Portland State Unversty PDXScholar Cvl and Envronmental Engneerng Faculty Publcatons and Presentatons Cvl and Envronmental Engneerng 2008 Combnng clmate, crash, and hghway data for mproved rankng of speed and wnter-weather related crash locatons n Oregon Chrstopher M. Monsere Portland State Unversty, monsere@pdx.edu Peter G. Bosa Portland State Unversty Robert L. Bertn Portland State Unversty Let us know how access to ths document benefts you. Follow ths and addtonal works at: http://pdxscholar.lbrary.pdx.edu/cengn_fac Part of the Cvl and Envronmental Engneerng Commons, and the Urban Studes and Plannng Commons Ctaton Detals Monsere, Chrstopher M.; Bosa, Peter G.; and Bertn, Robert L., "Combnng clmate, crash, and hghway data for mproved rankng of speed and wnter-weather related crash locatons n Oregon" (2008). Cvl and Envronmental Engneerng Faculty Publcatons and Presentatons. Paper 78. http://pdxscholar.lbrary.pdx.edu/cengn_fac/78 Ths Post-Prnt s brought to you for free and open access. It has been accepted for ncluson n Cvl and Envronmental Engneerng Faculty Publcatons and Presentatons by an authorzed admnstrator of PDXScholar. For more nformaton, please contact pdxscholar@pdx.edu.
Combnng Clmate, Crash, and Hghway Data for Improved Rankng of Speed and Wnter-Weather Related Crash Locatons n Oregon Chrstopher M. Monsere, P.E. 1, Peter G. Bosa 2, Robert L. Bertn, P.E. 3 Abstract: In recent years, the technques for screenng transportaton networks to dentfy hgh crash locatons have become more sophstcated. Many transportaton agences, however, lack suffcent data, ether n tmelness, completeness, or accuracy to mplement many of the recent advances. Ths paper presents the results of an emprcal analyss of screenng and rankng for specfc crash type (speed and ce related crashes) on rural 1.6 km (1 m) hghway sectons of Oregon hghways. The analyss ncludes data generated wth the extensve use of spatal technques and ncorporates clmate data to enhance envronmental consderatons. The paper compares the results of fve rankng methods crtcal rate (by functonal class), crtcal rate (by functonal class and clmate zone), potental for crash reducton, expected frequency (adjusted by emprcal-bayes), and frequency. For the emprcal- Bayes (EB) methods, safety performance functons were generated usng negatve bnomal regresson technques. The twenty top 1.6 km (1 m) sectons were dentfed for each method and compared. The results reveal that the frequency and expected frequency methods dentfed the most stes n common, followed by the rate-based methods. The potental for crash reducton method dentfed the most unque ranked lst. The results hghlght the dfferences n rankng methods and confrm that even wth sgnfcant aggregaton to mprove the rate-based methods they dd not dentfy segments smlar to the more sophstcated EB-technques. CE Database subject headngs: 1 Assstant Professor, Department of Cvl & Envronmental Engneerng, Portland State Unversty, P.O. Box 751, Portland, OR 97207-0751, E-mal: monsere@pdx.edu 2 Graduate Research Assstant, School of Urban Studes and Plannng, Portland State Unversty, P.O. Box 751, Portland, OR 97207-0751, E-mal: pbosa@pdx.edu 3 Assocate Professor, Department of Cvl & Envronmental Engneerng and Nohad A. Toulan School of Urban Studes & Plannng, Portland State Unversty, P.O. Box 751, Portland, OR 97207-0751, E-mal: bertn@pdx.edu
2 INTRODUCTION For state hghway safety mprovement programs, the dentfcaton of hgh crash locatons (HCL) s a crtcal tool for safety engneers and planners. In general, networks can be screened and locatons ranked based on frequency, severty, or rate of crashes, some combnaton of these factors n ndces, trend or pattern analyss, or an estmate of the excess crashes wth respect to the expected average for smlar stes (.e., potental for mprovement, potental accdent reducton). In recent years, technques for screenng transportaton networks to dentfy hgh crash locatons have become more sophstcated. These methods have more data requrements than the tradtonal methods and state departments of transportaton (DOTs) have been slow to adopt them. Many agences stll rely on tradtonal methods to dentfy canddate locatons for safety mprovement (frequency, rate, severty). For example, Hallmark, et al. (2001) recently surveyed 17 other state DOTs for the Iowa DOT and found that they all used frequency, rate, severty or a combnaton of methods. Each method has relatve advantages and dsadvantages but the fnal crteron for how well a screenng method does s, ultmately, how effectvely t dentfes locatons where correctable crashes are found (Cheng and Washngton, 2005a). Once an HCL has been dentfed, the typcal procedure s to perform detaled engneerng studes to dagnose potental countermeasures based on ste condtons, crash hstory, and other factors. Followng a beneft-cost analyss, the most promsng stes are usually programmed for mprovement. Thus, t s mportant to dentfy stes wth the most promse for mprovement snce engneerng studes are expensve, agences have lmted budgets, and f a ste wth potental s not dentfed, an opportunty to substantally mprove safety are mssed (Hauer, 2002). Whle networks are often screened for total crashes, another approach s to screen for unusual occurrences of a partcular crash type wth some countermeasure n mnd. Analyss of recent data n Oregon reveal that speedng (defned as a drver charged wth a speedng, racng, drvng too fast for condtons as ndcated by polce offcer, or exceedng the posted lmt) s a contrbutng factor n approxmately 31% of fatal crashes (Monsere et al., 2004). Speed has a strong correlaton wth hghway crashes for numerous reasons: t reduces a drver s ablty to safely navgate around curves or objects, ncreases the stoppng dstance of a vehcle, and decreases the tme avalable for collson avodance n dangerous stuatons. Safety professonals at the Oregon DOT are strongly nterested n managng vehcle speeds to mprove road safety through engneerng, enforcement, and educaton strateges. The current Oregon DOT HCL method s a sldng wndow approach that calculates a safety ndex as a functon of
