03.04.2023 Vortrag & Tutorial für Bayes'sche Meta-Analysen auf Youtube

Zwei Beiträge von Prof. Daniel Heck für die ESMARConf2023, einer Konferenz zu Meta-Analysen (https://esmarconf.org), sind jetzt auf Youtube verfügbar:

Vortrag: Bayesian Model Averaging for Meta-Analysis

Abstract:
Meta-analysis aims at the aggregation of observed effect sizes from a set of primary studies. Whereas fixed-effect meta-analysis assumes a single, underlying effect size for all studies, random-effects meta-analysis assumes that the true effect size varies across studies. Often, the data may not support one of these assumptions unambiguously, especially when the number of studies under consideration is small. In such a case, selecting one of the two models results in too narrow confidence intervals when assuming fixed-effects but in low statistical power when assuming random-effects. As a remedy, Bayesian model averaging can be used to combine the results of four Bayesian meta-analysis models: (1) fixed-effect null hypothesis, (2) fixed-effect alternative hypothesis, (3) random-effects null hypothesis, and (4) random-effects alternative hypothesis. Based on the posterior probabilities of these models, Bayes factors allow to quantify the evidence for or against the two key questions: "Is the overall effect non-zero?" and "Is there between-study variability in effect size?". Besides considering model uncertainty, Bayesian inference enables researchers to include studies sequentially in order to update a meta-analysis as new studies are added to the literature. The R package metaBMA facilitates the application of Bayesian model-averaging for meta-analysis by providing an accessible interface for computing posterior model probabilities, Bayes factors, and model-averaged effect-size estimates for meta-analysis.

Tutorial: The R Package metaBMA

Link: https://youtu.be/e68YX1VTe_A

GitHub repository: https://github.com/danheck/metaBMA

Abstract:
The metaBMA package implements Bayesian model-averaging for meta-analysis in R. Whereas fixed-effect meta-analysis assumes a constant, true effect size for all studies, random-effects meta-analysis assumes that true effect sizes vary across studies. Often, the data may not support one of these assumptions unambiguously. As a remedy, Bayesian model averaging combines the results of four meta-analysis models: (1) fixed-effect null hypothesis, (2) fixed-effect alternative hypothesis, (3) random-effects null hypothesis, and (4) random-effects alternative hypothesis. Based on the posterior probabilities of these four models, Bayes factors quantify the evidence for or against two key questions: "Is the overall effect non-zero?" and "Is there between-study variability in effect size?" Besides considering model uncertainty, Bayesian inference enables researchers to include studies sequentially in order to update a meta-analysis as new studies are added to the literature. In this tutorial, I provide a worked example of how to perform a Bayesian, model-averaged meta-analysis in R using the metaBMA package. I also explain how to specify prior distributions and how to interpret posterior model probabilities, Bayes factors, and model-averaged effect-size estimates.