Study on crash characteristics and injury severity at roadway work zones

Transcription

1 Unversty of South Florda Scholar Commons Graduate Theses and Dssertatons Graduate School 2009 Study on crash characterstcs and njury severty at roadway work zones Qng Wang Unversty of South Florda Follow ths and addtonal works at: Part of the Amercan Studes Commons Scholar Commons Ctaton Wang, Qng, "Study on crash characterstcs and njury severty at roadway work zones" (2009). Graduate Theses and Dssertatons. Ths Thess s brought to you for free and open access by the Graduate School at Scholar Commons. It has been accepted for ncluson n Graduate Theses and Dssertatons by an authorzed admnstrator of Scholar Commons. For more nformaton, please contact scholarcommons@usf.edu.

2 Study On Crash Characterstcs And Injury Severty At Roadway Work Zones by Qng Wang A thess submtted n partal fulfllment of the requrements for the degree of Master of Scence n Cvl Engneerng Department of Cvl & Envronmental Engneerng College of Engneerng Unversty of South Florda Major Professor: Jan Lu, Ph.D. Abdul Pnjar, Ph.D. Yu Zhang, Ph.D. Zhenyu Wang, Ph.D. Date of Approval: March 26, 2009 Keywords: ordered probt regresson, descrptve statstcs, age groups, traffc safety, margnal effects Copyrght 2009, Qng Wang

3 ACKNOWLEDGMENTS It s wth great prde that I thank the brllant mnds afflated wth the Department of Cvl and Envronment Engneerng at the Unversty of South Florda. I would lke to gve specal thanks to my major professor, Dr. Jan John Lu, for the gudance he has provded. In addton, I would lke to thank commttee members Dr. Abdul Pnjar, Dr. Yu Zhang, and Dr. Zhenyu Wang. Ths thess would not have been possble wthout your contrbutons.

4 TABLE OF CONTENTS LIST OF TABLES LIST OF FIGURES ABSTRACT v v INTRODUCTION Background Research Objectves and Approaches Organzaton 6 LITERATURE REVIEW Prevous Studes on Work Zone Crashes Prevous Studes on Crash Severty Model Log-lnear Model Logt Model Probt Model 16 DESCRIPTIVE STATISTICS ANALYSIS The Trend of Crashes Dstrbuton of Crashes by Drvers Age Dstrbuton of Crashes by Crash Severty Dstrbuton of Crashes by Clmatc Envronmental Condtons Dstrbuton of Crashes by Crash Types Dstrbuton of Crashes by Contrbutng Factors Predomnant Factors for Other Varables 31 CRASH SEVERITY MODEL Methodology Crash Severty Models Ordered Probt Regresson Crtera for Ordered Probt Models z - Test Pseudo - R Lkelhood Raton (LR) Test 40

5 4.1.4 Interpretaton of Model Coeffcents The Partal Change n y* Partal Change n Predcted Probabltes Data Collecton Data Base Data Descrpton Work Zone Crash Injury Severty Model Estmaton Procedure Cross Tabulatons between Explanatory Varables and Crash Severty Estmaton Results Interpretaton Sgns Magntude of Coeffcents Detaled Interpretatons Possble Countermeasures to Improve Work Zone Safety 58 SUMMARY Summary 60 REFERENCES 64 BIBLIOGRAPHY 67 APPENDICES 68 Appendx A: Varables and Codes of Work Zone Crash 69

6 LIST OF TABLES Table 4.1 Defnton and Descrpton of Crash Severty Level Table 4.2 Tables from Florda Traffc Crash Records Database Table 4.3 Descrpton of Selected Varables for Model Development Table 4.4 Descrpton Statstc of Contnuous Varables Table 4.5 Frequences of Dscrete Varables Table 4.6 Cross Tabulaton between explanatory Varables and Crash Severty Table 4.7 Estmaton of Ordered Probt Regresson for Work Zone Crash Severty Model Table 4.8 Partal Change n y * 53 Table 4.9 Partal Change n Predcted Probabltes Table 4.10 Analyss of the Coeffcent Sgns Table 4.11 Ranked Magntudes of Coeffcents

7 LIST OF FIGURES Fgure 1.1 Component Parts of a Work Zone Fgure 3.1 Work Zone Crashes and Work Zone Fatal Crashes n Florda Fgure 3.2 Dstrbuton of Work Zone and Non-work Zone Crashes by Age Group Fgure 3.3 Dstrbuton of Work Zone Crashes by Crash Severty Fgure 3.4 Dstrbuton of Work Zone Crashes by Lghtng Condtons Fgure 3.5 Dstrbuton of Work Zone Crashes by Weather Condtons Fgure 3.6 Dstrbuton of Work Zone Crashes by Road Surface Condtons Fgure 3.7 Dstrbuton of Work Zone Crashes by Crash Types Fgure 3.8 Dstrbuton of Work Zone and Non-work Zone Crashes by Crash Types Fgure 3.9 Dstrbuton of Work Zone Crashes by Contrbutng Factors Fgure 3.10 Dstrbuton of Work Zone Rear-end Crashes by Contrbutng Factors Fgure 3.11 Dstrbuton of Work Zone Angle Crashes by Contrbutng Factors Fgure 3.12 Dstrbuton of Work Zone Sdeswpe Crashes by Contrbutng Factors Fgure 3.13 Dstrbuton of Work Zone Crashes by Alcohol/Drug Involved Fgure 3.14 Dstrbuton of Work Zone Crashes by Heavy Vehcle Involved v 32

8 Fgure 3.15 Dstrbuton of Work Zone and Non-work Zone Crashes by Heavy Vehcle Involved Fgure 4.1 Latent Varables to the Observed Categores Fgure 4.2 Dstrbuton of y Gven x for the Ordered Regresson Model 35 v

