LONGITUDINAL METHODS FOR PRIORITIZING HIGHWAY SAFETY INVESTMENTS

ABSTRACT

 

This dissertation develops a statistical basis for the evaluation of crash costs associated with interstate travel in Washington State. The statistical basis consists of two components – a frequency component and a severity component. The two components are linked via a recursive modeling structure where the frequency evaluation is conducted at a segmental level. In particular, I demonstrate this evaluation at the interchange level of spacing. I evaluate the entire interstate network in Washington State, via a longitudinal analysis of crash histories for the period 1999-2007. For the nine year period, I collected geometric and traffic volume information, and evaluated crash cost bases for the most severe outcome of a crash.

The frequency component is motivated by the need to fill the gap in the extant literature. Current literature is sparse if not nonexistent in terms of insights on heterogeneity in the overdispersion parameter effect at the segmental level of analysis. The overdispersion parameter is modeled as a parametric function of highway geometry, resulting in a heterogeneous negative binomial model structure. Further, to improve insights into the effects of geometry from a design policy standpoint and enhance our understanding of the performance basis of geometric elements, I develop heterogeneous negative binomial models of heterogeneous geometry – based on the definition of segmental geometry that can consist of below-standard, at-standard and abovestandard design elements. By using this approach, enhanced insight is developed due to the fact that interactions between varying standards within a segment effectively capture statistical information observed from reported crashes. This is shown by improvements in likelihoods over contemporary models which employ a non-standards approach.

The second major contribution of this dissertation is the severity analytical component which is built as a second stage component. In the second stage analysis, the severity models are constructed using collision type information, as opposed to the traditional approach of involving

 

geometrics. This allows the simplification of severity models at the unconditional level to highly tractable multinomial type specifications – a computational advantage for computing efficiently the severity outcomes on entire networks.

The two-stage statistical basis thus developed in this dissertation offers a plausible and tractable methodology for determining accurately cost bases for the identification of high priority safety investment corridors. The fact that multi-year histories are used in both stages implies that the findings can be potentially robust and therefore contribute to stable parameter estimation and therefore increase our confidence in our understanding of the effect of geometry and collision type on crash frequency and severity.

 

 

 

