gms | German Medical Science

Introduction: A number of techniques have been proposed to handle missing binary outcome data (MOD) in systematic reviews with meta-analyses. However, none of these techniques have been evaluated empirically in a series of published systematic reviews with multiple interventions.

Methods: Using published systematic reviews with network meta-analysis (NMA) from a wide range of health fields, we evaluated comparatively the most frequently reported modeling techniques and missingness scenarios for binary MOD in order to elucidate their implications on core NMA components (i.e. basic parameters, between-study variance, inconsistency factor and intervention hierarchy). In particular, we aimed to investigate to what extend the NMA results differed when: (i) extreme scenarios about the missingness mechanism were adopted; (ii) uncertainty due to missingness was ignored; (iii) different structures were considered to model the missingness parameter [1]; and (iv) pattern-mixture model was used instead of selection model [2], [3], [4]. To illustrate the level of agreement between different techniques and scenarios, we used Bland-Altman plots. We extended the Bayesian random-effects NMA model to incorporate the informative missingness odds ratio (IMOR) parameter [1], [2], [3], [4], whereas we applied the node-splitting approach to investigate possible inconsistency locally.

Results: Imputation of MOD prior to analysis alongside clinically implausible scenarios led to systematically different and spuriously precise basic parameters and between-study variance compared to modeling MOD or considering the missing at random assumption. Hierarchical modeling of log IMORs reduced between-study variance, whereas assuming IMORs to be ‘intervention-specific’ inflated uncertainty around inconsistency factor and basic parameters. Pattern-mixture model yielded relatively larger basic parameters and between-study variance, whereas selection model reduced precision around log IMORs. Overall, different techniques agreed sufficiently in the light of low missingness in the network.

Discussion: Addressing MOD using clinically implausible scenarios or fixing the observations prior to analysis may seriously affect core NMA components. Choosing to model MOD can ensure valid conclusions and, furthermore, offer valuable insights on the underlying missingness mechanism.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.