Implications and Conclusion on Plausibility

Starting points are presented and two options are explicitly discussed as examples. The method delivers models with comparably high predictive power compared to the models in Sect. 5.3. The drawback of the method is a reduction in case numbers for the construction of the new models. However, in the examples displayed, this does not lead to a reduced ROC AUC for the models, but becomes evident in newly constructed ISS25+I16+ models, which do only consist of a constant and do not include any further variables. 

In general there are several implications for the practical use of injury probability models and consequently of the method as presented. The correct implementation of the models and the meaning of constraint 2 depend on the research question of the study  

In the example above. Option c was favored as starting model (ISS16+) due to the high explanatory value concerning number and kind of factors instead of the more generic Options a and b. Besides, other decision criteria for selecting the starting model could be ROC AUC values or explanatory factors included (or not) in the newly constructed models. 

The difficulties while working with empirical (observational) data are evident for the data used. Issues of potential confounding factors, multicollinearity, and selection effects are addressed. This approach is clearly limited by low case numbers. 

Previous findings documented in the literature are confirmed by the results. Impact speed is by far the most important predictor for injury severity and mortality (in both data sets). Pedestrian age is also included in every model. Vehicle characteristics (i.e., geometric quantities of the front end) as well as pedestrian physiology (e.g., BMI) are also significant in multivariate analysis. 

---