TABLE OF CONTENTS

List of Figures ………………………………………………………………………………………………………….. v

List of Tables …………………………………………………………………………………………………………… vi

Acknowledgements …………………………………………………………………………………………………… vii

Chapter 1  Introduction ……………………………………………………………………………………………… 1

1.1 Background …………………………………………………………………………………………………. 1

1.2 Objectives ……………………………………………………………………………………………………. 2

1.3 Scope of the Study ……………………………………………………………………………………….. 6

Chapter 2  Literature Review ……………………………………………………………………………………… 10

2.1 Previous Research ………………………………………………………………………………………… 10

2.1.1 Contemporary View on Geometric Effects on Crash Frequency ……………….. 10

2.1.2 Accident Severity ……………………………………………………………………………….. 13

2.2 Summary …………………………………………………………………………………………………….. 16

Chapter 3  Data Description ……………………………………………………………………………………….. 18

3.1 Data Collection Procedure …………………………………………………………………………….. 19

3.2 Segmentation Procedure………………………………………………………………………………… 20

3.3 Accident Data ………………………………………………………………………………………………. 22

3.4 Annual Average Daily Traffic Data ………………………………………………………………… 24

3.5 Roadway Lighting Data ………………………………………………………………………………… 26

3.6 Number of Lanes and Shoulder Width …………………………………………………………….. 27

3.7 Curve Alignment Data ………………………………………………………………………………….. 29

3.8 Descriptive Statistics for Roadway Infrastructure Data ……………………………………… 34

3.8.1 All Segments ……………………………………………………………………………………… 34

3.8.2 Interchange Segments …………………………………………………………………………. 36

3.8.3 Non-Interchange Segments ………………………………………………………………….. 37

3.9 Descriptive Statistics for Accident and AADT Data …………………………………………. 38

3.9.1 All Segments ……………………………………………………………………………………… 38

3.9.2 Interchange Segments …………………………………………………………………………. 39

3.9.3 Non-Interchange Segments ………………………………………………………………….. 40

Chapter 4  Methodology ……………………………………………………………………………………………. 41

4.1. Heterogeneous Negative Binomial Model of Crash Type Frequency …………………. 43

4.2. Multinomial Logit Model of Segmental Severity Proportions ……………………………. 48

4.3 A Heterogeneous Design Framework for Crash Analysis ………………………………….. 52

Chapter 5  Results …………………………………………………………………………………………………….. 58

5.1 Result of Accident Occurrence Model …………………………………………………………….. 64

5.2 Result of Accident Severity Model …………………………………………………………………. 72

5.3 Post-Model Evaluation of Heterogeneous Geometric Slopes Relative to Fixed

Geometric Slopes ……………………………………………………………………………………….. 82

5.3.1 Prediction Accuracy ……………………………………………………………………………. 83

5.3.2 Conditional Efficiency of Predictions ……………………………………………………. 84

5.3.3 Hausman Test for Evaluating the Conditional Efficiency Basis ………………… 88

5.3.4 Conditional Efficiency Regression Evaluation ……………………………………….. 89

Chapter 6  Conclusions and Recommendations …………………………………………………………….. 91

Chapter 1

 

Introduction

1.1 Background

As transportation systems have developed, traffic safety continues to remain a critical issue in modern society; much research has been conducted to improve traffic safety. Consequently, roadways have evolved to improve driver safety – advances in pavement design and construction, the addition and improvement in traffic signs to alert the driver, and defined design specifications on roadway alignment in an effort to reduce accident frequency and severity, all have enhanced driver comfort in the driving environment. However, as the ratio of vehicle ownership and exposure time on the road grows, the likelihood of experiencing an accident has increased. As a result, the number of segments where treatment is needed to reduce accidents has also increased as well.

However, budget is limited so that it is important to select the segments where treatment would be needed. One of the methods to identify segments is measuring the accident rate and sorting the segments with the highest value to the lowest value. By doing this process, budget could be allocated effectively.

For this process, until recently, most studies have attempted to find the relationship between accidents and geometric design by modeling a formula for each environmental situation – curved or straight segment with horizontal, or vertical alignment, or a combination of both. This only establishes a relationship with little consideration that the real situation was built upon. Based on the results from these types of studies, the spot or segment that has the highest number of accidents are selected as the most hazardous ones and funding is allocated to improve those specific spots or segments. Although this method can be considered as one solution, more

detailed analysis could be achieved with subdivided variables of geometrics and design criteria. Sometimes aggregated data have misleading results because a dominated value or part of variables in geometry design criteria disregards other important aspects, which do not address the budget allocation problem to effectively treat the chosen segments. Besides, being the heterogeneity on every single mile needs to be considered. This also misleads results with the same aggregation problem.

Based on the current methods of research, it is apparent that a more effective approach is needed that would contribute more of a benefit to society. By selecting segments prudently, the potential gains in highway safety could alter the manner in which roadways should be developed and treated. A method of investment based on need is emerging as an important finance, funding, and policy issue.

1.2 Objectives

The main objective of this research is to develop a performance-based methodology that would be the basis for identifying high-priority safety corridors on highways. There are multiple components that comprise this methodology: the crash occurrence component, the crash severity component and therefore a crash cost component. The crash cost component is usually determined on the basis of average societal costs incurred by the traveler as a result of the crash.  The average cost is the cost that is determined on the basis of averaging all crashes nationwide of the same reported severity. Crash severity therefore is fundamental to the determination of prioritization of safety corridors.

Basically, crash cost function for each segment which will be input and output values could be composed of frequency of crashes and cost of crashes.

