ivandifferencescausalityfactorssingle1999.md 25 KB
category: literaturenote citekey: ivandifferencescausalityfactorssingle1999 title: Differences in causality factors for single and multi-vehicle crashes on two-lane roads authors: "Ivan, John N.; Pasupathy, Raghubhushan K.; Ossenbruggen, Paul J." year: 1999 date: 1999-00-00 1999 doi: 10.1016/s0001-4575(99)00030-5 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/s0001-4575(99)00030-5" zotero_key: KEEZ8SAF zotero_storage: NG72WZ99 collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Ivan, John N."
status: converted
on two-lane roads
a *Department Ci*6*il and En*6ironmental Engineering U-37, *Uni*6ersity of Connecticut, Connecticut Transportation Institute, Storrs, CT 06269-2037, USA b Edwards and Kelcey, 529 Main Street, Suite 203, Charlestown, MA 02129, USA c *Department of Ci*6il Engineering, *Uni*6ersity of New Hampshire, Durham, NH 03824, USA Received 1 October 1998; received in revised form 9 February 1999; accepted 26 March 1999
Past research has found a non-linear relationship between traffic intensity or level of service (LOS) and highway crash rates. This paper investigates this relationship further by including the effects of site characteristics and estimating Poisson regression models for predicting single and multi-vehicle crashes separately. Analysis focuses on rural two-lane highways, with hourly LOS, traffic composition, and highway geometric characteristics as independent variables. The resulting models for single and multi-vehicle crashes have different explanatory variables. Single-vehicle crash rates decrease with increasing traffic intensity (lower LOS), shoulder width and sight distance. Multi-vehicle crash rates increase with the number of signals, the daily single-unit truck percentage, and the shoulder width, and decreased on principal arterials compared to other roadway classes. LOS does not significantly explain variation in the number of multi-vehicle crashes. Ongoing research by the authors is aimed at identifying other site factors, such as driveway density and intersection LOS, that can better explain the differing effects reported here and
John N. Ivan a,*, Raghubhushan K. Pasupathy b
predict crash rates of both types better. © 1999 Elsevier Science Ltd. All rights reserved.
Keywords: Highway crash rates; Level of service; Crash type; Two-lane roads; Daylight
One of many important aspects of highway safety research is developing crash prediction models to quantify the relationship between site characteristics and the number of crashes observed. These prediction models can help identify the site characteristics tending to cause highway crashes. Crash prediction models, till now, have tended to be macroscopic in nature, meaning that they predict using summary statistics on traffic such as annual average daily traffic (AADT) rather than microscopic measures such as hourly volume counts. These macroscopic models also tended to consider all crashes together rather than separately by single and multi-vehicle crash types. Studies that have considered hourly volumes have generally not consid-
* Corresponding author. Tel.: +1-860-486-0352; fax: +1-860-486-
0001-4575/99/\$ - see front matter © 1999 Elsevier Science Ltd. All rights reserved.
E-mail address: johnivan@eng2.uconn.edu (J.N. Ivan)
PII: S0001-4575(99)00030-5
, Paul J. Ossenbruggen c
ered geometric effects. In this paper, the significance of all of these factors—site location, highway geometry, daylight conditions, and traffic composition—are in-
The focus of this research was to find differences in the effects of both traffic condition and site characteristics as independent variables for the expected single and multi-vehicle crash rates, measured in crashes per year. Eight different rural two-lane state highway locations across Connecticut with hourly volume counts between September 1990 and October 1996 were studied. Poisson regression was used to estimate and evaluate models for predicting the expected number of single and
A number of researchers have investigated the effect of exposure on the number of crashes per year, or crash
vestigated together.
2. Previous work
multi-vehicle crashes at these sites.
Differences in causality factors for single and multi-vehicle crashes
Abstract
1. Introduction
2298.
