Robert L Grant1. 1. Honorary Research Fellow, Faculty of Health, Social Care & Education, Kingston and St George's, University of London, London, UK.
Abstract
AIM: Bayesian statistical methods can allow for more complete and accurate incorporation of evidence in meta-analyses. However, these methods remain under-utilized. METHODS: A scoping review was conducted to examine the proportion of biomedical meta-analyses that used Bayesian methods in the period 2005-2016. The review also examined the reproducibility of the work, the cited sources, the reasons for it, its success or failure, the type of model and prior distributions, and whether a mixture of Bayesian and frequentist methods were employed. RESULTS: We found that 1% of meta-analyses are Bayesian and that the reporting and conduct of these were often poor. Data were published in 41% of analyses, and programs to run the analysis in 18%. Network meta-analysis was the most common reason and became increasingly popular in recent years. In the majority of papers, models and distributions were either not reported or explained in such brief and ambiguous terms as to be uninformative. CONCLUSIONS: More use needs to be made of Bayesian meta-analysis, and reporting needs to be improved. Greater awareness of these methods and access to training in them is essential.
AIM: Bayesian statistical methods can allow for more complete and accurate incorporation of evidence in meta-analyses. However, these methods remain under-utilized. METHODS: A scoping review was conducted to examine the proportion of biomedical meta-analyses that used Bayesian methods in the period 2005-2016. The review also examined the reproducibility of the work, the cited sources, the reasons for it, its success or failure, the type of model and prior distributions, and whether a mixture of Bayesian and frequentist methods were employed. RESULTS: We found that 1% of meta-analyses are Bayesian and that the reporting and conduct of these were often poor. Data were published in 41% of analyses, and programs to run the analysis in 18%. Network meta-analysis was the most common reason and became increasingly popular in recent years. In the majority of papers, models and distributions were either not reported or explained in such brief and ambiguous terms as to be uninformative. CONCLUSIONS: More use needs to be made of Bayesian meta-analysis, and reporting needs to be improved. Greater awareness of these methods and access to training in them is essential.