Modeling School Bus Crashes Using Zero-Inflated Model

School bus crashes are rare, but their occurrence can have devastating effects on the school children involved. Such crashes are infrequent and random, and some roadway segments may not experience any school bus related crashes for a number of years (zero crashes). Despite the fact that no crashes may have occurred along particular stretches of road, these zerocrash road segments cannot be termed as safe sites, and they cause a dual state of crash experience (no crashes, but still at risk for crashes) compared to a single state of non-zero crash prone sections where risk is confirmed. Literature indicates that for extremely rare and random count data, such as school bus crashes, Poisson and Negative Binomial (NB) distributions become more applicable for modeling. Apart from Poisson and NB, there exists an alternative discrete distributional model that is used to model extra-zero discrete data, such as school bus crashes,that allows exploration of the impact of zero segments. This alternative modeling approach called zero-inflated negative binomial (ZINB) model is introduced in this study for evaluation of variables influencing school bus crashes. Although crash data rarely reveal variability, the ZINB model provides a more flexible modeling framework for school bus crashes. The study found that, ZINB yields better prediction (tight standard errors and higher z-statistics), compared to NB model though same variable coefficient signs. Presence of median and outside shoulders was found to have tendency of reducing school bus crashes. On the other end, wider medians, outside shoulders, inside shoulders, and lane widths were found to reduce the probability of these crashes. Presence of curb and gutter and two-way left turn lane (TWLTLL, high posted speed limits, multilane segments, and congested segments were found to increase the probability of school bus crashes. Language: en