Proposed Modification Factors for Roadside Slopes

Transcription

1 Proposed Modification Factors for Roadside Slopes Christine E. Carrigan, P.E., PhD RoadSafe, LLC Box Main Street Canton, Maine Phone: , Nauman M. Sheikh, P.E. Texas A&M Transportation Institute Texas A&M University System College Station, TX Phone: Ext , Submitted July 25, 2016 Resubmitted October 31, 2016 Word count Text = 4,500 Figures & Tables: 250 words each =2,750 Total number of words= 7,250 Paper prepared for consideration of presentation and publication at the 96th Annual Meeting of the Transportation Research Board, January 2017

2 Proposed Modification Factors for Roadside Slopes Christine E. Carrigan and Nauman M. Sheikh ABSTRACT Vehicle rollovers comprise twenty-eight percent of the roadway departure fatalities, representing the largest percentage of any roadway departure crash category. Understanding the probability of rollover can lead to crash prevention and is the initial action toward reducing rollover crashes. This study captured the effect of foreslope on rollover probability. The effect of foreslope width, encroachment speed and angle, and highway type were also studied to ascertain the influence each of these has on rollover outcome and crash severity. Crash modification factors were developed for use in the Highway Safety Manual s Roadside Model and trajectory adjustment factors were developed for incorporation into the AASHTO Roadside Design Guide s third version of the Roadside Safety Analysis Program. More than 200,000 rollover and non-rollover events in the state of Washington between 2002 and 2007 and Ohio between 2002 and 2010 were studied alongside 1,440 computer simulated MASH small car trajectories. This paper documents the collection, processing and assessment of these data sources and the results obtained. This research led to an improved understanding of the influence of slopes on rollover outcome and severity. It was found that a vehicle is approximately nine times more likely to rollover on a -2H:1V slope than a -10H:1V slope. It was also found that the crash severity is twice as high in a run-off-road terrain event when the vehicles rollovers over as compared to not rolling over.

3 INTRODUCTION Vehicle rollovers remain one of the major roadway departure crash types, representing the largest percentage of roadway departure fatalities (i.e., 28 percent).(1) Roadway departure rollover fatalities are more prevalent then tree-related fatalities. While there has been some recent progress with vehicle design countermeasures, crash prevention is necessary to reduce rollover crashes. Saravade recently synthesized the literature on rollover research, during the conduct of a study of fatal rollover crashes in Florida. (2) Viner also examined fatal rollover crashes in New Mexico. (3) Donelson et al. limited the study of rollovers to light-duty trucks. (4) Farmer et al. associated the driver characteristics with the probability of rollover. (5) While each of these studies have merit, understanding the probability of rollover, P(R), each roadside slope presents to all vehicle types and the potential increase in severity when a vehicle does rollover are the first steps toward affording roadside designers an opportunity to reduce run-offroad (ROR) rollover crashes. This study captured the effect of foreslope and studied how variations in vehicle type, highway type, foreslope width, encroachment speed and angle also influence P(R). This study also captured the effect rollover has on severity. The objective of this study was to develop crash modification factors (CMFs) for use in corridor planning and importation into the Highway Safety Manual s (HSM) Roadside Model and trajectory adjustment factors (TAFs) for incorporation into the third version of the Roadside Safety Analysis Program (RSAPv3). (6) Two different data sources were assembled for use in this study: simulated trajectories generated for a range of slopes and rollover crash data. Predicted probabilities of rollover by slope were developed using the first data set. The probabilities were used in conjunction with the crash data to develop both CMFs and TAFs. This paper documents the collection, processing and assessment of these data sources and the methodology used to develop both the CMFs and TAFs. This research led to an improved understanding of the influence of slopes on rollover outcome and severity SIMULATED TRAJECTORY DATA The Texas A&M Transportation Institute (TTI) conducted computer simulations for MASH test vehicles to identify the limits of recoverable, traversable, and critical slopes. Toward this objective, TTI simulated 1,440 MASH small car trajectories which interacted with foreslopes. Foreslope ratios (FS) studied included -10H:1V, -6H:1V, -4H:1V, -3H:1V, and -2H:1V on foreslope widths (FSW) of either eight or sixteen feet. These trajectories were simulated for a variety of encroachment speeds (i.e., 25, 30, 35, and 40 mph) and angles (i.e., 10, 20, and 30 degrees). The rollover and non-rollover events within the population are tabulated for each FS and FSW in Table 1. These data are shown aggregated across encroachment speed and angle for simplicity.

