Respiratory motion limits the use of focused ultrasound surgery and conformal radiation therapies for livers. Population motion models have shown their ability to predict unobservable motion from tracking results. These models require establishment of inter-subject correspondences, which is non-trivial. In this study we compare the use of a landmark-based method with a shape-based approach for this task. The influence of the inter-subject correspondences on the motion model predictions was assessed in leave-one-subject-out tests. For simulated 3D (2D) tracking results, the landmark-based motion models reduced the right liver lobe 95% motion from 12.4 (11.9) mm to 4.6 (4.5) mm on average. The shape-based models improved this further to 3.7 (3.8) mm and reduced the mean error by 13% (9%). While the landmarks were only defined for the right liver lobe, the shape-based approach enabled the motion prediction of the whole liver with a 95% error of less than 4.8 mm.