Flanders freight road network : multi-actor multi-criteria evaluation of alternative routes in the Flemish Ardennes West study region. Using crash data of the years 1997-1999, the Flemish government identified 1014 road locations as dangerous locations. This report presents a number of explanatory models, fitted on a dataset of 601 of these locations. The purpose of the report is to identify the most important underlying factors that determine the number of injury crashes at these dangerous intersections. Furthermore, the models in this report have been used to correct for the regression-to-the-mean effect in a before-after evaluation study of dangerous locations by the Flemish Policy Research Centre Mobility & Public Works, track Traffic Safety. Negative binomial models are used to deal with the overdispersion that is present in the dataset. The dependent variable is the number of injury crashes or severe injury crashes that has occurred at the location in the 2000-2003 period. The period has been chosen to account for the regression-to-the-mean effect that is known to be present in the 1997-1999 period that is used to select the locations. Crash data up to and including 2003 could be used because the first of these locations were reconstructed in 2004 in a large scale project of the Flemish government to reconstruct 800 dangerous locations to improve their safety performance. This way, full correspondence between the collected crash data and the intersection characteristics is assured. Information about 38 possible independent variables is collected. The models are fit using an iterative process to overcome missing data issues. The fit of the models is judged using the AIC-value. The final models are checked for multicollinearity using the type II tolerance test. A number of models have been fit. Two general models for all intersections have been fit; one using all injury crashes and one using only serious injury and fatal crashes. Submodels are fit for priority, signalized, four-leg and three-leg intersections, for intersections inside and outside built-up area, and for three different types of major road category intersections. There are a number of differences between the models, but a number of general conclusions can be drawn. A lot of crash prediction model studies indicate that exposure explains most of the structural variance. In this study, the exposure variables LOGVOLMAJTOTAL and LOGVOLMINTOTAL are also in this study highly important explanatory variables. The models in this study show a positive but inelastic correlation between the number of crashes and the vehicle volumes. Concerning the geometric variables, the presence of a median on the major or minor road tends to correspond with a higher number of crashes. The variable SIGNALS is only present in the general model for serious injury and fatal crashes. The variable indicates a higher number of serious crashes in case the intersection is signalized. Three-leg intersections on the other hand have a lower number of injury crashes than other types of intersections. Furthermore, intersections with a speed limit of 50 km/h tend to show a significantly higher number of crashes than the other categories. It is therefore recommended that these intersections receive special attention from road designers and policy makers. The presence of facilities for vulnerable road users at the intersections appears in some of the presented models. The patterns are however not consistent. Literature is however not conclusive on this subject either. Furthermore, some of the models indicate a safety benefit for non-perpendicular intersections compared to regular 90° intersections. Although this is rather unexpected, some literature comes to similar findings. The number of lanes of the intersection legs is present in some models, but is often not included in the end model due to multicollinearity issues. Functional road classification is a variable that is not often included in crash prediction model literature. The variables RDCATMAJ and RDCATMIN appear in quite a number of models in this study however. Intersections of primary roads tend to have a significantly higher number of injury crashes than the other categories. These intersections deserve special attention. Generally, the trend seems to be that a higher road category corresponds with a higher number of crashes. Only the category MAIN does not follow this pattern, which can be explained by the fact that these are actually intersections with on/off ramps of main roads. One submodel indicates a higher crash rate at intersections that belong to the cycle route network, but the variable could be a proximate for a higher volume of bicyclists. In a number of models, the land use variables PUBLIC, RESIDENTIAL and ECONOMIC indicate a higher number of injury crashes in case the land use is public facilities, residential or commercial activities respectively. These variables are however likely to function as a proxy for the exposure of vulnerable road users. BUILTUP is present in the submodel for secondary roads. The variable indicates a higher number of crashes when buildings are present, which stresses the safety issues of ribbon development. It is recommended that future research includes exposure data from other types of road users than only motor vehicles. Furthermore, more accurate motor vehicle counts could be collected, or a more sophisticated exposure measure such as the number of encounters. Also, future research should focus on establishing the causality of the correlations that have been found in the dataset. (Author/publisher)
Abstract