Kilovoltage imaging systems on linear accelerators are used for patient localization in many clinics. The purpose of this work is to assess on-board imaging (OBI) detection of systematic setup errors and in particular, the dosimetric consequences of undetected head roll in head-and-neck intensity modulated radiation therapy (IMRT) plans when using these systems. The system used in this study was the Trilogy linear accelerator and associated software (Varian Medical Systems, Palo Alto, CA). Accuracy of OBI localization was evaluated using an anthropomorphic head phantom. The head phantom is rigidly attached to a specially designed positioning device with 5 degrees of freedom, 3 translational and 2 rotational in the axial and coronal planes. Simulated setup errors were 3 degrees and 5 degrees rotations in the axial plane and displacements of 5 mm in the left-right, anterior-posterior, and superior-inferior directions. The coordinates set by the positioning device were compared with the coordinates obtained as measured by using the image matching tools of paired 2-dimensional (2D) orthogonal image matching, and 3D cone-beam computed tomography (CT) volume matching. In addition, 6 physician-approved IMRT plans of nasopharynx and tonsil carcinoma were recalculated to evaluate the impact of undetected 3 degrees and 5 degrees head roll. Application of cone-beam CT (CBCT) for patient localization was superior to 2D matching techniques for detecting rotational setup errors. The use of CBCT allowed the determination of translational errors to within 0.5 mm, whereas kV planar was within 1 to 2 mm. Head roll in the axial plane was not easily detected with orthogonal image sets. Compared to the IMRT plans with no head roll, dose-volume histogram analysis demonstrated an average increase in the maximal spinal cord dose of 3.1% and 6.4% for 3 degrees and 5 degrees angles of rotation, respectively. Dose to the contralateral parotid was unchanged with 3 degrees roll and increased by 2.7% with 5 degrees roll. The results of this study show that volumetric setup verification using CBCT can improve bony anatomy setup detection to millimeter accuracy, and is a reliable method to detect head roll. However, the magnitude of possible dose errors due to undetected head roll suggests that CBCT does not need to be performed on a daily basis but rather weekly or bi-weekly to ensure fidelity of the head position with the immobilization system.