9 Study on Crash Characterstcs and Injury Severty at Roadway Work Zones Qng Wang ABSTRACT In USA, despte recent efforts to mprove work zone safety, the number of crashes and fataltes at work zones has ncreased contnuously over several past years. For addressng the exstng safety problems, a clear understandng of the characterstcs of work zone crashes s necessary. Ths thess summarzed a research study focusng on work zone traffc crash analyss to nvestgate the characterstcs of work zone crashes and to dentfy the factors contrbutng to njury severty at work zones. These factors ncluded roadway desgn, envronmental condtons, traffc condtons and vehcle/drver features. Especally, specal populaton groups, whch dvded nto older, mddle Age, and young, were nspected. Ths study was based on hstory crash data from the Florda State, whch were extracted from the Florda CAR (Crash Analyss Reportng) system. Descrptve statstcs method was used to fnd the characterstcs of crashes at work zones. After then, an njury severty predct model, usng the ordered probt regresson technology, was developed to nvestgate the mpacts of varous factors on dfferent the njury severty at work zones. From the model, t can be concluded that some factors, v

10 ncludng the road secton wth curve, alcohol/drugs nvolved, a hgh speed, angle crash and too young or old drvers are more lkely to ncrease the probablty of angle crashes. Based on the magntudes of the varable coeffcents, the factor of maxmum posted speed have a great mpact to njury severty, whch shows restrcton to drvng speed s prncple countermeasure for mprovng work zone safety. v

11 CHAPTER ONE INTRODUCTION 1.1 Background In Hghway Capacty Manual 2002, the defnton of work zone s a segment of hghway n whch mantenance and constructon operatons mpnge on the number of lanes avalable to traffc or affect the operatonal characterstcs of traffc flowng through the segment. It should be typcally marked by sgns, channelzng devces, barrers, pavement markng, and/or work vehcles. It extends from the frst warmng sgn or hgh-ntensty rotatng, flashng, oscllatng, or strobe lghts on a vehcle to the End Road Work sgh or the last temporary traffc control devce. The Manual on Unform Traffc Control Devces lsts fve dstnct areas wthn a work zone. Each of these has a specfc purpose and may vary n sze and locaton dependng on the specfcs of each work zone. The fve areas are: advance warnng area, transton area, actvty area, buffer space, and termnaton area (Fgure 1.1). The advance warnng area s the secton of hghway where road users are nformed about the upcomng work zone or ncdent area. The transton area s that secton of hghway where road users are redrected out of ther normal path. Transton areas usually nvolve strategc use of tapers, whch because of ther mportance are dscussed 1

12 Fgure 1.1 Component Parts of a Work Zone 2

13 separately n detal. The actvty area s the secton of the hghway where the work actvty takes place. It s comprsed of the work space, the traffc space, and the buffer space. The work space s that porton of the hghway closed to road users and set asde for workers, equpment, and materal, and a shadow vehcle f one s used upstream. Work spaces are usually delneated for road users by channelzng devces or, to exclude vehcles and pedestrans, by temporary barrers. Typcally, the buffer space s formed as a traffc sland and defned by channelzng devces. When a shadow vehcle, arrow panel, or changeable message sgn s placed n a closed lane n advance of a work space, only the area upstream of the vehcle, arrow panel, or changeable message sgn consttutes the buffer space. The termnaton s the end area of work zone. Work zone safety has always been a hgh prorty ssue n hghway systems but remans unsatsfactory n USA. Based on the statstcs from FHWA (Federal Hghway Admnstraton), n 2007, there were 835 work zone fataltes, whch represent 2.0% of all roadway fataltes for the year. Over four out of every fve-work zone fataltes were motorsts. In addton, there are over 40,000 njures at work zones. The total cost of hghway work zone njures calculates to $9.25 bllon per year. The hghway work zone fataltes per bllon dollars spent, are at lst 4 tmes more than n total constructon (Maze et al., 2000). Estmatng between 1995 and 1997, the drect costs of hghway constructon zone accdents were as hgh as $6.2 bllon per year, and the average cost s $3687 per accdent (Mohan and Gautam, 2002) 3

14 To mprove work zone safety, four felds need to be approached contemporaneously: engneerng, educaton, enforcement, and coordnaton wth publc agences. Engneerng: Ths focuses on standardzaton and evaluaton. The standardzaton part s for traffc control and safety devces n work zone areas. The MUTCD (Manual on Unform Traffc Control Devces) s the natonal safety standards to control traffc through work zones, and the NCHRP350 (Natonal Cooperatve Hghway Research Program Report 350 Recommended Procedures for the Safety Performance Evaluaton of Hghway Features ) contans the federal standards and gudelnes for all work zone safety devces. The natonal gudelnes regardng plannng and mplementng work zones s keepng update to address the changng tmes of more traffc more congeston, greater safety ssues, and more work zones. Educaton: Publc awareness s mproved through a varety of actvtes lke clearnghouse webste ( tranng courses for federal, state, local and trbal hghway engneers; conferences, CDs; gudebooks; brochures (for the general publc and hghway practtoners); blngual safety publc outreach materals; and press events such as Natonal Work Zone Awareness Week. Enforcement: Engneers n federal hghway work closely wth state hghway to dentfy approprate engneerng safety countermeasures for hgh-rsk locatons new roads. They also work wth the enforcement communty such as the IACP (Internatonal Assocaton of Chefs of Polce). Speed enforcement s a top safety concern n work 4

15 zones snce t has crtcal relatonshp wth crash severty. In Maryland, Mchgan and Vrgna, VSL (Varable speed Lmts) demonstraton projects whch determne approprate speeds for work zones and change them when condtons change were to analyze varatons on speed and accompany drver behavor. Assocaton: Workng wth emergency medcal servces, polce and fre organzatons can ensure that publc safety s mantaned at hgh levels and access for emergency vehcles s possble durng work zone operatons. AASHTO (Amercan Assocaton of State Hghway and Transportaton Offcals), ATSSA (Amercan Traffc Safety Servces Assocaton) and FHWA found the Natonal Work Zone Awareness Week n Aprl every year to brng natonal attenton to motorst and worker safety and moblty ssues n work zones. Besde ths, lots of other publcatons lke Basc Traffc Control for Utlty Operatons manual and Strategc Hghway Safety Plan are the productons by more than one partner or sponsor. Researchng the characterstcs of crashes s the very frst step of learnng the defcences of work zone safety and countermeasures. In addton, studyng the characterstc dfferences between each crash njury severty level may cause the dscovery of factors nfluencng njury severty change, whch could beneft the development of traffc controls for reducng the proporton of hgh-severty crashes n total crashes. 5

