GENERALIZED LOGLINEAR MODELS OF TRUCK ACCIDENT RATES

Several methods for calibrating statistical models of truck accident rates are considered. A loglinear approach is suggested for assessing the effect of traffic environment on truck accident rates. A number of concerns associated with using a weighted least squares algorithm for estimating Beta parameters in the loglinear expression are noted, including the presence of reduced cell membership in the contingency tables of accidents and input variable incompatibilities between continuous exposure and categorical accident measures. An alternative form of generalized linear interactive model (GLIM) is proposed for calibrating loglinear expressions of truck accident rates. GLIM uses maximum likelihood techniques for estimating Beta parameters in loglinear expressions. As in the classical weighted least squares algorithm, this approach permits a stepwise statistical analysis of higher-order interactions in the traffic environment as related to accident frequencies, while adjusting directly for continuous measures of exposure. The results of a calibration of GLIM loglinear expressions are presented using 1983 truck accident and exposure data for Ontario as a basis.