A critical step in the clinical workflow for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) patients is lung disease severity assessment, providing valuable information to aid in effective patient care and management as well as treatment planning. Given the difficulty of performing such assessments by health-care workers and the necessity of expert radiologists who are al-ready burdened by the significant load caused by the pandemic,one promising direction is the use of computer-aided decision sup-port systems powered by deep learning. An important design consideration in the building of deep neural networks for SARS-CoV-2disease severity assessment is in the way severity scores are en-coded, as it can have a big influence on both the training and inference aspects of the neural network. In this study, we explore the performance impact of different severity encoding strategies for deep learning-based severity stratification of COVID-19 patients using chest x-rays (CXRs) on a clinical site cohort collected from the Stony Brook University Hospital. More specifically, we study the impact of different quantized severity encoding schemes, different granularity in the severity encoding, as well as compare quantized encoding vs. continuous encoding vs. hybrid centroid weighted en-coding.