16 1.2 Research Objectves and Approaches The man objectves of ths study are to nvestgate the characterstcs of accdents n work zones, to dentfy the factors contrbutng to njury severty levels, and to study how these factors nfluence njury levels. For more specfcally, ths study follows these steps: (1) Revew the prevous researches n the feld of work zone crash characterstcs and njury severty models. (2) Determne the most promsng model for model development part by comparng varous models n lterature revew part. (3) Investgate the dfferences of characterstcs such as crash severty, envronmental condtons, crash types and contrbutng factors among three drver age groups. (4) Develop a crash severty model for the dentfcaton of the most sgnfcant factors contrbutng to the njury severty levels. 1.3 Organzaton Ths thess conssts of fve chapters. Chapter 1 provdes a bref ntroducton of the research, ncludng the background of the research, research objectve and approaches. Chapter 2 dscusses the past studes n both work zone crash characterstcs and crash njury severty models, and chooses the most approprate model to develop the work zone njury severty model for ths study. Chapter 3 compares the descrptve characterstcs of 6

17 work zone crashes n three age groups, ncludng the crash severty, envronmental condtons and some other contrbutng factors. A crash njury severty model s produced and nterpreted; the factors that nfluence crash severty levels are found are gven n chapter 4. Fnally, chapter 6 provdes a summary and the concluson of ths research. 7

18 CHAPTER TWO LITERATURE REVIEW 2.1 Prevous Studes on Work Zone Crashes Many studes have been performed on accdent experence wthn work area n the Unted States. Most of them focus on the crash characterstc n dverse work zone types, crash severty levels, and dfferent locatons wthn work zone. Ullman et al. (2005) presented an analyss of the safety effects of nght work actvty upon crashes at two types of constructon projects n Texas. The frst project type nvolved both day and nght work, whereas the other project type nvolved pavement resurfacng actvtes performed only at nght. They found that crashes ncreased more sgnfcantly durng perods of work actvty than durng perods when the work zone was nactve. Overall, the ncrease durng work actvty was somewhat hgher at nght than durng the day. Researchers also found that crashes ncreased more at nght than durng the day at the hybrd projects even when the work zone was nactve, presumably reflectng a dsproportonate nfluence of the temporary geometrcs and traffc control upon nghttme travel at these stes. 77 fatal work zone crash stes throughout Texas from Feb to Apr were analyzed by Schrck (2004). Based on these nvestgatons, researchers concluded that 8

19 only 8 percent of the nvestgated crashes had a drect nfluence from the work zone, whereas 39 percent of the nvestgated crashes had an ndrect nfluence from the work zone. Researchers also concluded that 45 percent of the nvestgated crashes appeared to have no nfluence from the work zone (ncluded n ths subset are the 16 percent of the nvestgated crashes whch occurred n work zones that were work zones n name only, such as work zones that conssted only of project lmt sgnng). The characterstcs of hghway work zone collsons and ther detaled locatons wthn work zones were studed by Garber and Zhao (2002) to enhance the selecton of effectve countermeasures. The objectve was to determne the dstrbuton and characterstcs of crashes n specfc areas wthn a work zone and to compare selected characterstcs of work zone crashes wth those of non-work zone crashes. In ther study, the dfferent locatons n the work zone were referred to as the advance warnng area, transton area (taper), longtudnal buffer area, actvty area, and termnaton area. Based on the crash percentages regardng locaton, severty, and collson type, the researchers concluded several major fndngs. Frst, the actvty area had the hghest number of crashes and the hghest number of fatal crashes whle the termnaton area was the safest area n terms of numbers of crashes. Second, property-damage-only (PDO) crashes were the predomnant severty type, followed by the njury crashes. Thrd, rear-end crashes were predomnant for all areas and all road types except for the termnaton area, where all crashes were angle crashes. Fourth, as traffc moved from the transton area to the 9

20 work area, the proportons of rear-end and same-drecton sdeswpe crashes decreased and the proportons of fxed-object, off-road, and angle crashes ncreased, although rear-end crashes were stll predomnant. Last, most nghttme work zone crashes were n the actvty area and the severtes of nghttme and daytme work zone crashes were not sgnfcantly dfferent. In 2000, Danel et al. performed a study whch was expanded further to examne the dfference between fatal crash actvty wthn work zones compared wth fatal crashes n non-work-zone locatons. Usng data from three work zone locatons n Georga, fatal crash actvty wthn work zones also was compared wth nonfatal crashes wthn work zones. Fnally, fatal crash actvty was examned to determne the nfluence of the work zone actvty on the frequency of fatal crashes. The overall fndngs of the study ndcate that the work zone nfluences the manner of collson, lght condtons, truck nvolvement, and roadway functonal classfcaton under whch fatal crashes occur. The study also ndcates that fatal crashes n work zones are more lkely to nvolve another vehcle than non-work-zone fatal crashes, and fatal crashes n work zones are less nfluenced by horzontal and vertcal algnment than are non-work-zone crashes. Khattak et al. (2002) created a unque dataset of Calforna freeway work zones that ncluded crash data (crash frequency and njury severty), road nventory data (average daly traffc and urban/rural character), and work zone related data (duraton, length, and locaton). Crash rates and crash frequences were nvestgated n the pre-work zone and 10

21 durng-work zone perods. For the freeway work zones nvestgated n ths study, the total crash rate n the durng-work zone perod was 21.5% hgher (0.79 crashes per mllon vehcle km) than the pre-work zone perod (0.65 crashes per mllon vehcle km). Compared to the pre-work zone perod, the ncrease n non-njury and njury crash rates n the durng-work zone perod was 23.8% and 17.3%, respectvely. Next, crash frequences were nvestgated usng negatve bnomal models, whch showed that frequences ncreased wth ncreasng work zone duraton, length, and average daly traffc. Wang et al. (1996) dscussed the prmary questons that safety researches are attemptng to answer. The results were presented of an nvestgaton to (a) determned what s known about the magntude of hghway work zone crashes, (b) examned characterstcs of hghway work zone crashes usng the Hghway Safety Informaton System, (c) nvestgated how work zone accdents are reported on polce accdent report forms and wthn state accdent report systems, (d) dentfed crtcal vods n the knowledge of the relatve safety of work zones, and (e) examned possble ways to address unfulflled nformaton needs related to work zone safety. 2.2 Prevous Studes on Crash Severty Model Researchers have employed many statstcal technques to analyze crash severty level. Among these technques were log-lnear, logt, and probt models. 11

