12

There are now several different approaches to perform a network meta-analysis or mixed treatment comparison.

The most commonly used and accessible ones are probably the following:

  • in a Bayesian framework:

    • design-by-treatment interaction approach in WinBUGS (eg Jackson et al);
    • hierarchical arm-based Bayesian modeling in WinBUGS (eg Zhao et al);
    • hierarchical contrast-based (i.e. node-splitting) Bayesian modeling, either with WinBUGS or through gemtc and rjags in R (eg Dias et al or van Valkenhoef et al);
    • integrated nested Laplace approximations (INLA) in WinBUGS (eg Sauter et al);

  • in a frequentist framework:

    • factorial analysis-of-variance in SAS (eg Piepho);
    • multilevel network meta-analysis in SAS (eg Greco et al);
    • multivariate meta-regression with mvmeta in Stata or R (eg White et al);
    • network meta-analysis with lme and netmeta in R (eg Lumley, which is however limited to two-arm trials, or Rucker et al).

My question is, simply: are they roughly equivalent or is there one which is preferable in most cases for the primary analysis (thus reserving the others for ancillary ones)?

UPDATE

Over the time, there have been some comparative analyses on methods for network meta-analysis:

  1. Carlin BP, Hong H, Shamliyan TA, Sainfort F, Kane RL. Case Study Comparing Bayesian and Frequentist Approaches for Multiple Treatment Comparisons. Agency for Healthcare Research and Quality (US). 2013.
Giuseppe Biondi-Zoccai
  • 2,244
  • 3
  • 19
  • 48

1 Answers1

1

I think, the modeling approaches and estimation techniques should be viewed seperately. From modeling point of view, Lumley model only works for two-arm trials only. So it is not preferable. To my understanding, node-splitting approach, which you listed as Dias et al, is very intuitive. Also, I think you should add the design-by-treatment interaction approach (http://www.ncbi.nlm.nih.gov/pubmed/24777711). From estimation point of view, I dont know much about frequentist techniques, but one can use MCMC for almost all models for NMA. Lastly, there is a different technique (which is not widely known unfortunately) called INLA. You can use INLA from within R and fit NMA models, it is faster and no need to check convergence diagnostics. Here is the paper http://www.ncbi.nlm.nih.gov/pubmed/26360927. So, at the end I would prefer node-splitting and the design-by-treatment interaction approach using INLA.

Burak
  • 41
  • 5
  • 1
    You are asking which one is preferable: Bayesian or frequentist. But they are two different paradigms. And also this is beyond network meta-analysis, it is a general statistical inference question (or maybe even philosphical). So I dont think comparing Bayesian and frequentist approaches in the context of NMA reasonable. – Burak Oct 13 '16 at 13:09
  • 1
    Thanks for your perspective. There are of course key background and underlying differences, but my question is very practical. If I have to recommend a junior researcher which method is best for NMA, what should I pick? This might mean choosing between Bayesian and frequentist approaches, but the answer could even be more specific... – Giuseppe Biondi-Zoccai Oct 13 '16 at 19:50