Ranking of treatments in network meta-analysis: incorporating minimally important differences.

Background: In network meta-analysis (NMA), the magnitude of difference between treatment effects is typically ignored in the calculation of ranking metrics, such as probability best and surface under the cumulative ranking curve (SUCRAs). This leads to treatment rankings which may not reflect clinically meaningful differences. Minimally important differences (MIDs) represent the smallest value in a given outcome that is considered by patients or clinicians to represent a meaningful difference between treatments. There is a lack of literature on how MIDs can be incorporated into common NMA ranking metrics such as SUCRAs to give more clinically oriented treatment rankings.

Methods: Analogues to commonly available NMA ranking metrics that account for minimally important differences (MIDs) are provided. In particular, definitions are provided for MID-adjusted median ranks, MID-adjusted probability th best, MID-adjusted cumulative probability th best, and MID-adjusted SUCRA values. Since adjustment for MIDs allows for ties between treatments in a network, methods for handling ties in ranking are discussed, with it shown that the midpoint method for handling ties retains the property that the average value of all SUCRA values in a network is one half. Comparability of MID-adjusted P-scores and MID-adjusted SUCRA values is discussed, and a Bayesian software implementation of the MID-adjusted ranking metrics is provided.

Results: Two real-world applications of MID-adjusted ranking metrics are presented to illustrate their use. Specifically, NMAs are conducted based on published networks on treatments for diabetes and Parkinson's disease. To present the results, MIDs are selected from relevant literature to interpret MID-adjusted ranking metrics for these networks.

Conclusions: Failure to consider MIDs when ranking treatments can lead to ranking metrics which are not clinically relevant. Our proposed MID-adjusted Bayesian ranking metrics address this challenge. Further, we show that the use of the midpoint method for addressing ties ensures comparability between standard ranking metrics and MID-adjusted ranking metrics. The methods are easily applied in a Bayesian framework using the R package mid.nma.rank.