22 2.2.1 Log-lnear Model Usng 1994 and 1995 crash data from Florda, Abdel-Aty et al. (1998) used log-lnear technque to examne relatonshps between drver age and crash characterstcs. The three njury severtes n ther study were no njury, njury and fatalty, and ther results suggest that njury severty s postvely assocated wth age; they also concluded that mddle-age drvers are more lkely to be nvolved n some crashes, but older drvers are more lkely to be nvolved n fatal crashes. Km et al. (1995) used log-lnear models to predct automoble crash and njury severty. The results suggested that alcohol or drug use and lack of seat belt use ncrease the odds of more severe crashes and njures Logt Model Logstc regresson models were developed by Donnell and Mason (2004) usng both an ordnal and a nomnal response. The results ndcateed that modelng crash severty as an ordnal response provded approprate results for cross-medan crashes, whereas a nomnal response was more approprate for medan barrer crashes. Explanatory varables such as pavement surface condtons, use of drugs or alcohol, presence of an nterchange entrance ramp, horzontal algnment, crash type, and average daly traffc volumes affect crash severty. The analyss results mght be used by practtoners to understand the trade-off between geometrc desgn decsons and medan-related crash severty. Approxmately 0.7% medan barrer crashes on the 12

23 Interstate system resulted n a fatalty, whereas 43% were property-damage-only crashes and about 56% were njury crashes. More than 17% of cross-medan collsons were fatal, and 67% nvolved njury. Modelng severty as a dscrete outcome nvolves estmatng the probablty that a vehcular crash has a certan severty by determnng the lkelhood of outcomes gven that a crash has occurred. Lee and Chang (2002) estmated the severty of run-off-road crashes n the state of Washngton, agan by usng the nested logt model. Temporal, envronmental, drver, roadway, and roadsde characterstcs were used to estmate property damage and possble njury probabltes for rural run-off-road crashes condtoned on no evdent njury. The fndngs ndcated that wet pavement surfaces resulted n possble njury, drvers younger than 25 were more lkely to be nvolved n njury crashes, alcohol-mpared drvers were more lkely to be nvolved n njury crashes, and crashes n the presence of a horzontal curve were more lkely to nvolve an njury. Dssanayake and Lu (2002) used bnary logstc regresson model takes the followng form. Factors that prove most nfluental n predctng severty n young drver crashes ncluded nfluence of alcohol or drugs, ejecton n the crash, pont of mpact, crash locaton, exstence of horzontal curve or vertcal grades at the crash ste, speed of the vehcle, and restrant devce usage. Krull, Khattak, and Councl (2000) used logt models to analyze drver njury severty nvolved n a sngle-vehcle crash. Three-year crash data from Mchgan and 13

24 Illnos were analyzed to explore the effect of rollover, whle controllng for roadway, vehcle, and drver factors. Results showed that drver njury severty ncreases wth: (a) falure to use a seatbelt, (b) passenger cars as opposed to pck-up trucks, (c) alcohol use, (d) daylght, (e) rural roads as opposed to urban, (f) posted speed lmt, and (g) dry pavement as opposed to slppery pavement. Chang and Mannerng (1999) estmated a nested logt model to study the occupancy crash njury severty relatonshp. Crash data of prncple arterals, state hghways, and nterstates n Seattle, Washngton, durng 1994 were used n the analyss. The dependent varable was the crash severty, whch represents the most severe level of njury sustaned by any vehcle occupant nvolved n the crash. The occupancy can be sgnfcant because vehcles wth large occupances have an ncreased lkelhood of havng someone serously njured. Separate models were estmated for non-truck-nvolved crashes and for non-truck-nvolved crashes. Results showed that ncreased severty was more lkely for truck-nvolved crashes, hgh speed lmts, crashes occurrng when a vehcle s makng a rght or left turn, and rear-end types of collsons. Shankar, Mannerng, and Barfeld (1996) estmated a nested logt model to analyze crash severty of sngle-vehcle crashes on rural freeways. All possble nestng structures (whch examne possble correlaton among the choces) were consdered and statstcally tested by the lkelhood rato test. The authors found that a nested-logt model, whch 14

25 treated property damage only (no njury) and possble shared characterstcs of njury crashes, fts the data best. Shankar and Mannerng (1996) used a multnomal logt specfcaton for estmatng motorcycle rder crash severty lkelhood condtoned on the occurrence of a crash. Fve levels of severty are consdered: property damage only, possble njury, evdent njury, severe njury, and fatalty. Crash data were 5-year statewde data on sngle-vehcle motorcycle crash from the state of Washngton. Results showed that the multnomal logt formulaton s a promsng approach to evaluate the determnants of motorcycle crash severty. Nassar, Saccomanno, and Shortreed (1994) estmated a nested logt model to predct crash severty. Three separate models were calbrated for three crash stuatons: sngle-vehcle, two-vehcle, and mult-vehcle crashes. Factors that affect the level of damage experenced by ndvduals nvolved n traffc crashes nclude a crash dynamc term, seatng poston, seat belt use, vehcle condton, vehcle mass, drver condton, and drver acton. Road surface condton was nsgnfcant n the models. Bad weather condtons may prompt drvers to slow down and keep a safe dstance from other vehcles. 15

26 2.2.3 Probt Model Abdel-Aty and Keller (2005) produced ordered probt models for crash severty level and used the tree-based regresson to explore the factors whch affect njury level. The results of ths research showed that when attemptng to forecast the number of expected crashes of dfferent severty levels, t s mperatve that models are developed for each level of collson nstead of aggregatng crash types to predct the overall severty level. Whle the ordered probt model approach had been adopted, as dd many prevous researchers, usng the tree-based regresson for each severty level mproved our understandng of the specfc factors and ther mportance for each severty level. Furthermore, the results showed that crashes reported on short-forms are mportant and should therefore be retaned and ncluded n crash databases. Ignorng ths data could lead to basng the results by under reportng crashes of certan severty or type that could be related to specfc explanatory factors. Other crash types or severtes mght appear to have hgher percentages, and therefore, ther effect could be artfcally exaggerated. Khattak and Targa (2004), Khattak et al. (2002, 2003) used ordered probt models to predct the njury level for crashes occurrng at constructon zones and nvolvng trucks, to predct njury severty for sngle-vehcle truck rollovers, and to determne vehcle, roadway, drver, crash, and envronmental characterstcs that nfluence the severty level of older drvers nvolved n crashes, respectvely. 16

