SAFETY MODELING VIA SEGMENTATION OF TRANSPORTATION NETWORKS

  • : Ms Word Format
  • : 70 Pages
  • : ₦3000
  • : 1-5 Chapters
  •  
  • Click to DOWNLOAD Materials

SAFETY MODELING VIA SEGMENTATION OF TRANSPORTATION NETWORKS

ABSTRACT

Safety Modeling via Segmentation of Transportation Networks

 

 

This dissertation proposes a methodology to address a long-standing question in traffic safety relating to the evaluation of safety risk and the benefits associated with safety interventions.  Traditionally, safety risk has been assessed at the corridor level, with corridors being evaluated in terms of accident rates, as in accidents per million vehicle miles.  This measure allowed safety planners and engineers to look for correlations at the aggregate level using economic and sociodemographic data from counties and cities.  As roadway geometric data became more widely available, both in terms of general access to public agencies and in terms of measurement detail, statistical models for safety were developed incorporating correlations between safety outcomes at the roadway segment level and roadway geometrics.  This approach avoids the problem of ecological correlation that is likely to occur in modeling using county or city level independent variables.  The problem of ecological correlation occurs when correlations between safety outcomes in corridors are evaluated using mean accident rates and county or city level means for independent variables.  This approach assumes corridor and regional means reflect segment means accurately, an assumption that is not tenable, especially when segmental heterogeneity is significant, as has been shown to be, in the safety context.  Heterogeneity refers to the deviations in patterns of accident occurrences at individual roadway segments, and how these deviations are referred to an “average site.”  This average site can be a virtual site represented by the group mean.  Heterogeneity results in overdispersion of accidents, meaning that the variance of the accident distribution exceeds the mean.  This is due to the fact that the probability of accident occurrence is not uniformly distributed in space and time.  Therefore, one can expect accidents to cluster at various locations on the transportation network, such as at intersections, interchanges, lane drops or lane additions, horizontal or vertical curves, or at locations where decisions have to be made by drivers regarding lane changing, braking, speed reduction or acceleration, or route change.

Given this background, the problem of evaluating safety interventions is compounded by the challenge of modeling the effect of heterogeneity simultaneously alongside the modeling of the marginal effect of an intervention.  There are two primary contributors to this challenge.  The first contributor is selection bias, which arises when locations for safety interventions are not randomly chosen.  The second contributor is the scale of measurement of this bias.  It may be that selection bias at aggregate scales (for example, in instances where corridor length treatments are applied) is influenced by heterogeneity in a different manner compared to bias at smaller scales (for example, spot interventions).  The impact of this variation is that the assessment of safety interventions can be varied depending on the scale at which the evaluation is conducted.  Hence, the methodological problem of simultaneously addressing heterogeneity and selection bias is the objective of this dissertation.  This dissertation attempts to provide some perspective to this problem via multiple scales, by proposing a joint model of heterogeneity and selection bias using a discrete-count approach, and using this framework to address the following research questions:

 

  1. What is the impact of selection bias on safety intervention due to scale? In other words, if safety interventions are applied at locations where accident patterns are severe and frequent, how does one account for the lack of intervention at less problematic locations?  And how does a statistical methodology derived for selection bias provide inference across scales, as segments are scaled up from very small lengths to lengths of the order of corridors?
  2. How does one represent insights into the policy implications of selection bias in a manner that integrates context (i.e., roadway location and characteristics) and scale?

 

I use freeway roadway lighting as an example safety intervention to make these evaluations in this dissertation.  Roadway lighting is installed in order to improve traffic flow, thereby also contributing to improved roadway safety.  Roadway lighting is installed in various forms – as in median-side lighting, versus right-side lighting, versus tunnel lighting, versus ramp-mainline merge points, versus, installations on both sides of the traveled way.  This dissertation involved data collection on all 1,528 centerline miles of interstate freeway in Washington State and analyzed the correlation of accident frequencies with roadway lighting installation, after accounting for roadway geometrics and traffic flow levels.  It was determined that certain installations are more effective than others, when selection bias is taken into account.  For example, right-side lighting installation is found to be effective in reducing accident frequencies compared to other types of lighting installation, indicating a 30% reduction in accident frequencies compared to segments where there is no roadway lighting at the lighting segmentation scale.  Such a result appears to justify the installation of right-side lighting at critical locations such as ramp merge points or departure points.  The key phrase is “appears to justify”.  This dissertation explores the extent to which scale affects inferences such as the above.  With different scales of segmentation, such as interchange and noninterchange segments, one mile uniform length segments, or accident-cluster length segments, right-side installation has a smaller reduction of accident frequencies compared to accident reduction at the lighting segmentation scale.  In the case of accident-cluster level segmentation, right-side lighting installation is associated with an increase in accident frequency.  This example result demonstrates that the scale of data plays a very important role in safety inferences, especially when heterogeneity and selectivity bias are accounted for.

