Incident duration is one of the primary causes leading to traffic delay and secondary accidents. This study demonstrates an effort of collecting and fusing accident and duration data sets from various sources in Taiwan. We developed Accelerated Failure Time (AFT) models based on 4,074 records. The relationship between incident duration and three sets of explanatory variables-accident characteristics, road and environmental features, and spatial and temporal characteristics of accidents were investigated. The results show that the AFT model with a log-logistic distribution and inverse Gaussian-distributed frailty is the best fitted model, which implies the non-monotonic feature of incident duration in Taiwan and is different from the usually-seen monotonic Weibull model in the literature. The results also suggest further incident duration with the increase of accident severity and the number of involved vehicles, the involvement of large vehicles, fire accidents, rollover accidents and non-rear-end accidents. Significantly longer duration is also found during peak periods, nighttime, and the freeway segments in southern Taiwan. The analysis results could be considered as a preliminary study of incident duration in Taiwan. It could be used as a reference for estimating external cost due to traffic accidents and for the allocation of emergency response resources.