<p>Traffic crashes contribute significantly to non-recurrent congestion, thereby increasing delays. It is therefore no surprise that traffic managers are constantly assessing strategies to mitigate the impacts of traffic crashes. To this end, it is important to have tools that enable accurate prediction of incident duration. It is also necessary to understand factors that affect the incident duration of traffic crashes. To achieve these objectives, this study employed a complementary approach to predict and analyze factors that affect incident clearance time (ICT). Prediction models for crash-related ICT using different machine learning models were developed as a first step. The machine learning models utilized were Extreme Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP). The results of the analysis suggested that XGBoost outperformed the other machine learning models in this study based on several measures of prediction accuracy. Full parametric Weibull, log-logistic and lognormal Hazard-Based Duration Models (HBDMs) were also estimated to assess the influence of several factors on ICT. These models considered both fixed and random parameters. The results showed that the random parameter lognormal model had a better fit to the data in comparison to the other HBDMs considered. Variables related to heavy vehicles, road departure, and hospitalization of crash victims, rescue, and alcohol/drugs significantly impacted ICT. The results of this study are useful for traffic incident management agencies to develop and implement strategies that lead to reduced incident duration, congestion, pollution, and the risk of secondary crashes and economic losses.</p>