3 frequency, rate, and severty for all crashes and does not prortze hgh crash locatons by a specfc crash type (ODOT, 2006). The focus of ths paper s to compare fve screenng and rankng methods for one specfc crash type speed and ce related crashes. The paper compares crtcal rate (by functonal class), crtcal rate (by functonal class and clmate zone), potental for crash reducton, and emprcal-bayes adjusted frequency and frequency screenng methods. In the sectons that follow, the paper descrbes the data that were assembled for analyss and spatal technques used to create addtonal data. Each of the rankng methods s then descrbed, ncludng the development process of safety performance functons for use n the emprcal-bayes methods. Fnally, the results of the analyss are presented, followed by conclusons. Data Assembly Two prmary sources were used n generatng the data for analyss crash and hghway data. The Statewde Crash Data System (CDS) ncludes all legally reported crashes that occur on publc roads n Oregon. For ths analyss, three years of data (2000-2002) on state hghways were used. The crash data contans detals about the crash, envronmental and roadway condtons, and drver nformaton. For state hghways, crashes are located by a route-mlepost system. Crashes were selected from the crash data base where both cy road condtons were present and at least one vehcle was coded wth a speed nvolvement flag. The speed-nvolved flag s generated by the system and ncludes crashes where speed was too fast for condtons and where the vehcle was exceedng the posted lmt. Both of these felds are based on judgment and are known to be somewhat based (especally n cy condtons), however, t s the only current way to dentfy speed-nvolved crashes. Other hghway data were avalable n a spatal format such as number of lanes, total roadway wdth, shoulder wdth, average daly volumes, presence of medans, and functonal class. A geographc nformaton systems (GIS) representaton of the state hghway system contans a lnear referencng system whch allows these data to easly be attrbuted to any pont on the hghway network. Cty boundares were also used to attrbute sectons as rural or urban and as shown later, clmate data was used to attrbute envronmental factors. How these data were used for network screenng s descrbed n the followng sectons. Rankng Methods Consdered Ths emprcal analyss of rankng speed and ce crashes compares fve methods: crtcal rate (by functonal class), crtcal rate (by functonal class and clmate zone), potental for crash reducton, expected frequency (adjusted
4 by emprcal-bayes), and frequency. Whle severty data were avalable, rankngs that consder severty were not conducted. The rural state hghway system was the focus of ths analyss whch was categorzed accordng to functonal class (nterstate, prncpal arteral, and mnor arteral or collector). For use n ths network screenng, the hghways were dvded n to 1.6 klometer (1 mle) segments. A sldng wndow approach, whch can mprove screenng methods used n practce, was not used n ths comparson. There are 530 1.6 km (1 m) mle rural nterstate sectons, 2,224 rural prncpal arteral sectons, and 2,343 rural mnor arteral sectons. The followng sectons descrbe the methods used. Rate Qualty Control by Functonal Class The rate qualty control (RQC) method determnes a crtcal rate for a partcular segment usng the exposure (vehcle-klometers traveled VKT), a probablty constant, and the average rate for smlar sectons. The crtcal rate s calculated for each 1.6 km (1 m) secton as shown n equaton 1: R C R 1 A = RA + K + (1) M 2M where R C = crtcal rate for each 1.6 km (1 m) secton, R A = the average rate for smlar faclty, K = probablty constant based on desred level of sgnfcance (1.645 for 95%) and M = mllons of vehcle klometers traveled (VKT or enterng vehcles = (V*D*L)/1,000,000). In ths analyss, the average rate s for all roadway segments n the database sharng a smlar functonal classfcaton. The frst column (statewde) n Table 1 lsts the mean crash rates for each of the three functonal classfcatons. All rates were calculated for the total speed and ce crashes that occurred n the three year perod. The 2001 volumes were used for exposure and were assumed to be constant over the three year perod. If the observed crash rate for a segment exceeds the crtcal rate, the secton s consdered hazardous and then ranked by the percent that t exceeds the crtcal rate. Rate Qualty Control by Functonal Class wth Clmate Zones An mprovement n the rate qualty control method s to better defne peer stes for the crash type beng screened. In ths analyss, t s approprate to consder envronmental factors related to snow and ce when aggregatng roadway segments. Oregon s geography produces a dverse clmate whch means that wnter weather related crashes do not occur wth the same frequency n all regons of the state. Temperate regons such as the coast and central Wllamette valley tend to have much fewer ce related crashes than the mountanous regons of the Cascades and eastern Oregon. To capture ths, the state was parttoned nto nne clmate regons.
5 Based on data avalablty, two varables were consdered as contrbutng factors to cy road condtons the mean annual days of measurable snowfall exceedng 0.25 cm (0.1 n) and the mean annual days when the temperature fell below freezng. These data were obtaned from the Natonal Oceanc and Atmospherc Admnstraton s (NOAA) Clmate Atlas of the Unted States (NOAA, 2002). Coverage areas were presented as polygons, wth each polygon representng the mean numbers of days n each category over a 30 year perod. As shown n the upper left of Fgure 1, there are nne categores for mean annual days wth a temperature below 0 C (32 F). Note that the legend s not provded n the fgure but darker shades ndcate a greater mean number of days (the maxmum was 300 n the hghest elevatons, but the mean s 106 days). As shown n the upper rght of Fgure 1, there are nne categores for 29-year mean annual days wth measurable snowfall exceedng 0.25 cm (0.1 n) shown. The darker shades agan represent more days (the darkest color represents a maxmum of 90 days). Though ths resoluton s a bt coarse (0.8 to 1.6 km, or 0.5 to 1 mle) t was suffcent to estmate unque clmate zones. By makng one of the layers transparent, t was possble to vsually merge the two layers and hand-trace the zones based on the snow and freezng temperature layers and an understandng of the geographc features of the state. The result s a set of nne ndvdual clmate zones shown n the bottom half of Fgure 1. Usng these zones, hghways n them were attrbuted wth the clmate data usng a spatal query. The average crash rate for speed-ce related crashes for all roadway sectons n each clmate zone was determned by functonal classfcaton. It should be noted that the clmate zones 1 and 8 do not have any nterstate hghways. The results of ths analyss are shown n Table 1. The table reveals that, as expected, the clmate zones wth more wnter weather condtons have hgher average crash rates n each zone (.e. 5, 7, 8, and 9) wth the excepton of mnor arterals n the North Coast mountans. In fact, the crash rate for rural prncpal arterals sectons n the Cascade Mountans s 0.28 crashes per MVKT (0.45 per MVMT), more than 2 tmes the statewde average of 0.12 crashes per MVKT (0.20 per MVMT) for speed and ce crashes. Other mountanous clmate zones such the Northeast and Central Hgh Desert have smlarly hgh crash rates. The Oregon Coast and Valley & Foothlls clmate zones as well as the South Coast Mountans all have low wnter weather crash rates, due n large part to the temperate natures of these regons. In general, usng more aggregaton should mprove crtcal rate methods snce stes are compared more drectly to peer stes. Usng the average rate for each clmate zone and functonal class, a crtcal rate was calculated for each segment n the analyss regon as descrbed prevously. Sectons were agan ranked n descendng order by the percent that the observed crash rates for speed and ce crashes exceeded the crtcal rate.
