You are on page 1of 18

Paper No.

06-0017
Potential Real-Time Indicators of Sideswipe Crashes on Freeways

Chris Lee (Corresponding Author) Department of Civil and Environmental Engineering University of Central Florida Orlando, Florida 32816-2450 Phone: (407) 823-4902 Fax: (407) 823-3315 E-mail: cclee@mail.ucf.edu Mohamed Abdel-Aty Department of Civil and Environmental Engineering University of Central Florida Orlando, Florida 32816-2450 Phone: (407) 823-5657 Fax: (407) 823-3315 E-mail: mabdel@mail.ucf.edu Liang Hsia ITS Project Manager/Coordinator Florida Department of Transportation 605 Suwannee St. MS 90 Tallahassee, Florida 32399-0450 Phone: (850) 410-5615 Fax: (850) 410-5502 E-mail: liang.hsia@dot.state.fl.us

Words: 5,642 + 9 * 250 = 7,892 words

Transportation Research Board 85th Annual Meeting January 22-26, 2006 Washington, D.C.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

ABSTRACT This study investigates the real-time traffic factors that are associated with sideswipe crashes including a surrogate measure of lane change and compares conditions for sideswipe and rear-end crashes based on these factors. We extend the previous experimental study of lane change by Chang and Kao (1) by suggesting that a geometric mean of ratios of flows between adjacent lanes called the overall average flow ratio (OAFR) can be used to indicate the likelihood of sideswipe crashes given that a crash is likely to occur. OAFR was calculated for 5~10 minutes prior to crash occurrence at the location upstream of crash sites. Using a 4-year crash and loop detector data on a 36.3-mile section of Interstate-4 freeway in Orlando, Florida, the study found that OAFR is generally higher for sideswipe than rear-end crashes at a 95% confidence level. Analysis using logistic regression models showed that other trafficrelated factors such as the variation in flow and peak/off-peak periods are also important factors correlated with sideswipe crashes. The findings in this study suggest that the model is potentially capable of predicting the risk of specific type of crashes, not only generic crashes, using real-time traffic flow data on freeways.

Keywords: Sideswipe crashes, Lane change, Flow ratio, Loop detector, Freeway, Variation in flow

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

INTRODUCTION A majority of crashes on freeways are rear-end crashes due to relatively fewer interaction with vehicles approaching from other directions unlike intersections and arterial roadway sections. For this reason, most studies on freeway crashes have focused on frequency and injury severity of all crashes without considering the types of crashes. Thus, the identified causes of crashes are mainly associated with rear-end crashes. However, since there is also substantial number of other types of crashes (such as lane-change crashes), we need to account for the difference in causes between different types of crashes. In particular, sideswipe crashes are one of the major types of freeway crashes following rear-end crashes. Sideswipe crashes normally occur when vehicles change lanes. In this regard, a number of studies have analyzed lane-change crashes. Chovan et al. (2) classified the lanechange crashes into the two subtypes - proximity crashes and fast approach crashes. Proximity crashes occur when the subject vehicle changes lane and hits a vehicle in the adjacent lane traveling at similar speed. Fast approach crashes occur when the subject vehicle changes lane and hits a vehicle that approaches to the subject vehicle at higher speed in the adjacent lane. They reviewed the crash reports selected from the two U.S. crash databases in 1991 and 1992, and found that proximity crashes were more common than fast approach crashes and this subtype of lane-change crashes was mostly sideswipe crashes. Similarly, Sen et al. (3) observed about 539,000 two-vehicle lane-change crashes in the U.S. in 1996 and suggested that lane-change crashes most commonly occur when one vehicle intentionally changes lane and sideswipes or is sideswiped by a vehicle in the adjacent lane. Wang and Knipling (4) reported that although rear-end crashes can also occur during lane change, the proportion of rear-end crashes in the total lane-change crashes in the U.S. is relatively small (only 4% of total lane-change/merge crashes in 1991). Thus, it is reasonable to assume that most lane-change crashes are classified as sideswipe crashes. To identify the road geometric, traffic and environmental factors associated with sideswipe crashes, some researchers have investigated differential effects of the factors on the probabilities of sideswipe and rear-end crashes. Khattak et al. (5) found that sideswipe crashes are more likely to occur on curved roads with more number of lanes. They also observed that trucks are more involved in sideswipe than rear-end crashes due to trucks length (longer than passenger cars length) and truck drivers difficulty with seeing drivers in adjacent lanes. On the other hand, Shankar et al. (6) suggested that intensive rainfall creates water puddles that cause more lane changes and results in more sideswipe than rear-end crashes. Some researchers have conducted more detailed analysis investigating the relationship among lane change, short-term changes in traffic flow and risk of sideswipe crashes. For instance, using loop detectors and signature matching algorithms, Park and Ritchie (7) observed that lane change tends to increase the variation in speed and suggested that high variation in speed caused by lane change may also increase the likelihood of crash occurrence. However, they did not support their reasoning by relating the variation in speed to the frequency of crashes using real crash data. On the other hand, Golob et al. (8) directly analyzed the relationship between traffic factors and the types of crashes using loop detector and crash data on freeways in Orange County, California. Using median volume divided by occupancy, mean volume, the variation in speed, and the variation in volume in each lane during 30 minutes prior to the reported time of crashes, they identified eight traffic flow regimes; each regime represents typical traffic condition contributing to specific type of crashes. They found that lane-change crashes tend to occur when the variation in flow is low and the variation of speed is high whereas rear-end crashes tend to occur when both speed and the variation in speed are low. They explained that these results may be because when speed variation is high, drivers find more opportunities for improving speed and consequently they change lanes more frequently. Also, at low speed and low speed variation, drivers are more likely to be involved in stop-and-go condition that leads to high risk of rear-end crashes. This study provides important finding that lane-change crashes occur more frequently than rear-end crashes when there are more lane changes as indicated by high speed variation. However, it is not clear whether high speed variation indeed represents high lane change frequency. Also, traffic parameters in each lane were used independently rather than being compared to each other although the variation in traffic flow across lanes may be more important in estimating lane change. For instance, Gazis et al. (9) hypothesized that larger difference in density between neighboring lanes generally increases the rate of exchange of vehicles between the lanes (i.e. lane change frequency). Finally, given that crashes occur as a result of short-term variation in traffic flow, traffic conditions observed over a long term (30 minutes) prior to the occurrence of crashes

