Identifying the factors that significantly affect accident severity has become one of the many ways to reduce it. While many accident database studies have reported associations between factors and severities, few of them could assert causality, primarily because of uncontrolled confounding effects. This research is an attempt to resolve the issue by comparing the difference between what happened and what would have happened in different circumstances. Data on accidents were analyzed first with rough set theory to determine whether they included complete information about the circumstances of their occurrence by an accident database. The derived circumstances were then compared with each other. For those remaining accidents without sufficient information, logistic regression models were employed to investigate possible associations. Adopting the 2005 Taiwan single-auto-vehicle accident data set, the empirical study showed that an accident could be fatal mainly because of a combination of unfavorable factors instead of a single unfavorable factor. Moreover, the accidents related to rules with high support and those with low support showed distinct features.