While roadway lighting is used as an example for application of this dissertation’s analytical framework, it is expected that the full-purpose self-contained computational framework for analyzing safety outcomes will be of substantial interest to the safety community at large.  One can use this framework for the analysis of any safety intervention at any scale.  The framework incorporates the typical geometric design decisions used in practice, and therefore, analysts can use this framework to address selectivity bias arising from roadway improvement projects involving all geometric types.  In particular, the framework developed in this dissertation can also aid decision makers to conduct scenario testing.  One example of scenario testing would be to examine the impact of energy-conservation efforts on traffic safety patterns on urban and rural freeways.  Another would be to explore the design contexts associated with high levels of unobserved heterogeneity, where the discussion on the measurement of factors that do not currently exist in highway databases can be motivated.  Example factors relating to heterogeneity could involve measures of segment-level kinematics such as speed, speed dispersion, and headway following distances.  Or, they could involve microclimatic measurements such as pavement temperatures, determination of icing likelihoods, wind gust speeds and sun angles.

 

TABLE OF CONTENTS

LIST OF FIGURES ……………………………………………………………………………………….. ix

LIST OF TABLES …………………………………………………………………………………………. xi

Chapter 1  INTRODUCTION ………………………………………………………………………….. 1

Chapter 2  RELATED WORKS AND RESEARCH QUESTIONS ……………………… 6

2.1 Segmentation Methods ……………………………………………………………………….. 6

2.2 Selectivity Bias and Two Step Process ………………………………………………….. 13

2.3 Mixed Logit Model …………………………………………………………………………….. 21

Chapter 3  SAFETY ANALYSIS PROCESS AND EMPRIRCAL SETTINGS …….. 24

3.1 Safety Analysis Process ………………………………………………………………………. 243.2 Segmentation Data Setting ………………………………………………………………….. 28

3.3 Descriptive Statistics Results ……………………………………………………………….. 36

Chapter 4  STATISTICAL MODELING RESULTS FOR

INTERSTATE LIGHTING SEGMENTATION ………………………………………….. 44

4.1 Negative Binomial Model Results ………………………………………………………… 44

4.2 The Mixed Multinomial-Selection Negative Binomial Count Treatment Effects Model and Negative Binomial Model with Pre-Processing ………….. 49

4.3 Elasticity Estimation Result of Model Outputs ………………………………………. 65

4.4 Summary of Findings in Model Results ………………………………………………… 74

Chapter 5  CONCLUSION AND DIRECTIONS FOR

FUTURE RESEARCH …………………………………………………………………………….. 77

References …………………………………………………………………………………………………….. 93

Bibliography …………………………………………………………………………………………………. 96

Glossary ……………………………………………………………………………………………………….. 100

Appendix ………………………………………………………………………………………………………. 102

Chapter 1

 

INTRODUCTION

 

Since the development of the American Association for State Highway Transportation Officials (AASHTO) Strategic Highway Safety Plan (SHSP), several states have adopted a similar approach in developing their own highway safety plan.  The

Washington Department of Transportation (WSDOT) is a recognized leader in this area.  The WSDOT conducts statistical modeling and visualization analyses as part of their accident research initiative in order to establish systematic bases in its strategic safety plan.  This dissertation is an in-depth, original look at statistical approaches appropriate for public agency decision making; hence, the focus of this dissertation is empirical.  The main motivations for this dissertation are drawn from current and prior research conducted by the author for the Washington State Department of Transportation.

The WSDOT research effort was begun in 2006, with the author leading research activities in the area of data collection methods for freeway accident data systems.  Significant goals of the effort were: a) exchangeability of data in multiple formats, b) usability of data for the development of statistical models, c) post-processing of model outputs for visualization, and d) usability of the above components for integrated prioritization of freeway corridors. The author demonstrated the viability of components “a” and “b” via two bodies of work, namely his MS thesis (Oh 2006) and a research report published for the WSDOT in June 2008 (Shankar et al, 2008).

