...................... 58 4 3 Continuous variables included in the datab ase. ................................ ................. 59 5 1 Influencing variables 2 Way ANOVA statistics. ................................ .................. 67 5 2 Model coefficients and ANOVA statistics. ................................ ........................... 70 6 1 Results and significant variables for each model. ................................ ............... 80 6 2 Probabilities estimated for each non continuous variable. ................................ .. 82 6 3 Results and significant variables Alternative MV and P&B Models. ................. 88
9 LIST OF FIGURES Figure page 2 1 Existing density crash rate models and LOS thresholds. ................................ ... 22 3 1 Configuration, demand and initial signal timing on testing arterial. ..................... 36 3 2 Factorial effects for each input parameter and performance measure. .............. 44 4 1 Studied freeways, multilane highways and main urban areas. ........................... 52 4 2 Distribution of crashes by class and severity. ................................ ..................... 55 5 1 Relationship between traffic density and crash rate, by severity. ....................... 62 5 2 Relationships between traffic density and crash rate, by class. .......................... 63 5 3 Distribution of single vehicle crashes for LOS A E vs LOS F. ............................ 64 5 4 Relationships between density and crash rate, by influencing variable. ............. 68 5 5 Analysis of standardized residuals. ................................ ................................ .... 71 5 6 Application of the proposed model. ................................ ................................ .... 72 5 7 Comparison of results from the proposed E quation and the HSM. ..................... 74 6 1 Comparison between the marginal effect of each variable. ................................ 83 6 2 Relationships between traffic density and crash severity distribution. ................ 85 6 3 Variability of the relationship between speed and SV crash severity. ................. 86 6 4 Relationship between density differences and MV crash severity. ..................... 87
10 LIST OF ABBREVIATIONS AADT Annual average daily traffic AASHTO American Association of State Highway and Transportation Officials ATM Active t raffic m anagement EPA Environmental Protection Agency FEM Factorial Effect Method FHWA Federal Highway Administration FI Fatal and injury crashes GA Genetic a lgorithm HCM Highway Capacity Manual HCS Highway Capacity Software HSM Highway Safety Manual LOS Level of s ervice MNL Multinomial Logit Model MV Multiple v ehicle c rashes MVMT Millions of vehicle miles travelled NGSIM Next Generation Simulation NMA National Motorists Association OL Ordered Logit Model OP Ordered Probit Model P&B Pedestrians and bicycles PDO Property damage only crashes RTOR Right t urn on r ed SV Single v ehicle c rashes TRB Transportation Research Board
11 v / c Volume to capacity ratio VMT Vehicle miles travelled
12 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy RELATIONSHIP BETWEEN TRAFFIC OPERATIONS AND ROAD SAFETY By Gustavo Riente de Andrade Dec ember 2018 Chair: Lily Elefteriadou Cochair: Antonio C P Ferraz Major: Civil Engineering Since before the releas e of the Highway Safety Manual research ha s been indicating the need to incorporate mobility and control aspects to road safety analysis The first part of this work developed and implement in an existing computational engine a signal timing optimization method that consider s mobility, safety, and emissions measures simultaneously. A sensitivity analysis was conducted to provide insight on the practical effect s and order of relevan ce of 20 key input variables Mobility improvement performance usually coincides with emissions improvements, but sometimes at the expense of safety. The second part of this work investigated the relationship between hourly traffic density and crash rates on Brazilian expressways with different characteristics based on a database containing over 20,000 crashes and more than 35 million traffic volume observations and The resulting curves for urban expressways follow a U shape, with minimum values associate d with LOS B to C while the relationships for rural expressways were found to be continuously increasing, suggesting that low volume rural roads are safer than the higher volume ones The analysis of other influencing
13 factors revealed that nighttime condi tions weaving segments and u rban multilane highways could be related to higher crash rates The third part of the project extends the analysis to crash severity modeling using an ordered response choice model. The framework that better fit this database led to the development of two different models: single vehicle crashes (SV) and multiple vehicle crashes (MV), since the factors that explain the severity of crashes vari es widely between these models. For instance, guardrails and barriers proved to effectively reduce severity for SV crashes, for which run offs are the most severe crash type. The unique database used in this study also allowed for an investigation of the i nfluence of prevailing traffic conditions on crash severity, while still controlling for all other factors. The results suggested that multiple vehicle crash severity is nega tively related with traffic density, while single vehicle crashes are more closely related to speed. The findings of this work have implications to policy and design decisions, and the produced equation could be incorporated to active traffic management (ATM) and HCM reliability analysis.
14 CHAPTER 1 INTRODUCTION The Highway Safety Manual HSM (AASHTO, 2011) has been the most cited reference for road safety studies. The core of its methodology is based on Safety Performance Functions which relate the annual average daily traffic to the predicted num ber of crashes (Daniel and Maine, 2011) Similar to other existing models, the HSM models were developed using aggregate traffic measures and are more suitable for planning level analysis Since before the release of the HSM, some authors have been indicat ing the need to incorporate operational characteristics into road safety analysis making it possible to account for the effects of seasonal, daily or hourly flow and speed fluctuations on crash rates (Lord et al. 2005 ) The development of such relationship s can enable practitioners to account for both operations and safety aspects in designing and operating highway facilities. These relationships seem to be affected by hourly traffic parameters, and may influence a variety of policy assessments, ranging fro m Active Traffic Management strategies to capacity increase decisions In this context, the goal of this research is to model the relationship between cr ash rates and different operati n g conditions for two different components of a road network: corridors with signalized intersections and expressway (uninterrupted) segments. Chapter 2 contains a literature review on crash prediction models for signalized intersections and expressways, summarizing the main conclusions that are relevant to this research.
15 In C hapter 3 it is describe d the development, software implementation and testing of a method that accounts not only for mobility, but also safety and emissions, to optimize signal timings along corridors. Next, C hapte r 4 details the unique large and varied database that was used in the two subsequent stages of this dissertation. First, C hapter 5 describes the development of crash prediction models that relate crash rates to hourly traf fic density and other roadway an d environmental conditions, based on a large and diverse database. Chapter 6 further develops this research, presenting crash severity distribution models that are compatible with the crash prediction models described in Chapter 4, and incorporate the effe ct of a wide range of characteristics to the severity distribution functions. Chapter 7 closes this dissertation, highlighting the contributions of this research project and making recommendations for future work
16 CHAPTER 2 LITERATURE REVIEW Crash Prediction at Signalized Intersections A large amount of work has been conducted to investigate safety at signalized intersections, usually using Poisson or Negative Binomial distributions to describe the relationships between geometry and crashes. Since t incorporate signal timing control features to develop methods that are able to predict crashes as a function of both geometry and operational aspects (Poch and Mannering, 1996; Mitra et al. 2007; Lambert, 1992; Chin and Quddus, 2003; Wang and Abdel Aty, 2006; Agbelie and Roshandeh, 2015). These include cycle length, phasing, control types (pre timed versus adaptive), coordination, protected versus permitted left turns and yellow and all red intervals. Increasing the numbe r of phases was found to increase the number of crashes in general, but providing a protected left turn on the arterial could improve safety in some instances (Poch and Mannering, 1996, Chin and Quddus, 20037, Agbelie and Roshandeh, 2015), while increasing rear end crashes when used on minor streets (Wang and Abdel Aty, 2006). Coordination was found to improve both mobility and safety conditions (Chin and Quddus, 2007). Enhanced display types have also been related to safety improvements. Although not used in the USA, studies on traffic signal countdown timers (Spigolon, 2010; Islam et al. 2017) have been showing that this configuration is able to lead to safer responses by drivers by reducing the tion has been employed to assess the interactions of safety and operation at intersections, using conflict based surrogate
17 measures to estimate crashes ( Archer 2004; Ozbay et al., 2008; Gettman et al., 2008; Es sa and Sayed, 2015; Shahdah et al. 2015 ). The Highway Safety Manual HSM ( AASHTO, 2010 ) presents comprehensive models to evaluate safety on rural and urban intersections, using multiple explanatory variables and regression to predict crashes. Although comprehensive in nature, the HSM methodology inc ludes few signal control parameters. The work conducted by Turner for the New Zealand transport a uthority ( Turner et al. 2012 ) moved one step ahead by combining operational and geometry parameters to estimate crashes by type. In the resulting equations, a base local scaling factor is multiplied by a series of variables that modify the number of estimated crashes as a function of demand, operations, and geometry characteristics: ( 2 1 ) where A is the predicted rate for type j crashes in a 5 year period; b i are model coefficients; x i are the independent variables of the model; and b 0 is a local adjustment factor. This model was selected for this study T he median of the b 0 defaults calculated by Turner et al. (2012) for different cities was used in this study Table 2 1 shows all which variables and their respective coefficients that are used by the safety model to calculate the number of predicted crashes per intersection approach and crash type. For example, the number of rear end crashes at a small intersection approach with no exclusive right turn, no split phasing, a cycle lane and a nearby bus bay would be calculated as: ( 2 2 )
18 Note that the multiplicative structure of Equation 2 1 allows for the future inclusion of additional factors, obtained from other studies. Potential variables could be HSM Crash Modification Factors (AASHTO, 2001) the number of phases ( Chin and Quddus, 2007 ), maximum greens ( Agbelie and Roshandeh, 2015 ) or the presence of countdown timers (Spigolon, 2010; Islam et al. 2017) Table 2 1. Selected Safety Model Details Factors x i Right angle Left turn against Rear end Loss of control Other small medium large Through traffic x 0.311 x 0.447 x .496 x 0.985 x 0.541 x 0.262 Conflicting crossing traffic x 0.362 Left turning traffic x 0.155 Degree of saturation x 0.397 x 0.447 Total number of lanes e (.356 x ) e (.243 x ) e (.459 x ) e (.144 x ) Number of through lanes e (0.352 x ) Total width x 0.027 Left turn storage length (1+ x ) 0.124 (1+ x ) 0.259 (1+ x ) 1.142 Presence of shared turns 1.19 1.26 Shared left turn 0.72 Exclusive right turn lane 1.585 Intersection depth (m) x 0.602 High speed (> 35 mph) 1.449 0.985 1.57 1.98 Cycle time (sec) x 0.037 x 0.683 x 0.704 x 0.354 All red (sec) x 0.636 Free right turn on red 1.442 1.227 1.17 1.16 Lost time (sec) x 3.424 x 0.209 x 1.739 Split phasing 0.69 5.256 1.57 0.95 2.47 1.21 Full left turn protection 0.71 Coordinated 1.31 0.71 Advanced detector 2.06 0.44 Mast arm signal display 0.74 Median island 0.67 1.22 Merge on intersection exit 1.47 0.65 Cycle facilities 1.35 0.706 0.753 1.257 Approach bus bay 1.309 0.908 1.6 1.27 Upstream parking 0.58 0.7 CBD area 0.9 0.819 1.83 Residential area 0.75
19 Crash Prediction on Expressways The literature review first describes studies that sought to examine relationsh ips between road and operational characteristics on crash quantities or rates, normally by means of statistical analysis. Those studies contributed to the body of knowledge in the field by indicating the best methods to conduct such investigations and by suggesting candidate variables that could be added i n a model Following such recommendations, t he second subsection presents the evolution of studies that propose user ready models that relate operation al variables to crashes. Finally, the summary contrasts the main results and recommendations from the wor k described herein. Safety Operations Relationship operations and crashes faced the lack of information at the time of the crash, resulting in the use of averages that may mask t he effect of prevailing flow, speed and other contributing variables to crash risk handful of studies were able to show with statistical confidence that prevailing traffic and environmental characteristics influence crash occurrence Martin (2002) found that crash rates on French freeways were higher on weekends, when traffic is low and trucks are restricted. Severity was found to be greater when hourly traf fic is lower and increasing the number of lanes was related to fewer crashes However, by using relationships between the total prevailing traffic flow and the number of crashes, traffic speeds and densities were not the same within the samples segments with different number of lanes According to the auth ors, functional forms that include the prevailing density or v / c offer a better characterization of the crash process (Lord et al. 2005).
