Skip to main content

Languages

Evaluating the transitivity assumption when constructing network meta-analyses: lumping or splitting?

Date and Location

Session: 

O4.02.3

Date

Monday 23 September 2013 - 13:30 - 15:00

Location

Presenting author and contact person

Presenting author

Kristina Lindsley

Contact person

Kristina Lindsley
Abstract text
Background: In conducting a network meta-analysis of medical interventions for open-angle glaucoma including 311 trials and 139 interventions, our first task was construction of the network. The transitivity assumption, which says that the comparative effectiveness of treatments A vs. B can be derived from comparisons of treatments A vs. C and B vs. C, is key to network construction. In order for this assumption to hold, each treatment group, or node in the network, must be comparable within itself. Failure of the transitivity assumption manifests as statistical inconsistency. Objectives: To examine methodological and clinical factors that may affect the transitivity assumption when constructing a network of first-line drugs for open-angle glaucoma. Methods: We catalogued pre-specified factors that may affect the comparability of treatment groups (or nodes) in a network. We consulted a glaucoma specialist on how best to combine specific treatments within the network. We considered how decisions to group treatments could affect the transitivity assumption and interpretation of the analysis. This process is in many ways analogous to evaluating heterogeneity in a pair-wise meta-analysis. Results: Factors potentially affecting the transitivity assumption included class, type, concentration, frequency, and route of administration of drug. Medical interventions could be grouped into 4 classes, 11 drug types, and 24 concentrations (Table). We constructed two network graphs by splitting or lumping drug concentrations (Figure). Conclusions: By definition, a split network carries less inconsistency. However, lumping results in broader questions, which generally is the goal of network meta-analysis. It also allows for more use of direct comparisons which enhances the reliability of the findings. The clinical question of interest ultimately directs the construction of the network. Our next steps will encompass analyzing outcomes quantitatively, implementing various approaches for handling inconsistency, and performing sensitivity analyses to evaluate the impact of decisions made while constructing the network.
Attachments