At the TRRL/SWOV Workshop on Accident Analysis Methodology, heldin Amsterdam in 1988, the need to establish a methodology for the analysis of road accidents was firmly stated by all participants. Data from different countries cannot be compared because there is no agreement on research methodology, data collection, and analysis. Linear and log-linear models are regularly used for the analysis of suchdata. This paper discusses the background of these models and the model assumptions. It is stated that two relevant and fundamentally different classes of models exist. However, in most cases the difference between these classes remains implicit. These classes of models are called design oriented and object oriented models. Within the context of linear models, the basic assumptions deal with this distinction and the generalization of linear to log-linear models. It is argued that there are advantages and disadvantages for each choice andthat a check on the tenability of a particular model is to be recommended. Two different types of generalized linear models, both object-oriented, are used by SWOV and TRRL for the analysis of accidents of road users or on particular road locations. The qualitative data analysis (QDA) models used at SWOV are primarily descriptive and useful for model exploration. The generalized linear interactive modeling (GLIM) technique used at TRRL has stricter assumptions, but is better suited for model testing. In order to compare these models, a number of analyses has been carried out on roundabout data from TRRL.The outcomes are consistent and highly comparable. The results showthat for this kind of data these two types of models are to be preferred over classical multiple linear regression analysis (MLR) and log-linear analysis of contingency tables (LLA). QDA turned out to be the most powerful in data exploration. A QDA analysis showed that the dominant accident type of entering and circulating accidents hada different relation with the road and traffic characteristics of the roundabouts than the other accident types and should be analyzed separately. Furthermore, it was shown by a QDA analysis that the choice of a log-link function was preferable over an identity relation as used in MLR. This log-link function in combination with the choice of a Poisson error distribution results in a GLIM model with parameter estimates and errorbounds for these estimates. The errorbounds can be used as an indication of the reliability of the parameters. It is stated that the combination of QDA and GLIM is most powerful for the analysis of this type of problems. (A)
Abstract