4 FSW 8 feet 16 feet Table 1 Simulated Trajectory Study Population Outcome FS H:1V Rollover Non-rollover 1,414 1,371 1,304 1,259 1,168 Rollover Non-rollover 1,352 1,283 1, Trajectory Data Analysis The trajectory data include four categorical variables of interest for predicting rollovers on slopes: FS, FSW, encroachment speed, and encroachment angle. FS has five levels, FSW has two levels, encroachment speed has four levels, and encroachments angle has three levels. A factorial study of these factors can therefore be described as FS(5) x FSW(2) x Speed(4) x Angle(3), including 120 groups of rollover and non-rollover outcomes. The probability of rollover, P(R), on any roadside slope is simply the portion or percentage of all vehicles traversing the slope which rollover. Proportional data is strictly bounded between 0 and 100 percent. There will be no less than zero of the vehicles interacting with the slope rolling over and no more than 100 percent. A logistic curve asymptotes as zero and one, therefore, it prevents the model from fitting negative proportions and proportions greater than unity. All of the simulated trajectories in the dataset which interacted with the foreslope were considered either rollover events (R) or non-rollover events (NR). These definitions were used to conceptualize the relationships used in this analysis. The logit as a function of slope are 56 transformed back to probability of rollover using the original relationship: P(R) = eβ xh:1v 1+e β xh:1v 57 The statistical analysis and visual inspection of the data was completed using R. (7) 58 The main effects and the interaction of the FS, FSW, encroachment speed, and 59 encroachment angle were visually inspected. A two-way interaction between variables is said to 60 be present when the effect of one variable differs depending on the level of another variable. 61 The interaction between FS and FSW was visually apparent. The visual inspection also 62 indicated a complicated relationship between encroachment angle and P(R), having the greatest 63 effect at 20 degrees and lesser effect at both 10 and 30 degrees Statistical Analysis A factorial study of both the main effect of these four variables and the replication of the combination of factor levels was conducted. A negative binomial regression function from the MASS package available in R was used to fit the logit model discussed above. (8, 7) The main effects of each factor (i.e., FS, FSW, Speed, and Angle) averaged over the levels of the other factors addressing these questions: 1. FS: Does a change in foreslope impact the probability of rollover? 2. FSW: Does a change in foreslope width impact the probability of rollover? 3. Speed: Does a change in encroachment speed impact the probability of rollover? 4. Angle: Does a change in encroachment angle impact the probability of rollover? Some interaction of factors was anticipated based on the visual analysis of the data. For example, does the effect of FS on the probability of rollover differ depending on the FSW