27 Abdel-Aty (2003) appled the ordered probt models to predct crash njury severty on roadway sectons, sgnalzed ntersectons and toll plazas. Models explaned a drver s volaton was sgnfcant n the case of sgnalzed ntersectons. Alcohol, lghtng condtons, and the exstence of a horzontal curve affected the lkelhood of njures n the roadway sectons model. A varable specfc to toll plazas, vehcles equpped wth Electronc Toll Collecton (ETC), had a postve effect on the probablty of hgher njury severty at toll plazas. Other varables that entered nto some of the models were weather condton, area type, and some nteracton factors. Ths study llustrates the smlartes and the dfferences n the factors that affect njury severty between dfferent locatons. Kockelman and Kweon (2002) descrbed the use of ordered probt models to examne the rsk of dfferent njury levels sustaned under all crash types, two-vehcle crashes, and sngle-vehcle crashes. The results suggested that pckups and sport utlty vehcles are less safe than passenger cars under sngle-vehcle crash condtons. In two-vehcle crashes, however, these vehcle types were assocated wth less severe njures for ther drvers and more severe njures for occupants of ther collson partners. Other conclusons also were presented; for example, the results ndcated that males and younger drvers n newer vehcles at lower speeds sustan less severe njures. Toshyuk and Shankar (2002) used a bvarate ordered-response probt model to study drver and most severely njured passenger severty n collson wth fxed objects n Washngton State. Results showed that cy roadway surface and ran decrease the 17

28 probablty of more severe drver njury. The type of fxed objects sgnfcantly affects drver s njury severty. Guardrals have dfferent effects on drver s njury whether the collsons are wth ts face or wth ts leadng end. Proper use of a restrant system sgnfcantly decreases the probablty of more severe drver njury. Male and younger drvers have a lower probablty of more severe njury, probably because of ther physcal strength. Also, drver s unconscousness causes more severe drver njury. Duncan, Khattak, and Councl (1999) used ordered probt modelng to examne the occupant characterstcs and roadway and envronmental condtons that nfluence njury severty n rear-end crashes nvolvng truck-passenger car collsons. Two models were developed, one wth the basc varables and the other ncludng nteractons among the ndependent varables. Results revealed that an ncreased severty rsk exsts for hgher speed crashes, those occurrng at nght, for women, when alcohol s nvolved, and for crashes when a passenger car rear-ends a truck at a large dfferental speed between the two vehcles. Khattak (1999) appled the ordered probt model to examne the effect of nformaton (accuracy of nformaton conveyed by brake and turnng lghts) and other factors on rear-end crash propagaton and the propensty of drver njury n such crashes. Results on njury severty showed that n a two-vehcle crash, the leadng drver s more lkely to be njured, whereas, n a three-vehcle crash, the drver n the mddle s lkely to 18

29 be more severely njured. Furthermore, as rear-end crashes propagate from two-vehcles to three-vehcles the last drver s relatvely less severely njured. Klop (1998) examned the mpacts of physcal and envronmental factors on the severty of njury to bcyclsts n North Carolna. Usng the ordered probt model, the effect of a set of roadway, envronmental, and crash varables on njury severty was explored. Separate models were estmated for rural and urban locatons. Results ndcated that straght grades, curved grades, darkness, and fog sgnfcantly ncrease njury severty. Rensk, Khattak, and Councl (1998) estmated ordered probt models to explore the effects of polcy varables on njury severty. Results showed that hghway segments where speed lmts were rased by 10 mph resulted n a hgher probablty of ncreased severty than those rased by only 5 mph. No sgnfcant changes n njury severty were found for the hghway segments where speed lmts were rased from 65 to 70 mph. In assessng the probabltes of four levels of njury severty as a functon of drver attrbutes, O Donnell and Connor (1996) compared ordered logt and ordered probt specfcatons. Ther results suggest that njury severty rses wth speed, vehcle age, occupant age (squared), female gender, blood alcohol levels over 0.08 percent, non-use of a seatbelt, manner of collson (e.g., head-on crashes), and travel n a lght-duty truck. And, accordng to ther comparson of effects, seatng poston of crash vctms was most relevant (e.g., the left-rear seat of the vehcle was found to be most dangerous) and 19

30 gender least relevant. Many of ther results are echoed n the models presented here; the key dstncton s that here collson partners and crash-type are examned and emphaszed. Hutchnson (1986) developed an ordered probt model to study occupants njury severty when nvolved n traffc crashes. Brtsh crash data for had been processed to gve a cross-tabulaton of the severty of njury to the drver and to the front seat passenger n four types of sngle-vehcle crashes (overturnng and non-overturnng, each n rural and urban areas). Results showed that passengers tend to be more serously njured than drvers n nonoverturnng, but that there s no dfference n overturnng crashes. 20

31 CHAPTER THREE DESCRIPTIVE STATISTICS ANALYSIS 3.1 The Trend of Crashes The trend of work zone crashes and fatal crashes are ascendng contnuously from 2002 to 2006 n Florda (see Fgure 3.1). The average annual ncrease rate of work zone crashes s 18.8%, and the fatal crashes n 2006 are 64.4% more than one n Ths trend ndcates that the work zone safety n Florda remaned a serous concern. Fgure 3.1 Work Zone Crashes and Work Zone Fatal Crashes n Florda 21

32 3.2 Dstrbuton of Crashes by Drvers Age Fgure 3.2 shows the age dstrbuton of the at-fault drvers for work zone and non-work zone crashes. The drvers are dvded nto three age groups: Young Age (less than 25), Mddle Age (25 64) and Elderly Age (greater than 65). In work zone area, the mddle age drvers cause the hghest proporton (67%) of crashes, whle the elderly drvers are only responsble for 9% of the crashes. The drver group havng the second hghest crash rate (24%) s the young age drvers. Compared to work zone crashes, mddle age drvers n non-work zone area have a lower possblty of occurrng crashes (63%). Fgure 3.2 Dstrbuton of Work Zone and Non-work Zone Crashes by Age Group 22

