AN APPROACH TO EVALUATING THE EFFECT OF PRIORS IN NETWORK META-ANALYSIS
Author(s)
Rydevik G
Quantics Consulting ltd, Edinburgh, UK
OBJECTIVES: METHODS: Bayesian analysis is commonly applied for Network meta-analysis (NMA). In the UK, the National Institute for Health and Care Excellence (NICE) has issued guidelines for how to conduct NMA’s, including example code written in WinBUGS. The format of this code, including the use of vague priors, has become the de-facto standard for NMA of health technologies such as drugs and medical devices. A recent paper by Gabry et al [1] describes a workflow for evaluating the suitability of a Bayesian model by a series of visualisations of its predictive properties. In this work, we describe how this workflow can be applied to the NMA setting and evaluate the suitability of the priors recommended by NICE as compared to weakly informative priors. RESULTS: The results from the evaluation indicate that choosing uninformative priors can detrimentally affect the accuracy of the posterior estimates, in particular for sparse networks. CONCLUSIONS: The use of the visualisation framework of [1] can help in the process of deciding and defending the choice of priors in NMA and in medical statistics in general. [1] Gabry et al J. R. Statist. Soc. A (2019) 182, Part 1, pp. 1–14
Bayesian statistics differ from traditional frequentist statistics by requiring the specification of “priors” that describes the distribution of parameters of interest before any data is taken into account. The choice of priors is a controversial subject, in particular in medical and regulatory contexts. The established practice is to either use informative priors based on the results of previous studies, or to make use of vague or uninformative priors that attempt to minimise any effect on subsequent analysis. However, priors that fall in between these two extremes, such as weakly informative priors that attempt to capture the range of values that are reasonable given a problem setting, have traditionally been seen with skepticism.
Conference/Value in Health Info
2018-11, ISPOR Europe 2018, Barcelona, Spain
Value in Health, Vol. 21, S3 (October 2018)
Code
PRM163
Topic
Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference, Modeling and simulation
Disease
Multiple Diseases