20 A study that used data from California freeways (Golob and Recker, 2003) showed a connection between congested flow a nd increased crash risk. Fatal and injury crashes were grouped in one category labeled (FI), while the remaining incidents were grouped into a category labeled Property Damage Only (PDO) The authors found that FI crashes were influenced more by the prevai ling volume than by the prevailing speed Collisions were related to temporal variability in speed. Likewise, b y modeling the relationship between v/c and crash rates on freeways using the Poisson model Daniel and Maina (2011) showed that characteristics that affect capacity also affect safety Arona et al. ( 2015 ) went further, by investigating the possible effect of daylight, average traffic speed, occupancy, average headways and combinations of speed, relative speed and gaps in establishing relationship s between prevailing traffic conditions and crash rates on French urban freeways. The estimated performance functions for single vehicle crashes were found to be affected by speed, while multiple vehicle crashes are more closely related to occupancy, which is directly related to traffic density Safety Operations Modelling The first studies to model the relationship between traffic flows and crash rates fitted total accident rate hourly flow function followed the U shaped configuration. Kononov et al. (2011) proposed a sine function to model the crash rates in urban freeways in Colorado and California, which predicts slight increases in the number of crashes fo r low traffic flows and a sharp rise with increasing traffic U shaped functions were also used to describe the relationship between v/c and crash frequency on a freeway segment, a tunnel and a toll plaza in South Korea (Chang et al. 2000), as well
21 as an urban freeway segment in Detroit (Zhou and Sisiopiku, 2014) In these studies, the influence of ramps was not accounted for and was incorporated into the error term of the calibrated models Studies conducted more recently used models based on traffic dens ity for four urban regions of the USA (Harwood et al. 2013; Potts et al., 2014). Lower crash rates occur for moderate values of prevailing traffic density The resulting models describing the relationship between crash rates and v/c or traffic density are U s haped functions As in earl ier studies (Frantzeskakis and Iordanis, 1981; Ceder, 1982), this shape results from of a convex downward and a convex upward curve for single and multi vehicle accidents, respectively However, none of these studies propose models that can incorporate the statistically relevant factors to a single equation. Also, no study has investigated the to the present any effort to examine these relatio nships in the Brazilian context S ummary Studies that investigate the relationship between safety and operations suggest that congestion has an impact on crash rates. Several studies indicate that crash rates also increase under low volume condition s, resulting in U Shaped functions describing the relationship between crash rates and v / c or traffic density. Regarding severity, the literature recommend s grouping fatal and injury crashes (FI), the remaining incidents being Property Damage Only (PDO) (Golob and Recker, 2003). With respect to the best mobility measure to relate to crash rates, the literature indicates that the v/c or traffic density (Chang et al. 2 000; Daniel and Maina, 2011; Lord
22 et al. 2005) is preferable over flow. The use of traffic density has the additional advantage of better handling undersaturated and oversatured flows (Martin, 2002). Regarding best practices in collecting traffic data for studies to cor relate crash occurrence to prevailing traffic operational conditions, the use of 5 min intervals was found to be ideal by a previous study (Pande et al. 2005) Also, combining observations from all lanes was better than the use of lane by l ane data, because this not only captures across lane variation in speed or volume but also allows the use of a larger data set for analysis, since data from the specific crash lane is often missing. The only two USA practice ready models relating traffic d ensity in pc/mile/lane to crash rates in crashes per million of vehicle miles travelled (MVMT) are shown in Figure 2 1 The graph also shows the boundaries for the levels of service (LOS) for freeways and multilane highways, according to the Highway Capaci ty Manual (HCM) 6th Edition (TRB, 2016). Note that the order of magnitude of both curves is similar and that the lowest crash rates are associated with levels of service B to C. Figure 2 1 Existing density crash rate models and LOS thresholds
23 Table 2 2 provides an overview of the most relevant studies on the relationship between traffic operations and crash rates, listing both the dimensions of the databases that were used in each study and the variables considered. The la st column shows the corresponding values for this study, as a point of comparison Note that, while most studies used the relationship between measured flows and speeds to calculate the traffic density, Arona et al. (2015) obtained densities by measuring o ccupancies. Most of these studies relied on a limited amount of data and/or include a limited number of roadway and environmental characteristics. Some studies rely on data from a single facility, which provides valuable insight but is insufficient t o generate broader conclusions. As a result, in spite of the existent research on the effects on safety of many of the variables shown in Table 2 2 no study has incorporated statistically relevant factors into a single equation in the format presented in Fi gure 2 1. assessed. Finally, most models focused on urban freeways, and possible differences on su ch relationships for rural freeways remains to be investigated.
24 Table 2 2 D atabase size and characteristics of previous crash prediction work Martin (2002) Golob and Recker (2003) Lord, et al. (2005) Daniel and Maine (2011) Potts et al. (2014) Zhou and Sisiopiku (2014) Arona et al. (2015) This Study Location France Southern California Canada New Jersey 5 US Large Cities Detroit Area Marseille France So Paulo Brazil Facility Multilane Highways X Freeway Segments X X X X X X X X Ramps X Weaves X Traffic Stations 92 NA 9 987 1226 36 208 187 Crash Data Count 11,172 1,200 1,296 NA 42,740 7,731 292 21,778 Severity X X X X X X X Num of Vehicles X X X X Type X X X X Traffic Count 2,900,000 1,200 NA 97,834,433 3,983,401 18,220,800 36,830,195 Interval 6 min 30 sec 5 min 1 hour 15 min 15 min 6 min 5 15 min Method Negative Binomial PCA NLCCAA Poisson Negative Binomial / Poisson Regression Regression Logistic Regression Regression Traffic Volume X X X X X X Truck Percentage X X X X Free Flow Speed X X X Avg. speed X X X X X X Environment Lighting/Time X X X X X X Weekend X X X X X Work Zones X X X Rural Areas X X Urban Areas X X X X X X X X Geometry Number of Lanes X X X X X X Lane/Shoulder Width X X X Design Speed X X Toll Station X X Total Ramp Density X X X Guardrails / Barriers X X PCA Principal Component Analysis NLCCAA Nonlinear Canonical Correlation Analysis
25 Crash Severity Distribution on Expressways Several approaches have been used to investigate crash severity and the factors that affect it. While several nonparametric models have been used to model the discrete nature of the severity output, such as the boosted regression tree and other tree based models (Lee and Li, 2015), most studies have used discrete response models to address this topic. These are discussed in the next subsections. Studies on M odeling C rash S everity U sing D iscrete C hoice M odels The ordered probit (OP) and logit models (OL), already common in econometric and Connor, 1996; Kockelman and Kweon 2002; A bdel Aty, 2003; Yamamoto and Shankar, 2003), using da ta from roadway segments, intersections and toll plazas. In general, crashes were found to be more serious for older drivers, males, speeding drivers, and passengers without seat belt or involved in angle crashes. For the case of road segments, sharp curve s and poor lighting were found to contribute to more severe crashes. The authors of these studies suggested that the reason more serious crashes occur in rural areas may be due to the higher speeds, though speed or other prevailing traffic conditions were unknown. Other studies have used different types of discrete choice models, such as the mixed logit model (MLM), and the concept of random variables to account for unobserved effects caused by the lack of prevailing traffic data (Milton et al. 2007). Othe rs relied on the simpler multinomial logit model (MNL) (Geedipally et al., 2013; Penmetsa and Pulugurtha, 2018). Statistically relevant factors from these studies include the number and radius of curves, pavement friction, interchanges/intersection density rumble strips, barriers, underpass, on ramps, area type (rural/urban), and
26 weather. Traffic data used include daily traffic, average daily truck traffic, and truck percentages, however, prevailing traffic conditions were unavailable in these studies. The hypothesis that high speeds and congestion might affect crash severity has led to the inclusion of prevailing traffic conditions in the set of explanatory variables used in di screte choice modeling (Quddus et al. 2010). Crash and traffic data from London ring road freeway were used in an ordered response model to investigate the effect of congestion on road safety. Higher volumes were associated with less severe crashes, probably due to the lower speeds usually observed as demands increase. Other variable s found to be statistically significant include curve radius, number of lanes, day of the week, lightning, road surface condition, number of vehicles involved, and year. Many authors adjusted independent models for different classes of crashes. While Quddu s et al. (2010) proposed a model for single vehicle (SV) crashes only, Wu et al. (2014) conducted separate analyses for SV and multiple vehicle crashes (MV) on rural two lane highways in New Mexico. The authors concluded that the way the explanatory variab les affect each model varies significantly for SV or MV crashes. Eleven variables were found to significantly affect driver injury severities for the MV model, albeit only seven of them could be related to the SV crash model. Some studies focused on crash severity for one vehicle type. For motorcycles, Coutinho et al. (2015) used ordered response models to identify crash severity related factors for an urban environment, using data from the c ity of Fortaleza, Brazil. Chung et al. (2014) focused on Korean de livery motorcycles. The results highlighted that, for
27 motorcycle crashes, severity is more strongly related to human factors and nighttime conditions. The question of what model type would best predict crash severity as a function of different sets of data prompted several researchers to focus on model comparison studies. Xie et al. (2009) used Bayesian inference into an ordered probit model (BOP), concluding that this approach could produce better results for small samples. For rural single vehicle crashes in Florida, the latent class logit (LCL) was compared to a traditional multinomial logit as a benchmark (Xie et al. 2012). Both modeling alternatives were able to capture the impact of each significant explanatory variable. Simpler models such as the ord ered response models are preferable to more complex ones. From the 53 variables tested, 31 showed to be statistically significant at 95% confidence. Other studies (Yasmin and Eluru, 2013; elik and Oktay, 2017) tested several ordered response frameworks to the classical unordered response framework such as the multinomial logit for crash severity. Comparison of results indicated that the ordered model outperforms the multinomial logit. Similarly, Ye and Lord (2014) concluded that the unordered logit require s larger sample sizes to produce statistically reliable results, while the ordered, the smaller sample sizes. Summary There is no consensus among the authors of previous studies regarding the superiority of any specific discrete choice method over the othe rs, as the results of each study are dependent on the respective data used (Xie et al. 2012; Yasmin and Eluru, 2013; elik and Oktay, 2017; Ye and Lord, 2014). The ordered response model was therefore chosen for this work for it proved in previous work to provide reliable
28 responses for various types of datasets, while keeping a simple structure. Also, ordered response models have the advantage of naturally follow ing the logic of crash severity, in which the outcome is discrete and ranked in an order, in the casa, fatalities, injuries and property damage only crashes Differences between the severities of single vehicle versus multiple vehicle crashes were identifi ed in many studies (Quddus et al. 2010; Wu et al., 2014) While some treated this characteristic as an additional variable, others advocate that separate models should be developed for each case, given the distinct nature and possible causes for each of t hese types of crashes. Some studies focused on investigating crash severity for specific types o f vehicles, such as motorcycles (Chung et al. 2014; Coutinho et al. 2015). Table 2 3 lists the most relevant studies from the literature, the size of the cras h database used and period of study. The types of facilities included in each study is also shown, as well as the methods employed. The variables were grouped as crash characteristics, traffic and vehicle data, human fac tors, environment and geometry Most studies focused on generic roadways, and a few included intersection and control information as explanatory variable s or by creating models for intersections. Few studies tested for differences between freeways, ramps and weaving segments. Many studies ra ised the question whether speed or congestion (traffic density) influence crash severity, by analyzing indirectly related variables, such as area type, environment and visibility variables, estimates of speed at the time of the crash and annual average dai ly traffic (AADT). However, only Quddus et al. (2010) relied on a
29 database that could combine prevailing traffic conditions to other roadway and environmental data, for one facility in the UK. This research intended to analyze crash severity on several typ es of expressway segments in the light of a large database, covering an eight year period and counting on detailed prevailing traffic information and a comprehensive set of environment and geometry variables. As a limitation of this study, human factors we re not available in the database.
30 Table 2 3 Database size and characteristics of previous crash severity work and Connor (1996) Abdel Aty (2003) Milton et al. (2007) Kockelman and Kweon (2008) Quddus, Wang and Ison (2010) Geedipally et al. (2013) Coutinho et al. (2015) elik and Oktay (2017) Penmetsa and Pulugurtha (2018) This Study Location Aus F L WA State US A London UK CA M E and W A CE Brazil Turkey N C state S P Brazil Facility Road Segments X X X X X X X Local Streets X X Freeway Segments X X X Ramps X X Weaves X Intersections X X X X Crash Count 18 069 7891 22,568 NA 3 998 NA 3 232 11771 728801 21778 Period 91 96 97 90 94 98 2003 06 2004 07 2004 11 2008 13 2011 13 2005 13 Type X X X X Method OPM, OLM OPM MLM OP M OLM MNL, NMN L OPM, OLM OLM MNL OLM Traffic and Vehicle AADT X X X X Average Speed X X Truck Percent X X Num of Veh icle X X X Vehicle Type X Car X Moto. Heavy X Vehicle Age X X Private Vehicle X X Human Factors Driver Age X X X X X Gender X X X X Belt / Helmet X X X Alcohol Use X X X Speeding X X Environment Lighting X X X X X X X Weekend X X X X Weather X X X X X Work Zones X Urban/Rural X X X Geometry Num of Lanes X X X Lane/Shoulder X Grades X X X Curve X X X X X Speed Limit X X X Bridge /Tunnel X X Rumble Strips X Guardrails X X Note: MNL Multinomial Logit Model OPM Ordered Probit Model NMNL Nested Multinomial Logit OLM Ordered Logit Model MLM Mixed Logit Model
31 CHAPTER 3 EFFECTS OF MOBILITY, SAFETY AND EMISSIONS ON SIGNAL TIMING OPTIMIZATION The optimization of signal timings has been traditionally based on mobil ity performance measures only. There has been an increasing interest in utilizing signal timing optimization methods that can consider mobility, safety, and emissions measures simultaneously. The goal of this research is to develop and implement in an existing computational engine, the Highway Capacity Software (HCS), a signal timing optimization method that can consider mobility, safety, and emissions measures simultaneously. The HCS module Streets employs a genetic algorithm (GA) to optimize signal t imings, using a single objective function traditionally based on mobility measures. As part of that deployment, this research investigated forms of incorporating safety and emissions measures in the HCS GA optimization, while keeping the simplicity of its single objective function. This required the development and implementation of new models to estimate safety and emission measures to supplement the existing HCS procedures that estimate mobility measures. The introduction of the new models increased the c omplexity of the required inputs and the relationships between the inputs and outputs. This resulted in a large increase in input parameters to include factors related to geometric design, demand, control, optimization, and other features that impact the s ignal optimization results. It is important to understand the relative impacts of these features on the results of the signal timing optimization and the trade off s between the resulting mobility, safety, and emission measures. This understanding will allo w the designer, signal control engineer, and traffic analysts to consider these impacts when designing intersection geometry and when optimizing signal control. This research present s a method to optimize mobility,
