PurposeMixed beam electron–photon radiation therapy (MBRT) is an emerging technique that has the potential to reduce dose to normal tissue while improving target coverage for cancer sites with superficial tumors. Advances in optimization algorithms and robotic linear accelerators have made the creation and delivery of complex MBRT plans realistic without the need for special additional collimators, devices, or resetup of the patient. However, no study has been performed on the robustness of MBRT dose distributions to patient setup errors. Intensity‐modulated delivery of other charged particles such as protons have been shown to require robust planning techniques to maintain adequate target coverage under positioning errors. We therefore assess the sensitivity of MBRT treatment plans to positioning uncertainties when created under the traditional planning target volume (PTV)‐based planning paradigm and present a novel optimization model for the creation of robust MBRT plans.MethodsThe column generation method was applied to robust MBRT treatment planning by deriving the pricing problem for stochastic and “worst case” minimax optimization models, two common formulations of robustness. Robust treatment plans were created for two patient cases representative of the cancer sites which stand to benefit from MBRT: soft tissue sarcoma (STS) irradiation and chest wall irradiation with deep‐seated internal mammary, axillary, and supraclavicular nodes (CW‐N). For both patient cases, beamlet dose distributions for electrons and photons were generated for positioning shifts in six directions, in addition to a nominal unshifted scenario, for a total of seven sets of beamlets. Robust plans were created by specifying dose coverage constraints to the clinical target volume (CTV), as opposed to the PTV. Comparisons were performed against traditional PTV‐based plans created with a single set of unshifted beamlets.ResultsThe dose distributions of traditional PTV‐based MBRT plans showed significant degradation in target coverage homogeneity when patient positioning errors were considered. For both cancer sites, cold spots below 95% and hot spots above 108% of the prescription dose appeared within the CTV when shifting the patient by 5 mm, corresponding to the margin added to the CTV to form the PTV. In contrast, CTV‐based robust plans created with the new optimization model maintained target coverage within the 95%–108% limits, for all positioning errors.ConclusionThe quality of MBRT treatment plans created using a traditional PTV‐based optimization model was highly sensitive to patient positioning errors. For both patient cases, positioning errors resulted in perturbations to the nominal dose distributions which would have rendered PTV‐based plans clinically unacceptable. In contrast, CTV‐based robust plans were able to maintain adequate target coverage under all positioning error scenarios considered. We therefore conclude that to ensure the fidelity of the dose distribution delivered to the patient, robust optimization is critical when creating MBRT plans.