 

 

where, S : frequency of crashes i,             C : cost of crashes i, and             n : segments

In addition, frequency of crashes(S) could be the product of total number of crashes (λ) and probability (P) of severity (i) for segment (n), i.e . The national system of classification of crash severities includes the following five-category definition: a) property damage only, where vehicle only costs are incurred in excess of 500 dollars (costs below that are usually not reported on police forms), b) possible injury (which can include non-obvious injuries such as whiplash or sore neck, for example), c) evident injury which can include obvious external injuries that are incapacitating, d) disabling injury which includes very severe injuries that are incapacitating, and e) fatal injuries which include the death of at least one traveler at the scene of the crash or at the hospital. The nationally averaged costs for these categories are 7,400, 44,900, 79,000, 216,000, and 158,200 dollars respectively (AASHTO, 2010). For the optimization of societal costs over the entire network, the safety problem becomes at least a two-step problem: in the first step, the problem involves the computation of crash occurrence likelihoods of a certain underlying severity, which then is multiplied by the cost component at the segmental level and summed up across segments to yield a total expected societal cost over the transportation network. The second step involves the computation of safety investment alternatives based on the crash likelihood profiles developed in step 1. These alternatives at a minimum include a) a least cost alternative with minimal improvements such as traffic control adjustments, b) a medium cost alternative which includes for example, improvements such as lane additions for a portion of the corridor, and c) high-cost alternatives which includes for example, the construction of a new interchange to separate the movement of high-speed conflicting traffic. Once these alternatives are costed out, a benefit cost is determined for each alternative and the optimization problem is then viewed as a constrained problem with budget, time and availability of labor as major constraints and cross-programmatic constraints as additional constraints depending on the location of the project. Safety prioritization in the fullest sense involves the above steps. The above described process has been found to be sensitive to the computations of the first step where likelihoods are to be accurately determined (Milton et al. 2008). This dissertation appeals to the first step which in contemporary research is not built on longitudinal histories, but rather, simplified explanations (see for example, the HSM 2010) based on average daily traffic. A vast body of literature exists in this area in terms of calibration of these parsimonious methods.

However, little attention has been paid to the impact of “design standards,” the very foundation of highway planning and construction. A comprehensive framework that explicitly involves the integration of “design standards” from a performance based perspective into the computation of crash likelihoods is missing in the extant literature. Further, any existing methods addressing design standards assume fixed standards (prescriptive) to be input variables, rather than “soft standards” which would involve acceptable deviations from a prescriptive threshold for a particular design element. A “soft standards” approach explicitly allows for providing a finer resolution to the modeling spectrum in terms of variable behavior and interactions. Further, it helps us to explicitly address the well-known overdispersion in traffic safety in a flexible manner (Shankar et al. 1995).

Longitudinal methods involve the use of multi-year time histories of crashes, geometrics and traffic volumes. They can involve the development of likelihoods at the segmental level by accounting for within-segment variation across time and then computation of the overall likelihood across segments in a single step. Many of the methods involve non-closed form approaches such as simulated maximum likelihood involving quadrature or approximations using pseudo-random draws. In any event, the goal of this approach is to shed light on the comparative advantages of “soft standards” models in which segments which have standards varying within the segment are partitioned into multiple standard categories based on proportions by length.   The idea is that the combination of soft standards within certain ranges might yield acceptable safety thresholds even though the segment as a whole may not fulfill a prescriptive threshold.  This is the core motivation behind the modeling approach in this dissertation. I intend to explore this by developing two basic types of models: a) one where parameters are fixed, but standards vary, with the associated condition that the overdispersion effect is parameterized as a function of geometry, and b) where parameters are random and standards vary, but associated with the condition that parameter heterogeneity captures any overdipersion association with geometry and therefore does not require overdispersion parameterization.

 

 

Figure 1-1. Research Objectives.

1.3 Scope of the Study

The data used are accident history data in the Washing State area spanning 9-years

(1999~2007). Seven major Interstate highways in the Washington State area were analyzed: I-005 (276 centerline miles), I-082 (133 centerline miles), I-090 (297 centerline miles), I-182 (15 centerline miles), I-205 (11centerline miles), I-405 (30 centerline miles), and I-705 (1.5 centerline miles). Accidents have been analyzed to define the specific accident severity, type, and number of vehicles involved, that happened frequently or remarkably on each segment. Prioritizing segments based on the weighted value granted to each input-accident type and output-geometric was the principal method that was utilized. Figure 1-2 shows the distribution of accident severity types on the objective interstates. Specifically, the PDO severity type is the most frequently occurring accident on all of the interstates for the period 1999-2007.