33 3.3 Dstrbuton of Crashes by Crash Severty The dstrbuton of work zone crashes by crash severty s shown n Fgure 3.3, whch ndcates that the mddle age drvers nvolved the hghest percentage n the no njury crashes whch s 49%, and always has the lowest percentage n other severty levels. Whle n the more severe level crashes, elderly drvers contrbute more than the other two age groups (Incapactatng Injury: Old Drvers 9% and Fatal Injury: Old Drvers 2%). Fgure 3.3 Dstrbuton of Work Zone Crashes by Crash Severty 23

34 3.4 Dstrbuton of Crashes by Clmatc Envronmental Condtons Clmatc Envronmental condtons nclude lghtng condtons, weather condtons, and road surface condtons. Fgure 3.4 summarzes the dstrbuton of crashes by lghtng condtons. Most crashes occur when lghtng condton s good. Elderly drvers s most lkely to havng crashes under good lghtng condton (daylght), and only has 18% crashes under non-daylght condton ncludng dawn, dusk and dark condtons. In contrast, the dfference of crash rate between these two lghtng condtons n young drvers s not remarkable. Fgure 3.4 Dstrbuton of Work Zone Crashes by Lghtng Condtons 24

35 The results of analyss of the dstrbuton of work zone crashes by weather and road surface condtons are shown n Fgures 3.5 and 3.6 respectvely. The results ndcate that n all three age groups only a small proporton of work zone crashes occur n bad weather or bad road surface condtons. In contrast to the common sense, the adverse weather and road condtons do not have sgnfcant nfluence on the work zone fatal crashes. Fgure 3.5 Dstrbuton of Work Zone Crashes by Weather Condtons 25

36 Fgure 3.6 Dstrbuton of Work Zone Crashes by Road Surface Condtons 3.5 Dstrbuton of Crashes by Crash Types As llustrated n Fgure 3.7, the top three work zone crash types n all age groups are the same. There are rear-end, angle and sdeswpe whch are defned as the prncple crash types n ths study. In young and mddle age groups, the percentage of rear-end crashes s obvously hgher than angle and sdeswpe crashes. Elderly age group shows hgher rate n angle crashes than others. Compared work zone and non-work zone crash types n Fgure 3.8, read-end and sdeswpe crashes are more lkely to be occurred n work zone area. 26

37 Fgure 3.7 Dstrbuton of Work Zone Crashes by Crash Types Fgure 3.8 Dstrbuton of Work Zone and Non-work Zone Crashes by Crash Types 3.6 Dstrbuton of Crashes by Contrbutng Factors Fgure 3.9 represents the dstrbuton of contrbutng factors by all drvers and each age group. Among all drvers, careless drvng, the most predomnant contrbutng factor, s responsble for 43% of total crashes. Another predomnant factor s faled to yeld rght of way (11%) followed by no mproper drvng acton (10%) and mproper lane change (7%) respectvely. In young and mddle age group, the dstrbutons are bascally same as 27

38 whch of all drvers, except that young drvers show slghtly hgher rate n careless drvng (48%), and the second and thrd factors whch are not varant too much n rate. But n elderly age group, the rate of frst factor s just 34% and second one s more than 10% hgher than other two age groups. Fgure 3.9 Dstrbuton of Work Zone Crashes by Contrbutng Factors Fgure 3.10 to 3.12 express the dstrbuton of predomnant contrbutng factors over the prncpal crash types. The most predomnant contrbutng factor for rear-end crashes s careless drvng (average 74% n all three age groups). A dfference between elderly age group and the other two age groups s that mproper lane change s not a predomnant 28

39 contrbutng factor for older age drvers but t s for young age drvers and mddle age drvers. Fgure 3.10 Dstrbuton of Work Zone Rear-end Crashes by Contrbutng Factors Faled to yeld rght of way s the most predomnant contrbutng factors for angle crashes. In elderly age group, the rate of ths crash type s sgnfcantly hgher than young and mddle age groups; otherwse the rate of careless drvng s less than others. For sdeswpe crashes, the mproper lane change s the most frequent contrbutng factor n mddle (36%) and elderly (40%) age group, and second most one s careless drvng (19% for both groups). However, for young drvers, the top two factors have no much dfference (27% for mproper lane change and 30% for careless drvng). 29

40 Fgure 3.11 Dstrbuton of Work Zone Angle Crashes by Contrbutng Factors Fgure 3.12 Dstrbuton of Work Zone Sdeswpe Crashes by Contrbutng Factors 30

41 3.7 Predomnant Factors for Other Varables The dstrbutons of alcohol/drug nvolved and heavy vehcle (heavy truck and truck tractor) nvolved are gven n Fgure 3.13, 3.14 and Old drvers are seldom nfluenced by alcohol/drug (only 1% nvolved), and most work zone crashes for young age group s not ncluded by heavy vehcle. But heavy vehcle s more easly related to work zone crashes (14%) than non-work zone crashes (7%). Fgure 3.13 Dstrbuton of Work Zone Crashes by Alcohol/Drug Involved 31

42 Fgure 3.14 Dstrbuton of Work Zone Crashes by Heavy Vehcle Involved Fgure 3.15 Dstrbuton of Work Zone and Non-work Zone Crashes by Heavy Vehcle Involved 32

43 CHAPTER FOUR CRASH SEVERITY MODEL 4.1 Methodology As stated n prevous papers, In contrast to the multnomal models whch neglect the data s ordnarly and requre more parameters estmated and nested logt models that produce better results but have complexty n dentfyng the nestng structure, the ordered probt models wth a relatvely smple approach recognze the ndexed nature of varous response varables. They are recommended to analyze the crash severty levels Crash Severty Models The crash severty model n ths study was developed to nvestgate the factors that affect crash severty n work zone area. The dependent varable n the model s njury severty level, and the ndependent varables are the factors whch have sgnfcant nfluence on the crash severty. The crash njury severty s a typcal ordnal varable whch could be categorzed at fve levels from the least severe level to the most severe level (shown n Table 4.1). 33

