An empirical comparison of Bayesian modelling strategies for missing binary outcome data in network meta-analysis

Area of Science:

• Biostatistics
• Epidemiology
• Health Sciences

Background:

• Systematic reviews often encounter missing binary outcome data (MOD).
• Existing strategies for handling MOD in meta-analysis lack empirical evaluation.
• Network meta-analysis (NMA) is increasingly used, necessitating robust methods for MOD.

Purpose of the Study:

• To empirically evaluate Bayesian modeling strategies for MOD in published NMA.
• To compare different approaches for handling MOD, including pattern-mixture and selection models.
• To assess the impact of various informative missingness odds ratio (IMOR) structures and scenarios on NMA results.

Main Methods:

• Comparative evaluation of Bayesian modeling strategies for MOD using published NMA.
• Extension of Bayesian random-effects NMA model to incorporate the IMOR parameter.
• Application of node-splitting for local inconsistency assessment and Bland-Altman plots for agreement illustration.

Main Results:

• Extreme scenarios for MOD, ignoring uncertainty, yielded different and more precise log odds ratios (log ORs) than MAR assumption.
• Hierarchical IMOR structures reduced between-trial variance, especially with substantial MOD.
• Pattern-mixture and selection models showed agreement, but selection models reduced IMOR precision.

Conclusions:

• Ignoring MOD uncertainty or using extreme scenarios can negatively impact NMA estimates.
• Modeling MOD using the IMOR parameter provides bias-adjusted estimates and insights into missingness.
• Expert consultation is recommended for selecting IMOR structures and prior distributions; findings apply to pairwise meta-analyses.