A Comparison of Bayesian Models for Network Meta-analysis

Background
Since the emergence of evidence-based medicine movement, effective evidence synthesis becomes important for decision making for clinical researchers and policy makers. Meta-analysis of results from clinical trials is therefore an indispensable research tool for research synthesis. While traditional meta-analysis compares two treatment groups, network meta-analysis can compare more than two treatments within one statistical framework. The current Bayesian hierarchical model for network meta-analysis was first proposed by Lu and Ades with the use of the flexible statistical software WInBUGS.

Objectives
Because it is quite complex to set up Lu & Ades’s model, this research attempts to develop a new model which is simpler and more flexible. Consequently, the learning curve for clinical researchers to undertake network meta-analysis is less steep.

Methods
We propose a “Random Treatment Effects Model” and compare it to the Lu & Ades’s model and ”Contrast model” proposed by Piepho. We use a real data, which was from the AHCPR’s Smoking Cessation Guideline Panel by Fiore et al.,to illustrate the three models yields similar results, but our“Random Treatment Effects Model” is more intuitive and flexible.

Results
Two different random effect structures can be set up in Random Treatment Effects Model.There are no substantial differences in results between the four models, and the treatment effects in smoking cessation from high to low are group counseling、individual counseling、self-help and no contact.

Conclusions
The Random treatment effects model we proposed yields the same results as those from the Lu & Ades model. Furthermore, the model is more intuitive to understanding, and it has more flexibility to set up complex random effects structure, which is closer to the reality.