Improving the estimation accuracy of the Used Car Safety Rating component measures

A critical aspect of the UCSRs being useful for consumer information is producing ratings that are as accurate as possible and cover as many vehicle models as possible. Accuracy of the ratings is reflected by confidence limit widths on the ratings estimates with the general aim being to make the confidence limits as narrow as possible. Since vehicle model are filtered from presentation based on confidence limit width, a narrower confidence limit will allow more vehicle to pass the accuracy test for rating inclusion. In addition, since the confidence limit width reflect the range in which the true (rather than estimated) safety performance of the vehicle is likely to lie, narrower confidence limits will allow the vehicle rating to be placed in a star rating performance category with greater accuracy. As a result, ratings are less likely to move between categories in successive updates due purely to random variation in the rating estimate.

Developments in statistical methodology have provided additional analysis techniques that may assist in producing more accurate ratings estimates for each of the 3 base ratings that constitute the UCSRs. They are General Estimating Equations (GEEs), Multinomial Logistic Regression and Bootstrapping techniques for estimating variance. The aim of this project is to assess the potential use of these methodologies in producing ratings estimates and their standard errors to assess the benefits they offer over the current techniques in improving estimation accuracy. Provision of valid measures of variability is an essential part of a world’s best practice program.