32 safety, and emission measures simultaneou sly and the application of this method in the HCS. It also pr oposes sensitivity analysis method to provide insight on the effects of 20 key variables on mobility, safety, emissions and the ir trade offs. The next section summarizes the optimization function components, focusing on crash prediction models for intersections, the emissions model used, and the selected sensitivity analysis method selected for this work. Next, the test scenarios and the optimization procedure are described. The following section discusses the sensitivity analysis results. The last section summarizes the conclusions and limitations of the research, providing recommendations for future work Methodology Signal Timing Optimization The existing procedure in the Streets module of the HCS incorporates a GA optimization algorithm to determine the offset, optimal cycle length, phasing sequence and splits. GAs are adaptive methods widely used to solve optimization problems using a direct representing a possible solution to a given problem. Individuals are composed by genes, each representing an input variable of the model. Each individual is assigned a c indiv The least fitted members of the population are less likely to be selected for
33 reproduction, et al. 1993). Stevanovic et al. (2015) utilized GA in a signal timing optimization that simultaneously considers mobility, safety and emissions measures into a 3 dimentional objective fun ction (Pareto Front). The study concluded that changes in the weights assigned to these measures did not produce significant differences in the results. This study extended the GA optimization procedure of the HCS to consider all signal parameters in the o ptimization of the offsets, phasing sequence, split phasing use, Dallas phasing use and cycle length. Cycles between 60 and 160 sec were allowed in the search for the optimal solution. The maximum value of 160 is 33% higher than the default recommended by the Highway Capacity Manual (TRB, 2016) for major arterials with left turn phases in one street, allowing for a range of longer cycle lengths considered in the search region of the GA. For the GA, the utilized population size was 10, with 30% crossover pr obability, a minimum 4.0% mutation probability, and 1% convergence threshold, as suggested by the HCS user guide. A maximum of 2,000 generations was used, which led to the convergence to the global optimum solution in most cases While the existing GA proc edure in the HCS was kept, its objective function was modified to add the safety and emissions components to the optimization engine. Previous studies utilized a single objective function that combines multiple objectives in the GA optimization (Azar et al 1999; Yan et al., 2013; Stevanovic et al., 2015). The fitness function for this study, which was based on the method proposed by Azar et al. (1999) is:
34 ( 3 1 ) where w delay w safety and w emissi on are the weights to be used for the three performance measures ; is the estimated overall delay (sec/veh), c alculated by the HCM procedure; is the estimated total number of crashes over a five year period; is the estimated emissions (g of gases); is the minimum value for the performance measure i ; and is the maximum value for the performance measure i The HCM methodology was used for predicting delays. The crash prediction model by Turner et al. (2012) was selected The emissions equations are presented in the following subse ction Emissions Model This work used the model developed to estimate emissions by Chen et al. ( 2016 ) and Hadi et al. ( 2017 ). The model uses regression analysis, based on the outputs of the Environmental Protection Agency (EPA) MOVES, using information on the trajectories collected by the Federal Highway Administration (FHWA) as part of the Next Generation S IM ulation (NGSIM) program as inputs. More details about the development of the utilized model can be found in Hadi et al. (2017) The general function for emission estimation is: ( 3 2 ) where B 1 B 2 and B 3 are coefficients of the independent variables. In summary, the best prediction models for various environmental measures are presented in Equation 3 3 to Equation 3 6
35 ( 3 3 ) ( 3 4 ) ( 3 5 ) ( 3 6 ) Note that the VMT is given in mi les and the speed given in mph. Test Scenarios A series of scenarios that consist of a base condition and variations to each input that might affect the outcome of the optimization procedure were used to test the proposed approach At the end of each simulation, the following performance measures were obtained: (a) delays, to reflect mobility performance; (b) the number of predicted crashes to r eflect safety; and (c) grams of gases, to quantify emissions. Base Scenario A hypothetical three intersection arterial was used for testing. This layout was chosen to allow for the optimization of arterial coordination parameters while keeping an acceptable degree of simplicity. All three intersections have the same configuration ( Figure 3 1). Each arterial approach has two lanes with a 200 ft left turn pocket and a shared right and through lane. There are no merges on the intersection exit sides. The side streets have two shared lanes. The intersection spacing is 1,300 feet. The le ngth of all other segments is 1,000 feet. The speed limit of the arterial is 45 mph and that of the side streets is 30 mph. The base traffic demands are also shown in Figure 3 1. The demand was determined iteratively after a series of tests using the HCS G A optimization, so that the
36 level of service along the arterial ranged from C to E in the optimized base scenario, when considering all three objectives (m obility, safety and emissions). Figure 3 1 Configuration, d emand and i nitial s ignal t iming on t e sting a rterial The parameters of the control setup used for the base scenario are: Coordinated arterial, with semi actuated control and detectors present at each side street approach and protected left turn movements; Yellow time of 3 sec and red clearance of 2 sec for all movements; No Right Turn on Red (RTOR); Exclusive left turn phasing used for the main street and permissive only left turn phasing used for the side streets (no split phasing). The initial signal timing was the one optimized for the base scenario ( Figure 3 1 ). The following characteristics have significant impacts on the safety model only ( Turner et al. 2012 ) and were set as described below:
37 Intersection width of 110 feet: This is the distance between the stop bar on the approac h and the beginning of the downstream intersection leg. This variable defines intersection size, as: (a) Small: depth of 82 ft or less and 1 or 2 lanes per approach; (b) Medium: depth between 82 ft and 131 ft and 2 lanes per approach; and (c) Large: depth larger than 131 ft and 3 or more lanes for at least two approaches. Median islands along the arterial; Overhead mast arm signal displays at each intersection; No parking along any of the links; No bus bays or bicycle facilities. Other variables that affect safety, including degrees of saturation X and lost time per cycle L C are calculated iteratively as intermediate results and are accounted for in the final iteration. Alternative Scenarios For the creation of testing scenarios, three levels were considered per variable. T able 3 1 lists all variables included in the testing, along with their range and variation levels. The highlighted values are the inputs used in the base scenario. The following can be observed regarding the variations of these variables, a s listed in Table 3 1 : Volume Increments: The base volumes were decreased by applying multipliers of 0.9 and 0.8 to the base traffic flows entering from both the arterial and minor streets. As indicated earlier, oversaturated conditions we re not considered in this study; no protected phases are allowed; RTOR: Two RTOR scenarios were considered. In one of them, the t raffic volumes that could turn on red correspond to 25% of total right turning movements, while in t he other, 50% could turn on red; Bicycle Facilities, Parkings and Bus Bays: Indicates the existence of one of those elements within 330 ft from the stop b ar ; and
38 Weights: Three sets of scenarios were established for testing the weights in the optimization objective function. For each performance measure, these alternative schemes were tested: one variable dominance (50/25/25, 25/50/25 or 25/25/50), or two var iable dominance (40/40/20, 40/20/40 or 20/40/40). Table 3 1 Variables i ncluded in the s ensitivity a nalysis and v ariation r ange Domain N Variable Unit Levels Range 1 2 3 Design 1 Intersection spacing feet 700 1300 1900 1200 2 Intersection s ize 2 3 4 2 3 Left turn pockets feet 0 200 400 400 4 Shared right through lane No Yes 5 Shared left through lane No Yes 6 Arterial p osted s peed mph 35 45 55 10 Demand 7 Volume increments (multipliers) veh/h 0.8 0.9 1 0.2 8 Right turn percentages % 5% 10% 15% 0.1 9 Left turn percentages % 5% 10% 15% 0.1 Control 10 Protected l eft No Yes 11 Right t urn on r ed (% right turn flow) % 0 25% 50% 0.5 Safety factors 12 Mast arm signal display No Yes 13 Median island No Yes 14 Bicycle facilities No Yes 15 Bus bays No Yes 16 Parkings No Yes 17 Area type is residential No Yes Weight 18 Mobility % 33/33/33 50/25/25 40/40/20 25 19 Safety 33/33/33 25/50/25 20/40/40 25 20 Emissions 33/33/33 25/25/50 40/20/40 25 Sensitivity Analysis A sensitivity analysis was undertaken to gain a better understanding of the interactions among the independent variables on safety, mobility, and emissions and the trade offs between the three performance measures Although the inputs are known, the magnitude of the effect of each input for different scenarios is unknown. To categorize the degree of importance of each input of the model, sensitivity analysis is recommended ( Kuehl, 1999 ). Sensitivity analysi s examines to what extent the outputs
39 of a model depend on its inputs, by measuring the variation of the results as a function of different parameters ( Trucano et al., 2006 ). This type of analysis also helps identify interactions between variables of the m odel ( Saltelli et al., 2004 ). Different classes of sensitivity analysis methods exist for different purposes ( Santner et al., 2003 ). Local sensitivity analysis methods are used to identify trends for a specific variable within a given sample space, while g lobal sensitivity analysis considers the entire population. In this work, 20 variables that could affect the model outcome were identified eleven to be tested at two variation levels and nine to be tested at three variation levels In order to cover all c ombinations of values for the 20 inputs, (3 9 ) (2 11 ) = 4 10 7 scenarios would have to be generated which was unfeasible for the available computational resources. In this case when the sear ch area is too large to be tested in its entirety screening method s are recommended ( Saltelli et al., 2000). The best known screening method is the classical Ceteris Paribus in which one factor is varied within a range at a time, with all other variables held constant. A statistically significant sample is calculated fo r the number of scenarios to be constructed. Since only one variable is changed at a time, this method has limitations regarding the identification of interactions between variables, and restrictions in terms of possible scenarios. To remedy this situatio n, the Factorial Effect Method (FEM) was proposed by Morris ( 1991), further developed by Campolongo et al. (2007 ) and successfully applied in ano ther transportation engineering stud y ( Nunes, 2012 ). It is based on the concept of factorial effects d i (x) define d for each input and measured analogously to price elasticities:
40 ( 3 7 ) where d i ( x ) is the factorial effect of the variable i ; MOE(x k ) are the performance measure s for a vector on input data k ; and i is the v a riation of the factor i between vector x k and x k 1 generation. The first row of the matrix reflects an initial vector of inputs, generally corresponding to a base scenario. The second row is defined by changing one variable by i Likewise, the third row of the matrix is constructed by varying another variable This is done successively, never changing the same vari able twice, until all inputs have been shifted. The order in which each variable is changed is random for each trajectory, so that no two matrices are the same. A trajectory j of three variables model would be represented as: ( 3 8 ) For each trajectory, each factor is tested only once. To have statistically significant results, multiple trajectories must be constructed. From the previous experience by Morris (1991 ), a minimum of 10 trajectories is sufficient for a reliable analysis, while more than 30 trajectories di by Nunes (2012 ). When all vectors from all trajectories j are simulated, the factorial effects d i,j (x) of all variables are compiled, forming the distribution F i The absolute values of d i,j (x) form the G i distribution. Based on those distributions, sensitivity measures a re calculated for each variable :
41 ( 3 9 ) ( 3 10 ) ( 3 11 ) where r is the number of trajectories j The main advantages of FEM are: Sampling: it produces randomized samples of variables so that, differently from the pure Ceteris Paribus method, different combinations of variables are tested. This results in distribution of effects that allow the researcher to verify not only the mean effects of each variable, but also variance, biases and normality assumptions; By using a measure analogous to price elasticities, different variables are brought to a similar unit and can be compared and ranked in or der of importance to the model. Application The application of the FEM is based on the construction of trajectories, composed by vectors of input data. Each trajectory is represen ted by a matrix wi th 20 columns (number of variables) and 2 1 rows, in which the first row reflects the base scenario and all other rows represent variations of each variable within one of the levels ( Table 3 1 ). Table 3 2 was developed to generate random t rajectories. From the base scenario, each consecutive vector is composed by varying one single input (highlighted), to one of the possible levels. Selected levels and the order in which the variables are assorted are randomly defined. A second matrix was a lso constructed, containing unfeasible or unreasonable combinations of inputs. Examples of these are increasing left turn pockets while shared
42 left through lanes are in use, or the use of exclusive right and left turns when the arterial has two lanes, whic h would imply there are no through lanes. The trajectory generator ( Table 3 2 ) then constructs random scenarios until all vectors are valid. Table 3 2 Random t rajectory g enerator Intersection spacing Intersection Size Left turn pockets Shared right t urn Shared left turn Posted Speed Volume increments Right turn percent Left turn percent Protected Left Right turn on red Mast arm display Median island Bicycle facilities Bus bays Parkings Area is residential Mobility Safety Emissions Base V ector 2 2 2 2 1 2 3 2 2 2 1 2 2 1 1 1 1 1 1 1 Median is land 2 2 2 2 1 2 3 2 2 2 1 2 1 1 1 1 1 1 1 1 Left turn percent 2 2 2 2 1 2 3 2 3 2 1 2 1 1 1 1 1 1 1 1 Right turn percent 2 2 2 2 1 2 3 1 3 2 1 2 1 1 1 1 1 1 1 1 Shared right turn 2 2 2 1 1 2 3 1 3 2 1 2 1 1 1 1 1 1 1 1 Mast arm display 2 2 2 1 1 2 3 1 3 2 1 1 1 1 1 1 1 1 1 1 Right turn on red 2 2 2 1 1 2 3 1 3 2 3 1 1 1 1 1 1 1 1 1 Bicycle facilities 2 2 2 1 1 2 3 1 3 2 3 1 1 2 1 1 1 1 1 1 Posted Speed 2 2 2 1 1 3 3 1 3 2 3 1 1 2 1 1 1 1 1 1 Intersection spacing 3 2 2 1 1 3 3 1 3 2 3 1 1 2 1 1 1 1 1 1 Volume increments 3 2 2 1 1 3 2 1 3 2 3 1 1 2 1 1 1 1 1 1 Bus bays 3 2 2 1 1 3 2 1 3 2 3 1 1 2 2 1 1 1 1 1 Safety 3 2 2 1 1 3 2 1 3 2 3 1 1 2 2 1 1 1 2 1 Protected Left 3 2 2 1 1 3 2 1 3 1 3 1 1 2 2 1 1 1 2 1 Shared left turn 3 2 2 1 2 3 2 1 3 1 3 1 1 2 2 1 1 1 2 1 Left turn pockets 3 2 3 1 2 3 2 1 3 1 3 1 1 2 2 1 1 1 2 1 Area is residential 3 2 3 1 2 3 2 1 3 1 3 1 1 2 2 1 2 1 2 1 Parkings 3 2 3 1 2 3 2 1 3 1 3 1 1 2 2 2 2 1 2 1 Emissions 3 2 3 1 2 3 2 1 3 1 3 1 1 2 2 2 2 1 2 2 Intersection Size 3 1 3 1 2 3 2 1 3 1 3 1 1 2 2 2 2 1 2 2 Mobility 3 1 3 1 2 3 2 1 3 1 3 1 1 2 2 2 2 2 2 2 Each scenario (v ector) was simulated, using the levels shown in Table 3 2 and their correspondent values shown in Table 3 1 For each run, the resulting performance measures for mobility, safety and emissions were recorded. The factorial effect of each variable is then calculated as per Equation 3 7 and the three statistics given in Equations 3 9 to 3 11 were calculated. T he higher the the higher the indication that a variable has a significant impact on a performance measure.
