Modeling severity of single vehicle run-off-road crashes in rural areas: model comparison and selection

Run-off-road (ROR) crashes account for a large proportion of fatalities and serious injuries to vehicle occupants, especially in rural areas. While performing crash severity analysis using discrete choice models (DCMs), researchers may be confused by the following questions: first, should an ordered or unordered model structure be used and secondly, which modeling level is more appropriate, basic or advanced? A model selection framework is developed considering the following factors: (1) model structure-ordered or unordered; (2) intrinsic deficiency of each model; (3) computational burdens; (4) complexity of parameter interpretation; and (5) model fitness. Historical ROR crash data were utilized to illustrate how to choose an appropriate DCM based on the proposed framework. Using statistical tests and comparison of evaluation and validation measurements, both the mixed logit model and the partial proportional odds model yield a reasonable performance. All factors that significantly affect the severity level of a single-vehicle ROR crash were identified as well.