5 level? Three or four-way interactions are also plausible and were reviewed as well. A threeway interaction occurs when the two-way interaction is impacted by the level of the third variable. Likewise a four-way interaction occurs when the three-way interaction is impacted by the level of the four variables. Interaction Analysis Starting with the most complex model which included the four-way interactions, the model perfectly separated the probability of rollover into zeroes and ones (i.e., rollover or no rollover). While this might sound ideal, in fact it was a reflection of at least one zero in any case in each category of the constructed contingency table. The four-way interactions were therefore removed from the model. The same phenomenon was observed for the three-way interactions, therefore, the three-way intersections were removed from the model as well. The two-way interactions of FSxFSW, FSxSpeed, and FSxAngle were found to only be significant at the FS level of -2H:1V. It was found that when the FS is equal to -2H:1V and the width increased from eight to sixteen, the probability of a rollover increased by approximately twice (p< 2.53e-09). This is likely reflective of the increased exposure of the vehicle to the slope (i.e., more time on the slope). This analysis showed for the two-way interaction of FSxSpeed that as the encroachment speed increased and all factors are held constant, the probability of a rollover on a -2H:1V slope decreased. This interaction was only significant for the -2H:1V slope. The results showed a variable trend for the other FSxSpeed interactions. Specifically, as the encroachment speed increased, the probability of a rollover varied up and down by speed for different foreslopes. These findings are somewhat foreshadowed by the visual analysis, where all other factors were averaged, not held constant. The two-way interactions of FSWxSpeed and FSWxAngle were also found to be significant at limited levels. The two-way interaction of SpeedxAngle, however, was not significant at any level. Each of the interactions were removed from the model due to lack of statistical significance. Future analysis may find these interactions are statistically significant, however, that conclusion is not supported by these data. The main effects of each factor (i.e., FS, FSW, Speed, and Angle) remained in the model and were each highly significant while the interactions were not. Main Effects Upon fitting a binomial logit distribution for the main effects on the proportion of rollover and non-rollover data, the error was not binomial as assumed, the model was overdispersed. A quasibinomial model was fit to account for the overdispersion. The coefficients for the minimal adequate model are shown in Table 2. These coefficients are in logits. Changing from logits to probability was discussed above. 113

6 Table 2 Model Coefficients on Rollover Proportion Estimate Std. Error t value Pr(> t ) (Intercept) < 2e-16-6H:1V H:1V E-05-3H:1V E-09-2H:1V E-14 FSW= E-13 Speed = 30mph Speed = 35 mph E-06 Speed = 40 mph E-09 Angle = 20 deg E-13 Angle = 30 deg E-09 The predicted probability of rollover for each factor level was determined and tabulated in Table 3. The parameter values are different by different means, thus the average proportions are calculated by use of total rollovers and the total sample (i.e., rollovers plus non-rollovers). These predicted probabilities are used in conjunction with the crash data analysis to develop both crash modification factors (CMFs) and Trajectory Adjustment Factors (TAFs). The crash data analysis is presented immediately following this section. Table 3 Predicted Probability of Rollover for Population Factors a) Foreslope Factor b) Foreslope Width Factor 4 xh:1v P(R) Width P(R) c) Encroachment Speed Factor d) Encroachment Angle Factor MPH P(R) Degrees P(R)

7 CRASH DATA A case-control study was used. A case-control study is a study in which existing groups with differing outcomes are identified and compared on the basis of some supposed causal attribute. Sites and controls are identified by outcome. The treatment at each site is then determined. The World Health Organization (WHO) says the choice of controls and cases must not be influenced by exposure status, which should be determined in the same manner for both. (9) Lewallen and Courtright offer additional advice on selecting controls, stating [c]ontrols should be chosen who are similar in many ways to the cases. The selected control group must be at similar risk of developing the outcome. (10) Gross, Persaud, and Lyon, with respect to highway safety studies specifically, indicate [c]ase-control studies are based on crosssectional data. (11) This study isolated crashes on high-speed unrestricted environments (i.e., rural roads with posted speed limits of 50 mph or greater) from a cross-sectional database of crashes occurring in the state of Washington between 2002 and 2007 and Ohio between 2002 and Only ROR crashes where the sequence of events did not include longitudinal barriers were included (OC). The study population included more than 200,000 events. These data were used to determine the probability of rollover by vehicle type and establish the severity of rollover crashes. Probability of Rollover by Vehicle Type In this analysis of vehicle types, as shown in Figure 1 the cases are those vehicles which rollover and the controls are those vehicles which do not rollover. Both the cases and controls are exposed to the vehicle types traveling high-speed unrestricted roadways in Ohio or Washington. Notice that first the cases and controls are identified and then the vehicle types are identified. The results of this analysis are the probability of rollover by vehicle type. Vehicle A Vehicle B Vehicle A Vehicle B Time Direction of Inquiry Cases (Rollovers) Controls (Non-rollovers) Figure 1 Case-Control Study of Vehicle Types on Slopes Unprotected slopes The cross-sectional database described above for OC crashes occurring in high-speed, unrestricted environments was used as the starting population for the contingency table by rollover outcome and vehicle type shown in Table 4. Both divided and undivided highways were considered in the analysis of rollover as were the States. 5