Table 3 Preliminary models for multi-vehicle crashesa
| Variable | Model 1 | Model 2 | Model 3 | Model 4 |
|---|---|---|---|---|
| Intercept | −10.08 (−11.53) | −9.88 (−12.04) | −9.98 (−10.72) | −11.31 (−7.33) |
| Route 1 | Base | Base | Base | |
| Route 5 | −2.03 (−3.72) | −2.28 (−3.86) | −2.30 (−3.78) | |
| Route 7 | −1.86 (−4.19) | −1.84 (−4.12) | −1.81 (−3.99) | |
| Route 8 | −1.17 (−2.10) | −1.11 (−1.90) | −0.94 (−1.50) | |
| Route 12 | 0.88 (3.13) | 0.86 (3.00) | −1.01 (−3.18) | |
| Route 66 | −1.95 (−5.70) | −2.09 (−5.45) | −2.21 (−5.35) | |
| Route 81 | −2.59 (−3.15) | −2.67 (−3.22) | −2.63 (−3.05) | |
| Route 124 | −0.99 (−4.00) | −1.54 (−3.18) | −1.66 (−3.13) | |
| LOS A | Base | Base | Base | Base |
| LOS B | 0.64 (1.67) | 0.15 (0.38) | 0.14 (0.32) | |
| LOS C | −0.14 (−0.32) | 0.21 (0.49) | 0.33 (0.67) | |
| LOS D | −0.18 (−0.39) | 0.63 (1.16) | 0.76 (1.24) | |
| LOS E | 0.46 (1.02) | 0.83 (1.28) | 0.86 (1.21) | |
| Light | Base | |||
| Dark | 0.26 (0.79) | |||
| Dusk | 0.45 (1.27) | |||
| Exponent on exposure | 0.88 (12.20) | 0.76 (10.37) | 0.86 (10.73) | 0.96 (7.96) |
| Dispersion | 1.37 | 1.44 | 1.38 | 1.40 |
| Null deviance | 788.97 | 788.97 | 788.97 | 788.97 |
| Residual deviance | 373.28 | 504.44 | 370.26 | 367.97 |
| AIC value | 391.28 | 516.44 | 396.26 | 397.97 |
a Values in parentheses are t-statistics after correction for dispersion.
variables, is clearly the best. It has the smallest AIC value, and none of the LOS or light condition coefficients in the other models are significantly different from zero at 95% confidence. It seems that segment LOS and light conditions are not useful for predicting multi-vehicle crashes. This point will also be discussed later.
5.2. Site characteristics
The preliminary analysis clearly showed that each site has a different average crash rate for both crash types. The next step is to replace the site variable with the actual site characteristics in the model estimations. These site characteristics are represented by a linear combination of specific continuous variables, as represented in the initial model form. The LOS and light condition variables were retained as categorical variables. The list of site variables and their abbreviations as used in the models are shown in Table 4. The base model (with respect to LOS and light conditions) was defined as before. All site characteristic, traffic, and light condition variables, and all two-way interactions between the categorical variable levels, were considered as predictors in estimating models for both types of crashes.
Table 5 shows the results from estimating models for single-vehicle crashes. Model 5 includes the five LOS categories, along with three site-related variables: the percentage of the section with a sight distance of at least 1500 feet, the right shoulder width (feet), and the percentage of single-unit trucks in the peak traffic stream. All coefficients are significantly different from zero with at least 95% confidence; however, paired t-tests show that the coefficients on LOS D and LOS E are not significantly different from each other. Consequently, model 6 was run, which combines LOS D and E into a single category. The resulting increase in residual deviance is offset by the reduction in model complexity (one less parameter), so this model has a better (smaller) AIC value.
Table 4 Site characteristic variables
| Variable name | Type of variable | Observed values |
|---|---|---|
| Functional classifica- tion |
Categorical | Principal arterial— othera ; minor arterial; major collector |
| Right shoulder width (feet) |
Continuous | 0–8 |
| Sight distance over 1500 feet (% of sec tion) |
0, 10, 80 | |
| Number of signals | Continuous | 0–1 |
| % Single unit trucks (daily) |
Continuous | 0–4 |
| % Single unit trucks (peak) |
Continuous | 0–2 |
| Speed limit (mph) | Continuous | 35, 40, 45 |
a Non-freeway.