43 Result s Figure 3 2 graphically shows the results obtained from a total of 30 trajectories ( 30 trajectories 20 variables = 6 0 0 individual scenar ios), in terms of the distribution of the factorial effects across all trajectories. A positive sign indicates a positive relationship between the variable and the performance measure (dime nsionless, analogously to price elasticities). For instance, an inc rease in traffic volume is associated with increases in delay, crashes and emissions. For practical purposes, variables with a negative sign for the mobility and emissions indicate improvements in all performance measures (less overall delay, emissions and fewer crashes). For example, for the intersection size, the higher the number of lanes, the lower the delay and emissions. The degree of improvement for these variables oscillated between scenarios, as reflected by the dispersion of the results (higher va riability). use of a protected left phase increased the delay at a rate ranging from 2.5 to 7 sec/veh (median of 5). Demand level variables, such as the volume multipliers, are directly related to all performance measures. For most scenarios, through movements were the critical ones, and the use of a lane for shared turns and through movements enhanced mobility and emissions, albeit often at the expense of safety. The emissio n results followed a similar trend to that observed for the mobility domain. For instance, RTOR leads to a slight reduction of emissions that could be considered statistically more significant than at least ten other variables, at 95% confidence.
44 Fi gure 3 2 Factorial e ffects for e ach i nput p arameter and p erformance m easure.
45 Intersection spacing was found to increase emissions more than other variables. This is expected, since longer segment lengths result in more driving and higher emissions. Giving more weights to specific performance measures (variables Mobility, Safety and Emissions, respectively) had a limited effect on the results. This suggests that the optimum solutions attained by the proposed algorithm are only marginally affected by t he user defined weights w delay w safety and w emission Discussion on the Trade off b etween Measures While the impact of some variables, such as demand levels, is straightforward for all performance measures, for many others the factorial effects deviate ar ound 0. This suggests that the influence of that particular variable on a specific performance measure depends on the combination of other variables (i.e., the particular scenario analyzed), and a trade off among different measures may exist. For example, for the case of permissive versus exclusive left turns, two different effects were identified: the use of permissive left turns creates an additional conflict that increases the risk of left turn against crashes. However, depending on the traffic flows per approach/movement, t his scheme has the potential to reduce the degrees of saturation, phase complexity and cycle time, improving the intersection safety. Due to this trade off, depending on the scenario, there were overall increases/decreases in the numbe r of crashes. Larger intersections were associated with more angle an d rear end crashes when no left turn pockets were present. Shared Left Turns increase the number of predicted crashes when used to replace an exclusive left turn lane with a pocket. On th e other hand, for the case of large intersections (4+ lanes on the arterial) with short/no pockets provided for the left turn,
46 the use of shared left turns was found to be a better design The effect of shared right turns versus exclusive right turns is si mpler. In this case, the use of shared right turns decreases delay at the cost of safety for all simulated scenarios. Emissions are a function of three main factors: distance vehicles travelled (VMT), number of stops, and average speed. The intersection si ze was defined by the number of lanes, which in turn affects the road width. Depending on the scenario (combination of variables), better intersection performance is obtained with the addition of extra lanes and that causes a decrease in the number of stops and an increase in the average speed, reducing emissions. On the other hand, when the extra capacity is unnecessary because of lower degrees of saturation, the additional lanes result in extra distance to be covered between the approaches (road width ), increasing VMT and, thus, emissions. Ranking of Influencing Variables by Measure Although the central and dispersion measures shown in Figure 3 2 provides some insight on the magnitude of the impact of each variable, the dispersion of some results sugg ests that one to one comparisons of variables with similar effects is not straightforward. For that reason, a multiple pairwise comparison statistical test was performed. The analysis used the absolute factorial effects ( ) of each variable, which is cons idered to be the most statistically significant measure in the method ( Campolongo et al., 2007 ). In a preliminary analysis, it was noted that the distribution of did n o t follow a normal (Gaussian) pattern. This happens because the interaction between va riables generates scenarios for which the results diverge from the mean unexpectedly, as discussed in the previous subsection. For that rea son, the non parametric Kruskal Wallis test was chosen, which accounts for the frequency that each variable assumes a
47 specific position in the overall ranking. The results for the test, at the 95% confidence level, are shown in Table 3 3 The variables are ranked in order of importance, per the Kruskal Wallis test criteria, according to the number of times each one of th em figured among the most relevant to the model outcomes. The higher value of the test statistic H as compared to the critical 2 indicates that H 0 can be rejected for all performance measures, and at least two variables would differ from each other. To sp ecify which ones, pairwise comparisons were made. Table 3 3 shows which groups of variables could be ranked higher relatively to the others. Table 3 3 Kruskal Wallis t est for all v ariables and p erformance m easures Delay Number of Crashes Emissions 2 H 2 H 2 H 42. 6 271 42.6 157 42.6 421 Variable Rank Variable Rank Variable Rank Protected Left 1 Volume increments 1 Intersection spacing 1 Intersection Size 1 Left turn pockets 1 Intersection Size 1 Volume increments 1 Intersection Size 1 Shared right through 1 Shared right through 1 Mast arm signal display 1 Shared left through 1 Intersection spacing 2 Median island 1 Volume increments 1 Shared left through 2 Shared right through 1 Left turn percentages 2 Area type is residential 2 Bus bays 1 Right turn percentages 2 Left turn percentages 2 Parkings 1 Arterial Posted Speed 3 Mobility 2 Right turn on Red 1 Protected Left 4 Right turn percentages 2 Shared left through 1 Right turn on Red 4 Emissions 2 Bicycle facilities 2 Area type is residential 4 Safety 3 Area type is residential 3 Mobility 5 Arterial Posted Speed 3 Protected Left 4 Emissions 6 Right turn on Red 3 Arterial Posted Speed 4 Safety 6 Left turn pockets 4 Left turn percentages 4 Bicycle facilities 6 Median island 4 Mobility 4 Median island 6 Mast arm signal display 4 Emissions 5 Mast arm signal display 6 Bicycle facilities 4 Right turn percentages 5 Left turn pockets 6 Bus bays 4 Intersection spacing 5 Bus bays 6 Parkings 4 Safety 5 Parkings 6
48 Impact of Weight Schemes on Signal Timing The final stage of this research assessed the impacts of weight schemes on the signal timings. Using the base scenario (Figure 3 1), three additional scenarios were generated, giving all the weight to one performance measure at a time. The resulting performance per intersection and weight scheme are shown in Table 3 4. As expected, providing all weight to the mobility measure resulted in the best overall operational performance. When the weig ht is assigned to either safety or mobility, the movements with the highest traffic demands (usually the main street) have longer green times and the minor streets and left turn movements have the shortest green times. The phasing sequence was kept unalter ed by the algorithm in all cases, but the cycle length varied significantly with the changes in the weights on the three measures: 65 sec for the 100% mobility weight, 160 sec for the safety scheme and 150 sec for the emissions scheme. The selection of sh orter cycles for optimal mobility performance can be explained by the lower degree of saturation of the intersection movements. Congested conditions were not examined in this work and thus longer cycles were not selected to provide more capacity to the mov ements. Regarding safety, longer cycle lengths were selected since longer green times allow more gaps in the opposing through traffic for permitted left turn movements and thus reduce the potential for accepting shorter gaps that potentially result in cras hes. Also, by reducing the number of times of phase changes, longer cycles further reduce rear end, left turn and loss of control crashes that mainly occur during the phase change periods. Regarding emissions, longer cycles tended to reduce the number of s tops, which is related to the increase in all types of emissions.
49 Table 3 4 Results for d ifferent o ptimization s chemes, per i ntersection Approach Movement EB L EB T EB R WB L WB T WB R NB L NB R SB L SB R Intersection 1 Demand (veh/h) 108 860 108 76 608 76 20 20 20 20 Mobility Optimization Level of Service (LOS) C C C C B B C C C C Intersection Delay / LOS 21.3 C Num. of Crashes (5 years) 6.25 Emissions (g) 480.6 Safety Optimization Level of Service (LOS) E B B E A A E E E E Intersection Delay / LOS 30 C Num. of Crashes (5 years) 4.55 Emissions (g) 470.2 Emissions Optimization Level of Service (LOS) F B B E A A E E E E Intersection Delay / LOS 28.6 C Num. of Crashes (5 years) 4.65 Emissions (g) 461.0 Intersection 2 Demand (veh/h) 90 720 90 90 720 90 20 20 20 20 Mobility Optimization Level of Service (LOS) C A A C A A C C C C Intersection Delay / LOS 9.9 A Num. of Crashes (5 years) 5.2 Emissions (g) 465.8 Safety Optimization Level of Service (LOS) F A A E A A E E E E Intersection Delay / LOS 21.8 C Num. of Crashes (5 years) 4.2 Emissions (g) 470.2 Emissions Optimization Level of Service (LOS) E A A E A A E E E E Intersection Delay / LOS 21.2 C Num of Crashes (5 years) 4.3 Emissions (g) 458.3 Intersection 3 Demand (veh/h) 76 608 76 108 860 108 20 20 20 20 Mobility Optimization Level of Service (LOS) C A A C A A C C C C Intersection Delay / LOS 13.9 B Num. of Crashes (5 years) 5.15 Emissions (g) 465.8 Safety Optimization Level of Service (LOS) F B B E B B E E E E Intersection Delay / LOS 26.9 C Num of Crashes (5 years) 4.25 Emissions (g) 474.3 Emissions Optimization Level of Service (LOS) F A A E B B E E E E Intersection Delay / LOS 23.4 C Num. of Crashes (5 years) 4.35 Emissions (g) 458.3
50 Final Remarks A methodology was developed and implemented in a computational engine (the Highway Capacity Software HCS), employing a genetic algorithm now capable of optimizing signal timing and coordinatio n on arterials, allowing the optimization to be based not only on mobility but also safety and emissions. The resulting methodology is available to the users as part of the HCS Streets module. To attain this goal, an emissions model and a set of crash pr ediction equations were adapted and incorporated to the HCS existing single optimization function. The selected safety model has the advantage of considering geometry and control aspects while its multiplicative structure allows for local calibration of e xisting parameters and future inclusion of additional factors, such as HSM Crash Modification Factors, the number of phases, maximum greens or the presence of countdown timers, for example. This research presents a new method, the results from its implemen tation and sensitivity analysis conducted to provide insight on the effects and order of relevance of offs between mobility, safety, and emission. This insight will help the designer, signal control engineer, and traffic analyst when designing intersection geometry and signal control. By using the Factorial Effects Method (FEM), it was concluded that, while the impacts of groups of variables on the mobility, safety, and emission measures are noticeably more significant than others, the effect of several of them is highly dependent on the combination with other variables, so that the individual degree of importance of each variable to the model cannot be fully ranked, at the 95% confidence. Demand level variables and the size of the intersection, defined by the number of lanes on the arterial, were found to largely affect all three performance measures. Large
51 intersections with no pockets were associated with a much higher number of angle an d rear end crashes. This effect can be mitigated with either the use of exclusive left turn pockets or by allowing the left turn movement on this lane to be shared with the through movement. Protected left turns increase delay, but only improve safety when left turn volumes are significant, imposing extra delay and crash risk. Right turn on red were associated with a few more crashes, while slightly improving the overall mobility and reducing emissions. Weight schemes for the performance measures had a limi ted impact on the optimal results, suggesting that the single objective optimization algorithm converges for a similar optimum solution regardless of the user defined weights. A limitation of this study is that some of the safety related variables could no t be fully tested for their effect on other performance measures. The presence of bus bays, parking or bicycles facilities were analyzed in terms of safety impacts only, due to the large variability and uncertainties of operation effects of these variables Finally, future research should incorporate pedestrian mobility and safety to the optimization method.
52 CHAPTER 4 CRASH AND TRAFFIC DATABASE FOR EXPRESSWAYS This study used a database built from different sources, covering a period from 2005 to 20 13, comprising 81,484 crashes and observations of traffic volumes and speeds by vehicle class collected at 5 to 15 minute intervals by 187 traffic monitoring stations on 13 toll express ways in So Paulo state, Brazil ( Figure 4 1 ) These expressways consist of 634 km and 818 km of directional freeways and divided multilane highways respectively They serv e over 22 million people in their vicinities. Figure 4 1 Studied f reeways m ultilane h ighways and m ain u rban a reas
53 All traffic monitoring stations are located on segments with 2 to 5 lanes with a 3.5 m (11.5 ft) width, and 2.5 m (8.2 ft) wide shoulders. All multilane highway segments are divided and have no at grade crossings or traffic signals, but can be accessed by regular driveways (not necessarily ramps) Shoulders can be used by cyclists and pedestrians in paved roads in the absence of bike facilities, particularly in urbanized areas. Automobile posted speeds range from 80 to 120 km/h (50 75 mph); the limits for trucks are lower on most highways, similarly to the policies in force in the states of California, Indiana, Montana, Oregon and Washington (NMA, 2018). The crash database includes the type, severity, number, and types of vehicles involved. The date, time and location of each crash are also specified. This information includes not only the milepost, but also whether the crash occurred on the main lanes, shoulders, service roads, intersections, tunnels, or toll plaza booths. The data were supplemented by information regarding work zones, peri ods with special operation schemes and enforcement. This made it possible to: a) exclude periods when road work zones were active, or when police or automated enforcement were in operation; and b) with the support of historical Google Earth images, identif y major changes on the infrastructure (opening or closure of ramps, number of lanes etc.), helping the classification and characterization of each segment over time. Each crash was assigned to a segment monitored by a corresponding traffic s tation for the same period. Those seg ments were classified as basic segments, weaves and segments in the influence area of on ramps or off ramps, as defined in the Highway Capacity Manual (TRB, 2016). S egments of the studied highways that could not be classified in one o f those types were not included in the study. Likewise, crashes
54 outside the boundaries of monitored segments, within the influence of toll plazas or during excluded periods were excluded from the data. As a result, 1039 km of the original 1452 km of highwa ys and 21,969 of 81,484 crashes were included in this study. Table 4 1 shows the frequency of roadway characteristics that might affect operations and safety, the total exposure (mi of veh km) and the number of crashes for the study period. W hile the relat ive length of urban segments is lower compared to rural segments, the number of stations for both categories is similar. In some cases crash occurrence may b e associated to higher exposure In other s c rash frequency is high in contrast to the exposure. I n example, five control stations on weaves cover 0.5% of the total length correspond ing to 1% of the exposure and 2% of the number of crashes. Table 4 1 Frequency of m ain r oad c haracteristics, n umber of c rashes and e xposure Variable Categories Traffic Monitoring Stations Freq. (%) Length (km) Freq. Exposure (veh km/10 6 ) Freq. (%) Crashes Freq. (%) Road Type Freeway Segment 65 34.8 425.3 44.4 8917 43.8 9306 42.4 Multilane Highway 87 46.5 512.9 53.5 10802 53.1 11621 52.9 OnRamp 18 9.6 10.2 1.1 329 1.6 413 1.9 OffRamp 12 6.4 5.4 0.6 102 0.5 186 0.8 Weaving 5 2.7 4.4 0.5 206 1.0 443 2.0 Area Rural 94 50.3 691.7 72.2 12627 62.0 11179 50.9 Urban 93 49.7 266.5 27.8 7729 38.0 10790 49.1 Num of Lanes 2 99 52.9 567.8 59.3 8722 42.8 10268 46.7 3 59 31.6 302.4 31.6 6680 32.8 7663 34.9 4 25 13.4 78.1 8.2 4429 21.8 3438 15.6 5 4 2.1 9.8 1.0 525 2.6 600 2.7 Posted Speed 80 16 8.6 33.9 3.5 567 2.8 1245 5.7 90 4 2.1 12.4 1.3 88 0.4 68 0.3 100 80 42.8 212.3 22.2 8059 39.6 9528 43.4 110 31 16.6 237.9 24.8 1688 8.3 1581 7.2 120 56 29.9 461.8 48.2 9949 48.9 9526 43.4
55 Figure 4 2 shows the distribution of severity among the three levels considered in this study, for the whole database [81,484 crashes, Figure 4 2 (a)] and for the used dataset [21,969 crashes, Figure 4 2 (b)] (a) Whole database (b) Used dataset Figure 4 2 Distribution of crashes by class and severity
56 The crashes were classified as: Property damage only (PDO), Injury (I), Fatal (F), for SV crashes, MV crashes MV and pedestrian and bicycle crashes (P&B). The percentages shown on the top reflect the severity distribution within each class, while the numbers in brackets re present the overall proportions of each crash class and severity relative to the total. Fatal crashes correspond to 2.2% of the total. Note that there are no PDO crashes for the P&B class, and fatalities for these can be as high as 39.1% of the total numb er of fatal crashes. All variables of the database are shown in Table 4 2. For variables with multiple categories, binary variables were created, while continuous variables were kept unchanged. In addition to these variables, the year of each crash occurre nce was included, in order to avoid any bias from temporal trends in the data. The proportion of each crash class per category is also shown, allowing for the assessment of each class can be more associated to a particular characteristic category. In examp le, regarding crash type, runoffs and hit animal are typical single vehicle crashes, while rear end, angle, sideswipe and head on must involve two or more vehicles. The discrete variables can be described as: Area Type: urban areas include urban and suburb an abutting land of cities with more than 5,000 inhabitants, as well as industrial areas. Crash Types: collision with objects, animals and run offs are single vehicle crashes, while head on, sideswipe, angle and rear end crashes may involve two or more vehicles. Weekend: weekend days comprises Saturday and Sunday, not including Friday nights. Other days are labeled as weekdays.
