Voice over IP is one of the most popular applications in broadband access networks. It is anticipated that the characteristics of call holding times (CHTs) for VoIP calls will be quite different from traditional phone calls. This article analyzes the CHTs for mobile VoIP calls based on measured data collected from commercial operation. Previous approaches directly used the Kolmogorov-Smirnov (K-S) test to derive the CHT distributions, which may cause inaccuracy. In this article we propose a new approach to derive the CHT distributions for mobile VoIP calls and other call types. Specifically, our approach uses hazard rate to select an appropriate distribution, and then utilizes the K-S test to validate our selection. We show that the mobile VoIP CHT distribution can be accurately approximated by a mix of two lognormal distributions. Based on the derived distributions, we compare the mobile VoIP CHTs with those for non-VoIP calls and fixed-network VoIP calls. Our study indicates that the characteristics for mobile VoIP calls are quite different from those of the non-VoIP mobile phone calls and are more close to those of fixed-network phone calls.