Node-Splitting Generalized Linear Mixed Models for Evaluation of Inconsistency in Network Meta-Analysis

Network meta-analysis for multiple treatment comparisons has been a major development in evidence synthesis methodology. The validity of a network meta-analysis, however, can be threatened by inconsistency in evidence within the network. One particular issue of inconsistency is how to directly evaluate the inconsistency between direct and indirect evidence with regard to the effects difference between two treatments. A Bayesian node-splitting model was first proposed and a similar frequentist side-splitting model has been put forward recently. Yet, assigning the inconsistency parameter to one or the other of the two treatments or splitting the parameter symmetrically between the two treatments can yield different results when multi-arm trials are involved in the evaluation.

We aimed to show that a side-splitting model can be viewed as a special case of design-by-treatment interaction model, and different parameterizations correspond to different design-by-treatment interactions.

We demonstrated how to evaluate the side-splitting model using the arm-based generalized linear mixed model, and an example data set was used to compare results from the arm-based models with those from the contrast-based models.

The three parameterizations of side-splitting make slightly different assumptions: the symmetrical method assumes that both treatments in a treatment contrast contribute to inconsistency between direct and indirect evidence, whereas the other two parameterizations assume that only one of the two treatments contributes to this inconsistency. With this understanding in mind, meta-analysts can then make a choice about how to implement the side-splitting method for their analysis.