Table 5 Final models for single vehicle crashesa
| Variable | Model 5 | Model 6 | Model 7 | Model 8 | Model 9 |
|---|---|---|---|---|---|
| Intercept | −14.78 (−11.74) | −14.88 (−12.22) | −13.06 (−10.48) | −13.06 (−11.84) | −8.91 (−4.13) |
| LOS A | Base | Base | Base | Base | Base |
| LOS B | −1.65 (−5.22) | −1.66 (−5.30) | −1.66 (−4.89) | −1.65 (−4.91) | −1.59 (−4.90) |
| LOS C | −2.81 (−6.27) | −2.83 (−6.31) | −2.53 (−5.36) | ||
| LOS D | −3.90 (−5.72) | ||||
| LOS E | −3.70 (−6.31) | ||||
| LOS (D, E) | −3.77 (−7.01) | −2.59 (−5.90) | |||
| LOS (C, D, E,) | −2.56 (−6.57) | −2.76 (−6.86) | |||
| Sight distance over 1500 feet | −0.02 (−2.30) | −0.02 (−2.30) | −0.02 (−2.31) | −0.02 (−2.44) | −0.02 (−2.49) |
| Right shoulder width (feet) | −0.27 (−4.23) | −0.27 (−4.36) | −0.12 (−2.32) | −0.12 (−2.33) | −0.15 (−2.84) |
| % Single-unit trucks (peak) | 1.47 (3.68) | 1.49 (3.80) | |||
| % Single-unit trucks (daily) | −0.01 (−0.06) | ||||
| Speed limit (mph) | −0.09 (−2.10) | ||||
| Exponent on exposure | 1.18 (11.71) | 1.19 (12.13) | 1.16 (11.26) | 1.16 (11.50) | 1.13 (11.53) |
| Dispersion | 0.64 | 0.64 | 0.75 | 0.75 | 0.71 |
| Null deviance | 363.50 | 363.50 | 363.50 | 363.50 | 363.50 |
| Residual deviance | 185.95 | 186.01 | 194.94 | 194.96 | 191.44 |
| AIC value | 203.95 | 202.01 | 210.94 | 206.96 | 205.44 |
a Values in parentheses are t-statistics after correction for dispersion.
The peak hour truck percentage variable seemed inappropriate, since each case potentially represents conditions aggregated over periods of time other than the peak. We therefore ran model 7, which replaces the peak truck variable with a corresponding daily variable; this model performs much worse; the coefficient on the daily truck percentage is not significantly different from zero at even 90% confidence. Furthermore, a paired t-test comparing the coefficients on LOS C and the combined LOS D/E category showed them to be statistically identical at 95% confidence (t*=1.96).
In consideration of these findings, we then combined LOS C, D and E into one category and removed both truck variables: this was run as model 8. Because other research by the senior author (Ivan and O'Mara 1997) has found the posted speed limit to be a significant variable for predicting crash rates, model 9 was then run adding 'speed limit' to model 8. Model 9 performs better than model 8, having a smaller AIC, but not as well as model 6, which included the suspect peak singleunit truck percentage variable.
Similar analysis was performed with the number of multi-vehicle crashes as the dependent variable; these results are shown on Table 6. In the first model (number 5), the best predictor variables were found to be the functional classification of the road, the number of traffic signals, the daily truck percentage, and the right shoulder width. Only the coefficient on the principal regional arterial roadway class was significantly different from the others (minor arterial or major collector) at 95% confidence; this class has a significantly lower average crash rate than the others. The positive coefficients on both the number of signals and the daily single unit truck percentage make intuitive sense; one would expect both of these variables to increase the intensity of conflicts between vehicles.
However, the positive coefficient on right shoulder width is troubling; one normally expects a wider shoulder to be a safety feature. Consequently, we examined the data surrounding this variable more carefully. Figure 3 shows how the multi-vehicle crash rate varies with the right shoulder width: note how there is no clear linear relationship between the two quantities. We therefore examined several categorical representations of right shoulder width; those that performed best are shown as models 6, 7 and 8. These three models each divide the right shoulder width into two categories, with the lower category having shoulder widths up to 2, 3 and 4 feet, respectively. None of these models perform as well as model 5. The issue of right shoulder width will be discussed more in the next section.