6 Potental for Crash Reducton To apply the potental for crash reducton method, safety performance functons were estmated for the three functonal classes of roadways. The best of these models were then combned wth observed crash counts for three years n an emprcal-bayes approach to estmatng the excess crashes on each 1.6 km (1 m) secton. These models were also used to generate a fourth rankng based on an adjusted frequency descrbed n the next secton. Predctve crash models are used to estmate the frequency of crashes based on a set of explanatory varables. In most recent research, safety performance functons have been estmated wth Posson and negatve bnomal (NB) regresson models. Safety performance functons (SPFs) developed usng these technques have been advanced by Hauer (1997), Poch and Mannerng (1996), Shankar et al. (1995), and Maou and Lum (1993) and Hauer (1998), Vogt and Bared (1998), and Persaud (1993), Harwood et al. (2002) and many others. As suggested n Lord et al. (2005), the Posson and NB models are theoretcally appealng representatons of the crash occurrence process. The Posson model s suted for crash count data snce the dstrbuton approxmates rare events; however, t requres the mean of the count process equal ts varance. Evdence n a large body of lterature suggests that most crash data wll lkely be overdspersed (.e., the varance wll be sgnfcantly greater than the mean) and the restrcton can be relaxed n the negatve bnomal regresson method. The Posson model n terms of the probablty of havng y number of crashes per year at locaton s gven by equaton (2) as: P ( y ) e = (2) λ λ y y! where P (y ) = probablty of a crash at locaton havng y crashes per year, λ s the expected number of crashes per year for locaton.e., E (y ). Posson regresson models estmate λ usng standard maxmum lkelhood methods as a functon of X explanatory varables usng the log-lnear model n equaton (3) (Washngton et al., 2003). ln λ = βx (3) The NB model adds an ndependent dstrbuted error term ε n the parameter equaton (3) whch allows relaxes the assumpton the mean equal the varance as shown n (4). The NB model s estmated from equaton (5) usng standard maxmum lkelhood technques where α = a measure of dsperson (Washngton et al., 2003). ln λ = βx + ε (4)
7 L ( λ ) = ((1/ α) + y ) ((1/ α) y!) Γ Γ 1/ α 1/ α + λ 1/ α λ 1/ α + λ y (5) In ths research, crash models were developed usng 2001 data. As one would expect gven that the more populated areas of the state do not experence sgnfcant wnter weather, there s a large percentage of 1.6 km (1 m) segments wth zero crashes n the secton for the year 2001 (72% of nterstate, 85% prncpal arteral, 94% mnor arteral). The mean number of crashes per km per year for nterstates (0.35, or 0.56 crashes per mle), and prncpal arterals (0.14, or 0.23 crashes per mle), and rural mnor (0.04, or 0.07 crashes per mle) are low. The followng varables were selected from the analyss data set for ncluson n the model. A summary of the response varable and the fve predctve varables used n the model s shown n Table 2. TOTALSI_01: total speed and ce crashes on the 1.6 km (1 m) segment for 2001 AADT: average daly traffc for 2001 (vehcles per day) SNWDYS: 30 year mean number of days wth snowfall greater than 0.25 cm (0.1 n), (days) TOT_SURF_W: total wdth of the roadway (ncludng shoulders), (feet) WIDTH_LANE: total wdth of the roadway (ncludng shoulders) per lane, m (ft). TOT_SURF_W: total wdth of the roadway (ncludng shoulders), m (ft) SP_CD: posted speed lmt, km/h (m/h) The counts are clearly over-dspersed and the negatve bnomal regresson models were chosen to model speed and ce crashes. Predctve varables for horzontal curvature were also desred but were not avalable. For the TOT_SURF_W varable on nterstates, the maxmum value for both roadways was used. The WIDTH_LANE varable was created and ntended to capture the amount of roadway allocated per lane. For example, a two-lane road wth 3.6 m lanes and 1.8 m shoulders would have a 5.4 m WIDTH_LANE varable. Sx negatve bnomal regresson models were estmated for 2001 speed and ce crashes (TOTALSI_01) as a functon of exposure, road geometry, and weather varables n Table 2 usng the glm.nb functon n the statstcal software R (R, 2005). The followng models were specfed (note that exposure was transformed wth the natural logarthm consstent wth other approaches to model counts). Model 1: TOTALSI_01 ~ LOGAADT Model 2: TOTALSI_01 ~ LOGAADT + SNWDYS Model 3: TOTALSI_01 ~ LOGAADT + SNWDYS + WIDTH_LANE Model 4: TOTALSI_01 ~ LOGAADT + SNWDYS + TOT_SURF_W Model 5: TOTALSI_01 ~ LOGAADT + SNWDYS + TOT_SURF_W + SP_CD Model 6: TOTALSI_01 ~ LOGAADT + SNWDYS + WIDTH_LANE + SP_CD
