OBJECTIVES: Ordinal models are popular for synthesis in disease areas where outcomes can be ordered in terms of severity or where they arise from discretization of an underlying continuous scale. These models typically assume a single treatment effect across all levels. In this abstract we review methods to relax this assumption.

METHODS: Narrative review with code examples.

RESULTS: Models included in the overview include the simple proportional odds model in addition to fixed and random effect models that take various approaches to relaxing the this assumption: (1) treatment specific distances between categories; (2) a partial proportional odds model that assumes a distinct treatment effect for each threshold which may or may not exchangeable across categories; (3) a model that assumes a linear trend (increasing/decreasing) across thresholds; (4) a model that assumes a polynomial trend across thresholds. Model 1 has been used in recent NMAs conducted within PsO and has nice properties including guaranteed maintenance of threshold ordering but models 2-4 are more established within pairwise meta-analysis and individual participant data analysis and may have advantages in terms of interpretability and efficiency.

CONCLUSIONS: When conducting pairwise or network meta-analysis of ordinal outcomes, there are various alternatives for exploring the validity of the proportional odds assumption. The relative performance of these methods has not been evaluated, and no method has been implemented in commonly used software.