|This report is an archived publication and may contain dated technical, contact, and link information|
Publication Number: FHWA-RD-98-133
Date: October 1998
Accident Models for Two-Lane Rural Roads: Segment and Intersections
A variety of modeling techniques - Poisson, negative binomial, extended negative binomial, and logistic - have been applied in this chapter, along with measures of overdispersion, goodness-of-fit, and concordance. In general the Poisson models, negative binomial, and extended negative binomial models give mutually consistent values for regression coefficients. The T1 statistic indicates that overdispersion is present and thus that negative binomial models are to be preferred. The logistic models are not particularly satisfactory, perhaps because of the relative infrequency of serious accidents and the relatively greater importance of missing variables.
The segment models - our final model is in Table 27 - support the assertion that most of the variables in the study are significant. Some variables that correlate with accidents (e.g., commercial traffic percentage T) are omitted because they are not as significant as competing variables. However, the chief variables - exposure, lane and shoulder width, Roadside Hazard Rating and driveway density, and the alignment variables - are all represented. Differences between States appear to be genuine and are captured by the variable STATE. When we pass to the negative binomial and the extended negative binomial, the coefficient estimates are reapportioned somewhat as overdispersion and localized vertical and horizontal measures make their contribution to the variation in accident counts.
With regard to intersections, the final models are presented in Table 35. Minnesota data are taken as fundamental because the Washington intersection data are non-random and less reliable. Furthermore, the criteria for significance are relaxed so that "best guess" coefficients for alignment design variables can be presented. The effects of number of driveways, Roadside Hazard Rating, the angle variables, and channelization show notable variation between the three-legged intersections and the four-legged. Number of driveways has unexpected sign (negative) on three-leggeds in both States. Roadside Hazard Rating has unexpected sign (negative) on four-leggeds in both States. The acute/obtuse angle variable HAU behaves as expected on four-leggeds but not on three-leggeds, but another angle variable, deviation DEV from 90°, is more significant on four-leggeds. The presence of major road turning lanes increases accidents on three-leggeds but decreases them on four-leggeds. In the final models of Table 35 number of driveways (wrong sign) is omitted from the three-legged intersections, while Roadside Hazard Rating (wrong sign) and right turn lanes (insignificant) are omitted from the four-legged intersections.
Some noteworthy differences also appear between the Minnesota and Washington models, for example, the insignificance of Roadside Hazard Rating in Minnesota segments (due perhaps in part to less variation), the anomalous sign of lane width in Washington segments (perhaps related to design differences), differences in the commercial traffic percentage variable T between the two States, and insignificance of most variables on the Washington three-legged intersections.
The combined segment model (Table 27) and the Minnesota intersection models (Table 35) exhibit the effects of the chief variables, while minimizing anomalies found in some variables and in Washington intersection data.
Topics: research, safety, rural roads, interchanges, intersections, two-lane highway
Keywords: research, safety, rural roads, interchanges, intersections, two-lane highway, Minnesota, traffic accidents, crash data, mathematical models
TRT Terms: Traffic accidents--Minnesota, Rural roads--Minnesota, Roads--Minnesota--Interchanges and intersections, Traffic accidents--Washington (State), Rural roads--Washington (State), Roads--Washington (State)--Interchange and intersections, Two lane highways, Mathematical models, Accident data