Principal Component Analysis Applied to Crash Data on Multilane Roads

During the last few years, a lot of road-accident-predictive models have been developed by using Multiple Linear Regression and Poisson or Negative Binomial Distribution. More innovative methodologies based on fuzzy logic and neural networks have also been used. The application of these methodologies is not easy when a large number of variables is considered. Moreover, the influence of some variables on road accidents might not be equally significant. It would thus appears useful to have an analysis tool primarily in order to remove the redundant variables for accident-predictive models. Even if under-used in crash data, Principal Component Analysis (PCA) may be suitable for this purpose. PCA is a form of analysis used for extracting a reduced number of factors, called principal components, from a set of original variables, discarding as little of the information as possible. Our objective is to verify PCA potentiality for removing redundant variables in accident analysis. For this purpose a five-year monitoring period was carried out on a four-lane median divided road. A database was subsequently created with the surveys regarding the type and number of accidents, traffic flow, horizontal and vertical alignment, sight distances and pavement surface characteristics. PCA was applied to homogeneous sections having constant horizontal curvature, separated into tangents and curves. By means of the correlation matrix the results indicate that the number of accidents on curves increases with the length (L) of the homogeneous sections, the curvature radius (1/R), the average daily traffic (TGM) and the design speed change (ΔV) between tangents and curves; whereas there is a negative correlation between these crashes and the longitudinal slope (i%), the sight distance (vis) and the pavement friction defined in terms of CAT (Side Friction Coefficient measured by means of a SCRIM equipment). Thus the results obtained prove the knowledge about this subject. Six principal components were found to account for about 90% of the variance in the original eight variables. The multiple correlation coefficient (ρP2P) between the original variables and the principal axes shows that the least significant variable is ΔV. In keeping with the literature, the correlation matrix for tangents indicates that road accidents are positively correlated to the length of the homogeneous sections (L) and the average daily traffic (TGM), and negatively correlated to the pavement friction (CAT) and the longitudinal slope (i%). Four principal components were found to account for over 90% of the variance in the original five variables. The multiple correlation coefficient for tangents (ρP2P) shows that the variables examined are all equally significant.