8 For each model, estmates of the varable coeffcents were generated, dagnostc plots were created, correlatons of varables were determned, and measures of model ft generated. The estmated coeffcents and sgnfcance for each of the models are shown n Table 3. In all models, the exposure varable s sgnfcant. For all models of nterstate functonal class the coeffcent for exposure s counterntutve n that the sgn s negatve. Ths s because the hgher volume sectons are n clmate zones wth very lmted wnter weather condtons sectons (Valley & Foothlls). The weather varable (SNWDYS) s also sgnfcant n all models and has the expected sgn n all models. The geometry varables are only sgnfcant n about half of the models. The posted speed code s sgnfcant n the nterstate and prncpal arteral models (but has the opposte sgn) and n model 6 for the mnor arterals. In Table 3, measures of model ft are presented. R 2 values (as n least squares lnear regresson) are not avalable for negatve bnomal regresson but a number of pseudo-r 2 values have been proposed. In Table 3, the pseudo-r 2 R 2 DEV presented n Cameron and Trved (1998) s shown (1-Resdual Devance/Null Devance). In addton, Akake s Informaton Crtera (AIC) are shown. Models wth lower AIC crtera are preferred and t does not necessarly ncrease wth addtonal regressors. Usng the dagnostcs, the sgnfcance of the coeffcents for each model, and judgment, a model was chosen for each functonal class. The selected models are shaded n Table 3. For nterstates, model 5 was chosen and model 6 was chosen for prncpal arterals. In both cases, both the R 2 DEV and AIC were the best for these models. For mnor arterals, model 2 was chosen snce the addtonal predctve varables n model 4, 5, 6 were not sgnfcant. Two dmensonal plots (volume and crash) of the chosen models are shown on left of Fgure 2 wth plots of observed and ftted values on the rght. Usng these models, the expected number of crashes for smlar segments (λ ι ) can be made. The predctve equaton takes the form n (6) where β s the coeffcent values estmated by the model and X n s the predctve varable. λ = EXP β + β X + β X +... β X ) (6) ( 0 1 1 2 2 n n Usng an emprcal-bayes approach dscussed n Hauer et al. (2002), the expected number of crashes can be combned wth the observed count of crashes at a locaton to produce an mproved estmate of the expected number of crashes. The observed crash counts (K) can be combned wth the expected value of λ from equaton (6) n the followng equaton:
9 λ = E ( λ) α + (1 α)k (7) where α s the weght and s calculated n equaton (8) α 1 1+ ( E( λ)* Y ) φ = (8) where Y s the years of crash counts K and φ s the overdsperson parameter estmated from the SPFs and shown n Table 3. For a full dervaton and justfcaton, see the dscusson by Hauer (1997, pp. 193 194). When subtractng the expected number of crashes estmated n equaton (7) from the observed number of crashes, values resultng n a postve value ndcate the excess crashes above the average. The segments are then ranked by excess crashes n descendng order. Fgure 3 shows graphcally the results of the screenng and rankng method. The predcted 3 year snow and ce crashes for each segment (from equaton 7) are plotted aganst the observed number of crashes. An equvalent lne s also shown n the Fgure. Ponts that are above the equvalent lne have excess crashes, those below have fewer crashes than expected and the magntude of the excess was used to rank these segments. Frequency and Adjusted Frequency The fnal rankng method nvolves a smple rank of the three-year counts (frequency) and the rank based on the EB estmate n equaton (7). These rankng are ncluded for addtonal comparson to the other rankng methods. RESULTS The rankng results of all fve methods are shown n Table 4. Any segment that was dentfed n the Top 20 has been ncluded wth rankngs hgher than 20 beng shaded for better vsual dentfcaton. The lst s sorted n descendng order for the Potental for Crash Reducton method. The table ncludes values for all predctve varables. A total of 54 segments were dentfed as top 20 by at least one method. The top 20 stes by each method dentfed segments wth total speed and ce crashes of 224, 198, 163, 268, and 287 crashes, respectvely. For reference there were 3,026 crashes on all 5,097 1.6 km (1 m) segments n the database. A smple vsual nspecton of Table 4 reveals that the rankng methods dentfy dfferent stes. A tabulaton was made of the number of common segments dentfed by each pared comparson and shown n Table 5. There are no stes that were dentfed by all fve methods. From the table one can see that the crtcal rate methods dentfy 9 common segments. However, the stes ranked 1 and 3 by the crtcal rate wth clmate zones (also ranked by the potental for crash reducton) were ranked 287 and 31 by the crtcal rate wth only functonal class aggregaton.
