PURPOSE: To quantify the interfractional variability in lung tumor trajectory and mean position during the course of radiation therapy. METHODS AND MATERIALS: Repeat four-dimensional (4D) cone-beam computed tomography (CBCT) scans (median, nine scans/patient) routinely acquired during the course of treatment were analyzed for 56 patients with lung cancer. Tumor motion was assessed by using local rigid registration of a region of interest in the 3D planning CT to each phase in the 4D CBCT. Displacements of the mean tumor position relative to the planned position (baseline variations) were obtained by using time-weighted averaging of the motion curve. RESULTS: The tumor trajectory shape was found to be stable interfractionally, with mean variability not exceeding 1 mm (1 SD) in each direction for the inhale and exhale phases. Interfractional baseline variations, however, were large, with 1.6- (left-right), 3.9- (cranial-caudal), and 2.8-mm (anterior-posterior) systematic variations (1 SD) and 1.2- (left-right), 2.4- (cranial-caudal) and 2.2-mm (anterior-posterior) random variations. Eliminating baseline variations by using soft-tissue guidance decreases planning target volume margins by approximately 50% compared with bony anatomy-driven protocols for conventional fractionation schemes. CONCLUSIONS: Systematic and random baseline variations constitute a substantial portion of the geometric variability present in the treatment of patients with lung cancer and require generous safety margins when relying on accurate setup/immobilization or bony anatomy-driven correction strategies. The 4D-CBCT has the ability to accurately monitor tumor trajectory shape and baseline variations and drive image-guided correction strategies that allows safe margin reduction.