	All Accidents	PDO	Disabling Injuries	Evident Injuries	Possible Injuries	Fatalities
I – 005	71,238	44,647	735	5,806	19,809	241
I – 082	4,198	2,504	175	809	649	61
I – 090	18,611	11,872	351	2,506	3,739	143
I – 182	717	449	22	125	117	4
I – 205	1,184	751	14	108	306	5
I – 405	16,267	9,924	105	952	5,261	25
I – 705	294	168	2	38	85	1

Figure 1-2. Accident Severity Distribution across the Seven Interstates in Washington State in 1999-2007.

In Figure 1-3, the distribution of the number of vehicles involved in accidents on the objective interstates is shown. Among the five types of vehicle involvement, one or two vehicles involved in accidents are the most frequently occurring types on all of the interstates for the period 1999-2007. Two vehicles involved accidents on I-005 are remarkably higher than the others; two vehicles related accidents on I-405 are high as well. One vehicle involved accident seems a little higher on I-005, I-082, and I-090 than other types of vehicle involvements for the objective interstates.

	All Accidents	One Vehicle	Two Vehicles	Three Vehicles	Four Vehicles	Five or More Vehicles
I – 005	71,238	15,534	41,900	10,403	2,556	845
I – 082	4,198	2,883	1,197	104	8	6
I – 090	18,611	9,751	7,189	1,275	296	100
I – 182	717	342	335	33	6	1
I – 205	1,184	332	700	122	21	9
I – 405	16,267	2,015	10,475	2,895	685	197
I – 705	294	103	156	26	8	1

Figure 1-3. Number of Vehicles Involved Accident Distribution across the Seven Interstates in Washington State in 1999-2007.

Figure 1-4 presents the distribution of seven accident types on the objective interstates for the period 1999-2007. In particular, the accident type of rearends occurred extraordinarily frequently on I-005 and I-405, similarly as two vehicles related accidents in Figure 1-3. The next frequent accident types which have high proportions are fixed objects and sideswipe. More specific accident data will be discussed in Chapter 3.

	All Accidents	Sideswipe	Headon	Rearends	Overturns	Fixed Objects	Same Direction	Others
I – 005	71,238	11,629	59	38,061	1,859	11,880	3,470	4,280
I – 082	4,198	382	10	505	1,270	1,126	228	77
I – 090	18,611	1,896	20	4,862	2,689	5,702	1,185	2,257
I – 182	717	95	1	171	158	125	61	106
I – 205	1,184	217	0	508	75	220	75	89
I – 405	16,267	2,188	4	11,174	170	1,837	550	344
I – 705	294	43	5	122	7	95	12	0

Figure 1-4. Accident Type Distribution across the Seven Interstates in Washington State in 19992007.

In addition, data description along with other types of data, such as a range of design perspective, annual averaged daily traffic, and lighting types, were used in each analysis and more specific information about this data will be shown later. To achieve the goal of this study, the dissertation is organized as follows: I begin with a literature review discussing prior studies to establish the state of knowledge and knowledge gaps this thesis aims to fill, followed by a description of the empirical context of this dissertation, with attention to crash histories, geometries and traffic volumes, and methods available to determine segmental definitions used in this dissertation. The fourth chapter discusses the methodologies used in the estimation of the two basic types of statistical models and the associated estimation procedures. These techniques are tested in chapter 5 which describes the results of the modeling efforts, a comparative analysis of the two basic approaches in terms of variable findings, differences in marginal effects and elasticities where applicable. In chapter 6, I conclude by visiting the major findings of this dissertation, including the strengths and weaknesses of this study as well as directions for future study.

It is hoped this dissertation will fill a major void in the extant safety literature via this comparative analysis of longitudinal methods. This is an emergent field with vast application and as more fine resolution data becomes available, the prospects of extending methods in this dissertation are believed to be fertile.

LONGITUDINAL METHODS FOR PRIORITIZING HIGHWAY SAFETY INVESTMENTS

Sharing is caring!