10 These two stes are outlers for ther clmate zones but not when compared aganst the statewde average. The remanng stes dentfed by both methods are farly close n rank. The potental crash reducton dentfed only 3 stes n common wth the smple frequency and no segments n common wth EB adjusted frequency. The EB adjusted frequency and smple frequency produced the most smlar rank ordered lst wth 16 segments n common. Ths s perhaps as expected, as the count of crashes s ncluded n the EB adjustment. The segment ranked frst by the potental for crash reducton method, s ranked 5 by the crtcal rate wth clmate zones and frequency, but 287 by the crtcal rate wth functonal class and 169 by the adjusted EB method. Ths segment s a rural freeway (I-5) south of Salem that experences relatvely few snow days (4) but recorded 15 speed and ce crashes n the three year perod. The predctve models estmate a mean crash occurrence of essentally 0 (0.09 crashes) and the crtcal rate for other nterstate secton n ts clmate zone s 0.02/MVKT (0.01/MVMT). Only 1 crash occurred n 2001 (used n the modelng effort). That 15 ce related crashes occurred here over 3 years appears unusual but could be a result of the weather-related condtons that occurred n 2000 (5 crashes) and 2002 (9 crashes). A modelng approach that used annual weather data rather than 30-year averages mght be able to better capture these unusual weather events. Wthout havng knowledge of the true safety of segments, t s dffcult to say whch rankng method s superor. In most recent work, however, rate-based methods have been dscouraged n favor of EB-based methods. It was thought that better aggregaton mght better approxmate the potental for crash reducton methods but ths appears to not be the case. Ths analyss shows that these methods wll dentfy dfferent segments than the EB-based method. The frequency method appears to dentfy smlar segments as the EB-based methods. Recent research by Cheng and Washngton (2005) suggest that the EB approach produces a superor dentfcaton of truly unsafe stes by lmtng the number of falsely dentfed stes. The soon to be released Safety Analyst software wll use the potental for crash reducton method as one network screenng method. In practce, a number of states use rate, or crtcal rate. What s a useful from ths analyss s that t s clear that rate-based methods wll produce dfferent lst even wth more detaled aggregatons. One ssue that deserves further analyss s the safety performance models generated for use n the analyss. A smplfcaton was made n adoptng consstent 1.6 km (1 m) sectons for modelng rather than varyng secton lengths to capture consstent varables (.e. breakng segments by changes n lane wdth). There were hghway data that were developed n Oregon by Strathman et al. (2001) but the segment lengths were very short. In Strathman s models of manlne crashes, zero-nflated models were ft. Wth speed and ce crashes beng a subset of total crashes,
11 the negatve bnomal models dd not ft well. An alternatve approach that was consdered was to model total crashes, then estmate the expected number of speed and ce crashes by assumng an average percentage of total crashes for each of the functonal classes and clmate zones. Whle the models of total crashes had better fts, the usefulness of the modelng approach to capture other varables nfluences on speed and ce crashes would be dmnshed. Fnally, none of these SPF models ncluded any measure of ntersectons or drveways. These varables should be added to future models. The resoluton and comparablty of the weather data (30 year averages) mght hghlght segments as above average when they may have just been msclassfed. CONCLUSIONS Ths emprcal analyss of rankng speed and ce crashes compared fve rankng methods: crtcal rate (by functonal class), crtcal rate (by functonal class and clmate zone), potental for crash reducton, expected frequency (adjusted by emprcal-bayes), and frequency. For each method, the top 20 1.6 km (1 m) sectons were dentfed. A comparson of the methods showed that rate-based methods dentfed smlar rank-order segments and that the EB-adjusted frequency compared well wth the smple frequency method. The potental for crash reducton screenng method (wth EB-adjusted expected crash frequences) dentfed the most unque lst. A defntve answer on whch method s superor s dffcult, snce that answer would requre a detaled analyss of each ste dentfed. However, wth other research confrmng the advantages of EB-based methods, one should consder these results to support the use of count-based methods snce even wth sgnfcant aggregaton to mprove the rate-based methods, a dfferent set of segments were dentfed. The SPFs used n ths screenng and rankng method could be mproved. Other predctve varables could be ncluded