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

may dilute actual impact of the variation in traffic flow on lane-change crashes. Thus, there is a need to develop a new surrogate measure of lane change that considers the short-term variation in traffic flow across lanes. More than a decade ago, Chang and Kao (1) developed a measure of lane change called the average flow ratio (AFR) which accounts for imbalance of lane flow across neighboring lanes during short time periods (5 minutes). From the field observation, they empirically found that AFR is closely related with lane change frequency. In this study, we investigate whether the modification of AFR can be used as a surrogate measure of lane change to investigate the impact of lane change on sideswipe crashes. We also investigate other traffic-related factors that may be associated with sideswipe crashes. The objectives of this paper are to explore the use of the modified AFR as a potential indicator of sideswipe crashes on freeways, identify the other factors associated with sideswipe crashes and investigate the method of classifying into sideswipe or rear-end crashes based on these factors. This paper comprises six sections. The second section briefly reviews the past studies on lane change and explains the details of Chang and Kaos work. The third section modifies the expression of AFR for the use of the factor in assessing the impact of lane change on sideswipe crashes. The fourth section describes the data used in this study. The fifth section analyzes the relationship between the odds of sideswipe relative to rear-end crashes and a number of other traffic/road geometric factors including the modified AFR. The last section draws conclusions based on findings and recommends the future studies. REVIEW OF LANE CHANGE STUDIES A number of researchers have examined lane change behavior. For instance, Gipps (10) suggested that the objectives of lane change are 1) to maintain drivers desired speed and 2) to be in the correct lane to turn to other roads (e.g. right-turn to cross streets at intersections, exit from freeways to arterials via off-ramps). He also suggested six driver, traffic and environmental characteristics that affect drivers lane-change decision. Among them, from a safety point of view, drivers tend to change lanes only when they perceive that there is a sufficient gap in the target lane so that they can avoid a collision with other vehicles. Hidas (11) classified lane change into three types: free lane change, forced lane change and cooperative lane change. In the free lane change, the subject vehicle does not change lane until there is enough space gap and follower vehicles do not need to reduce speed. In the forced lane change, the subject vehicle forces follower vehicles to reduce speed so that the vehicle can change lane. In the cooperative lane change, follower vehicles voluntarily reduce speed to allow the subject vehicle to change lane. He suggested there is more interaction between drivers in the forced and cooperative lane change. Clearly, it is expected that the forced and cooperative lane change are more likely to cause crashes than the free lane change. Brackstone et al. (12) concluded based on the past studies that the lane change rate increases with flow under uncongested condition and the rate decreases with flow under congested condition as shown in Figure 1. It should be noted that uncongested condition does not always imply free-flow condition because the rate of change in flow is constrained at near-capacity condition. This is because as flow increases under uncongested condition, more drivers will take advantage of higher speed by changing lanes. However, as flow reaches capacity, acceptable gaps between vehicles for lane change consequently decrease and the lane change rate decreases. After congestion is formed, it becomes even harder to find acceptable gaps and the lane change rate continues decreasing. This suggests that the level of congestion is closely related with lane change. However, it is uncertain how to estimate the intensity of lane change based on traffic condition. Chang and Kao (1) conducted field studies to identify macroscopic traffic factors related to lane change behavior at two test sites on freeway segments of I-20 between Dallas and Arlington using a video recorder during a 1-hour uncongested period. They found that lane change fraction (the number of vehicles changing lanes divided by total number of vehicles) and frequency are strongly correlated with mean headway, variance of headways, average speed and the average flow (or density) ratio between neighboring lanes measured every 5 minutes. Among all these traffic factors significantly associated with lane change, the average flow ratio (AFR) is the only variable that accounts for the variation in traffic conditions across lanes. It is generally known that this cross-sectional variation in traffic flow has a significant effect on lane change (9). More specifically, they observed that the ratio of lane flow better reflects the number of lane changes than the difference in lane flow. They observed that as AFR increased, lane change frequency also increased. AFR can be calculated using the following expression:

