Severity Analysis of Crashes on Multilane Arterials Using Conditional Inference Forests.

The study is aimed at identifying traffic/highway design/driver-vehicle information significantly related with fatal/severe crashes on urban arterials for different crash types. Since the data used in this study is observational (i.e., collected outside the purview of a designed experiment); aninformation discovery approach is adopted for this study. Random Forests,which are ensembles of individual trees grown by CART (Classification andregression tree) algorithm, are applied in numerous applications for thispurpose. However, recent research has shown that the æGini indexÆ criterion used in the algorithm is biased towards continuous variables or categorical variables with more levels. Since the æGini indexÆ criterion searchesfor a favorable split, the chance of finding one increases if the variable satisfies the above criteria. The fallout is that a variable could be selected even though itÆs not particularly informative in terms of its relationship with the target variable. To overcome this issue conditional inference forests have been implemented. In each tree of the conditional inference forest, splits are based on how good the association is, i.e., the resulting daughter node should have a higher association with the binary dependent variable. Chi-square test statistics are used to measure the association. Apart from identifying the variables which improve classification accuracy, the methodology also clearly identifies the variables which are neutral to accuracy and also those which decrease it. The methodology is quite insightful in identifying the variables of interest in the database. For example, the usual suspects like alcohol/ drug use and higher posted speed limits are contributing to severe crashes. Failure to use safety equipment by all passengers and presence of driver/passenger in the vulnerable age group (more than 55 years or less than 3 years) increased the severity of crashes. A new variable, æelementÆ has been used in this study which assigns crashes to segments, intersections or access points based on the information from site location, traffic control and presence of signals. The authors were able to identify roadway locations where severe crashes tend to occur. For example, segments and access points were found to be riskierfor single vehicle crashes. Higher skid resistance and k-factor also contributed towards increased severity of crashes.