44 Table 4.1 Defnton and Descrpton of Crash Severty Level Level Defnton Descrpton 1 No Injury there s no reason to beleve any person receved bodly harm from the crash 2 Possble Injury No vsble sgns of njury but complant of pan or momentary unconscousness 3 Non-ncapactatng Vsble njures from the such as bruses, abrasons, Injury lmpng, etc. 4 Incapactatng Any vsble sgns of njury from the crash and person(s) Injury had to be carred from the scene. 5 Fatal Injury an njury sustaned n a motor vehcle crash that results n death wthn 90 days Ordered Probt Regresson The ordered probt model s as followed: y = α + β + ε x (4.1) where y s the latent and contnuous measure of crash njury severty; s the number of crashes faced by ths severty level; x s a vector of parameters to be estmated; ε s a random error term whch assumed to follow a normal dstrbuton wth mean 0 and varance 1. The pdf (Probablty Densty Functon) s 2 1 ε φ (4.2) ( ε ) = exp 2π 2 and the cdf (Cumulatve dstrbuton Functon) s 2 ε 1 ε ( ) Φ ε = exp dt 2π 2 (4.3) The observed and coded dscrete crash njury severty varable y s determned from the model as follows: 34

45 35 < < < < < = If If 1 If 3 If 2 If n n n y n y n y y y y τ τ τ τ τ τ τ τ (4.4) Ths mappng from the latent varable to the observed crash njury severty class s llustrated n Fgure 4.1. Fgure 4.1 Latent Varables to the Observed Categores Fgure 4.2 Dstrbuton of y Gven x for the Ordered Regresson Model

46 Consder Fgure 4.2 whch shows the dstrbuton of y for four values of x. The errors are dstrbuted normally around the regresson lne E( y x) = α + βx. The Probablty of outcome m corresponds to the area of the error dstrbuton between the cutponts τ m 1 and τ m. Ths area s computed as follows. Frst consder the formula for the probablty that y = 1. We observe y = 1 when y falls between Pr τ = and τ 1. Ths mples that 0 ( = 1x ) = Pr( τ y < τ x ) y 0 1 (4.5) Substtutng Pr y = α + β + ε, x ( 1x ) = Pr( τ α + x β + ε < τ x ) y 0 1 Then, subtractng = (4.6) x β wthn the nequalty, ( y 1 x ) = Pr( τ α x β ε < τ α x β x ) = 0 1 Pr (4.7) The probablty that a random varable s between two values s the dfference between the cdf evaluated at these values. Therefore, Pr ( y 1x ) = Pr( ε < τ α x β x ) ( ε < τ α x β x ) = 0 1 Pr ( τ α x β ) ( τ α x β ) = 1 F 0 F (4.8) These steps can be generalzed to compute the probablty of any observed outcome y = m gven x: ( y m x ) = F( τ α x β ) F( τ α x β ) Pr (4.9) = m m 1 When computng Pr ( y = 1x), the second term on the rght-hand sde drops out snce ( τ β ) = F( xβ ) 0 F 0 = x ; when computng Pr ( y = J x), the frst term equal 1 snce 36

47 ( τ xβ ) = F( xβ ) 1 F = J. Thus, for a model wth four observed outcomes, such as shown n Fgure 4.2, the formulas for the ordered probt model are Pr Pr ( y 1x ) = Φ( τ α βx ) = 1 ( y 2 x ) = Φ( τ α βx ) Φ( τ α βx ) = 2 1 ( y 3 x ) = Φ( τ α βx ) Φ( τ α βx ) = 3 2 Pr (4.10) Pr Pr ( y n 1 x ) = Φ( τ α βx ) Φ( τ α βx ) = n n 1 ( y n x ) = 1 Φ( τ α βx ) = n 1 where s an ndvdual; 1, 2, 3 n-1, n are response alternatves; Φ ( ) s the standard normal cumulatve dstrbuton functon. Snce y s latent, ts mean and varance cannot be estmated. The varance s dentfed by usng that Var ( ε x) = 1. Whle these assumptons dentfy the varance, the mean of y s stll undentfed. The consequences of ths can be seen by consderng the model y = α + βx + ε wth cutponts τ m. Thnk of α and the τ s as the true parameters n the sense that they were used to generate the observed data. Defne an alternatve set of parameters: α = α δ and τ m = τ m δ (4.11) where τ s an arbtrary constant. The probablty that y = m s dentcal, whether the true or alternatve parameters are used: 37

48 Pr ( y m x) = F( τ α xβ ) F( τ α xβ ) = m m 1 ([ τ δ ] [ α δ ] βx) F( [ τ δ ] [ α δ ] βx) = F m m 1 ( τ α x β ) ( τ α x β ) = F (4.12) m 1 F m 1 Snce both sets of parameters generate the same value for the probablty of an observed outcome, there s no way to choose between the two sets of parameters usng the observed data: a change n the ntercept n the structural model can always be compensated for by a correspondng change n the thresholds. That s to say, the model s undentfed. Whle there are an nfnte number of assumptons that could be made to dentfy the model, only two are commonly used: (1). Assume that τ 1 = 0. Ths nvolves settng δ = τ 1 n Equaton (2). Assume that α = 0. Ths nvolves settng δ = α n Equaton Both assumptons dentfy the model by mposng a constrant on one of the parameters. The dfferent dentfyng assumptons lead to what are known as dfferent parameterzatons of the model. The choce of whch parameterzaton to use s arbtrary and does not affect the β s (except for β 0 ) or assocated sgnfcance tests. Further, as known by Equaton 4.12, the probabltes are not affected by the dentfyng assumpton. However, understandng the dfferent parameterzatons s mportant snce dfferent software uses dfferent parameterzatons. Programs such as LIMDEP uses the frst assumpton, whle programs such as Markov, SAS s LOGISTIC, and Stata use the second 38

49 one. The choce of parameterzaton does not affect estmates of the slopes, but does affect the estmates of β 0 and the τ s Crtera for Ordered Probt Models z - Test z - Test s used to test the statstcal sgnfcance of ndvdual estmated coeffcent n ordered porbt models. Maxmum lkelhood estmators possess a number of desrable propertes when certan general condtons apply. Independent and dentcally dstrbuted observatons, and ndependence of the x and the model errors (the ε ) are all that s requred. Wth these condtons satsfed, the maxmum lkelhood estmator s asymptotcally unbased (consstent), s normally dstrbuted, and has the smallest varance among all consstent and asymptotcally normal estmators. The t ratos for the null hypothess H 0 that β = 0, and the test statstc s ˆ β z = (4.11) ˆ σ k where βˆ s the estmator of β ; and β s the th coeffcent of the model; σˆ s the estmator of standard devaton of the coeffcent β ; s number of observatons. If H 0 s true, the coeffcent β of the model s not statstcally sgnfcant. If H 0 s rejected at a confdence level (usually s 0.05), the coeffcent β s sgnfcant to the response. 39

