The purpose of this paper is to identify, through empirical validation, the best analytical model for QoS and capacity estimation in mobile data networks. We first present two different sets of models that have been proposed in the literature: the infinite source, Erlang-like models and the finite source, Engset-like models. While the former models are widely developed and adapted to mobile networks, the latter are less used in the mobile context. We thus derive novel analytical models to complete the finite source theory before moving to the comparison of the different models. We make use of network measurements originating from an HSPA network and compare the modeled performance with the observed one, for each of the available analytical models. Our results show that a Processor Sharing analysis that takes as cell capacity the harmonic average of the achievable throughputs and as traffic inputs the volumes generated in each CQI, gives results that are close to the field measurements. As of the finite source models, we observed a good match on some cells and a mismatch on other cells. This is due to the difficulty of extracting traffic parameters from the field. However, finite source models have a better predictive power than infinite source models, as their traffic characteristics can more easily be linked to changes in the network such as the introduction of new services or new devices.