Specification and diagnosis of multivariate model equations in road safety studies.

摘要

Multivariate regression method is now commonplace in road safety statistical modelling studies because it is acknowledged to hold promise to improve and produce more accurate and reliable estimation results of the effect(s) of accident countermeasures. However, the method comes with its problems in respect of appropriate means to specify and diagnose the model equations. Model specification involves the tasks of identifying the “important” variables to be included in the model equation and selecting their suitable functional forms. Model diagnosis is the task of ascertaining the correctness of the specified model equation. Two tools are proposed in this dissertation for accomplishing these tasks.; The I&barbelow;ntegrate-D&barbelow;ifferentiate (ID) method is the demonstrated promising tool for model specification . The C&barbelow;u&barbelow;mulative R&barbelow;e&barbelow;siduals (CURE) method is the demonstrated promising tool for model diagnosis. The former is based on the idea that the cumulative (integral) form of a formless accident scatterplot reveals a more definite pattern. A couple of candidate integral functions can then be suggested. The derivative of the best fitted integral function then yields the sought functional form. The graph of the CURE method is more revealing and informative than that of the ordinary residuals. Observations and inferences can be made about deviations of the cumulative residuals with respect to the zero residual line.; The applications of both tools are illustrated using a Highway Safety Information Systems accident database on rural road segments of the same length in Maine. The aim with this database is to ascertain whether a particular specified univadate accident model equation that relates expected single-vehicle accidents to traffic volume can be found that is plausible for the different populations of road segments of the same length. The answer is No. However, the Unconstrained Hoerl function is demonstrated to be superior in terms of estimation capabilities. The disregarded Cubic Polynomial function is a good competing functional form. The often assumed Log-linear (or Exponential) models in modelling accident counts is inappropriate. It is an over-approximation of the Cubic Polynomial function.; The (de)merits of the ID and CURE methods are highlighted. The merits of both methods outweigh the demerits.