ABSTRACT
In safety analysis, we are oen interested in nding the relationship between crashes and roadway characteristic variables such as number of lanes, lane width, shoulder width, median type, median width, etc., and trac volumes. Crashes are oen examined in terms of crash counts, and because of the nature of crash counts, they are discrete and rare. ey usually do not follow a normal distribution and, in general, it has been found that the variance increases as the crash count increases. Consequently, it is usually inappropriate to model untransformed crash count as a linear regression model with a normal error distribution because some of the underlying model assumptions such as normality and constant variance are seriously violated. Note that, as discussed in Chapter 10, it is
possible that the crash counts can be transformed so that the distribution of the transformed variable becomes closer to a normal distribution with constant variance. In this situation, a linear regression model may be employed to develop a prediction equation. However, a transformation does not always provide a satisfactory result; namely, transformed crash counts may still not be close enough to a normal distribution.
