The purpose of this study was to assess three-dimensional (3D) set-up errors using megavoltage computed tomography (MVCT) during image-guided intensity-modulated radiation therapy (IMRT) for supine craniospinal irradiation (CSI) on helical tomotherapy (HT). Patients were immobilized in a customized 4-clamp thermoplastic head mask with or without whole-body vacuum cradle. Set-up was based primarily on a set of cranial fiducial markers. MVCT scans were acquired and co-registered with planning scan separately at three different levels (brain, upper, and lower spine) at every fraction. Only translational displacements were analysed, wherein positive sign denotes deviation in anterior, left, and superior direction; while negative sign denotes deviation in posterior, right, and inferior direction. Mean displacements, systematic, and random errors of the study population were calculated at all three levels separately. Local residual uncertainty of the upper and lower spine was also derived assuming perfect co-registration of the skull. Set-up margins for clinical target volume (CTV) to planning target volume (PTV) were derived at these three levels separately using published margin recipes. Data from 1868 co-registrations in 674 fractions on 33 patients was included. The mean displacements in the lateral, longitudinal, and vertical directions were -1.21, -1.36, and 1.38 mm; -1.25, -0.34, and 0.65 mm; and -1.47, -2.78, and 0.22 mm for the brain; upper spine; and lumbar spine respectively. The corresponding 3D vector of displacement was 2.28; 1.45; and 3.15 mm respectively. There was a distinct systematic trend towards increasing inaccuracy from the brain towards the lower spine. Using Stroom's formula, the minimum recommended CTV to PTV margins in absence of daily image-guidance were 6.5; 7.0; and 9.5 mm for the brain; upper spine; and lower spine respectively. This increased to 7.5; 8.5; and 11.5 mm using van Herk's formula. Subset and sensitivity analyses could not identify any factor predictive of increased inaccuracy. Residual uncertainty of the spinal column was lesser after daily co-registration referenced to the skull, suggesting that smaller set-up margins maybe appropriate while using daily image-guidance with an online correction protocol. Daily MVCT imaging during supine CSI on HT provides volumetric verification of the set-up process. There is substantial site-dependent variability in translational displacements that increases systematically from brain towards the lower spine with implications for differential set-up -margins for the brain, upper, and lower spine.