Although in recent years considerable progress has been made to establish r
elationships between accidents and highway characteristics, no specific acc
ident models are widely accepted by the highway engineering agencies. Four
commonly used models are developed in this study, that is, two conventional
linear regression models and two Poisson models. Accident data monitored b
y the National Freeway Bureau in Taiwan are used to model statistically rel
ationships between accident types and highway features. Traditional models
are demonstrated to possess unsatisfactory statistical properties that cann
ot adequately describe discrete, non-negative, random and sporadic accident
events along a highway. Based upon statistic estimates, the Poisson models
are shown to be more appropriate, accurate and reliable than the conventio
nal linear regression models.