Heinz Hautzinger

• Empirical vehicle crashworthiness studies are usually based on national or in-depth traffic accident surveys: Data on accident-involved cars/drivers are analysed in order to quantify the chance of driver injury and to assess certain risk factors like car make and model. As the cars/drivers involved in the same accident form a "cluster", where the size of the cluster equals the number of accident-involved parties, traffic accident survey data are typical multi-level data with accidents as first-level or primary and cars/drivers as secondlevel or secondary units (car occupants in general are to be considered as third level units). Consequently, appropriate statistical multi-level models are to be used for driver injury risk estimation purposes as these models properly account for the cluster structure of traffic accident survey data. In recent years various types of regression models for clustered data have been developed in the statistical sciences. This paper presents multi-level statistical models, which are generally applicable for vehicle crashworthiness assessment in the sense that data on single and multiple car crashes can be analysed simultaneously. As a special case of multi-level modelling driver injury risk estimation based on paired-by-collision car/driver data is considered. It is demonstrated that assessment results may be seriously biased, if the cluster structure inherent in traffic accident survey data is erroneously ignored in the data analysis stage.