57 Nighttime: the sunrise and sunset times were obtained for each day of the year from data by the U.S. Naval Observatory, availabl e at IAG/USP (2018), so that each crash could be assigned to the daytime or nighttime periods accounting for the effect of the seasons. Visibility: perceived overall visibility by the driver and officer during crash reporting Weather: drizzle was defined as rain that is sufficient to cause the pavement to pavement surface and visibility, while fog does not affect the surface and heavily impacts visibility Vertical Alignment: a ll crashes happening in grades steeper than 3% were of traffic. Horizontal Alignment: defined according to the segment where the crash happened, being curvature (deg/km) used as a measure. Values under 100 deg/km were associated to straight segments. From 100 500 deg/km, segments were labeled as smooth curved, while segments with a curvature value above 500 deg/km were considered sharp curves. Segment Type: the presence of ramps or weaves was considered as an explanatory variable and their influence areas were defined per the HCM criteria (TRB, 2016). Bridge: this variable refers to bridge segments themselves, and do not include underpass areas. Guardrails: this variable was assigned to all crashes in which vehicles hit a median or roadside barrier or metallic guardrail. Road works: this included warning and speed reduction areas from work zones.
58 Table 4 2 Discrete v ariables included in the database Variable Categories Freq uency Proportion by crash class Total Relative MV P&B SV Area Type Rural 5722 44.8% 52.9% 2.3% 44.8% Urban 7049 55.2% 65.5% 3.0% 31.5% Crash Type Other 1097 8.6% 14.6% 0.0% 85.4% Collision w/ object 779 4.3% 61.5% 0.0% 38.5% Hit Animal 229 1.0% 3.5% 0.0% 96.5% Run off 4340 23.6% 8.2% 0.0% 91.8% Head on 131 1.3% 100.0 % 0.0% 0. 0 % Sideswipe 1978 16.4% 100.0% 0.0% 0.0% Angle 257 2.6% 100.0% 0.0% 0.0% Rear end 4340 32.4% 100.0% 0.0% 0.0% Weekend No 9101 71.3% 64.0% 2.5% 33.4% Yes 3670 28.7% 49.5% 3.1% 47.3% Nighttime No 7276 57.0% 63.2% 1.5% 35.3% Yes 5495 43.0% 55.4% 4.2% 40.3% Visibility Good 10912 85.4% 61.9% 2.5% 35.6% Average 1762 13.8% 48.6% 3.9% 47.5% Poor 25 0.2% 56.0% 0.0% 44.0% Weather Condition Good 10973 85.9% 37.5% 6.3% 56.3% Drizzle 294 2.3% 58.2% 0.3% 41.5% Rain 1396 10.9% 37.0% 7.4% 55.6% Fog 27 0.2% 61.9% 2.9% 35.2% Vertical Alignment Level 2992 23.4% 55.1% 2.7% 42.2% Upgrade 2156 16.9% 59.6% 3.1% 37.3% Downgrade 2512 19.7% 65.7% 2.7% 31.6% Horizontal Alignment Tangent 8486 66.4% 43.8% 1.3% 54.9% Smooth Curve 1390 10.9% 54.7% 3.7% 41.7% Sharp Curve 625 4.9% 62.3% 2.7% 35.0% Segment Type Segment 12209 95.6% 64.0% 2.2% 33.8% Ramp 305 2.4% 56.9% 3.1% 40.0% Weave 257 2.0% 61.9% 1.2% 36.9% Bridge No 12749 99.8% 59.9% 2.7% 37.4% Yes 22 0.2% 63.6% 0.0% 36.4% Guardrail No 10882 85.2% 65.7% 3.0% 31.3% Yes 1402 11.0% 9.1% 0.0% 90.9% Road works No 6451 50.5% 59.7% 3.0% 37.3% Yes 2494 19.5% 57.6% 2.3% 40.1%
59 Other variables were treated as continuous variables, as shown in Table 4 3. The number of vehicles involved in a MV crash, the posted speed and number of lanes used integer values, while other traffic variables were defined by real numbers. The density variation variable was calculated as the difference between the density right prior to the crash and at the time interval immediately before. Note that, information such as substance abuse, age or gender of the drivers involved in the crash were not avail able. Table 4 3 Continuous v ariables included in the database Variable Mean Std. Deviation Min Max Number of Vehicles 1.75 0.83 1 16 Traffic Density (pc/km/lane) 9.59 6.98 0 32 Average Speed (km/h) 89.18 16.02 0 174 Percent of Trucks 22.87 15.20 0 100 Density Variation (pc/km/lane) 1.11 4.98 32 66 Posted Speed (km/h) 107.65 11.46 80 120 Number of Lanes 2.76 0.82 2 5
60 CHAPTER 5 RELATIONSHIP BETWEEN CRASH RATE AND OPERATIONS ON EXPRESSWAYS Studies that investigated crashes as a function of hourly flow rates rather than as a function of AADT focused on urban freeway segments, and generally concluded that crash rates are higher for both low and high traffic flow conditions, reaching a minimum when traffic density is moderate (Frant zeskakis and Iordanis, 1981; Ceder, 1982; Chang et al., 2000; Martin, 2002; Zhou and Sisiopiku, 2014; Harwood et al., 2013; Potts et al., 2014). Those studies often counted on a limited amount of data and could account for the influence of a restricted num ber of road and environmental characteristics. Some rely on data from a single facility, that provides valuable insight but is insufficient to happened on different road types such as basic segments, weavings or ramps. Finally, few data from rural highways were available. This work was motivated by the availability of a large and diverse database containing about 35 million traffic observations and over 20,000 crashes from urb an and rural multilane highways, freeway segments, merges, diverges and weaves with different geometry and environmental conditions, that allows for in depth investigation of the relationship between traffic density and crash frequency for different situat ions. The object ives of this stage of the research were to: 1. M odel the relationship between traffic density and crash rate in expressways and compare the results to previous findings in the literature; 2. D evelop a novel model that incorporates significant roa d and traffic characteristics, capable of predicting hourly crash rates based on several roa dway and traffic variables; and 3. D iscuss the implications of the research findings for highway design and operations.
61 The next section describes the metrics and the relationship between density and crash rates derived for the whole dataset, which confirmed trends found in previous studies. The fourth section explains the methodology used to build the model proposed in this paper, which was comprised of selecting statistically significant variables and fitting equations by using regression modeling. Finally an analysis on the consistency of the proposed method to traditional crash prediction methods is conducted along with a discussion on the impl ications of the findings and final remarks. Relationship b etween Density a nd Crash Rate This research uses traffic density (pc/km/lane) to model the relationship between crash rate and traffic operational quality. Density ( D ) was calculated for each 5 to 15 minute interval as: ( 5 1 ) where S is the average speed (km/h) measured in the field at the stations indicated, and v is the equivalent flow in passenger cars per lane per hour. Truck volumes were converted to passenger car flows by using equivalents calculated by a previous study (Piva, 2015) from a sample of the same database. The database was then divided in density bins of 3 pc/km/lane. Each bin was assigned the total number of crashes, considering the density value of the tim e interval immediately prior to the crash, as well as the corresponding exposure, which is the traffic volume during the same interval multiplied by the segment length (veh km). The crash rate (crashes per million veh km) for each density bin b is then cal culated as: ( 5 2 )
62 where Crashes i and V i are the total number of crashes and the total number of vehicles respectively, on segment i within the density b ; L i is the length of each segment i ; n is the total number of segments. The aggregate result for the whole dataset confirmed findings from previous studies (Frantzeskakis and Ior danis, 1981; Ceder, 1982; Chang et al. 2000; Zhou and Sisiopiku, 2014; Harwood et al. 2013; Potts et al., 2014), a nd is shown in Figure 5 1 The relationship can be roughly described by a cubic polynomial, with the minimum rate of crashes at LOS B and C. For the same LOS, the proportion of FI and PDO crashes are similar. For more congested conditions, however, PDO cra shes become more prevalent. Figure 5 1 Relationship b etween t raffic d ensity and c rash r ate, by s everity
63 Figure 5 2 shows the relationships by crash class. As in previous studies, the combination of the higher frequency for single vehicle crashes at low density and the rise of multiple vehicle crashes at higher densities explains the U shape form of the relationship. Crashes with pedestrians and bicycles follow a similar relationship, with noticeable higher frequencies at low density. A better underst anding of this type of crash, which accounts for 39% of fatal crashes in the database, is recommended, and would require the incorporation of additional factors, such as pedestrian and bicycle volumes. (a) Multiple vehicle crashes (b) Single vehicle crashes (c) Pedestrian and bicycle crashes (d) Total Figure 5 2 Relationships b etween t raffic d ensity and c rash r ate, by c lass
64 Based on this analysis, it can be concluded that for this database the overall relationship between density and crash rate is consistent with the relationship found in other similar databases in the US. Differently from previous studies however a slight increase of single vehicle crashes beyond the boundary of the level of service E was observed. To understand this phenomenon the graph shown in Figure 5 3 was constructed, compar ing the proportion of single vehicle crash types for undersaturated (LOS A E) and oversaturated (LOS F) conditions. As expected, cra s h types related to high speeds, those encompassed: fires/explosions, flooding, landslides and vehicles being hit by objects/debris. These crash subtypes have a strong component of randomness so that, the more vehicles travelling on the road, the more likely it was that an unexpected adverse event happened to one of them. Figure 5 3 Distribution of s ingle v ehicle c rashes for LOS A E vs LOS F
65 Modelling t he Relationship Between Density a nd Crash Rate This section focuses on the development of a model to predict crash rate as a function of traffic density and other influencing variables. The first step consists of a statistical analysis on the variables available in the database to determine which ones should be included in the model, and at what levels or categories. The second step is focused on building regression models using the selected variables. The model developed is examined for consistency and accuracy relative to the traditional HSM methodology. Finally, a discussion on the implications of the findings is provided. Influencing Factors The principle of the model building process proposed is to find a balance between having as many policy sensiti ve explanatory variables as possible and keeping the model consistent and functional. The process used to seek this balance is described as follows. The data were subdivided into 3 pc/km/lane traffic density bins according to the levels or categories of th e candidate variables to be added to the model, so that each bin reflects a unique combination of density range and other characteristics. For each bin, the number of crashes, exposure and crash rates were computed as shown in Equation 5 2 If variables an d levels of variation are added to the model, more combinations of characteristics are created. This increases the number of individual data bins, at the expense of reducing the sample for each. The required sample n (number of observed crashes for one bin ) within each density range for reliable modeling purposes would be: ( 5 3 )
66 where Z 0.025 for a desired confidence of C = 1 0.05 is 1.96; is the standard deviation of the crash rates within each density range; and E is an accepted margin of error, established to be 1 crash per million veh km in this research To select which variables should be considered for inclusion in the model, the following procedure was followed, for each variable : Table 5 1 shows the ANOVA results, as well as the means and 95% confidence intervals of each category and variable that remained in the model: Segment Ty pe, Area Type, Posted Speed and Nighttime. The differences between those categories can be also observed in Figure 5 4 Regarding the segment type, only weaves could be considered different from other types of segments at 95% confidence. The relationship for weaving segments, although somewhat fuzzy, also follows the U shape and has a higher crash rate as compared to other segment types especially for high density conditions.
