Modeling missing binary outcome data while preserving transitivity assumption yielded more credible network meta-analysis results.
OBJECTIVES:The objectives of this study were to elaborate on the conceptual evaluation of transitivity assumption in the context of binary missing participant outcome data (MOD) in network meta-analysis (NMA) and to emphasize on the importance of statistical modeling as a mean to address MOD. STUDY DESIGN AND SETTING:We designate the notion of transitivity assumption in the context of binary MOD and indicate scenarios that compromise transitivity in complex networks. We propose a modification of these scenarios that preserves transitivity assumption. Using a published NMA, we indicate the implications of excluding or imputing, rather than modeling MOD, on NMA findings. RESULTS:Arm-specific scenarios for MOD, as commonly applied in conventional meta-analysis, compromise the validity of transitivity assumption in complex networks. The motivating example reveals that imputation of those scenarios yields estimates in the opposite direction for the basic parameters with narrower credible intervals and inflates between-trial variance. Contrariwise, modeling MOD after modification of the scenarios yields robust estimates for the basic parameters but wider credible intervals and reduces between-trial variance. CONCLUSION:Application of arm-specific scenarios for binary MOD requires modification in complex networks to ensure valid transitivity assumption. Analysts should model, rather than exclude or impute MOD, to provide bias-adjusted results.