OBJECTIVES Network meta-analysis (NMA) techniques have been developed to study relative treatment effects for several outcome types (e.g. time-to-event outcomes). No literature exists comparing models of NMA for ordered categorical data, though models are available with different characteristics. This study compared the proportional odds (PO) and multinomial logistic (ML) model for NMA in ordered categorical datasets based on model fit and qualitative characteristics. METHODS To contrast model performance, two extreme datasets were simulated, one which exactly satisfied the PO assumption (POA dataset), and one which did not (nPOA dataset). The models were also tested in a clinical dataset including ordered response categories for four different treatments in psoriasis patients. Both fixed and random effects models were studied. RESULTS In the POA dataset, the PO fixed effects model had the lowest residual deviance (54.8 versus 58.9 for the ML model) and uncertainty of treatment effects (49% lower standard error (SE)). In the nPOA dataset, the predictions of the PO model were biased, and the ML model had the lowest residual deviance (52.7 versus 271.0 for the PO model). Visual inspection indicated a partial violation of the PO assumption in the psoriasis data. Analyses of the psoriasis data, showed that the PO fixed effects model had the lowest residual deviance (18.1 versus 20.9) and uncertainty (62% lower SE). However, PO model predictions were biased for treatment responses which violated the PO assumption. CONCLUSIONS Statistical selection of NMA models for ordered outcomes should be based on the PO assumption and deviance measures. If data satisfies the PO assumption, the PO model differentiated treatment effects better as a result of lower uncertainty. In terms of flexibility, the PO model can handle data from studies that use different cut-offs for response categories and the ML model can be applied to datasets violating the PO assumption.