67 Overall, urban sites could not be considered different from rural sites (F obs = 4.07 < F 0.05,1,10 = 4.96). However, an interaction to density was observed (F obs = 23.73 < F 0.05,10,1118 = 1.84), noticea bly for low traffic density ranges. For that reason, this variable was considered for inclusion in the model. Table 5 1 Influencing v ariables 2 Way ANOVA s tatistics Variable Category Mean W Tukey 95% Confidence Interval LB UB Segment Type Expressway Segment 1 00 0.97 0.0 3 1.9 7 Weavings 2.68 1.71 3.65 Area Type Rural 0.92 0.26 0.6 6 1.18 Urban 1.1 9 0.9 3 1.4 5 Posted Speed (km/h) 80 90 1. 60 0.38 1. 22 1.9 8 100 120 0.9 9 0. 61 1.3 7 Nighttime No 0.94 0.15 0.79 1.10 Yes 1.16 1.00 1.31 Source Df SS MS F F crit p value Segment Type 1 854.61 854.61 11.62 4.96 0.007 Density 10 9669.71 966.97 13.14 2.98 0.000 Interaction 10 735.72 73.57 886.30 1.84 0.000 Error 1188 98.62 0.08 Total 1209 11358.65 Area Type 1 21.48 21.48 4.07 4.96 0.071 Density 10 1932.06 193.21 36.57 2.98 0.000 Interaction 10 52.82 5.28 23.73 1.84 0.000 Error 1188 264.42 0.22 Total 1209 2270.79 Posted Speed 1 111.29 111.29 9.70 4.96 0.011 Density 10 4218.50 421.85 36.78 2.98 0.000 Interaction 10 114.69 11.47 50.58 1.84 0.000 Error 1188 269.36 0.23 Total 1209 4713.84 Nighttime 1 14.12 14.12 7.47 4.96 0.021 Density 10 2006.53 200.65 106.10 2.98 0.000 Interaction 10 18.91 1.89 8.05 1.84 0.000 Error 1188 279.08 0.23 Total 1209 2318.64
68 (a) Merges and Diverges (b) Weaving Segments (c) Rural Highways (d) Urban Highways (e) Posted Speed 100 120 km/h (f) Posted Speed 80 90 km/h Figure 5 4 Relationships b etween d ensity and c rash r ate, by i nfluencing v ariable Concerning posted speeds, after the process described in this subsection, two functional groups could be identified. The 100 120 km/h segments comprise mostly
69 rural expressways with modern design, or urban freeways segregated from the local road system and only accessible through interchanges spaced at long distances. The 80 90 km/h segments encompass urban multilane highway segments. For these, an overall higher crash rate was observed for all traffic densities, so that H 0 could be re jected, and the two classes could be considered statistically different at 95% confidence. Nighttime conditions were found to be related to slightly higher rash rates for all density intervals. T here was not sufficient evidence to support that the number o f lanes, percent of trucks or weekend traffic affect the studied relationships. Proposed Model Regression analysis was used next to build model alternatives, based on the candidate variables described in the previous subsection. Each observation of the model corresponded to a data bin for which the sample size was considered statistically significan t according to Equation 5 3 To use polynomial relationship between crash rates and density in a linear regression method, two transformed variables were created : cubic and quadratic densities (Density 3 and Density 2 respectively). Interaction variables we re also tested, by relating the original variables to density variables. The criteria used to select the proposed model were: (a) F statistics as compared to the simplified model; (b) all model variables must be different from zero and statistically signif icant with the possible exception of the model constant at 95% confidence (t value > 1.96); c) low collinearity with other variables in the model.
70 The proposed regression equation (F 7, 62 = 3 4 089 p < .000) ha s an adjusted R 2 value of 0. 763 Table 5 2 sh ows the model coefficients and ANOVA statistics for the selected model. Apart from traffic density, weaves, lower standard urban multilane highways (posted speed 80 90 km/h) and nighttime conditions were found to increase crash rates. By adding these expla natory variables to the model, much of the asymmetry between low and high density ranges that was described by the cubic term of the equation in previous work (Figure 2 1) could be explained by the interaction variables (ex.: Density 2 Urban). Table 5 2 Model c oefficients and ANOVA s tatistics Variable B Std. Error t Sig. (Constant) .25504 .134 1.908 .061 Density 2 .00179 .000 4.321 .000 Density 3 X Weaving .00023 .000 10.599 .000 Urban 1.87078 .206 9.101 .000 Density X Urban .22427 .024 9.340 .000 Density 2 X Urban .00538 .001 6.402 .000 Density X Posted Speed 80 90 km/h .02701 .007 3.756 .000 Nighttime .31655 .095 3.338 .001 ANOVA Source df SS MS F p value Regression 7 36.729 5.247 34.089 0.000 Error 65 10.005 .154 Total 72 46.733 The proposed model to predict total crash rate, in 1 million veh km, is as follows: ( 5 4 ) where D is the traffic density; W = 1 when the segment is a weaving and 0 otherwise; and PS = 1 when the segment is an urban multilane highway (posted speed
71 80 90 km/h), and 0 otherwise; (urban freeways or rural expressways with posted speed higher than 90 km/h); U = 1 when abutting area is urban and 0 otherwise; and N = 1 for nighttime conditions and 0 for daytime. Finally, Figure 5 5 shows the analysis of the standardized residuals of the model. The Kolmogorov Smirnov test was performed to check the assumption that the residuals follows a normal distribution. The test statistic obtained was 0.096, with a p value of 0.095, indicating that H 0 cannot be reject, and that the residuals follow roughly a normal distribution, with a few values over 2.5 [Figure 5 5 (a)]. The absence of evident bias es can be confirmed in Figure 5 5 (b), in which no trend could be identified. (a) Histogram (b) Scatterplot Figure 5 5 Analysis of s tandardized r esiduals Figure 5 6 presents the application of the model, highlighting the effect of some characteristics to the predic ted crash rates.
72 Figure 5 6 Application of the proposed model Figure 5 6 (a) shows that, unlike previous findings, rural expressway segments associated with higher crash risk for both low and high density conditions, reaching
73 minimum cr ash rates at LOS B C [Figure 5 6 (b)] This difference may be associated with less forgiving roadway and roadside designs to accommodate higher speeds that are t ypical at low density condition and can be better handled on the rural freeway environment. In urban areas, upgrading from a multilane to freeway design could improve safety for any density level N ighttime is associated with slightly increased risk s Comparison Between the Proposed Model and the HSM By multiplying the crash rate obtained by Equati on 5 4 by the traffic volume in each interval and segment length, the predicted number of crashes is calculated. The sum of the predicted number of crashes for all time intervals within a year yields the annual predicted number of crashes, which is the sam e output produced by the Highway Safety Manual (HSM) methods. In order to compare the proposed E quation to the HSM, the annual number of crashes for each segment was calculated using both methods and compared to field values. AADT values ranged from 3,680 to 93,070 veh/day (average of 30,980). The HSM method for multilane highways (AASHTO, 2001) was used for all segments. The model was calibrated for the studied region, yielding an adjustment factor C r = 2.1. Figure 5 7 presents the results of this analysis Figure 5 7 (a) compares the annual number of crashes estimated by the calibrated HSM and the E quation developed in this research while Figure 5 7 (b) shows the distribution of errors relative to the field data for both methods. Although the proposed mode l produces estimate s that d eviate somewhat from the calibrated HSM to an extent the distribution of errors follows a similar pattern.
74 These results suggest that the proposed E quation is capable of producing estimates on an hourly basis that are still consistent when computed annual estimates. (a) Number of Predicted Crashes (b) Distribution of Errors Figure 5 7 Comparison of results from the proposed E quation and the HSM Discussion The findings of this work and the proposed model have applications for various types of safety and mobility studies. Differently from planning level methods, i n which the number of crashes is predicted in an annual basis, this work aims to provide relationships to hourly traffic densities, thus allowing for the assessment of crash risk. From the design perspective, minimum crash rates are associated to LOS A for rural expressways, and B to C for urban expressways. Also, the risks associated with weaving segments and the be nefits from the improvement of an urban multilane highway to a freeway can be estimated. At the operations level, the proposed model can be of use in applications related to ATM. Strategies such as ramp metering could be adjusted to find best balances
75 betw een delays and crash risk. This is particularly true when weaving segments exist along the mainline freeway. For planning studies using the HCM (TRB, 2016), Equation 5 4 could be used to estimate crash rates to be used in the generation of sce narios for reliability analysis, still accounting for factors such as are type and day and night conditions. Final Remarks This research developed a novel model to describe the relationship between hourly traffic density and crash rates on expressways, composed of r ural and urban multilane highways, basic freeway segments, ramp influence areas and weaving segments, based on a database containing over 35 million traffic observations and more than 20,000 crashes. Preliminary analysis confirmed that the curve for the en tire database follows a U shape, which is similar to previous studies. This shape is the consequence of a convex downward and a convex upward curve for single and multi vehicle crashes, respectively. Regarding crash severity, the proportion of fatal and injury crashes (FI) and PDO crashes were similar for LOS A to C, reaching a minimum between LOS B and C. For congested conditions, PDO crashes are more prevalent. The analysis of influencing factors revealed that weaves, urban areas, nighttime conditions a nd urban multilane highways (segments with 80 90 km/h posted speed) are related to higher crash rates, at 95% confidence. Those factors also explained much of the asymmetry of the curves, described by interaction variables. Using the selected variables, a regression equation was fitted. The application of the proposed model showed that, unlike urban expressways, the density crash rate
76 that low volume rural roads are safer than the higher volume ones. In urban areas, the upgrade from a multilane to a freeway design could improve safety for all density levels. The comparison with the HSM and field data suggests that the proposed E quation is capable of producing estimates on an hourly basis that are still consistent when computed year round projections. The findings of this work have implications to policy and design decisions, and the produced E quation could be incorporated to active traffic management (ATM) and HCM reliabili ty analysis.
77 C HAPTER 6 ESTIMATION OF CRASH SEVERITY ON EXPRESSWAYS Although studies on crash severity concentrate in developed countries, the majority of deaths occur in low and middle income countries (Malveira et al., 2015; elik and Oktay, 2017), spe cially on highways and rural roads, even while carrying less traffic than urban streets (ROSPA, 2017). The severity of a crash defines the consequence of a traffic incident, ranging from property damage to fatalities in the worst case. In order to understa nd the factors that affect crash severity and to design countermeasures to prevent injuries and fatalities, researchers and engineers have been using of multivariate statistics to account for the interactions between the several variables that might play a role in the gravity of a road crash. While many studies focused on geographic, socioeconomic and other human factors as the main cause of crashes (Quddus, 2015), more recent research recognize the interactions between more factors, including policy sensit ive roadway and environmental aspects that can be accounted for when designing safety countermeasures. However, only Quddus, Wang and Ison (2010) relied on a database that could combine prevailing traffic conditions to other roadway and environmental data, for one facility in the UK. This work was motivated by the absence of stablished models capable of predicting severity distributions simultaneously as a function of prevailing traffic conditions and other explanatory variables. The achievement of such goa l was made possible by the availability of the database described in Chapter 4 This will allow for better quantification of the impacts of safety countermeasures and other policy sensitive variables to crash severities and frequencies, which could support transportation
78 feasibility and cost benefit studies, warranting investment strategies based not only in mobility, but also safety criteria. In this context, the specific objectives of this stage of the research are to: (a) develop crash severity distribut ion models for expressways as a function of prevailing traffic conditions and other explanatory variables, using a modeling framework adequate to the crash severity analysis problem; (b) compare the results to a generic crash severity distribution model, i n order to verify the consistency of the proposed model; (c) discuss the marginal impact of the significant variables and the implications for safety analysis practices. Method ology Overview In this work, the Ordered Multinomial Logit Model is used to investigate the relationship between crash severity probabilities and the explanatory variables available in the database. As indicated in the literature, this model is typically applied to problems in which the outcome is discrete and ordered in a specifi c sequence, in the case, fro m fatal (F), to injury (I), to property damage only (PDO) crashes. The concept of utility used in other types of choice models is replaced by an c q a s et of thresholds n is defined equal to the number of alternatives k minus 1: ( 6 1 ) The thresholds in this case are such that: ( ) The probability that an injury crash happens is given by: ( )
79 where V q is the propensity function given the attributes q To define the propensity functions, the process described above was followed. First, the database was cleaned, checked for consistency and divided in three subsets: (a) SV crashes; and (b) MV crashes. Only variables that could possibly be associated with each subset were used. For example, crash types head on, sideswipe, angle or rear end are only possible for events involving more than one motorized vehicle and are therefo re discarded from the analysis of single vehicle crashes. The models were fitted using variables that describe prevailing traffic conditions, such as density and average speed, not individual vehicles types. The intention was to develop models that could b e used to generate severity distribution functions based on characteristics that can be observed in the field. The statistical analysis itself was made separately for SV and MV crashes, using a n ordinal regression technique. In the first step, all explanat ory variables in the database were considered. A 95% confidence was used to determine the statistically significant variables (p value < 0.05). In each subsequent step of the procedure, one variable with the highest p value was eliminated, until all remain ing variables could be considered statistically significant. The statistical significance of the model as a whole and of each thresholds was also verified in each step R esults Table 6 1 shows the results of the statistical analysis, for SV and MV crash es, listing all variables considered statistically significant for at least one of the subsets. The parameters are described only for the variables that remained in the final models, for each type of crash. Table 6 1 also shows the resulting Log li kelihood of the models. The three severity categories are divided by two threshold s (Equation 6 2 )
80 Table 6 1 Results and significant variables for each model Variable Single vehicle crashes Multiple vehicle crashes B Std. Error Wald p value B Std. Error Wald p value Crash Type Collision with object 1.948 .150 168.004 .000 Hit animal 2.952 .215 187.899 .000 Runoff 1.098 .090 150.145 .000 Head on 1.171 .194 36.583 .000 Rear end .228 .067 11.469 .001 Sideswipe .208 .076 7.566 .006 Time and Environment Year .041 .018 5.338 .021 Weekend .211 .057 13.778 .000 Night .182 .064 7.974 .005 .211 .051 16.974 .000 Drizzle .401 .198 4.101 .043 Rain .868 .093 86.439 .000 .438 .090 23.632 .000 Urban area .173 .073 5.566 .018 .257 .052 24.692 .000 Geometry and design Number of Lanes .122 .043 8.252 .004 Sharp Curve .251 .123 4.186 .041 .788 .148 28.291 .000 Weave .607 .157 14.845 .000 Bridge segment 1.259 .553 5.179 .023 Guardrails .708 .077 84.823 .000 Work zone .345 .083 17.286 .000 Traffic Variables Traffic density .030 .004 59.505 .000 Traffic density variation .009 .004 4.174 .041 Average speed .007 .003 7.273 .007 Number of vehicles .144 .033 18.685 .000 Percent of Trucks .006 .002 8.913 .006 (Injury threshold) 1.737 .306 32.164 .000 .519 .105 24.416 .000 (Fatal threshold) 2.908 .326 79.505 .000 4.173 .135 953.854 .000 Model Fitting Information N 4,782 7,648 2 Log likelihood at convergence 6392.373 10599.182 2 Log likelihood at equal shares 7158.493 10933.718 Chi Square 766.119 334.536 p value 0.000 0.000
81 Runoffs (SV) and head on (MV) were considered the most severe types of crashes. Lower visibility conditions, reflected by nighttime and rain affected all models. Regarding the geometry aspects, weaves and bridges were related to crash severity for MV crashes, while the presence of work zones, guardrails and barriers affected the outco me of SV crashes, which is consistent with runoffs being the gravest type of single vehicle crash. The assessment of the influence of prevailing traffic conditions revealed that while the severity of MV crashes is linked to the traffic density and the vari ability of density through time, the severity of SV crashes is more closely related to average speed. Geometry variables that were analyzed but not found to be statistically significant were the presence of upgrades/downgrades. This may be due to two facto rs: (a) there are not enough crash data in these segments to draw statistically significant conclusions; (b) the studied expressways have high design standards, and thus these segments provide the users practically the same level of safety and comfort as t he remaining segments. Regarding environment variables, the no association was made regarding visibility classification in the field by crash victims and officers. Margin al Effects The magnitude of the coefficients B shown in Table 6 1 provides a preliminary indication on the importance of each variable in the models. However, contrary to traditional linear regression modeling, neither the magnitude nor the sign of the variables can be directly used to make inferences on the marginal ef fect of each variable to the discrete choice model response (Greene and Hensher, 2009).
