OBJECTIVES: Interpretation of network meta analyses in areas of research with many interventions can be complex and difficult to communicate. We explore the use of sparsity inducing Bayesian priors in the case where there is prior belief that a large number of interventions have similar effects.

METHODS: We use an in progress review of pharmacological and non-pharmacological interventions for pain relief to compare a traditional NMA based on NICE TSD 2 code to one that induces sparsity in treatment effects. To induce sparsity we use a single common treatment effect with a an additional deviation term for each therapy. Terms for the additional deviation are parameterized using stochastic search variable selection based on the assumption of a fixed number of treatments that are meaningfully different from the pooled average across all therapies. The calibration of the model is verified using simulated data.

RESULTS: The original NMA featured 36 unique interventions, leading to 630 pairwise comparisons. The reduced model shrunk most treatments together, resulting in five interventions with treatment effects that were different than the average pooled effect. The reduced model had a DIC of 1232 compared to 1236 for the baseline model, indicating meaningfully better fit. The model was able to accurately capture treatment effects/differentiate between data simulated from a reduced vs full model.

CONCLUSIONS: SSVS can be used as an exploratory tool to reduce the complexity of large NMAs. This can be helpful in mature therapeutic areas, particularly when dominated by inexpensive/simple therapies with limited theoretical differentiation.