8 Table 4 Tabulation for Case-Control Analysis for Rollover by Vehicle Type State Highway Cases and Type Controls PC PU SUT TT MC Bus Other WA Divided Rollover 1,467 2, WA Divided Non-Rollover 3,849 2, WA Undivided Rollover 3,066 4, WA Undivided Non-Rollover 7,475 5, OH Divided Rollover 1,784 2, OH Divided Non-Rollover 21,544 10, , OH Undivided Rollover 9,010 8, ,459 2, OH Undivided Non-Rollover 61,317 31,479 1,783 4,715 1, ,984 6

9 The crash was considered a rollover event if the vehicles rolled over anywhere in the event sequence. Event code 1 was referenced for rollovers in Ohio and 11 in Washington. The passenger car (PC) was considered to be code 1 in Washington and codes 1 through 5 in Ohio. The large passenger vehicles such as a pickup (PU) includes codes 2 in Washington and codes 6 through 8 in Ohio. The Single Unit Truck (SUT) includes code 3 and 4 in Washington and 9 and 10 in Ohio. The truck-trailer (TT) includes codes 5 through 7 in Washington and 11 through 17 in Ohio. The motorcycle (MC) was considered to be codes 12, 13, and 15 in Washington and 18 and 19 in Ohio. The Bus included codes 10 and 11 in Washington and 20 through 23 in Ohio. Other vehicles codes in each state where included in the Other group. The statistical analysis and visual inspection of the main and interactive effects was completed using R. (7) The visual inspection showed motorcycles to have a higher probability of rollover on divided highways (div) where SUTs appeared to have a higher probability of rollover on undivided highways (undiv). It was unclear if there was an interaction of highway type and vehicle type from the visual analysis. There also appeared to be some difference between the two states. A factorial study of both the main effect of vehicle type and highway type and the interaction of vehicle type, highway type, and State was conducted. A negative binomial regression function from the MASS package available in R was used to fit the logit model. (8, 7) The two-way interactions of vehicle type x highway type, vehicle type x State, and highway type x State were removed from the model because the findings were insignificant at all levels. This resulted in a model which considered only the main effects of vehicle type, highway type, and State. This model with only main effects was overdispersed. After correcting for overdispersion using quasibinomial model, the main effect of the motorcycle level of vehicle type remained significant, however other vehicle type levels became insignificant. The main effect of State and Highway type were found to be significant. The model coefficients are shown in Table 5. Table 5 Model Coefficients on Rollover Proportion Estimate Std. Error t value Pr(> t ) (Intercept) Motorcycle Other Passenger Car Pickup SUT TT Undiv Washington E-11 The predicted probability of rollover for each factor level was determined and tabulated in Table 6. The data support the thesis that the variation in vehicle type affects the rollover probability though the effect is not statistically significant. It appears that pickups and TTs may be 50 percent more likely to rollover than passenger vehicle and SUTs may be twice as likely as passenger vehicles. Motorcycles could be as much as four times more likely to rollover than 7