82 Table 6 2 scenario reflects the average probability of each severity level obtained by applying t he models proposed in this study to all crashes in the database. Each row, labelled after one variable tested, represent one scenario created by changing only that variable while esent the average probabilities of each severity level in nighttime conditions. By comparing the altered probabilities to the ones from the base scenario, it is possible to assess the impact of each variable on crash severity. Table 6 2 Probabilities es timated for each non continuous variable Variable Single vehicle crashes Multiple vehicle crashes PDO Injury Fatal PDO Injury Fatal Base 52.4% 46.4% 1.2% 60.8% 37.4% 1.8% Crash Type Collision with object 70.2% 29.3% 0.3% Hit animal 86.4% 13.4% 0.0% Runoff 56.7% 42.5% 0.8% Head on 31.1% 62.4% 6.5% Rear end 62.2% 36.2% 1.6% Sideswipe 62.9% 35.5% 1.6% Time and Environment Weekend 53.8% 43.8% 2.4% Night 56.4% 42.5% 1.0% 55.4% 42.4% 2.2% Drizzle 62.2% 37.2% 0.7% Rain 69.5% 30.0% 0.3% 69.5% 29.3% 1.2% Urban area 47.8% 50.7% 1.5% 59.4% 38.7% 1.9% Geometry and design Sharp Curve 52.9% 45.9% 1.2% 76.0% 23.1% 1.0% Weave 47.4% 49.6% 3.0% Bridge segment 28.2% 64.1% 7.6% Guardrails 66.1% 33.4% 0.5% Work zone 54.4% 44.5% 1.1%
83 Figure 6 1 graphically presents this information, making it possible to visually compare the different severity levels. The further the lines are from the center, the higher the probability of that severity level as compared to the other levels. The following subsec tions discuss each variable group in more detail. (a) Single vehicle PDO and I crashes (b) Single vehicle F crashes (c) Multiple vehicle PDO and I crashes (d) Multiple vehicle F crashes Figure 6 1 Comparison between the marginal effect of each variable Crash Type As mentioned before, different crash types were associated to the severity of single vehicle or multiple vehicle crashes. In no model collisions with objects in the road
84 were considered to be statistically different from any other type of crash. This may be due to the varied nature of objects that can be possibly hit, from large animals to stationary road work equipment. Runoffs, on the other hand, proved to be a more severe type of crash, esp ecially for single vehicle crashes. For multiple vehicle crashes, angle and head on crashes were considered more severe than other accidents, at 95% confidence. Although head on accidents in multilane highways are not common, they are a possibility when t here is no physical barrier on the median, and significantly more severe than the average, according to the analysis. Finally, sideswipe and rear end accidents could not be considered any different than other types of crashes. Time and Environment Weekend was considered statistically significant for the MV model, slightly increasing the probability of more severe accidents; similarly, crashes in urban areas were found to be more severe for both SV and MV crashes. Although the cause of this is unclear, it mi ght be related to substance abuse or conflicts between highway and urban traffic. While nighttime increased crash severity for MV crashes, this condition was associated to less severe crashes for the SV model. Similarly, rainy conditions were linked to les s severe crashes, which could be explained by the extra caution that drivers use under low visibility conditions. Geometry and Design With respect to geometry, it was not possible to affirm that grades affect the severity of accidents with statistical conf idence. On the other hand, curved segments were associated with less severe MV crashes. Note that these segments are usually part of expressways on rolling terrain with consistent design standards, so that lower severities could be explained by lower speed
85 Weaves and bridge segments were found to significantly increase crash severity for MV crashes, while the guardrails and barriers proved effective in reducing the probability of severe SV crashes. This finding r einforces the role of guardrails and other types of physical barriers as a countermeasure of single vehicle, runoff accidents. Although these cannot eliminate crashes, a well designed device may potentially reduce the consequences of the accident, from FI to PDO. Traffic Characteristics Figure 6 2 presents the relationships between traffic density and crash severity for SV and MV crashes. Each small dot represents the estimate for one observation in severity distribution for the whole database, for each density interval of 3 pc/km/lane obtained from the models calibrated in Chapter 5 (Figure 5 1) (a) Single vehicle crashes (b) Multiple vehicle crashes Figure 6 2 Relationships between traffic density and crash severity distribution The variability of the results reflects the effect of the multiple factors accounted for by the fitted models. It can be noticed that, for MV crashes, there is strong evidence
86 of the relationship between traffic de nsity and crash severity, evidenced by the lower deviation from the average values. The relationship between density and severity of SV crashes can be roughly observed until the range between 5 10 pc/km/h, after which results become much more dependent on other explanatory variables, such as the area type. Figure 6 3 shows the relationships between average travel speed and severity distribution. F atal and injury crashes are grouped as FI crashes. In the range of 80 100 km/h, the distribution of PDO and FI crashes is more homogeneous. For both lower and higher speed ranges, the variabity is greater, with a higher proportion of PDO crashes for high speed. This phenomenon may be explained by the fact that higher speeds are possible in higher des ign standards expressways. Conversely, lower speeds are more frequent on more heterogeneous highways, including urban segments. Figure 6 3 Variability of the relation ship between speed and SV crash severity. Figure 6 4 illustrates the variability of t he crash severity distributions for different values of density difference, defined as the variation, in pc/km/lane between the density right prior the crash and the moment immediately before it. The proportion of PDO and
87 FI crashes is the most balanced w hen the density variation is lower, near zero. Negative density differences (left side of the graph) reflect discharge conditions, resulting in proportionally less FI crashes P ositive values of density variation (right side of the graph) represents conges tion formation. For this situation, the variability of crash severity distributions increase and are more difficult to predict Figure 6 4 R elation ship between density differences and M V crash severity. Alternative Model Following the practice from many previous studies, two alternative models were created, in which the specific vehicle types involved in the crash are taken as explanatory variables. In this study, however, the variables that describe prevailing traffic con ditions are maintained in order to study the possible interactions between the surrounding traffic and vehicles involved in the crash. Table 6 3 shows the results f or these models the MV and P&B. The parameters for MV crashes are similar to the ones previously observed, though part of the phenomenon is explained according to the vehicles involved in the crash especially
88 when for crashes involving at least one motorcycle For the P&B model, the analysis of the data was not sufficient to draw broader conclusions, but indicated that crashes involving motorcycles tend to be less severe, and crashes at night, more severe. Table 6 3 Results and significant variables Alternative MV and P&B Models Variable Multiple vehicle crashes Pedestrian & Bicycles B Std. Error Wald p value B Std. Error Wald p value Crash Type Head on 1.197 .197 36.965 .000 Rear end .188 .071 7.037 .008 Sideswipe .173 .080 4.634 .031 Time and Environment Year .054 .013 18.315 .000 Weekend .251 .060 17.380 .000 Night .225 .054 17.098 .000 .982 .256 14.675 .000 Rain .368 .094 15.269 .000 Geometry and design Sharp Curve .654 .161 16.428 .000 Weave .469 .164 8.242 .004 Bridge segment 1.239 .561 4.886 .027 Traffic Variables Traffic density .046 .004 106.631 .000 Average speed .012 .002 47.176 .000 Number of vehicles .152 .035 18.847 .000 Involved Vehicles Auto .559 .053 112.488 .000 Bus .311 .144 4.661 .031 Truck .235 .063 13.774 .000 Motorcycle 2.411 .096 632.009 .000 1.176 .538 4.778 .029 (Injury threshold) .680 .219 9.608 .002 4.029 .528 58.220 .000 (Fatal threshold) 3.416 .236 208.638 .000 1.090 .223 23.933 .000 Model Fitting Information N 7,648 343 2 Log likelihood at convergence 10004.808 24.031 2 Log likelihood at equal shares 11328.486 45.826 Chi Square 1323.678 21.795 p value 0.000 0.000
89 Final Remarks This stage of the research modeled the impact of a series of explanatory variables to predict crash severity, using the ordered response choice model and the database described in Chapter 4 The framework that better fit this database led to the development of two different models: single vehicle crashes (SV) and multiple vehicle crashes (MV). The analysis of the results showed that the factors that explain the severity of crashes varies widely between these models, as the causes for SV and MV crashes are different. Head on crashes were found to be the most severe crash type for the MV model, while run offs were more significant for the SV model. Consequently the use of guardrails and barriers as a safety countermeasure proved to effectively reduce severity for SV crashes. Other des ign features found to be statistically significant were weave and bridge segments, which were associated with more severe MV crashes. Finally, conditions that impair visibility, such as rain and nighttime were linked to more severe MV crashes, but slightly less severe SV crashes. The unique database used in this study allowed the investigation of the influence of prevailing traffic conditions on crash severity, while controlling for other factors. The results suggested that MV crash severity is nega tively r elated with traffic density, while SV crashes are more closely related to speed. T he variability of the crash severity estimates for low speed and high speed ranges proved to be high, reflecting the impact of other explanatory variables on low and high spe ed roads. For future studies, it recommended that the results obtained in this research are translated into crash severity functions that can be integrated in crash prediction models
90 which are able to predict crash frequency as a function of roadway and en vironmental conditions, as well as prevailing traffic conditions.
91 CHAPTER 7 CONCLUSIONS T h e goal of this research wa s to model the relationship between crash rates and different operating conditions for two different components of a road network: corridor s with signalized intersections and expressway (uninterrupted) segments. The first part of this project produced a new methodology capable of combining mobility, safety and emissions measures to the objective function of a genetic algorithm now capable of optimizing signal timing and coordination on arterials. The resulting methodology is available to the users as part of the HCS Streets module. The selected safety model has the advantage of considering geometry and control aspects, while its multiplicative structure allows for local calibration of existing parameters and future inclusion of the effects of additional factors, such as HSM Crash Mod ification Factors, the number of phases, maximum greens or the presence of countd own timers, in example. It was also present ed the results from a sensitivity analysis conducted to provide insight on the effects and order of relevance of 20 key variables o outcomes and the associated trade offs between mobility, safety, and emission. This insight will help the designer, signal control engineer, and traffic analyst when designing intersection geo metry and signal control. T he subsequent parts of this research project were part of an effort to develop innovative models able to predict crash frequencies and severities as a function of prevailing traffic conditions and other explanatory variables. This goal was only possible due to the availabilit y of a unique database o f Brazilian expressways, composed of rural and urban multilane highways, basic freeway segments, ramp influence areas and
92 weaving segments, with over 35 million traffic observations and more than 20,000 crashes. In the second part o f the research, it was developed a novel model to describe the relationship between hourly traffic density and crash rates Preliminary analysis confirmed that the curve for the entire database follows a U shape, which is similar to previous studies. This s hape is the consequence of a convex downward and a convex upward curve for single and multi vehicle crashes, respectively. Further analysis of the influencing factors revealed that weaves, urban areas, nighttime conditions and urban multilane highways (se gments with 80 90 km/h posted speed) are related to higher crash rates, at 95% confidence. Using the selected variables, a regression e quation was fitted. The application of the proposed model showed that, unlike urban expressways, the density crash rate r elationship for rural rural roads are safer than the higher volume ones. In urban areas, the upgrade from a multilane to a freeway design could improve safety for all density l evels. In the last part of the research, it was modeled the impact of a series of explanatory variables to predict crash severity, using the ordered response choice model the database described in Chapter 4. The framework that better fit this database led to the development of two different models: single vehicle crashes (SV) and multiple vehicle crashes (MV). The analysis of the results showed that the factors that explain the severity of crashes varies widely between these models, as the causes for SV and MV crashes are different.
