Inconsistency identification in network meta-analysis via stochastic search variable selection.

Area of Science:

• Biostatistics
• Evidence Synthesis
• Health Research Methodology

Background:

• Network meta-analysis (NMA) reliability hinges on the transitivity assumption, requiring similar effect modifier distributions across comparisons.
• Statistical consistency, where direct and indirect evidence align, is a key manifestation of transitivity.
• Existing methods for evaluating consistency often involve adding inconsistency factors to NMA models.

Purpose of the Study:

• To propose a novel method, stochastic search inconsistency factor selection (SSIFS), for evaluating the consistency assumption in network meta-analysis.
• To assess inconsistency both locally and globally using variable selection techniques.
• To develop an informative prior for network consistency based on historical NMA data.

Main Methods:

• SSIFS describes inconsistency factors using candidate covariates selected via variable selection techniques.
• The method applies stochastic search variable selection to determine the inclusion of inconsistency factors.
• Posterior inclusion probabilities quantify the likelihood of inconsistency in specific comparisons, with decisions aided by posterior model odds or the median probability model.

Main Results:

• The proposed SSIFS method effectively evaluates network meta-analysis consistency.
• It quantifies the likelihood of inconsistency for specific treatment comparisons.
• The method incorporates differences between direct and indirect evidence and utilizes an informative prior based on 201 published NMAs.

Conclusions:

• SSIFS offers a robust approach to assessing network meta-analysis consistency, enhancing the reliability of synthesized evidence.
• The method provides a data-driven way to identify and quantify inconsistencies.
• The SSIFS methodology is available as an R package on CRAN.