10 passenger vehicles but, unfortunately, the data are inconclusive. This experiment should be repeated again possibly using simulated trajectories for other vehicle types. If the results are significantly different, the findings should be considered to allow these results to be applicable across the entire fleet. If the effect is found to be not significant again, than one can conclude the vehicle type need not be considered and these findings are relevant to the entire fleet. Table 6 Predicted Probability of Rollover for Population Factors a) Vehicle Type b) Highway Type c) State Vehicle P(R) Bus Motorcycle Other Passenger Car Pickup SUT TT Type P(R) Divided Undivided State P(R) Washington Ohio Severity of Rollover Crashes The same cross-sectional database described above for OC crashes occurring in highspeed, unrestricted environments was used to populate the contingency table for this analysis of variations in crash severity for rollover crashes. The tabulation for the analysis is shown in Table 7. The tabulation of crashes by severity in Table 7 references the KABCO injury scale where a K represents a fatal injury, an A represents an incapacitating injury, a B represents a serious injury, a C represents a minor injury and an O represents a property damage only crash. The letter U is used for crashes of unknown severity. The case control study of crash severity considers the crash resulting in a severity of a particular level or greater as the cases and all lesser crash severity outcomes as the controls. Both the cases and controls are exposed to the roadside slopes on high-speed unrestricted roadways in Ohio or Washington. First the cases and controls are identified by severity and then the vehicles which rollover or do not rollover are identified. The results of this analysis are the probability of a crash severity for rollover and non-rollover crashes. This process of identifying cases and controls first then collecting the exposure of each agrees with the methodology described in the literature. (9, 10, 11)

11 Table 7 Tabulation for Case-Control Analysis for Severity by Rollover State Hwy Outcome KA BCOU KAB COU KABC OU WA Div R 345 3,889 1,776 2,458 2,554 1,680 WA Div NR 226 6,942 1,104 6,064 2,213 4,955 WA Undiv R 889 8,254 3,825 5,318 5,398 3,745 WA Undiv NR 1,014 13,740 3,584 11,170 5,821 8,933 OH Div R 857 4,664 3,020 2,501 3,620 1,901 OH Div NR 1,972 35,985 7,871 30,086 11,403 26,554 OH Undiv R 4,268 18,703 13,245 9,726 15,813 7,158 OH Undiv NR 8,617 94,214 29,392 73,439 39,425 63,406 The question addressed in this section is does a rollover lead to a significant increase in the proportion of KA, KAB, or KABC crash outcomes in the population sample of OC ROR crashes? The results are used to consider rollover severity alongside the simulated trajectory results which do not consider crash severity. The response variable for this analysis was a matched pair of counts (e.g., KA and BOUC; KAB and COU; or KABC and OU). A negative binomial regression function from the MASS package available in R was used to fit a logit model for each of these matched pairs of counts using rollovers as the explanatory variable. (8, 7) One model was fit for each severity proportion. After correcting for overdispersion using quasibinomial model, the model coefficients for severity proportions were found, as shown in Table 8. Returning to the original question, rollover does lead to a significant increase in the proportion of KA, KAB, and KABC crash severity. Table 8 Model Coefficients on Severity Proportions Std. Estimate Error t value Pr(> t ) Proportion of KA/KABCOU (Intercept) E-06 Rollover Proportion of KAB/KABCOU (Intercept) E-05 Rollover Proportion of KABC/KABCOU (Intercept) Rollover The predicted probability of KA, KAB, and KABC for rollover crashes is shown in Table 9. These findings are used in conjunction with the results obtained for the probability of rollover to finalize the rollover CMF. 9

12 Table 9 Predicted Probability of Crash Severity for Rollovers and Non-Rollovers Rollover Non- Rollover KA KAB KABC RESULTS The results from the crash data analysis and those obtained using the trajectory data were used to develop a CMF for use with the HSM high-speed unrestricted roadside model and a TAF for use with RSAPv3. The probability of rollover by slope levels shown in Table 3a were used to develop the CMF. The unprotected roadsides in the crash dataset were found to be traversable, essentially flat slopes, thus the base condition for this CMF was set to -10H:1V or flatter. These results can be easily adjusted for localized base conditions, due to the simulated nature of the data, if a State is knowledgeable of their own roadside slopes. These data provided no evidence that different vehicles have different probabilities of rollover, therefore, this CMF is proposed for use across the vehicle fleet. There was evidence of variation in crash severity as the rollover frequency increased. The CMF recommended, considering these findings, varies for different crash severity levels as shown in Table 10. Furthermore, the CMFKABCOU also serves as a TAF. When all crash severity levels are considered, it can be used as a TAF in RSAPv3. RSAPv3 has an independent model for rollover crash severity, therefore, it is unnecessary to separately consider severity. Table 10 Proposed Slope CMF and TAF CMFKABCOU xh:1v and TAF CMFKABC CMFKAB CMFKA CONCLUSIONS The objective of this study was to develop CMFs and a TAF for slopes. The results discussed above can be used by any jurisdiction. A base condition of -10H:1V was set, however, any jurisdiction can update this base condition to local conditions. The two levels of FSW show width has a highly significant effect and the effect size is quite large. Sweeping conclusions relative to effect of FSW from this two level factor, however, are not justified, but insight into the future direction of slope-based research can be inferred. The effect of FSW increasing from eight to sixteen approximately doubles the probability of a rollover. The interaction of FSW and FS was not found to be significant, however, the increase