6. Discussion
Several general observations may be made about these results:
- 1. The coefficient on shoulder width was *negati*6e for predicting single vehicle crashes, but *positi*6e for multi-vehicle crashes.
- 2. For single-vehicle crashes, the most important predictor variables are the LOS and the percent sight distance over 1500 feet.
- 3. For multi-vehicle crashes, the most important predictor variables are the class of roadway, number of signals and daily single unit truck percentage.
Single vehicle crashes seem to occur when traffic volume is low and on roads with less forgiving geometry. This makes sense, since a single vehicle crash involves a driver losing control and colliding with a fixed object off the roadway (or an object on the roadway). Under low traffic volumes, there is less likely to be another vehicle to hit, and the less than ideal geometry (sharp horizontal curves and narrow shoulders) reduces the maneuvering room for the driver to recover without leaving the pavement, for example, if the driver had been speeding.
Conversely, multi-vehicle crashes appear to occur in the presence of intense traffic conflicts such as would be observed at a traffic signal, or on roads serving local trips, or when there are many trucks in the traffic stream. Traffic signals tend to be placed in locations where intersecting streets have high traffic volumes. Minor arterials and major collector roads tend to serve local trips and thus have more driveways and local street intersections, increasing the potential for vehicle conflicts.
Clearly, more analysis is required to properly interpret these findings. For one thing, the higher levels of LOS—which correspond to the higher numbers of single-vehicle crashes—are observed more typically at night rather than during the day. Consequently, it is not immediately clear from this analysis whether it is the low volumes, the darkness, or drowsiness of late night drivers that results in greater numbers of singlevehicle crashes. Each of these factors needs to be accounted for explicitly in order to fully understand this phenomenon.
A second issue with these findings is that the LOS was computed for the analysis segment, not for the intersections on the segment. This might explain why LOS was not significant for multi-vehicle crashes, which occur where there is potential for vehicle conflicts, such as at intersections. If LOS were computed for the intersections on the segment, then it might become significant for predicting multi-vehicle crashes.
Finally, it is difficult to explain the positive coefficient on right shoulder width in the multi-vehicle crash models. One interpretation is that with wider shoulders drivers are more likely to drive too fast and collide with other vehicles exiting or entering driveways. Another is that drivers might be tempted to use a wide shoulder to pass drivers waiting to make left turns, and thus not see a hazardous situation beyond the stopped vehicle. Both of these explanations are related to the presence of intersections on the segment; this issue is addressed in the next section.
7. Conclusions
In response to these findings, ongoing research by this research team is aimed at including new variables which attempt to clarify the effects of the predictive factors used in this study. Specifically, intersection volumes and the potential for conflicts leading to angle and turning collisions are incorporated into the analysis by considering the AADT on roads intersecting each analysis section. Second, the number of driveways of different types (e.g. residential, service station, retail, etc.) is included to account both for the land use intensity and for conflicts associated with driveway intersections. Third, the potential effect of time of day is being investigated by incorporating it as an additional variable along with daylight conditions.
