Evaluation of the passive safety of vehicles from accident data has been attempted by different institutions. Researchers try to distinguish the relative importance of different factors associated with accidents often by means of regression models. But the decisive problem in evaluating the passive safety of vehicles from accident data by regression models is the assumption of linearity. Till now all applied procedures combine the crash parameters linearly either to calculate the injury severity (linear regression) or to calculate the logit of the injury risk (logistic regression). These methods are not an appropriate tool for evaluating the passive safety of vehicles since the influence of the crash parameters is obviously not linear. An increase in the change of velocity from 10 kmlh to 40 kmlh has a different effect than an increase from 80 km/h to 110 km/h. Most of the crash parameters exhibit nonlinear behaviour but it gets "linearized" by the above mentioned linear regression models. Therefore, the influence of the crash variables is described wrongly and one gets falsified results for the evaluation of passive safety of vehicles.
To avoid the linearity of these models we introduce a nonlinear nonparametric additive model. Here the crash parameters are combined additively after an appropriate transformation on each variable: Y=c+f1(x1)+…+fk(xk) with Y being the injury severity, x1,…,xk the crash parameters, c the average injury severity calculated from the accident data and f1,…,fk the unknown functions which have to be estimated nonparametricaily from the data. Nonparametricaily means that we do not assume the functions to be polynoms or to have another a priori given structure. The estimated functions show the effect of each of the variables, i.e. the function fj represents the effect of the parameter xj. It can be seen within which ranges of the parameter the contribution to the injury severity increases or decreases and within which areas the contribution to the injury severity remains constant.
This new approach is shown by an example from a real world data base.