50 Pseudo - A Pseudo - 2 R 2 R s often used as a goodness-of-ft measure n non-lnear models. They look lke 2 R n the sense that they are on a smlar scale, rangng from 0 to 1, but they cannot be nterpreted as one would nterpret an ordnary least squares (OLS) 2 R and dfferent Pseudo - 2 R can arrve at very dfferent values. Here, the Pseudo - R 2 Lˆ ( M full ) ( M ) nt ercept 2 R s provded as ln = 1 (4.12) ln Lˆ where M full s the model wth predctors; M nt ercept s the model wthout predctors; Lˆ s the estmated lkelhood. A lkelhood falls between 0 and 1, so the log of a lkelhood s less than or equal to zero. If a model has a very low lkelhood, then the log of the lkelhood wll have a larger magntude than the log of a more lkely model. Thus, a small rato of log lkelhoods ndcates that the full model s a far better ft than the ntercept model Lkelhood Raton (LR) Test The lkelhood rato test s a statstcal test of the goodness-of-ft between two models. It reles on a test statstc computed by takng the rato of the maxmum value of the lkelhood functon under the constrant of the null hypothess to the maxmum wth that constrant relaxed. The null hypothess s H : β 0, where β s the ntercept. Ths statstc s gven as 40 0 =

51 G 2 [ L( M ) ln L( M )] = 2 ln (4.13) constraned unconstraned where L ( M constraned ) s the lkelhood of the constraned model; ( ) lkelhood of the unconstraned model. L s the M unconstraned Ths LRT statstc approxmately follows a ch-square dstrbuton. The degree of freedom s equal to the number of addtonal parameters n the unconstraned model. If the null hypothess s rejected (the confdence level s usually 0.05), t can be concluded that at least one ndependent varable has sgnfcant nfluence for the dependent varable Interpretaton of Model Coeffcents The Partal Change n y In the ordered regresson model, y = xβ + ε (4.16) and the partal change n y wth respect to x k s y x k = β k (4.17) Snce the model s lnear n y, the partal change can be nterpreted as: for a unt ncrease n x k, y s expected to change by β k unts, holdng all other varables constant. Because the varance of y cannot be estmate from the observed data, the meanng of a change of β k unts n y s unclear. Interpretatons should be based on y -standardzed coeffcents. 41

52 If σ s the uncondtonal standard devaton of the latent y y, then the y -standardzed coeffcent for x k s Sy β k β k = (4.18) σ y whch can be nterpreted as: for a unt ncrease n x k, y n expected to ncrease by Sy k β standard devatons, holdng all other varables constant. y -standardzed coeffcents ndcate the effect of an ndependent varable n ts orgnal unt of measurement. Ths s sometmes preferable for substantve reasons and s necessary for bnary ndependent varables. The varance of 2 σ = ˆ β Vˆ ar y x y can be estmated by the quadratc form: ( ) ˆ β + Var( ε ) (4.19) where ar( x) Vˆ s the covarance matrx for the x s computed from the observed data; βˆ contans Maxmum Lkelhood (ML) estmates; and Var ( ε ) = 1 n the ordered probt model Partal Change n Predcted Probabltes The predcted probablty that y = m gven x s ( y = m x) = F( τ xβ ) F( τ xβ ) Pr (4.20) m m Takng the partal dervatve wth respect to x k, 42

53 ( y = m x) F( xβ ) F( τ x ) Pr τ m m 1 β = x x x k k = β k f m 1 ( τ xβ ) β f ( τ xβ ) [ f ( τ xβ ) f ( τ xβ )] k k m = β 1 (4.21) k m The partal change or margnal effect s the slope of the curve relatng m x k to Pr ( y = m x), holdng all other varables constant. The sgn of the margnal effect s not necessarly the same as the sgn of β, snce f ( τ xβ ) f ( τ xβ ) m 1 m can be negatve. Indeed, t s possble for the margnal effect of x k to change sgns as x k changes. In general, the margnal effect does not ndcate the change n the probablty that would be observed for a unt changes n x k. However, f an ndependent varable vares over a regon of the probablty curve that s nearly lnear, the margnal effect can be used to summarze the effect of a unt change n the varable on the probablty of an outcome. 4.2 Data Collecton Data Base The dataset used to ft the ordered probt model was extracted from the Florda Crash Analyss Reportng (CAR) system. CAR system s a relatonal database for State System crashes consstng of nne tables whch contan dfferent data relevant to a certan facet of a traffc crash (Table 4.2). It mantans electronc crash records based on crashes reported on the long-form crash report. That the varable FIRST ROAD CONDITION 43

54 CRASH COD s equal to 04 (road under repar/constructon) s used as the ndcator of work zone crashes. In ths study, the work zone crash dataset contaned all the work zone crashes from 2002 to Some varables n the database were selected for modelng. They may nclude ordnal varables, nomnal varables, or contnuous varables. In order to get better result performance all categorcal varables should be purposely converted to bnary ones (dummy varable). The contnuous varables need to be normalzed (by dvdng by each maxmum value) to have values whch le between 0 and 1. The reason for ths s that the dummy varables have means between 0 and 1, and ordered multple choce models are almost never estmable f the varables are of very dfferent magntudes (Greene 1993). All the mssng values are deleted from database. Appendx A lsts the descrpton of every orgnal varable n ths work zone crashes database. Fle Name Events Drvers Passengers Pedestrans Property Vehcles Table 4.2 Tables from Florda Traffc Crash Records Database Descrpton Contans nformaton about the crash event (.e. date, tme, harmful events, etc.). Ths s the "parent fle" of the database. Contans nformaton about each drver nvolved n the crash demographc and causal). Contans nformaton about each passenger nvolved n the crash (demographc and causal). Contans nformaton about each pedestran nvolved n the crash (demographc and causal). Contans nformaton about property (other than vehcles) damaged n the crash. Contans nformaton about each vehcle nvolved n the crash. 44