n the models and addtonal years of data could be ncluded n the modeled data. Whle ths research effort focused on dentfyng segments related to speed and ce condtons, the methodology could be appled to any number or combnatons of crash varables. For example, the analyss of wet weather crashes could nclude precptaton, pavement roughness, and the usual hghway geometry n generatng a network screenng approach. Fnally, the treatment of severty or crash type n network screenng should be explored further. ACKNOWLEDGMENT The authors gratefully acknowledge the Oregon Department of Transportaton for sponsorng ths research as well as the Department of Cvl & Envronmental Engneerng n the Maseeh College of Engneerng & Computer Scence at Portland State Unversty.
12 REFERENCES Cameron, C.A. and Trved.P. K. (1998). Regresson Analyss of Count Data. Econometrc Socety Monogrpahs No. 30, Cambrdge Unversty Press, Cambrdge, Unted Kngdom. Cheng, W. and S. Washngton. (2005) Expermental evaluaton of hotspot dentfcaton methods. Accdent Analyss and Preventon. Vol. 37. pp 870-881 Clmate Atlas of the Unted States (2002), Natonal Oceanc Atmospherc Admnstraton. Accessed http://www.ncdc.noaa.gov/oa/about/cdrom/clmatls2/datadoc.html Hallmark, S.L., R. Basavaraju, and M. Pawlovch. (2002) Evaluaton of the Iowa DOT s Safety Improvement Canddate Lst Process. Center for Transportaton Research & Educaton. Iowa State Unversty. Ames, IA. 2002. Harwood, D.W., Bauer, K.M., Potts, I.B., Torbc, K.R., Rchard, K.R., Kohlman Rabban, E.R., Hauer, E., and Elefteradouos, L. (2002) Safety Effectveness of Intersecton Left- and Rght-Turn Lanes. FHWA, U.S. Department of Transportaton. Hauer, E. (1996). Identfcaton of stes wth promse. In Transportaton Research Record: Journal of the Transportaton Research Board 1542, TRB, Natonal Research Councl. Washngton, D.C., 1996. pp. 54-60. Hauer, E. (1997) Observatonal Before After Studes n Road Safety. Pergamon, Oxford, U.K., 1997. Hauer, E., Harwood, W.D., Councl, F.M., and Grffth, M.S. (2002) Estmatng Safety by the Emprcal Bayes Method. In Transportaton Research Record: Journal of the Transportaton Research Board 1784, Natonal Research Councl, Washngton, D.C., pp. 126-131. Hauer, E., J. C. N. Ng, and J. Lovell. Estmaton of Safety at Sgnalzed Intersectons. In Transportaton Research Record: Journal of the Transportaton Research Board 1185, TRB, Natonal Research Councl. Washngton, D.C., 1988, pp. 48 61. Hauer, E., J. Kononov, B. Allery, and M. S. Grffth. (2002) Screenng the Road Network for Stes wth Promse. In Transportaton Research Record: Journal of the Transportaton Research Board, No. 1784, TRB, Natonal Research Councl, Washngton, D.C., 2002, pp. 27 32. Lord, D., Washngton. S., Ivan. J. (2005). Posson, Posson-gamma and zero-nflated regresson models of motor vehcle crashes: balancng statstcal ft and theory. Accdent Analyss and Preventon. Vol. 37. pp 35-46
13 Maou, S.P. and Lum, H. (1993) Modelng Vehcle Accdents and Hghway Geometrc Desgn Relatonshp. Accdent Analyss and Preventon, Vol. 25, No. 6, p. 689-709. Monsere, C. Bertn, R.L., Bosa, P., Ch, Dela. (2006). Comparson of Identfcaton and Rankng Methodologes for Speed-Related Crash Locatons. Oregon Department of Transportaton, FHWA-OR-RD-06-14, Salem Oregon, June 2006. Monsere, C., Dll, J., Newgard, C., Rufolo, T., Wemple, E., Mllken, C. and Bertn, R.L. (2004) Impacts and Issues Related to Proposed Changes n Oregon s Interstate Speed Lmts. Portland State Unversty, Center for Transportaton Studes, Research Report, September 2004. Oregon Department of Transportaton (ODOT). (2006) Traffc Manual. Salem, Oregon. May 2006. Persaud, B. N., and L. Dzbk. Accdent Predcton Models for Freeways. In Transportaton Research Record: Journal of the Transportaton Research Board 1401, TRB, Natonal Research Councl. Washngton, D.C., 1993, pp. 55 60. Poch, M. and Mannerng, F. (1996) Negatve Bnomal Analyss of Intersecton-Accdent Frequences. Journal of Transportaton Engneerng, Vol. 122, No. 2, p. 105-113. R Development Core Team (2005). R: A language and envronment for statstcal computng. R Foundaton for Statstcal Computng, Venna, Austra. ISBN 3-900051-07-0, URL http://www.r-project.org. Shankar, V., Mannerng F., and Barfeld, W. (1995) Effects of Roadway Geometrcs and Envronmental Factors on Rural Freeway Accdent Frequences. Accdent Analyss and Preventon, Vol. 27, No. 3, p. 371-389. Strathman, J.G., Duecker, K.J., Shang, J., and Wllams, T. (2001) Analyss of Desgn Attrbutes and Crashes on the Oregon Hghway System. Oregon Department of Transportaton Research Group and FHWA, U.S. Department of Transportaton Vogt, A., and J. Bared. Accdent Models for Two-Lane Rural Segments and Intersectons. In Transportaton Research Record: Journal of the Transportaton Research Board 1635, TRB, Natonal Research Councl. Washngton, D.C., 1998, pp. 18 29. Washngton, S.P., Karlafts,.M.G., and Mannerng, F.L., (2003) Statstcal and Econometrc Methods for Transportaton Data Analyss. Chapman & Hall/CRC New York. Washngton. S. and Cheng, (2005) W. Hgh Rsk Crash Analyss - Fnal Report 558. Arzona Department of Transportaton. FHWA-AZ-05-558. Phoenx, Arzona.
