Estimating the contributions of speeding and impaired driving to insurance claim cost.

INTRODUCTION Logistic regression modelling of crash counts likely associated with speeding and impaired driving was earlier found [Zheng, Y., Cooper, P. J., & Dean, C. B. (2007). Modeling the contribution of speeding and impaired driving to insurance claim counts and costs when contributing factors are unknown. Journal of Safety Research, 38(1)] to successfully predict known outcomes in a series of simulations, but the same types of models did not accurately predict average crash costs. The aim of the study reported here was to develop a means to adjust classification model results that would improve their cost-predicting efficiency. METHOD A classification modelling process was adjusted at the back-end using non-linear optimization to rationalize the classified proportions with the true proportions when the model was applied to representative subsets of the training data. Corrections were developed to account for cost (severity) differences arising from the classification process that were not due to true variations. The process was then applied to insurance claim test data where crash contributors were unknown. RESULTS The optimization and severity correction procedure resulted in substantial improvement in average crash cost prediction for both impaired and unsafe speed collision involvements. The error measured against true values in 20 simulations was about half for the adjusted classification model of what it was for either unadjusted classification or logistic regression models. CONCLUSIONS Non-linear optimization of classification matrices appears to be a workable tool for improving the predictive efficiency of models where desired outcomes represent average characteristics of records as compared to simple counts or proportions. IMPACT ON INDUSTRY Using the methodology on a full-year of insurance claim data indicated that reliance on police-reported records alone would have underestimated the total cost of unsafe speed and impaired crashes by about 40%. Since most jurisdictions use police data to base policy decisions and set program spending around such safety issues, this finding has important implications.