 

The earlier work in 2006 reproduced accident, geometric, and traffic flow data by direction for 124 centerline miles in a consistent, complete record format.  The latter work in 2008 extended this reproduction to the entire interstate system consisting of 1,528 centerline miles.  To the author’s knowledge, consistent and complete database compilation involving over 100 accident related variables, traffic flow, and geometrics at a statewide scale has not been done in the nation.  While the lack of such efforts may sound surprising, potential reasons do exist.  Some challenges occurred when creating record consistency and completeness in the proper format.  For example, some highway log information such as number of lanes, shoulder widths, and presence of median barrier by type, were available as text documents in their original form.  Traffic flow data such as annual average daily traffic (AADT) was available in electronic form at 0.1-mile or 1mile intervals.  The 0.1-mile data were interpolations of AADT measured using loop detectors which are not necessarily regularly placed at 0.1-mile or 1-mile intervals on the state interstate system.  Interpolations were produced by WSDOT in-house through a feedback algorithm that ensures consistency with neighboring AADT computations.   Accident data was available for multiple years in the form of detailed accident reports that contained information by severity type (property damage only, possible injury, evident injury, disabling injury and fatality), type of collision such as entering at angle, sideswipe, same direction, fixed object, overturn or headon, vehicle involvement, driver related factors such as alcohol or drug involvement, seat belt use, age, gender, occupant information including factors similar to that for the driver and in addition, occupant position in vehicle, and environmental factors such as occurrence of snowy, icy, rainy or dry driving conditions, as well as presence or absence of roadway lighting.  It should be noted here that this database is event-specific.  In order to construct segment-level decision frameworks, which is the primary objective of this dissertation, event-specific information needs to be aggregated to appropriate scales.  The appropriateness of scale depends on the level and nature of the research questions being asked.  In this dissertation, the following research questions are asked:

  1. What is the impact of selection bias on safety intervention due to scale? In other words, if safety interventions are applied at locations where accident patterns are severe and frequent, how does one account for the lack of intervention at less problematic locations?  And how does a statistical methodology derived for selection bias provide inference across scales, as segments are scaled up from very small lengths to lengths of the order of corridors?
  2. How does one represent insights into the policy implications of selection bias in a manner that integrates context (i.e., roadway location and characteristics) and scale?

 

Given this background and objectives, the remainder of this dissertation is organized as follows.  I review literature of direct relevance to the dissertation and in addition provide a bibliography relevant to the dissertation itself.  The review includes segmentation and selection bias research related to transportation applications.  In this sense, the technical benefit of the work will be to provide flexibility, and as a result, scalability in modeling safety.

It was noted that the marginal effects of key infrastructure variables are of interest.  From a policy standpoint, this is definitely a substantial motivation, because cost effectiveness is a major factor driving prioritization schemes for safety, mobility, and accessibility related infrastructure improvements.  Herein lies an issue of statistical significance; typically, transportation improvements are applied at locations where a need is determined to exist.  Therefore, the application of improvements is not random; rather, it follows a selection rule.  In the transportation case, a selection rule may be based on ordering of need.  A decision making framework that is empirically based uses accident data observed at either the selection locations alone, or also at locations where improvements are not applied.  In either case, some accounting needs to occur for the selection bias associated with the marginal effect of the improvement in question.  For example, if I consider roadway lighting as the variable of interest, then it can be argued that the marginal effect of roadway lighting can be expected to decrease accident propensity at locations where lighting is installed.  A policy based on the examination of just lighting-only locations may estimate the effectiveness of roadway lighting with bias. Statistical and econometric methods involving the treatment of selectivity bias are documented, and the literature review in this dissertation addresses that.

Following the literature review, I present the methods employed in this dissertation.  A description of data collection and segmentation methods to obtain structured datasets at multiple scales for accident analyses is provided.  Statistical  modeling designs for this research are also presented.  I discuss the results to demonstrate the viability of the methods proposed and conclude with major findings and recommendations for future work.

SAFETY MODELING VIA SEGMENTATION OF TRANSPORTATION NETWORKS

Sharing is caring!

Leave a Reply