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

v i (t ) v (t ) NLi ,i 1 (t ) + i NLi ,i +1 (t ) v i 1 (t ) v i +1 (t ) AFRi (t ) = NLi ,i 1 (t ) + NLi ,i +1 (t )

(1)

where, AFRi(t) vi(t) vi-1(t) = average flow ratio in subject lane i during time interval t (representing lane change from lane i to adjacent lanes i-1 and i+1); = average flow in subject lane i during time interval t (vehicles/unit time); = average flow in inner lane (i-1) adjacent to lane i, if lane i-1 exists, during time interval t (vehicles/unit time); = average flow in outer lane (i+1) adjacent to lane i, if lane i+1 exists, during time interval t (vehicles/unit time); = the number of lane changes from lane i to i+1, if lane i+1 exists, during time interval t.

NLi,i-1(t) = the number of lane changes from lane i to i-1, if lane i-1 exists, during time interval t; vi+1(t) NLi,i+1(t)

MODIFICATION OF AVERAGE FLOW RATIO Although AFR was found to be a significant factor correlated with lane change, there exist some limitations of applying the factor to investigate the effect of lane change on crashes. It should be noted that Chang and Kaos work only relates the number of lane changes in specific lane to AFR in the corresponding lane but does not consider the total number of lane changes in all lanes in the form of overall AFR. However, the total number of lane changes is particularly important in representing general traffic stability on freeways and its consequent impact on crash risk. For instance, even if a crash occurs in the left lane while the vehicle changes the lane from the center lane to the left lane, the lane change that actually causes the crash may not originate from the center lane the original cause may arise from high frequency of lane change from the right lane to the center lane that force vehicles in the center lane to change lanes to inner lanes. For this reason, AFR calculated for each subject lane must be combined to reflect the total number of lane changes in all lanes and explain the impact of lane change on crash risk. However, it was found from sample crash reports that most lane-change crashes occurred in the target lane for the drivers at fault who made improper lane change. The target lane can be inferred from the reports that describe the drivers lane-changing action (i.e. changing from which lane to which lane) before the crash occurrence. Thus, crash risk in the subject lane must be related to lane change to the subject lane (target lane) from the adjacent lanes rather than lane change from the subject lane to the adjacent lanes. This is more clearly explained in the example of a 3-lane freeway segment as shown in Figure 2. First, sideswipe crashes that occur in lane 1 will be mostly associated with lane change from lane 2 to lane 1 (target lane) (Figure 2(a)). Similarly, sideswipe crashes in lane 2 will be associated with either lane change from lane 1 or lane 3 to lane 2 (target lane) (Figure 2(b)) and sideswipe crashes in lane 3 will be associated with lane change from lane 2 to lane 3 (target lane) (Figure 2(c)). Consequently, to represent lane change associated with crash risk in the subject lane, we re-defined AFR as the ratio of the flow in the adjacent lanes to the flow in the subject lane as follows:

AFRi (t ) =

NLi 1,i (t ) NLi +1,i (t ) vi 1 (t ) v (t ) + i +1 vi (t ) NLi 1,i (t ) + NLi 1,i 2 (t ) vi (t ) NLi +1,i (t ) + NLi +1,i + 2 (t )

(2)

In Equation 2, flows and the numbers of lane changes from/to the adjacent lanes (i-2, i-1, i+1 and i+2) can be calculated only if the adjacent lanes exist. In combining the above lane-by-lane AFRs, since AFR is the factor in

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

ratio, a geometric mean better represents an average of AFRs than an arithmetic mean (i.e. sum of AFRs divided by the number of lanes). A geometric mean is generally useful for calculating average rate of change. Thus, overall AFR (OAFR) can be described as a geometric mean of AFRs as follows:

n OAFR(t ) = n AFR1 (t ) AFR2 (t ) L AFRn (t ) = AFRi (t ) i =1

1/ n

(3)

where, OAFR(t) = overall AFR in all lanes during time interval t; n = number of lanes.

