This paper presents a new approach to modeling call holding time of VoIP traffic on tandem networks. Call holding time is a key variable of traffic engineering models, and the traditional Erlang-B model uses the negative exponential function to model call holding time. Our study of hundreds of millions of telephone calls shows that the exponential assumption is not valid for modern large-scale VoIP networks. We propose to use time-to-event analysis which consists of fitting a parametric model to the call cease rate. Then we study several probability distribution functions and compare their capability to model the VoIP call departure rate. We find that both the log-logistic and the generalized gamma distributions provide a good fit for the data. Our statistical analysis shows that the approach of call cease rate provides more accurate results than the traditional exponential and log-normal holding time models.