A Multivariate Analysis of Road Severity Accident Index

Safety performance models are commonly used to correlate explanatory variables in the form of
geometric, operational, environmental and human c
haracteristics to estimate road accident
frequency and severity. Generally these models
employ multivariate relationships between
variables requiring local calibration. The selection of independent variables normally follows
traditional tests of significance but lacks a real test of the ability of each factor to explain the
observed response and the nature of the variance-covariance structure. This paper uses
principal components analysis and correlation matrices to identify and retain significant factors,
find linear codependences and cluster similar explanatory factors. Nonlinear regression
methods for safety severity were examined by looking at their ability to develop safety
performance functions for a case study of r
egional highways in New Brunswick. Linear
dependency between shoulder and lane width was founded. Lighting and road surface condition
were not relevant in explaining accident severity. It was found that vertical alignment, vehicles
running of the road and colliding with obstacles and AADT were of intermediate relevance, of
high importance; speed, percentage of trucks, intensity of intersections per kilometer, horizontal
alignment and type of facility (divided or undivided). A clear cluster of geometric characteristics
and another of operational characteristics was obser
ved. The analysis aims to serve as a guide
for practitioners in need to develop locally calibrated safety performance functions able to
explain locally observed road accidents by severity