14 Table 1 Average Crash Rate per MVKT (MVMT) by Clmate Zone and Functonal Classfcaton for Speed and Ice Related Crashes on Rural Hghways, 2000-2002. Functonal Oregon Coast Valley & Foothlls N. Coast Mountans S. Coast Mountans Cascade Mountans Columba Basn Southeast Central Hgh Desert Northeast Class Statewde 1 2 3 4 5 6 7 8 9 Interstate 0.11 (0.18) 0.01 (0.02) 0.03 (0.04) 0.04 (0.06) 0.08 (0.12) 0.27 (0.44) 0.37 (0.59) Prncpal Arteral 0.12 (0.20) 0.03 (0.05) 0.05 (0.08) 0.07 (0.11) 0.03 (0.05) 0.28 (0.45) 0.07 (0.12) 0.11 (0.18) 0.08 (0.13) 0.14 (0.23) Mnor Arteral 0.09 (0.15) 0.07 (0.11) 0.05 (0.08) 0.11 (0.17) 0.03 (0.04) 0.17 (0.27) 0.08 (0.13) 0.06 (0.1) 0.11 (0.17) 0.16 (0.25)
15 Table 2 Summary of Varables Used n Model Fttng. Interstate (n=530) Prncpal (n=2,224) Mnor (n= 2,343) Varable Mn Max Mean Std Dev Mn Max Mean Std Dev Mn Max Mean Std Dev TOTALSI_01 0 13 0.67 1.36 0 9 0.23 0.70 0 4 0.07 0.3 (crashes) AADT (vpd) 7,700 137,800 22,859 19,740 220 110,300 5,224 6,944 30 29,700 1,994 3,340 SNWDYS (days) 0 46 12.58 13.59 0 90 17.37 18.89 0 90 14.19 16.07 WIDTH_LANE m (ft) TOT_SURF_W m (ft) SP_CD km/h (m/h) 3.66 (12) 7.92 (26) 80.47 (50) 8.23 (27) 20.12 (66) 104.6 (65) 5.92 (19.43) 12.11 39.72) 103.9 (64.56) 0.38 (1.26) 1.2 (3.94) 3.36 (2.09) 2.44 (8) 6.71 (22) 40.23 (25) 21.95 (72) 29.87 (98) 104.6 (65) 5.45 (17.88) 11.69 (38.36) 87.48 (54.36) 1.82 (5.96) 4.02 (13.2) 4.55 (2.83) 2.74 (9) 5.18 (17) 32.19 (20) 13.72 (45) 25.6 (84) 88.51 (55) 4.33 (14.19) 8.43 (27.65) 85.02 (52.83) 1.37 (4.49) 2.62 (8.59) 11.68 (7.26)
16 Table 3 Estmates of Coeffcents for Models of Speed and Ice Crashes on Rural Hghways, 2001. Intercept LOG(AADT) SNWDYS WIDTH_LN TOT_SURF_W SP_CD Number of days wth > Average 0.25 cm (0.1n) Average road Total surface Posted speed daly traffc, snow wdth allocated wdth of lmt kph vpd accumulaton per lane, m (ft) roadway, m (ft) (mph) Pseudo R2 AIC φ Overdsperso n 1 Interstate 1 14.95 *** -1.62 *** 0.24 1050.2 1.26 2 12.99 *** -1.44 *** 0.015 ** 0.26 1046.7 1.19 3 10.65 *** -1.46 *** 0.016 ** 0.430 (0.131) * 0.26 1045.0 1.14 4 10.73 *** -1.71 *** 0.018 ** 0.404 (0.123) *** 0.29 1033.1 1.07 5 18.27 *** -1.65 *** 0.020 *** 0.266 (0.081) * -0.063 (-0.101) ** 0.30 1028.3 1.04-0.061-0.092 6 23.48 *** -1.51 *** 0.020 *** (-0.018) (-0.147) *** 0.29 1033.3 1.07 1-4.45 *** 0.36 *** 0.04 2470.3 4.22 Prncpal Arteral 2-7.11 *** 0.59 *** 0.033 *** 0.20 2281.9 1.83 3-7.09 *** 0.64 *** 0.034 *** -0.084 (-0.026) * 0.21 2278.7 1.80 4-7.17 *** 0.60 *** 0.033 *** -0.006 (-0.002) 0.20 2283.8 1.83 5-12.72 *** 0.63 *** 0.033 *** -0.002 (-0.004) 0.060 (0.096) *** 0.21 2272.0 1.75-0.083 0.060 6-12.69 *** 0.68 *** 0.033 *** (-0.025) * (0.096) *** 0.22 2267.0 1.73 Mnor Arteral 1-6.72 *** 0.56 *** 0.09 1098.5 3.47 2-7.78 *** 0.63 *** 0.028 *** 0.15 1055.8 1.84 3-7.77 *** 0.69 *** 0.028 *** -0.096 (-0.029) 0.16 1055.6 1.85 4-7.82 *** 0.66 *** 0.028 *** -0.017 (-0.005) 0.16 1057.5 1.88 5-6.56 *** 0.66 *** 0.026 *** -0.015 (-0.004) -0.011 (-0.024) 0.16 1056.2 1.63-0.102-0.011 6-6.45 *** 0.69 *** 0.028 *** (-0.032) (-0.025) * 0.16 1053.7 1.60 *** Sgnfcant > 0.001, ** Sgnfcant at 0.01, * Sgnfcant at 0.05, # Sgnfcant at 0.1 1 Estmates of overdsperson parameter produced by R s the recprocal of the parameter n Hauer et al. (2002) The values n the table are shown consstent wth Hauer s.
17
18 Table 4 Results of Network Screenng and Rankng Model Varables Ranks Total Crashes (00-02) Total S&I Crashes (00-02) Total S &I Crashes (01) AADT (2001) ID Functonal Class Clmate Zone 249 Interstate Valley 44 15 1 58,500 104.7 (65) 14.6 (48) 4 4.9 (16) 287 3 1 169 5 4722 Pr. Arteral Valley 14 11 2 6,100 88.6 (55) 9.1 (30) 4 4.6 (15) 31 1 2 181 14 4644 Pr. Arteral Cascade Mts 13 11 4 5,000 88.6 (55) 9.8 (32) 13 4.9 (16) 20 54 3 162 14 4645 Pr. Arteral Cascade Mts 13 8 2 5,000 88.6 (55) 19.5 (64) 13 9.8 (32) 45 113 4 288 42 5431 Pr. Arteral Cascade Mts 11 7 0 4,600 64.4 (40) 14.9 (49) 26 4.9 (16) 56 141 5 566 59 6156 Pr. Arteral N. Coast Mts 9 8 7 8,900 88.6 (55) 6.7 (22) 4 6.7 (22) 108 22 6 222 42 5372 Pr. Arteral Valley 14 10 4 15,700 88.6 (55) 20.7 (68) 8 5.2 (17) 166 17 7 104 22 248 Interstate Valley 28 7 1 58,500 104.7 (65) 14.6 (48) 4 4.9 (16) 287 52 8 338 59 5430 Pr. Arteral Cascade Mts 7 6 0 4,200 64.4 (40) 14.9 (49) 26 4.9 (16) 69 185 9 715 81 5368 Pr. Arteral Valley 15 9 2 16,600 88.6 (55) 20.7 (68) 4 5.2 (17) 206 27 10 133 29 4754 Pr. Arteral Cascade Mts 