Online Program

A Bayesian hierarchical model for network meta-analysis with selection bias

Bradley P. Carlin, Division of Biostatistics, University of Minnesota School of Public Health 
Haitao Chu, Division of Biostatistics, University of Minnesota School of Public Health 
Hwanhee Hong, Division of Biostatistics, University of Minnesota School of Public Health 
James D. Neaton, Division of Biostatistics, University of Minnesota School of Public Health 
Lei Nie, Division of Biometrics IV, Office of Biometrics/OTS/CDER/FDA 
Guoxing Greg Soon, Division of Biometrics IV, Office of Biometrics/OTS/CDER/FDA 
Beth Virnig, Division of Health Policy and Management, University of Minnesota School of Public Health 
*Jing P. Zhang, Division of Biostatistics, University of Minnesota School of Public Health 

Keywords: network meta-analysis, missing mechanisms, bayesian hierarchical models, selection bias

Network meta-analysis (NMA), a meta-analytic statistical method, expands the scope of a conventional pairwise meta-analysis to simultaneously multiple treatments comparisons, synthesizing both direct information and indirect information. The typical data in NMA has an incomplete-blocks structure with heavy missing data problem. It is common to assume that the data are missing at random (MAR). However, sometimes the highly missing data may be due to deliberate choice, for example, clinicians tends to select more effective treatments in RCTs based on previous RCTs, which leads to nonignorable missingness, or missing not at random (MNAR). Sensitivity analysis is usually used to evaluate the effect of varying assumptions on study conclusions. We propose a Bayesian Hierarchical Model for NMA and perform sensitivity analysis using selection models. We also talk about heterogeneity and inconsistency, which are two key issues in NMA. We apply our model to a smoking cessation data and evaluate our model through a simulation study under various missing mechanisms (Complete, MCAR, MAR, MNAR).