Service failure is an incident when a defect of size over threshold value is noticed and the track is taken out of service. They are a major concern for the railroads due to their impact on economic and safety aspects of the rail company. Despite the effective inspection strategies developed by the railroad, service failure may still occur in their network because of the growth of an undetected defect or the defect that is initiated after an inspection. Various factors influence the growth rate of a defect, causing service failure. Identifying and understanding these factors will assist in accurately predicting the risk of a service failure. This paper develops a data-driven approach based on accelerated failure time model using the data collected from one of the Class-I US railroads during the period from 2011 to 2016. Different distributions of the hazard function are evaluated, and the log-normal distribution is found to be the best fit. This study identifies the following factors: past rail defects, past geometry defects, days from the last grinding, product of rail age and degrees of curvature and infrastructural data such as the number of turnouts, grade of rail and rail segment length to have an influence on failure time of the rail. Temporal transferability model validation and comparison with other machine learning approaches are performed using the data collected in the year 2016. The results of this study are useful for railroads to develop effective strategies for rail inspections and preventive maintenance.