Poisson basic assumption of equidispersion is often too much restrictive for crash count data; in fact this type of data has been found to often exhibit overdispersion. Underdispersion has been less commonly observed, and this is the reason why it has been less convenient to model directly than overdispersion. Overdispersion and underdispersion are not the only issues that can be a potential source of error in specifying statistical models and that can lead to biased crash-frequency predictions; these issues can derive from data properties (temporal and spatial correlation, time-varying explanatory variables, etc.) or from methodological approach (omitted variables, functional form selection, etc.). This article focuses on the potential of the Conway-Maxwell (COM-Poisson) model in handling underdispersion that arose in the development of a Safety Performance Function for urban four-leg signalized intersections; other issues, as temporal data correlation, have been intentionally eluded to test the best way of handling underdispersion. Results confirmed that the COM-Poisson model properly handled crash data set for which neither Poisson nor negative binomial model were able to account for dispersion phenomenon; they also showed that the COM-Poisson model provided a good statistical performance and a better goodness-of-fit than the quasi-Poisson and the traditional Poisson model.
|Number of pages||15|
|Journal||JOURNAL OF TRANSPORTATION SAFETY & SECURITY|
|Publication status||Published - 2011|
All Science Journal Classification (ASJC) codes
- Safety Research