DESCRIPTION OF DATA To estimate the relationship between the proposed OAFR and sideswipe crashes, this study used crash and traffic data collected from a 36.3-mile stretch of Interstate-4 freeway (I-4) in Orlando, Florida over four years, 1999-2002. A total of 3,124 crashes occurred on this stretch of I-4 during the four-year period. When crashes are classified by type of collision, rear-end, sideswipe and angle crashes account for 52%, 11% and 10% of the total crashes, respectively. However, it was verified from police crash reports that some sideswipe crashes that occurred in the left and center lanes of mainline traffic were not angle crashes they were sideswipe crashes. More specifically, due to a little interaction with vehicles approaching from other directions on freeways, angle crashes rarely occur in the inner lanes in the mainline. Typical angle crashes on freeways occur in the right lane where ramp traffic interacts with mainline traffic (e.g. merge from on-ramps). Thus, these misclassified crashes were re-classified into sideswipe crashes and the percentage of sideswipe crashes increased to 16%. To collect traffic information associated with crashes, the traffic data were obtained from 138 loop detector stations in eastbound and westbound directions on I-4 as shown in Figure 3. From these detector stations, 30-second average speed, volume and occupancy in the left, center and right lanes were extracted. The average spacing between detectors is approximately 0.5 miles. As mentioned previously, the lane change rate varies with the level of congestion. Thus, it is worthwhile to observe the general relationship between the level of congestion and lane change (more precisely, sideswipe crashes caused by lane change) using these traffic data. However, there exists an uncertainty associated with objectively determining congested and uncongested conditions. From a theoretical perspective, the onset of congestion is normally determined as the state when flow reaches capacity (maximum flow). When this occurs, occupancy reaches a critical level and this occupancy is often referred to as critical occupancy. The value of critical occupancy can be determined from an analysts visual inspection of a flow-occupancy graph (13) as shown in Figure 4. Thus, it is expected that lane change frequency increases as occupancy approaches critical occupancy and lane change frequency decreases when occupancy exceeds critical occupancy. It was found that the actual flow-occupancy relationship obtained from loop detector data is very similar to the relationship shown in Figure 4. However, when the flow-occupancy relationships for sideswipe and rear-end crashes were compared, an interesting trend was observed as shown in Figure 5. Each data point of flow and occupancy in the figure represents 5-minute average values in three lanes at the location upstream of crash site during 5~10 minutes prior to the time of crash. While most sideswipe crashes occurred under uncongested condition (i.e. when occupancy is less than critical occupancy) (Figure 5(a)), rear-end crashes almost equally occurred under both congested and uncongested condition (Figure 5(b)). This trend is similar to the findings in Golob et al. (8) that large proportion of lane-change crashes occurred under free-flow condition whereas rear-end crashes occurred under both free-flow and congested-flow conditions. The result indicates that since drivers can find gaps for lane change more easily under uncongested condition (i.e. they are more likely to change lanes), sideswipe crashes generally associated with lane change are also more likely to occur under such condition. On the other hand, since

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

rear-end crashes are not greatly associated with lane change, they occur under both uncongested and congested conditions. Similar trend was observed when the speeds at the time of crashes were compared between the two types of crashes. It was found that approximately 80% of sideswipe crashes occurred at high speed (greater than 50 miles/hour) whereas only 40% of rear-end crashes occurred at the same speed regime. From these results, it can be speculated that the factor distinguishing uncongested condition from congested condition can be a potential indicator of sideswipe crashes. In addition to the level of congestion, as discussed in the previous section, OAFR needs to be considered to estimate the intensity of lane change. To calculate OAFR for each crash, the traffic data at the detector station immediately upstream of the crash site some time prior to the crash time were obtained. The location of crash site is specified in the milepost which is the basis of determining the nearest upstream detector station. The reported crash time at the study site is considered as a good estimate of the crash time because drivers normally report crashes to the traffic control center using mobile phones immediately after they witness the occurrence of crashes. The time of reporting by mobile phones is verified by traffic control operators visual inspection through video surveillance. Then Florida Highway Patrol officers enter this verified time as the time of crash in the crash report. The traffic data were collected prior to this time of crash. The time period of observation was determined to be 5~10 minutes prior to the time of crash for the purpose of prediction (i.e. the prediction of the traffic condition that is more likely to lead to crashes in the short run). To calculate OAFR, we must have complete lane-by-lane traffic data as suggested in Equations 2 and 3. However, due to malfunction of detectors in some lanes, traffic data were not available in all lanes for some crashes. Thus, OAFR could not be calculated for these crashes and were not included in the analysis. Since there are normally three lanes on I-4, AFR in each lane can be calculated as follows:
NL2,1 (t ) v 2 (t ) NL (t ) + NL (t ) v1 (t ) 2,1 2 ,3
v1 (t ) v3 (t ) + v2 (t ) v2 (t ) NL2,3 (t ) v 2 (t ) NL (t ) + NL (t ) v3 (t ) 2 ,1 2,3

AFR1 (t ) =
AFR2 (t ) = AFR3 (t ) =

(4)

where, AFR1(t) AFR2(t) AFR3(t) = average flow ratio in lane 1 (left lane) during time interval t; = average flow ratio in lane 2 (center lane) during time interval t; = average flow ratio in lane 3 (right lane) during time interval t;

v1(t), v2(t), v3(t) = average flow in lane 1, 2 and 3, respectively, during time interval t; NL2,1(t), NL2,3(t) = the number of lane changes from lane 2 to 1 and from lane 2 to 3, respectively, during time interval t. In the above equation, since the fractions of lane changes from lane 2 to lanes 1 and 3 are unknown in this study, they are assumed to be equal (i.e. NL2,1/(NL2,1+ NL2,3) = NL2,3/(NL2,1+ NL2,3) = 0.5). In case of AFR in lane 2, since there is only one way of lane change from lanes 1 and 3, there is no need to estimate the fractions of lane changes. In short, OAFR can be calculated using the following expression:

v (t ) v (t ) v (t ) + v3 (t ) 0.5 2 OAFR(t ) = 3 0.5 2 1 v (t ) v (t ) v2 (t ) 1 3

