Title: Modeling baseline treatment effects in Bayesian network meta-analysis of disconnected networks
Authors: Audrey Beliveau - University of Waterloo (Canada) [presenting]
Sergiu Pocol - University of Waterloo (Canada)
Abstract: Network meta-analysis is a set of statistical tools conventionally used to establish comparative efficacy/safety of more than two interventions using data extracted from a systematic literature review of randomized controlled trials. In recent years, Bayesian implementations of network meta-analysis have been prominent since they lend themselves naturally into ranking of treatment efficacy and into health-economic decision modeling. Disconnected networks arise when some of the treatments of interest are not compared head-to-head within a study nor indirectly through studies with treatments in common. The analysis of disconnected networks is typically frowned upon due to the lack of a trusted gold-standard method. The standard contrast-based model for the network meta-analysis of connected networks treats baseline treatment effects as fixed but treating these as random makes the estimation of contrasts between disconnected treatments possible. Using a normal distribution on the baseline treatment effects inherently assumes that the baseline treatment effects are exchangeable across studies and that a normal distribution represents accurately the variation across studies. We explore empirically to what extent this assumption of normal and exchangeable baseline treatment effects could be a problem in real applications and whether alternative distributions such as the Student t distribution could help mitigate problems.