Presentation (CTI)

OBJECTIVES: In network meta-analysis (NMA), between-study heterogeneity can arise from differences in study design, biases, or random variation. When heterogeneity is anticipated, random-effects models are often employed, however, these models can become impractical due to the inability to reliably estimate heterogeneity in the presence of a sparse network or weak data. Since heterogeneity contributes to the uncertainty in treatment effect estimates, unreliable heterogeneity estimates can artificially inflate credible intervals. Incorporating informative priors for the between-study heterogeneity parameter offers a potential solution to this challenge in the Bayesian framework. This work aims to demonstrate the impact of using informative priors for between-study heterogeneity on NMA model conclusions.
METHODS: Random-effects models NMAs were conducted on simulated datasets under various networks and data specifications. The impact of incorporating informative priors was assessed and the results were compared to those obtained using non-informative priors.
RESULTS: Informative priors for the between-study heterogeneity parameter provided more stable estimates in the presence of sparse networks leading to tighter credible intervals when compared to results obtained from models employing non-informative priors. The influence of prior choice diminished as the amount of available data increased.
CONCLUSIONS: In NMAs with few studies, the information can fail to overcome vague or non-informative priors, leading to analyses being disproportionately influenced by the chosen distribution. Informative priors offer a practical solution by incorporating external scientific knowledge, ensuring that results remain robust even with sparse data. Defining how few studies is too few is inherently challenging, though the distinction may ultimately be unnecessary as informative priors are unlikely to negatively impact analyses with extensive evidence, underscoring the importance of carefully selecting and justifying the prior distribution for all NMAs, irrespective of size.