Arm-based versus contrast-based methods for network meta-analyses: radical differences or misunderstood nuances?

Tags: Oral
Shrier I1, Schnitzer M2, Steele RJ3
1Centre for Clinical Epidemiology, Lady Davis Institute, McGill University, Canada, 2Faculty of Pharmacy, Université de Montreal, Canada, 3Department of Mathematics and Statistics, McGill University, Canada

Background: Network meta-analyses have traditionally estimated exposure effects by modeling contrasts (e.g. risk ratios or risk differences). Although some have recently argued that modeling arm-specific risks directly is also possible, this 'radical' suggestion has met considerable resistance from certain segments of the research synthesis community. The most commonly used argument against arm-based methods is that arm-based methods 'break randomization'. Interestingly, there are very few papers that explicitly discuss and compare the underlying assumptions of these two methods.

Objectives: The objectives of this presentation are to review the differences in the approaches at a conceptual level, and explain the challenges and benefits associated within each under different contexts. We posit that the usual goal of ranking treatments is an arm-based objective.

Methods: We use a causal inference approach and simulation studies.

Results: The essential difference is that arm-based methods rank exposure arms directly, whereas contrast-based approaches must convert estimated contrast effects into arm-based rankings afterwards. We show that the differences between the two methods can be defined in terms of the weights associated with the study treatment arms and the resulting variances of the estimators of the arm-specific parameters. When all studies include only two exposure arms, one arm-based analysis will produce identical point estimates to the contrast based method, but power is reduced. The variances are likely to be similar when the network meta-analyses is based on three-arm studies, and the variance will generally be less for the arm-based approach when the network meta-analysis is based on four-arm studies. More generally, both approaches require appropriate modeling of the causes driving particular participants into different studies. Failure to do this will result in biased estimates early in the process, or late in the process, but biased estimates nonetheless.

Conclusions: Our preliminary work suggests that arm- and contrast-based approaches yield unbiased estimates when done appropriately; variances depend on the number of study arms.