15 9 6 2,900 88.6 (55) 13.4 (44) 46 6.7 (22) 13 28 11 118 29 4526 Pr. Arteral Central Hgh Dst 6 6 2 3,100 88.6 (55) 11 (36) 13 5.5 (18) 48 11 12 358 81 5414 Pr. Arteral Cascade Mts 11 7 4 1,300 88.6 (55) 11 (36) 46 5.5 (18) 8 12 13 215 59 4753 Pr. Arteral Cascade Mts 14 9 8 2,900 88.6 (55) 13.4 (44) 46 4.6 (15) 13 28 14 101 29 4752 Pr. Arteral Cascade Mts 10 9 6 2,900 88.6 (55) 13.4 (44) 46 4.6 (15) 13 28 14 101 29 4646 Pr. Arteral Cascade Mts 12 7 2 7,700 88.6 (55) 11 (36) 13 5.5 (18) 123 283 16 199 59 4435 Pr. Arteral Valley 34 5 1 10,500 72.5 (45) 9.4 (31) 2 4.9 (16) 287 78 17 600 107 2047 Pr. Arteral Cascade Mts 11 8 2 8,400 88.6 (55) 12.2 (40) 26 6.1 (20) 98 283 18 129 42 4643 Pr. Arteral Cascade Mts 10 6 1 5,000 88.6 (55) 9.8 (32) 13 4.9 (16) 89 236 19 264 81 180 Interstate Valley 24 5 0 37,400 104.7 (65) 11.6 (38) 2 5.8 (19) 287 69 20 493 107 5415 Pr. Arteral Cascade Mts 6 6 3 1,300 88.6 (55) 11 (36) 46 5.5 (18) 16 23 22 244 81 5417 Pr. Arteral Cascade Mts 8 6 2 1,500 88.6 (55) 11 (36) 46 5.5 (18) 19 35 23 230 81 4735 Pr. Arteral Valley 9 5 2 4,300 88.6 (55) 9.8 (32) 8 4.9 (16) 117 16 25 371 107 2826 Interstate Northeast 24 21 13 8,900 104.7 (65) 11.6 (38) 26 5.8 (19) 4 41 26 1 1 4738 Pr. Arteral Valley 6 4 0 2,900 88.6 (55) 16.8 (55) 13 8.5 (28) 121 19 38 613 158 5401 Pr. Arteral Cascade Mts 26 13 7 1,800 88.6 (55) 11.6 (38) 90 5.8 (19) 1 2 40 12 11 2810 Interstate Northeast 23 14 7 10,400 104.7 (65) 11.3 (37) 26 5.8 (19) 33 179 42 9 7 2780 Interstate Colb. Basn 24 14 3 9,900 104.7 (65) 11.3 (37) 26 5.8 (19) 30 4 45 8 7 2825 Interstate Northeast 24 17 7 8,800 104.7 (65) 11.6 (38) 26 5.8 (19) 12 85 50 4 4 2822 Interstate Northeast 32 20 7 8,900 104.7 (65) 12.5 (41) 26 6.4 (21) 6 57 56 2 2 2808 Interstate Northeast 17 12 1 10,200 104.7 (65) 11.3 (37) 26 5.8 (19) 39 248 71 14 12 5400 Pr. Arteral Cascade Mts 17 11 4 1,800 88.6 (55) 11.6 (38) 90 5.8 (19) 3 5 72 23 14 2807 Interstate Northeast 19 11 3 10,200 104.7 (65) 11.3 (37) 26 5.8 (19) 51 283 94 22 14 4757 Pr. Arteral Cascade Mts 21 14 2 2,900 88.6 (55) 9.8 (32) 90 4.9 (16) 2 6 99 7 7 2864 Interstate Northeast 12 11 4 9,700 104.7 (65) 11.6 (38) 26 5.8 (19) 49 280 135 17 14 Posted Speed km/h (m/h) Total Surface Wdth m (ft) Snow Days Total Surface Wdth/Lane m (ft) Crtcal Rate (FC) Crtcal Rate (FC w/cz) Potental Crash Reducton Frequency EB Adjusted Frequency
19 5402 Pr. Arteral Cascade Mts 18 9 5 1,800 88.6 (55) 11.6 (38) 90 5.8 (19) 5 10 145 38 29 2776 Interstate Colb. Basn 13 10 3 10,100 104.7 (65) 11.6 (38) 26 5.8 (19) 64 18 163 27 22 2827 Interstate Northeast 14 11 5 8,900 104.7 (65) 11.6 (38) 26 5.8 (19) 38 235 209 15 14 5399 Pr. Arteral Cascade Mts 14 8 3 1,800 88.6 (55) 11.6 (38) 90 5.8 (19) 10 14 221 46 42 5405 Pr. Arteral Cascade Mts 13 7 4 1,300 88.6 (55) 9.8 (32) 90 4.9 (16) 8 12 237 70 59 2884 Interstate Southeast 12 11 7 8,000 104.7 (65) 12.2 (40) 13 6.1 (20) 35 111 256 13 14 5398 Pr. Arteral Cascade Mts 12 8 1 1,800 88.6 (55) 13.4 (44) 90 4.6 (15) 10 14 276 42 42 5407 Pr. Arteral Cascade Mts 7 6 1 1,300 88.6 (55) 12.2 (40) 90 6.1 (20) 16 23 277 92 81 5393 Pr. Arteral Cascade Mts 27 18 9 7,300 88.6 (55) 15.2 (50) 90 5.2 (17) 7 20 278 3 3 2781 Interstate Colb. Basn 15 12 4 9,500 104.7 (65) 11.6 (38) 46 5.8 (19) 36 7 310 10 12 2868 Interstate Northeast 13 10 3 7,700 104.7 (65) 11.6 (38) 26 5.8 (19) 40 229 424 20 22 2821 Interstate Northeast 21 10 3 8,900 104.7 (65) 12.5 (41) 26 6.4 (21) 54 283 430 19 22 2823 Interstate Northeast 13 10 5 8,800 104.7 (65) 12.5 (41) 26 6.4 (21) 50 283 462 18 22 2786 Interstate Northeast 12 10 5 9,200 104.7 (65) 11.6 (38) 46 5.8 (19) 57 283 599 16 22 5391 Pr. Arteral Cascade Mts 25 15 5 7,700 88.6 (55) 20.1 (66) 90 5.2 (17) 18 51 606 5 5 5392 Pr. Arteral Cascade Mts 28 14 6 7,700 88.6 (55) 17.4 (57) 90 5.8 (19) 22 65 667 6 7 7 Interstate S. Coast Mts 46 8 3 14,700 88.6 (55) 13.4 (44) 4 6.7 (22) 193 9 771 35 42 6 Interstate S. Coast Mts 26 8 5 13,900 88.6 (55) 13.4 (44) 4 6.7 (22) 185 8 878 33 42 2824 Interstate Northeast 18 11 4 8,800 104.7 (65) 12.5 (41) 46 6.4 (21) 37 232 895 11 14
20 Table 5 Number of Top 20 Segments Commonly Identfed By Each Screenng Method Crtcal Rate (FC) Crtcal Rate (FC w/cz) Potental for Crash Reducton Frequency EB Adjusted Frequency Crtcal Rate (FC) - 9 4 7 9 Crtcal Rate (FC w/cz) 9-4 5 8 Potental for Crash Reducton 4 4-0 3 Frequency EB Adjusted 7 5 0-16 Frequency 9 8 3 16 -
Fgure 1 Spatal Clmate Data Combned to Defne Clmate Zones. 21
22 RURAL MINOR ARTERIAL RURAL PRINCIPAL ARTERIAL RURAL INTERSTATE Fgure 2 Two-dmensonal Plots and Ftted vs. Observed Values of Selected Models for Each Functonal Class.
23 20 18 16 14 Observed Crashes (3yrs) 12 10 8 6 4 2 0 0 2 4 6 8 10 12 14 16 18 20 Predcted Crashes (3yrs) Fgure 3 Results of the Potental for Crash Reducton Screenng.