(5)

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

STATISTICAL ANALYSIS This section compares the effect of a number of factors associated with lane change (including OAFR developed in the previous section) on the types of crashes using statistical analysis techniques. Two major types of freeway crashes, sideswipe and rear-end, were compared in this analysis. For this purpose, seven traffic-related factors and one road geometric factor were considered as shown in Table 1. Traffic-related factors include average speed, flow and occupancy - lane average of 30-second speed, flow and occupancy computed over 5-10 minutes prior to the reported crash times. Coefficient of variation in speed (CVS) represents standard deviation of 30-second average speeds within a 5-minute period divided by average speed in order to account for the tendency of increased variation with increased average value. Coefficient of variation in flow (CVF) is calculated in the same manner. Peak/off-peak period (PEAK) was also considered because it represents the typical pattern of traffic conditions in daily traffic. From the distribution of hourly traffic volume at the study site, peak periods were determined to be 7:00-10:00 am in the morning and 4:00-7:00 pm in the afternoon. The time periods other than peak periods were assumed to be off-peak periods. Road geometric factor includes the curvature of road section where the detector station is located. Road sections are classified into 1) the curved road section (0 < radius of curvature 3000 feet) and 2) the straight road section (otherwise). To identify the factors contributing more to sideswipe than rear-end crashes, Table 1 compares the average values (or percentages) of each factor between sideswipe and rear-end crashes. The results of t-test showed that the differences in average values between sideswipe and rear-end crashes were statistically significant for all factors except the curvature of road section. Insignificance of the curvature is supported by the fact that the percentages of crashes on curved/straight road sections were almost the same for both sideswipe and rear-end crashes. Thus, the curvature of road section was not considered for further analysis. It was found that OAFR was generally higher for sideswipe than rear-end crashes as expected. This trend was also observed from the comparison of distributions in OAFR between sideswipe and rear-end crashes as shown in Figure 6. On the other hand, average speed and CVF were higher and, average flow, occupancy and CVS were lower for sideswipe than rear-end crashes. Also, substantially higher percentage of sideswipe crashes occurred during off-peak period whereas almost equal percentages of rear-end crashes occurred during off-peak and peak periods. Among statistically significant factors, high average speed (i.e. low occupancy) and low flow seem to represent uncongested traffic condition where lane change frequency is high and sideswipe crashes occur more frequently as shown in Figure 5(a). However, care must be taken to interpret the effect of average speed, flow and occupancy because of their confounding effects under different traffic conditions. For instance, increased flow under uncongested condition indicates more vehicles that potentially change lanes and lane change frequency is more likely to increase (refer to Figure 1). On the other hand, increased flow under congested condition indicates less gaps available for lane change and lane change frequency is more likely to decrease. Similar explanation can be applied to the effects of average speed and occupancy. Thus, these three variables must be considered in two separate traffic regimes (congested and uncongested conditions). However, due to a small number of sideswipe crashes under congested condition (average occupancy > 0.13), the separate analyses could not be performed in the study. Thus, these three traffic parameters were not considered for further analysis. Although average speed, flow and occupancy were not included, it should be noted that peak/off-peak period partially reflects the effect of these parameters. In other words, congested condition during peak period is represented by low speed, high flow and high occupancy whereas uncongested condition during off-peak period is represented by high speed, low flow and low occupancy. On the other hand, a negative effect of the increased variation in speed (CVS) on sideswipe crashes contradicts the findings of the previous studies (7,8) that there may be more opportunities of improving speed by changing lanes (i.e. increased risk of sideswipe crashes) when the variation in speed is high. This result may be because high variation in speed not only reflects high lane change frequency but also reflects large difference in speed between leading and following vehicles, that contributes more to the occurrence of rear-end crashes. Clearly, this combined effect complicates identifying differential effects of CVS on sideswipe and rear-end crashes. Thus, CVS was not considered for further analysis. Given that the three remaining variables, OAFR, CVF and peak/off-peak period, are potential indicators of sideswipe crashes, the quantitative relationship was identified using the statistical modeling technique. In this study, logistic regression was used because the response variable is binary in nature (sideswipe or rear-end crashes). The

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

theoretical background of logistic regression is briefly described in this section. Logistic regression describes the probability of response variable being classified into one of pre-defined categories (not the value of response variable as in linear regression) in the function of explanatory variable(s) as follows:
P(Y = i ) = + 'x 1 P(Y = i )

ln

(6)

where, P(Y = i) = the probability of response variable y being classified into the category i (in a binary model, i = 0 or 1);

= the constant; = the vector of coefficients; = the vector of explanatory variables.