93 Head on crashes were found to be the most severe crash type for the MV model, while run offs were more significant for the SV model. As a consequence, the use of guardrails and barriers as a safety countermeasure proved to effect ively re duce severity for SV crashes. The investigation of the influence of prevailing traffic conditions on crash severity revealed that multiple vehicle crash severity is positively related with traffic density, while single vehicle crashes are more clos ely related to speed. Still, the variability of the crash severity estimates for low speed and high speed ranges proved to be high, reflecting the impact of other explanatory variables on low and high speed roads. The findings of this work have implication s to policy and design decisions, and the produced E quation could be incorporated to active traffic management (ATM) and HCM reliability analysis still accounting for factors such as are type and day and night conditions For futu re work, it is recommende d that the research done on freeway systems and multilane highways to be implemented and tested in a tool, as done for the methodology developed for signalized arterials. This could allow users to benefit from the findings of this research, as well as give insight of the impacts of the variables considered in the modelling to transportation planning and operations. Also, the models developed for expressways could potentially be extended to other types of facilities, such as two lane highways. This would all ow the assessment of the ideal traffic and design standards that justify the improvement of a two lane highway to a multilane highway. Also, other variables that could be incorporated to the crash prediction and crash severity models would require more det ailed analysis on the
94 roadway design, and could include median types and widths, rumble strips, ramp/interchange density and detailed horizontal and vertical alignment information, such as the precise curve radii and grade magnitudes and lengths along all studied network. Finally, it is recommended that the relationships for pedestrian and bicycles to be further investigated, given the high fatality rate of these crash types. This will require not only motorized vehicle data, but also non motorized traffic on curbsides and crossings.
95 LIST OF REFERENCES Abdel Aty, M. (2003) Analysis of driver injury severity levels at multiple locations using ordered probit models. Journal of Safety Research 34, 597 603. Agbelie, B., and Roshandeh, A. (2015) Impacts of Signal Related Characteristics on Crash Frequency at Urban Signalized Intersections. Journal of Transportation Safety & Security 7, 199 207. AASHTO (2010) Highway Safety Manual American Association of Stat e Highway and Transportation Officials Washington, D.C. Abdel Aty, M. (2003) Analysis of driver injury severity levels at multiple locations using ordered probit models. Journal of Safety Research 34, 597 603. Archer, J. (2004) Methods for the Assessment and Prediction of Traffic Safety at Urban Intersections and their Application in Micro S imulation Modeling PhD dissertation, Division of Transport and Logistics, Royal Institute of Technology, Stockholm, Sweden. Arona, M., Billot R., el Faouzi, N. E., a nd Seidowsky, R. (2015) Traffic indicators, accidents and rain: some relationships calibrated on a French urban motorway network. Transportation Research Procedia 10, 31 40. Azar, S., Reynolds, B., and Narayanan, S. (1999) Comparison of Two Objective Opti mization Techniques with and within Genetic Algorithm. Proceedings of the ASME Design Engineering Technical Conferences Beasley, D., Bull, D. and Martin R. (1993) An Overview of Genetic Algorithms: Part 1, Fundamentals. University Computing 15 (2) 58 69. Campolongo, F., Cariboni, J., and Saltelli, A. (2007) An effective screening design for sensitivity analysis of large models. Environmental Modelling & Software 22 ( 10 ) 1509 1518. elik, A. K., and Oktay, E. (2017) A comparison of ordered and unord ered response models for analyzing road traffic injury severities in the North Eastern Turkey. Periodica Polytechnica Transportation Engineering 45 ( 3 ) 119 132. Chang, J., Oh, C. Chang M. (2000) Effects of Traffic Condition (v/c) on Safety at Freeway F acility Sections. Transpor tation Research Circular E C018, 200 208. Chen, X., Hadi, M., Xiao, Y., and Elefteriadou, L. (2016) Development of Macroscopic Emission Estimation Model Based on Microscopic Operating Modes Transportation Research Record 25 70, 3 9 4 7.
96 Chin, H., and Quddus, M. (2003) Applying the random effect negative binomial model to examine traffic accident occurrence at signalized intersections Accident Analysis and Prevention 35, 253 259. Coutinho, C. H. L. Cunto, F. J. C.; and Ferreira, S. (2015) Anlise da severidade dos acidentes com motocicletas utilizando modelos probit e logit ordenados. Transportes 23, 60 66. Daniel, J. R., and Maina, E. (2011) Relating Safety and Capacity on Urban Freeways Procedia Social and Behavioral Sciences 16, 317 328. Essa, M., and Sayed, T. (2015) Simulated Traffic Conflicts. Do They Accurately Repres ent Field Measured Conflicts? Transportation Research Record 2514 48 57. Frantzeskakis, J. M., and Lordanis, D. I. (1981) Volume to Capacity Ratio and Traffic Accidents on Interurban Four Lane Highways in Greece. Transportation Research Record 1112, 29 38. Geedipally, S. R. Bonneson, J. A. Pratt, M. P. and Lord, D. (2013) S everity Distribution Functions for Freeway Segments. Transportation Research Record 2398, 19 27. Gettman, D., Pu, L., Sayed, T., and Shelby, S. (2008) Surrogate Safety Assessment Model and Validation Final report. FHWA HRT 08 051. FHWA, U.S. Department o f Transportation. Golob, T. F., and Recker, W. W. (2003) Relationships Among Urban Freeway Accidents, Traffic Flow, Weather, and Lighting Conditions Journal of Transportation Engineering 129 (4), 342 353. Greene, W. H. and Hensher, D. A. (2009) Modeling Ordered Choices: A Primer DOI: 10.1017/CBO9780511845062. Hadi, M., Elefteriadou, L., Chen, X., Wang, T., and Xiao, Y. (2017) Signal Optimization with Consideration of Environmental and Safety Impacts, Part A STRIDE Final Report, Gainesville, FL. Harwood D W ., Bauer, K. M. and Potts I. B. (2013) Development of Relationships Between Safety and Congestion for Urban Freeways. Transportation Research Record 2398, 28 36. Instituto de Astronomia, Geofsica e Ci ncias Atmosfricas (2018) Nascer e ocaso do S ol (2011 2025). U niversity of So Paulo. Retrieved on July 2018 from http://www.iag.usp.br/astronomia/nascer e ocaso do sol
97 Islam, M. R., Wyman, A. A., and Hurwitz, D. S. (2017) Safer driver responses at intersections with green signal countdown timers. Transportation Research Part F 51, 1 13. Kockelman, K. M. and Kweon, Y. J. (2002) Driver injury severity: An application of ordered probit models. Accident Analysis and Prevention 34 ( 3 ) 313 321. Kononov, J., Lyon, C., Allery, B. K. (2011) Relating Flow, Speed and Density of Urban Freeways to Functional Form of an SPF. 90th Transportation Research Board Annual Meeting Washington, D.C. Kuehl, R. (1999) Design of Experiments: Statistica l Principles of Research Design and Analysis 2 nd ed Duxbury Press. Lambert, D. (1992) Zero Inflated Poisson Regression with an Application to Defects Manufacturing. Technometrics 34 ( 1 ) 1 14. Lee, C. and Li, X. (2015) Predicting Driver Injury Severity in Single Vehicle and Two Vehicle Crashes with Boosted Regression Trees. Transportation Research Record 2514, 138 148. Lord, D., Manar, A., Vizioli, A. (2005) Modeling crash flow density and crash flow VC ratio relationships for rural and urban freeway se gments. Accident Analysis & Prevention 37, 185 199. Malveira, D. M. de Sousa, D. L. Ribeiro, P. C. M. and Leite, A. C. S. (2015) Contribuio metodolgica para anlise dos benefcios na segurana do transito devido a modificaes virias. XIII Rio de Transportes Rio de Janeiro, Brazil. M artin, J. L. (2002) Relationship between crash rate and hourly traffic flow on interurban motorways Accident Analysis and Prevention 34, 619 629. Mensah A., and Hauer, E. ( 1998 ) fr om the estimation of the relationship between accidents and traffic flow. Transportation Research Record 1635, 37 43. Milt on, J. C., Shankar, V. N., and Mannering, F. L. (2008) Highway crash severities and the mixed logit model: An exploratory empirical a nalysis. Crash Analysis and Prevention 40, 260 266. Mitra, S., Chin, H. C., and Quddus, M. (2007) Study of Intersection Accidents by Maneuver Type. Transportation Research Record 1784, 43 50. Montgomery, D. C., and Runger, G. C. (2014) Applied Statistics and Probability for Engineers Wiley, USA. Morris, M. D. (1991) Factorial Sampling Plans for Preliminary Computational Experi ments. Technometrics 33(2), 161 174.
98 National Motorists Association (2018) State Speed Limit Chart. Retrieved on May 2018 from https://www.motorists.org/issues/speed limits/ Nunes, D. F. (2012) Procedimento para Anlise de Sensibilidade do Programa HDM 4 Master thesis, So Carlos School of Engineering, University of S o Paulo, Brazil. and Connor, D. H. (1996) Predicting the severity of motor vehicle accident injuries using models of ordered multiple choice. Accident Analysis and Prevention 28 ( 6 ) 739 753. Ozbay, K., Yang, H., Bartin, B., and Mudigonda, S. (2008) Derivation and validation of new simulation based surrogate safety measure. Transportation Research Record 2083, 105 113. Pande, A. Abdel Aty, M. and Hsia, L. (2005) Spatiotemporal variation of r isk preceding crashes on freeways. Transportation Research Record 1908, 26 36. Penmetsa, P. and Pulugurtha, S. S. (2018) Modeling crash injury severity by road feature to improve safety. Traffic Injury Prevention 19 (1), 102 109. Piva, F. (2015) Fatores de E quivalncia de V eculos P esados em R odovias de P ista D upla Master thesis, So Carlos School of Engineering, University of So Paulo, Brazil. Poch, M., and Mannering, F. (1996) Negative Binomial Analysis of Intersection Accident Frequencies. Jo urnal of Transportation Engineering 122, 05 113. Potts, I. B., Harwood, D. W., Fees, C. A., and Bauer, K. M. (2014) Further development of the safety and congestion relationship for urban freeways SHRP 2 Report S2 L07 RR 3. Transportation Research Board, Washington, D.C. Quddus, M. A. Wang, C. and Ison, S. G. (2010) Road traffic congestion and crash severity: Econometric analysis us ing ordered response models. Journal of Transportation Engineering 136, (5), 424 435 Quddus, M. ( 2015) Effects of geodemographic profiles of drivers on injury severity from traffic crashes using multilevel mixed effects ordered logit model. Transportation Research Record 2514, 149 157. ROSPA (2017) Rural Road Safety. The Royal Society for the Prevention of Accidents. Retrieved on July 2017 from https://www.rospa.com/road safety/advice/road users/rural/ Santner, T. J., Williams, B. J., and Notz, W. I. (2003) The Design and Analysis o f Computer Experiments Spring Verlag, New York, NY. Saltelli, A., Chan, K., and Scott, M. (2000) Sensitivity Analysis Wiley, New York, NY.
99 Saltelli, A., Tarantola, S., Campolongo, F., and Ratto, M. (2004) Sensitivity analysis in practice: A guide to asse ssing scientific models John Wiley & Sons Ltd, The Atrium, Southern Gate, Chichester, UK. Shahdah, U., Saccomanno, F., and Persaud, B. (2015) Application of traffic microsimulation for evaluating safety performance of urban signalized intersections Trans portation Research Part C 60, 96 104. Spigolon, L. M. G. (2010) Semforo : Grupo focal convencional x grupo focal com informao do tempo de verde / vermelho restante Master thesis, So Carlos School of Engineering, University of So Paulo, Brazil. Stevanovic, A., Stevanovic, J. So, J. and Ostojic. M. (2015) Multi criteria optimization of traffic signals: Mobility, safety and environment. Transportation Research Part C 55, 46 68 TRB. (2010) Highway Capacity Manual 2010. Transportation Research Bo ard of the Nati onal Academies, Washington, D.C TRB (2016) Highway Capacity Manual, Sixth Edition: A Guide for Multimodal Mobility Analysis. Transportation Research Board, Washington, D.C. Trucano, T. G., Swiler, L. P., Igusa, T., Oberkampf, W. L., and Pil ch, M. (2006) Reliability Engineering & System Safety 91, 1331 1357. Turner, S., Singh, R., and Nates, G. (2012) Crash prediction models for signalized intersections: signal phasing and geomet ry NZ Transport Agency research report 483. Wang, X. S., and Abdel Aty. M. (2006) Temporal and spatial analyses of rear end crashes at signalized intersections. Accident Analysis and Prevention 38, 1137 1150. Wu, Q. Chen, F. Zhang, G. Liu, X. C. Wang H. and Bogus, S. M. (2014) Mixed logit model based driver injury severity investigations in single and multi vehicle crashes on rural two lane highways. Accident Analysis and Prevention 72, 105 115. Xi, Y. Zhang, Y. and Liang, F. (2009) Crash Injury Severity Analysis Using Bayesian Ordered Probit Models. Journal of Transportation Engineering 135 ( 1 ) 18 35. Xie, Y. Zhao, K. and Huynh, N. (2012) Analysis of driver injury severity in rural single vehicle crashes. Accide nt Analysis and Prevention 47, 36 44. Yamamoto, T. and Shankar, V. N. (2004) Bivariate ordered response probit model of Accident Analysis and Prevention 36 ( 5 ) 869 876.
100 Yang, W., L. Zhang, L. Zhang, and H. Shi. (2013) A Golden Ratio Based Genetic Algorithm and Its Application in Traffic Signal Timing Optimization for Urban Signalized Intersections Yasmin, S. and Eluru, N. (2013) Evaluating alternate discrete outcome fra meworks for modeling crash injury severity. Crash Analysis and Prevention 59, 506 521. Ye, F., and Lord, D. (2014) Comparing three commonly used crash severity models on sample size requirements: Multinomial logit, ordered probit and mixed logit models. A nalytic Methods in Accident Research 1, 72 85. Younshik, C., Song, T. J., and Yoon, B. J. (2014) Injury severity in delivery motorcycle to vehicle crashes in the Seoul metropolitan area. Accident Analysis and Prevention 62 (5), 79 86. Zhou, M., and Sisio piku V. P. (2014) Relationship Between Volume to Capacity Ratios and Accident Rates. Transportation Research Record 1581, 47 52.
101 BIOGRAPHICAL SKETCH Bachelor s in c ivil e ngineering from Federal University of Minas Gerais, Brazil (2008) and m Engineering EESC/USP. Works in transportation related field since 2004, having served as director for Transitus Trans portation Engineering, of consulting group Tectran, between 2011 and 2015. Has focused on the following areas of knowledge: traffic engineering, road safety, PPPs and cost benefit studies, demand forecasting and transportation planning. Authored over 30 pu blications on conferences proceedings and scientific journals.