Table 6 Final models for multi-vehicle crashesa
| Variable | Model 5 | Model 6 | Model 7 | Model 8 |
|---|---|---|---|---|
| Intercept | −15.87 (−14.69) | −14.72 (−14.07) | −15.32 (−11.13) | −18.57 (−12.89) |
| Minor arterial or major collector | Base | Base | Base | Base |
| Principal arterial | −5.79 (−9.83) | −3.34 (−8.79) | −4.05 (−7.55) | −7.98 (−7.09) |
| Number of signals | 2.31 (7.78) | 1.50 (4.40) | 2.53 (6.39) | 3.55 (8.53) |
| Daily single-unit truck percentage | 1.26 (6.84) | 1.01 (5.39) | 1.40 (4.92) | 2.24 (8.17) |
| Right shoulder width (feet) | 0.44 (5.15) | |||
| Right shoulder width (52 feet) | Base | |||
| Right shoulder width (\2 feet) | 1.16 (5.07) | |||
| Right shoulder width (53 feet) | Base | |||
| Right shoulder width (\3 feet) | 0.71 (1.84) | |||
| Right shoulder width (54 feet) | Base | |||
| Right shoulder width (\4 feet) | 3.79 (4.48) | |||
| Exponent on exposure | 0.88 (13.83) | 0.90 (13.57) | 0.87 (11.67) | 0.91 (13.22) |
| Dispersion | 1.30 | 1.32 | 1.69 | 1.33 |
| Null deviance | 788.97 | 788.97 | 788.97 | 788.97 |
| Residual deviance | 370.88 | 374.07 | 400.26 | 384.67 |
| AIC value | 382.88 | 386.07 | 412.26 | 396.67 |
a Values in parentheses are t-statistics after correction for dispersion.
Fig. 3. Single and multi-vehicle crashes per million vehicle miles by right shoulder width.
Intersection traffic volumes and driveway counts are not as readily available as the other data used in this analysis. The permanent count stations only provide data for the main road; consequently, the precise exposure and volumes gathered for each analysis section are not available for the roads intersecting them. Instead, in this new research, these intersecting road volumes are estimated using an AADT observed for each road and the observed hourly volumes. Similarly, the HPMS and other roadway inventory datasets do not provide detailed information about the driveways along the section. This information is being collected using the ConnDOT photolog library, a unique data source comprised of driver's eye view images of the entire length of every state highway in Connecticut, updated once each year. Using the photolog library permits expanding our analysis beyond HPMS sites; consequently, we have identified ten new analysis sites in the vicinity of two-lane ATR stations, each 0.5 mile in length.
This newly expanded dataset permits us to estimate additional models for predicting single and multi-vehicle crashes which consider these new variables. In addition, in this analysis we are including the time of day (divided into five categories), permitting us to isolate time of day, daylight and LOS effects and identify the value of each. We are also computing new variables for measuring exposure to multi-vehicle crashes which consider not only the traffic volume on the section being analyzed, but also the traffic volume on intersecting roads. We expect these new models to explain more variation than the models presented in this paper.
Acknowledgements
The research described in this paper was funded by a grant from the United States Department of Transportation through the New England (Region 1) University Transportation Center and was conducted at the Connecticut Transportation Institute of the University of Connecticut.
References
Agresti, A., 1990. Categorical Data Analysis. Wiley, New York.
Bozdogan, H., 1987. Model selection and Akaike's Information Criterion (AIC): the general theory and its analytical extensions. Psychometrika 52, 345–370.
Frantzeskakis, J., 1983. Accident analysis on two non-controlled access national highways in Greece. ITE Journal 65 (10), 32–34. Gwynn, D., 1967. Relationship of accident rates and accident involvements with hourly volumes. Traffic Quarterly 21, 407–418.
TRB, 1994. Highway capacity manual. Special Report 209. National Research Council, Washington, DC.
Ivan, J., O'Mara, P., 1997. Prediction of traffic accident rates using Poisson regression. Presented at 76th Annual Meeting of the Transportation Research Board, Washington, DC, paper no. 970861.
Persaud, B., Mucsi, K., 1995. Microscopic accident potential models for two-lane rural roads. Transportation Research Record 1485, TRB, National Research Council, Washington, DC, pp. 134–139. S-PLUS User's Manual, 1995. Version 3.3 for Windows. Statistical
Sciences Division, Mathsoft, Seattle. Valavanis, S., 1959. Econometrics: An Introduction to Maximum
Likelihood Methods. McGraw-Hill, New York.
Vogt. A., Bared, J., 1998. Accident Models for Two-Lane Rural Segments and Intersections. Transportation Research Record 1635, TRB, National Research Council, Washington, DC, 18–29.
Zhou, M., Sisiopiku, V., 1997. Relationship Between Volume to Capacity Ratios and Accident Rates. Transportation Research Record 1581, TRB, National Research Council, Washington, DC, pp. 47–52.