The term in the left side of Equation 6 is defined as the natural logarithm of the odds (i.e. the ratio of the probability that Y = i to the probability that Y i), or logit(Y) (14). Thus, in binary logistic regression, the odds that Y = i is equal to exp( +x). The probabilities that Y = i and Y i are as follows:

P(Y = i ) =

exp( + ' x) 1 , P(Y i ) = 1 + exp( + ' x) 1 + exp( + ' x)

(7)

In this study, sideswipe crashes are defined as category 1 and rear-end crashes category 0. Thus, the probability of a crash being classified into sideswipe crashes can be expressed as P(Y = 1) and the odds that a crash is a sideswipe crash rather than a rear-end crash (i.e. dependent variable of a logistic regression model) are ln(P(Y = 1)/P(Y = 0)). As a result of model calibration, all three factors (OAFR, CVF and peak/off-peak period) were found to be significant when they were used as an only explanatory variable as shown in Table 2. Positive coefficient indicates that the factor contributes to higher odds of sideswipe crashes and negative coefficient indicates that the factor contributes to lower odds of sideswipe crashes. The results indicate that higher OAFR, higher CVF and off-peak period generally increase the odds of sideswipe crashes. This clearly shows that high OAFR represents traffic condition with high lane change frequency that is more likely to lead to sideswipe than rear-end crashes. Similarly, it was found that the increased variation in flow (CVF) increases the odds of sideswipe crashes. This result suggests that high flow variation better reflects the traffic instability reflecting high lane change frequency associated with sideswipe crashes than speed variation. Also, the model suggests that sideswipe crashes are more likely to occur during off-peak than peak period. This reflects that traffic condition is mostly uncongested during off-peak period when more gaps for lane change are available. In this study, the models with different combinations of two or more factors were also compared as shown in Table 2. Compared to the single-factor models, the effects of individual factors in the multiple-factor models (i.e. signs of coefficients) were the same. There are no strong correlations among the factors (coefficients of correlation < 0.3) and the factors are statistically significant at a 90% confidence level (p-values < 0.10). Among the models, the best model was chosen based on the statistical significance (indicated by log-likelihood ratio chi-square) and model fit accounting for the number of explanatory variables (indicated by Akaike Information Criterion, AIC (15)). It was found that all models were statistically significant at a 95% confidence interval using log-likelihood ratio chisquare test (p-value < 0.05). It was also found that the AIC statistic was the lowest (963.83) for the model with the three factors. Therefore, the following model was selected as the best model:

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

10

logit (Y ) = 3.8292 + 1.8865 OAFR + 4.1822 CVF 0.4897 PEAK (p-values = 0.0657) (< 0.0001) (= 0.0065)

The above results indicate that sideswipe crashes are more likely to occur as a result of complex interaction of traffic-related factors associated with higher lane change frequency. Given that OAFR, CVF and peak/off-peak period are independent (i.e. no interaction between factors), the elasticity of the odds of sideswipe crashes to each variable was analyzed separately. Elasticity analysis was performed such that the effect of variation in one variable on the odds of sideswipe crashes was observed while the two other variables remain constant. As shown in Figure 7, the odds exponentially increase with an increase in OAFR and CVF during both peak/off-peak periods. It was observed that the elasticity is higher for CVF than OAFR as the range of its marginal effect on the odds is greater.

CONCLUSIONS AND RECOMMENDATIONS

This study identified the factors associated with the odds of sideswipe relative to rear-end crashes. Among many factors, the study developed a surrogate measure of lane change called overall average flow ratio (OAFR). The study modified the original expression of average flow ratio between adjacent lanes suggested in the previous empirical study (1) by 1) considering the target lane of lane change more importantly and 2) calculating a geometric mean of lane flow ratios to represent the total number of lane changes in all lanes. The values of OAFR were calculated during the time period 5~10 minutes prior to the time of crashes to provide enough time for real-time mitigation as part of an ATMS crash prediction system. The results of the analysis showed that the values of OAFR were generally higher for sideswipe than rearend crashes. This indicates that higher variation of flows across lanes contributes more to the occurrence of sideswipe than rear-end crashes. Thus, given that flow ratio is closely correlated with lane change (1), the result supports our reasoning that frequent lane changes increase the risk of sideswipe crashes. The study also found that the occurrence of sideswipe crashes is closely related to the level of congestion sideswipe crashes are more likely to occur during off-peak period when traffic condition is generally uncongested. This may be because it is easier for drivers to find gaps for lane change under uncongested than congested condition. The study also found that high variation in flow (CVF) contributes more to sideswipe than rear-end crashes. Finally, the study obtained theoretically valid relationship among the odds of sideswipe relative to rear-end crashes, OAFR, CVF and peak/offpeak period using a logistic regression model - the odds of sideswipe crashes increases as the values of OAFR and CVF increase and when the time period is off-peak period. The proposed model can be used to estimate the intensity of lane change and also the likelihood of sideswipe relative to rear-end crashes. Thus, it is recommended that traffic control strategies should aim at reducing the values of OAFR and CVF by balancing flow across lanes and allowing smooth lane change over longer distance. This may help reduce high risk of sideswipe crashes during particularly uncongested period. In future works, more specific strategies of traffic flow control (e.g. lane control signals) need to be investigated for reducing the risk of sideswipe crashes.

ACKNOWLEDGEMENTS

The authors would like to thank Anurag Pande and Prof. Nizam Uddin at the University of Central Florida for their invaluable comments. The authors also acknowledge the financial support from the Florida Department of Transportation.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

11

REFERENCES

1. 2.

