The Exchangeability Assumption in Network Meta-Analysis: Its Meaning and Evaluation.

OBJECTIVES: The validity of network meta-analysis (NMA) relies on 3 assumptions, namely, homogeneity, consistency, and transitivity. A recent article argued that all 3 assumptions can be reduced to the exchangeability assumption. However, it is not entirely clear what is exchangeable in NMA and how the exchangeability assumption can be assessed. In this article, we demonstrate that the contrast-based model for NMA assumes that contrasts between pairs of treatments (treatment contrasts) are exchangeable across studies.

METHODS: We propose using a simple graphical approach to assessing the validity of exchangeability in NMA by plotting the absolute treatment effects against potential effect modifiers. We also use this graphical approach to show the distinction between absolute and relative effect modifiers. We demonstrate how the concept of exchangeable treatment contrasts is related to consistency and transitivity and the consequence of violating the exchangeability assumption.

RESULTS: We use a real NMA from dentistry to illustrate how to apply our method to evaluate the extent to which exchangeable treatment contrasts is a valid assumption. Our method provides valuable insights into the correct interpretation of NMA results when the exchangeability assumption is met or questionable.

CONCLUSIONS: Our approach demonstrates not only the assumption of exchangeable treatment contrasts but also raises an important issue about designing an NMA: whether a strict selection of studies for NMA should be used to ensure the exchangeability but limited generalizability of results, or a broad selection of studies to explore the conditions under which the results of NMA are applicable.