Reliability analysis of motorcycle crash severity outcomes: Consideration of model selection uncertainty

While a large amount of work has been conducted on different types of crash injury severity models, model selection uncertainty remains a critical issue in traffic safety research. The objective of this study is to handle model selection uncertainty by combining multiple models.

Motorcycle crashes in Michigan from 2010 to 2014 are collected for the analysis. A model averaging approach is used to integrate useful information from three commonly used crash injury severity models: multinomial logit model, ordered logit model, and ordered probit model to deal with the situation where the model selection uncertainty exists in crash data analysis. The ratios of model posterior probabilities between models are used to quantify the model selection uncertainty. In addition, the effectiveness of the method is illustrated by comparing it with the single-best model.

The ratios of model posterior probabilities among models approximate to 1. It means that three models have the same importance in statistical analysis of motorcycle injury severity, resulting in model selection uncertainty. The comparison between the results of model averaging approach and single-best model shows that the single-best model tends to overestimate the effects of risk factors on motorcycle injury severities because of ignoring the model selection uncertainty; parameter errors and confidence intervals of model averaging are greater and wider than those of the single-best model due to between-model uncertainty included in the model averaging; some risk factors are significant in the model averaging approach while not in the single-best model. Results from model averaging approach reveal that drunk or riding under influence, angle/sideswipe/head on crashes, speed limit of 35 mph or higher, and signal control play significant roles in the motorcycle crashes.

The study contributes to the existing crash injury-severity literature by developing a model averaging approach to explore the relationship between motorcyclist's injury-severity and its contributing factors. The model averaging approach overcomes the limitations of the current crash injury-severity modeling approaches by (1) revealing the potential model selection uncertainty among injury-severity models with model posterior probabilities; (2) more reliably accounting for the effects of risk factors on motorcyclist' injury severities through integrating all information from the candidate models; and (3) better presenting the underlying unreliability of the analysis results from each individual model.