Chang, G. and Y. Kao. An Empirical Investigation of Macroscopic Lane-Changing Characteristics on Uncongested Multilane Freeways. Transportation Research Part A 25, No. 6, 1991, pp. 375-389. Chovan, J. D., L. Tijerina, G. Alexander and D. L. Hendricks. Examination of Lane Change Crashes and Potential IVHS Countermeasures. U.S. Department of Transportation, National Highway Traffic Safety Administration, Report No. DOT-VNTSC-NHTSA-93-2, 1994, 60 pages. Sen, B., J. D. Smith and W. G. Najm. Analysis of Lane Change Crashes. U.S. Department of Transportation, National Highway Traffic Safety Administration, Report No. DOT-VNTSC-NHTSA-02-03, 2003, 49 pages. Wang, J. and R. Knipling. Lane Change/Merge Crashes Problem Size Assessment and Statistical Description. U.S. Department of Transportation, National Highway Traffic Safety Administration, Report No. DOT HS 808 075, 1994, 61 pages. Khattak, A. J., P. Kantor and F. M. Council. Role of Adverse Weather in Key Crash Types on Limited-Access Roadways. In Transportation Research Record 1621, TRB, National Research Council, Washington, D.C., 1998, pp. 10-19. Shankar, V., F. Mannering and W. Barfield. Effect of Roadway Geometrics and Environmental Factors on Rural Freeway Accident Frequencies. Accident Analysis and Prevention 27, 1995, pp. 371-389. Park, S. and S. G. Ritchie. Exploring the Relationship Between Freeway Speed Variance, Lane Changing, and Vehicle Heterogeneity. Presented at 83rd Transportation Research Board Annual Meeting, 2004. Golob, T. F., W. W. Recker and V. M. Alvarez. Freeway Safety as a Function of Traffic Flow. Accident Analysis and Prevention 36, 2004, pp. 933-946. Gazis, D., R. Herman and G. H. Weiss. Density Oscillations Between Lanes of a Multilane Highway. Operations Research 10, 1962, pp. 658-667.

3. 4.

5.

6. 7. 8. 9.

10. Gipps, P. G. A Model for the Structure of Lane-Changing Decisions. Transportation Research Part B 20, No. 5, 1986, pp. 403-414. 11. Hidas, P. Modelling Vehicle Interactions in Microscopic Simulation of Merging and Weaving. Transportation Research Part C 13, 2005, pp. 37-62. 12. Brackstone, M., M. McDonald and J. Wu. Lane Changing on the Motorway: Factors Affecting Its Occurrence, and Their Implications. Road Transport Information and Control, Conference Publication No. 454, 1998, pp. 160-164. 13. Chen, C., Z. Jia and P. Varaiya. Causes and Cures of Highway Congestion. IEEE Control Systems Magazine, 21(6), 2001, pp. 26-33. 14. Menard, S. Applied Logistic Regression Analysis, 1995, Sage Publications. 15. Akaike, H. Information Theory and an Extension of the Maximum Likelihood Principle. The Second International Symposium on Information Theory. Akademiai Kiado, Budapest, Hungary, 1973, pp. 267-281.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

12

TABLE 1 Variables Associated with Lane Change

Type Traffic-related Factors

Variables Overall average flow ratio Average speed (mph) Coefficient of variation in speed (=Variation of speed / Average speed) Average flow (vehicles/30 second) Coefficient of variation in flow (=Variation of flow / Average flow) Average occupancy Dummy for peak/off-peak period 0= off-peak period 1= peak period Dummy for curvature of road section 0= straight section 1= curved section

Symbol OAFR AS CVS AF CVF AO PEAK CURVE

Average value or percentage Sideswipe crashes Rear-end crashes (No. of obs.=220) (No. of obs.=725) 0.83 0.81 53.03 41.34 0.10 9.28 0.34 0.09 72% 28% 87% 13% 0.18 11.39 0.27 0.16 53% 47% 89% 11%

Road geometric Factors

TABLE 2 Estimated Parameters in Logistic Regression

Parameters OAFR

Logistic regression model logit(Y) = -4.2544 + 3.7507 OAFR

Log-likelihood ratio chi-squarea

Akaike Information Criterion (AIC)b 1016.62 971.15 1002.93 969.39 995.24 965.28 963.83

12.98 (p = 0.003) CVF 58.45 logit(Y) = -2.7551 + 5.0120 CVF (p < 0.0001) PEAK 26.66 logit(Y) = -0.8817 0.8393 PEAK (p < 0.0001) OAFR, CVF 62.21 logit(Y) = -4.2774 + 1.9655 OAFR + 4.8364 CVF (p < 0.0001) OAFR, PEAK logit(Y) = -3.5155 + 3.2074 OAFR 0.7950 PEAK 36.36 (p < 0.0001) CVF, PEAK 66.32 logit(Y) = -2.3634 + 4.4289 CVF 0.4982 PEAK (p < 0.0001) OAFR, CVF, logit(Y) = -3.8292 + 1.8865 OAFR + 4.1822 CVF 69.77 PEAK (p < 0.0001) 0.4897 PEAK a The p-value less than 0.05 indicates that the model is statistically significant at a 95% confidence level. b Smaller AIC value represents better model with accounting for the number of explanatory variables.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

13

Lane Change Rate

Uncongested

Congested

Flow
FIGURE 1 Relationship between lane change rate and flow based on conjecture by Brackstone et al. (12).

