Background: Bayesian network meta-analyses (NMA) are an extension of classic meta-analysis, but where a meta-analysis includes only direct evidence, NMAs draw on both direct and indirect evidence. The lack of head-to-head trials for newer drugs has seen an increase in the use of NMAs. When appropriate trials are to be synthesised, model choice for NMA is extensive. Researchers should not be limited to commonly reported methods, but should select models which appropriately capture the relationship of the outcomes being synthesised.

Objectives: To demonstrate the appropriate application of a hierarchical multinomial ordered logit model for a network meta-analysis in the context of a recent systematic review and decision model.

Methods: A NMA was used to estimate the probability that patients have a Psoriasis Area and Severity Index (PASI) 50/75/90 response or American College of Rheumatology (ACR) 20/50/70 response. These outcomes represent percentage reduction, have an order and are clearly related. The models were implemented in WinBUGS. The hierarchical multinomial ordered logit model, allowed an estimation of the probability of achieving one level of response compared with another as a linear function of the explanatory variables. Two separate analyses were undertaken for the outcomes of PASI 50/75/90 response and ACR 20/50/70 response. We estimated the probability that patients have a PASI 50/75/90 or ACR20/50/70 response by means of a cumulative logistic model. All priors were non-informative. Alternative modelling scenarios/assumptions were assessed. Results and Conclusions: The models presented provided flexible methods to ensure that the relationship between the ordered outcomes was appropriately maintained and synthesised. In the absence of head-to-head trials, and outcomes that are clearly related, researchers need to look beyond the standard models in our toolkits and use methods that deal appropriately with the clinical outcome of interest.