Many respiratory motion compensation approaches in radiation therapy of thoracic and abdominal tumors are guided by external breathing signals. Patient-specific correspondence models based on planning 4D data are used to relate signal measurements to internal motion. The motion estimation accuracy of these models during a treatment fraction depends on the degree of inter-fraction motion variations. Here, we investigate whether motion estimation accuracy in the presence of interfraction motion variations can be improved by (sub)population models, which incorporate patient-specific motion information and motion data from selected additional patients. A sparse manifold clustering approach is integrated into a regression-based correspondence modeling framework for automated identification of subpopulations of patients with similar motion characteristics. In an evaluation with repeated 4D CT scans of 13 patients, subpopulation models, on average, outperform patient-specific correspondence models in the presence of inter-fraction motion variations.