Issues in performing a network meta-analysis

Abstract

The example of the analysis of a collection of trials in diabetes consisting of a sparsely connected network of 10 treatments is used to make some points about approaches to analysis. In particular various graphical and tabular presentations, both of the network and of the results are provided and the connection to the literature of incomplete blocks is made. It is clear from this example that is inappropriate to treat the main effect of trial as random and the implications of this for analysis are discussed. It is also argued that the generalisation from a classic random-effect meta-analysis to one applied to a network usually involves strong assumptions about the variance components involved. Despite this, it is concluded that such an analysis can be a useful way of exploring a set of trials.