Handling missing continuous outcome data in a Bayesian network meta-analysis

Background: A Bayesian network meta-analysis (NMA) model is a statistical method aimed at estimating the relative effects of multiple interventions against the same disease. The method has recently gained prominence, leading to the synthesis of the evidence regarding rank probabilities for each treatment. In several cases, an NMA is performed excluding incomplete data of studies retrieved through a systematic review, resulting in a loss of precision and power. Methods: There are several methods for handling missing or incomplete data in an NMA framework, especially for continuous outcomes. In certain cases, only baseline and follow-up measurements are available; in this framework, to obtain data regarding mean changes, it is necessary to consider the pre-post study correlation. In this context, in a Bayesian setting, several authors suggest imputation strategies for pre-post correlation. In other cases, a variability measure associated with a mean change score might be unavailable. Different imputation methods have been suggested, such as those based on maximum standard deviation imputation. The purpose of this study is to verify the robustness of Bayesian NMA models concerning different imputation strategies through simulations. Results: Simulation results show that the bias is notably small for every scenario, confirming that rankings provided by models are robust concerning different imputation methods in several heterogeneity-correlation settings. Conclusions: This NMA method seems to be more robust to missing data imputation when data reported in different studies are generated in a low-heterogeneity scenario. The NMA method seems to be more robust to missing value imputation if the expectation of the prior distribution, defined on the heterogeneity parameter, approaches the true value of the variability across studies