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

14

v1

v2

v3

Lane 1 Lane 2 Lane 3

(a) Sideswipe crashes in lane 1 associated with lane change from lane 2 to lane 1 (represented by v2/v1)

v1

v2

v3

v1

v2

v3

Lane 1 Lane 2 Lane 3

Lane 1 Lane 2 Lane 3

(b) Sideswipe crashes in lane 2 associated with lane change from lane 1 or lane 3 to lane 2 (represented by v1/v2 + v3/v2)

v1

v2

v3

Lane 1 Lane 2 Lane 3

(c) Sideswipe crashes in lane 3 associated with lane change from lane 2 to lane 3 (represented by v2/v3)

FIGURE 2 Sideswipe crashes in a three-lane freeway segment.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

15

71 0.67 13 16 71

70 0.44 18 24 70

69 0.52 49 20 69

68 0.48 17 17 68

67 0.39 7 2 67

66 0.63 3 1 66

65 0.51 7 5 65

64 0.45 6 3 64

63 0.42 4 1 63

62 0.62 2 4 62

61 0.56 2 3 61

60 0.59 18 14 60

59 0.51 7 8 59

58 0.75 4 11 58

57 0.47 22 11 57

56 0.54 16 6 56

55 0.54 6 4 55

54 11 8 54

Seminole 54 0.51 11 16 54

Orange 53 0.43 14 24 53 26 20 52 52 0.56 12 17 51 51 0.45 10 2 50 50 0.40 47 1 49 49 0.50 24 5 48 48 0.44 47 3 47 47 0.41 30 1 46 46 0.57 30 4 45 45 0.50 36 3 44 44 0.49 44 14 43 43 0.48 66 8 42 42 0.52 59 11 41 41 0.53 56 11 40 40 0.51 26 6 38 38 0.35 48 4 37 37 0.52 78 8 36 36

36 0.51 78 48 36

35 0.45 57 48 35

34 0.54 38 24 34

33 0.35 63 55 33

32 0.53 54 8 32

31 0.51 23 40 31

30 0.72 65 24 30

29 0.55 29 9 29

28 0.55 28 43 28

27 0.49 24 23 27

26 0.55 14 50 26

25 0.40 26 17 25

24 0.58 15 27 24 Orange

23 0.55 28 18 23 Osceola 6 0.63 0.66 8 7 6

22 0.49 43 17 22

21 0.59 6 13 21

20 0.53 17 21 20

19 4 12 19

19 0.50 4 12 19

18 0.57 6 14 18

17 0.64 11 40 17

16 0.55 14 25 16

15 0.56 15 14 15

14 0.57 5 8 14

13 0.77 4 4 13

12 0.61 6 12 12

11 0.57 30 29 11

10 0.62 29 12 10

9 0.58 5 6 9

8 0.68 15 12 8

7 6 9 7

5 0.53 10 9 5

4 0.48 10 22 4

3 0.53 5 7 3

2 7 12 2

Westbound Eastbound

FIGURE 3 Location of loop detector stations on Interstate-4 freeway.

Note: 1. The arrows pointing outward indicate off-ramps and the arrows pointing inward indicate on-ramps. 2. The numbers inside the squares denote the station numbers. 3. The numbers shown between two successive detectors are the distance in miles. 4. Shaded detector stations are the stations located on curved road sections (0 < radius of curvature 3000 feet). 5. The numbers above and below the center lines are total number of crashes over 4 years at the nearest station.

Flow

Uncongested Maximum Flow

Congested

Critical Occupancy

Occupancy

FIGURE 4 Relationship between flow and occupancy.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

16

25 20 15 10 5 critical occupancy = 0.13 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 5-min. Average Occupancy uncongested congested

5-min. Average Flow (vehicles/30 sec)

(a) Sideswipe crashes

25 uncongested 5-min. Average Flow (vehicles/30 sec) 20 15 10 5 critical occupancy = 0.13 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 5-min. Average Occupancy congested

(b) Rear-end crashes


FIGURE 5 Comparison of flow-occupancy relationships between sideswipe and rear-end crashes.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

17

0.40 Proportion of Total Crashes 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 Overall Average Flow Ratio Sideswipe Crashes Rear-end Crashes

FIGURE 6 Comparison of OAFR between sideswipe and rear-end crashes.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

Lee, Abdel-Aty and Hsia

18

2 Odds of Sideswipe Crashes

1.5 Off-peak period Peak period 1

0.5

0 0.6 0.7 0.8 0.9 OAFR 1 1.1 1.2

Odds of Sideswipe Crashes

1.5

Off-peak period Peak period

0.5

0 0.1 0.2 0.3 0.4 CVF 0.5 0.6 0.7

FIGURE 7 Elasticity of odds of sideswipe crashes to OAFR and CVF.

TRB 2006 Annual Meeting CD-ROM

Paper revised from original submittal.

You might also like