PhD Seminar: Paul Morris, "Bayesian Model Selection for Component Network Meta-Analysis Models"
Speaker: Paul Morris, PhD Candidate, Department of Statistics, Iowa State University
Title: Bayesian Model Selection for Component Network Meta-Analysis Models
Abstract: Network meta-analysis (NMA) is used to simultaneously compare multiple interventions by synthesizing direct and indirect evidence from a network of randomized clinical trials. Component network meta-analysis (CNMA) enables the modeling of the combinative comparative effects as a function of the effects of the individual components. Explicit modeling of the comparative effects comes with a risk of severe model misspecification. Thoughtful model selection and assessment of model fit are therefore important aspects of conducting a CNMA. In this paper, we propose using posterior model probabilities to compare Bayesian CNMA models for binary outcomes, as this approach has several desirable properties: the posterior model probabilities implicitly penalize model complexity, and they select the best model asymptotically under mild conditions. Bridge sampling is used to estimate the marginal likelihoods, as it is relatively straightforward to implement for high-dimensional models and has been found to provide accurate estimates. In addition, hierarchical priors are employed for the trial-specific baseline and component effect parameters as a means to improve the robustness of the model selection procedure to the choice of parameter priors. The proposed model selection approach is then applied to a network of interventions aimed at preventing the formation of liver abscesses in beef cattle with the purpose of deciding between a restricted model under which the effects associated with one component are fixed at zero and an unrestricted full model. In a simulation study based on this analysis, the proposed model selection approach was found to select the data generating model at a high rate, particularly when the size of the network was large. Posterior model probabilities therefore appear to be a viable model selection approach for CNMA models despite the difficulties involved in their calculation.