13 in width has a significant effect on slopes of -2H:1V. Future research should be considered to provide guidance on limiting the width of slopes, particularly steep slopes. Much of the Roadside Design Guide uses guidance based on the encroachment probability model, therefore, encroachment speed and angle are directly considered in that model. Encroachment speed and encroachment angle were included at their respective levels and the effect assessed herein. This analysis showed this interaction between encroachment speed and angle is not significant except on steep slopes (i.e., -2H:1V). Individually, however, these factors were highly significant and the results reported above. As the angle increases from 10 to 20 degrees, the probability of rollover increases. Most interestingly, however, this relationship does not continue as the angle increases from 20 to 30 degrees. These findings provide solid evidence of the probability of rollover on slopes which can be used to update the Roadside Design Guide. These findings also show that -2H:1V slopes have an increased probability of inducing rollover for every interaction considered. Guidance for the protection of steep slopes should be considered against the risks the barriers present ACKNOWLEDGMENTS The simulations described were conducted under ongoing NCHRP project for which the Texas A&M Transportation Institute is the lead investigator. The crash data collection and analysis described was conducted under NCHRP for which RoadSafe is the lead investigator. The authors wish to thank the research panels for each of these NCHRP projects for their thoughtful comments and allowing the research teams to collaborate to produce these much needed results REFERENCES 1 Roadway Departure Safety, U.S. Department of Transportation, Federal Highway Administration, accessed June 25, Saravade, S.B., A Study of Fatal Rollover Crashes in the State Of Florida: A Study of Fatal Traffic Crashes in Florida From , Florida State University College of Engineering, 2005, accessed online accessed July 1, Viner, J.G., Rollovers on Sideslopes and Ditches, Accident Analysis and Prevention, Vol 27, No. 4, Donelson, A.K. Ramachandran, K. Zhao and A. Kalinowski, Rates of Occupant Deaths in Vehicle Rollover: Importance of fatality risk factors, Transportation Research Board, Washington, D.C. 1999, 5 Farmer, C.M. Adrian K. Lund, Rollover Risks of Cars and Light Trucks after Accounting for Driver and Environmental Factors, Accident Analysis and Prevention, Vol 34, No 4, Ray, M. H., Carrigan, C. E., Plaxico, C. A., Johnson, T. O., Engineer s Manual: Roadside Safety Analysis Program (RSAP) Update, Roadsafe LLC, Canton, ME, December 2012.

14 R Core Team (2016). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL 8 Venables, W. N. & Ripley, B. D. (2002) Modern Applied Statistics with S. Fourth Edition. Springer, New York. ISBN R. Bonita, R. Beaglehole, T. Kjellström, Basic epidemiology, 2nd edition, World Health Organization, page 45, Lewallen S, Courtright P. Epidemiology in Practice: Case-Control Studies. Community Eye Health. 1998;11(28): Gross, F., Persaud, B., and C. Lyon, A Guide to Developing Quality Crash Modification Factors, 27 Vanasse Hangen Brustlin, Inc., Report No. FHWA-SA , Final Report Submitted to Federal 28 Highway Administration, Washington, D.C., December 2010.