Article type
Year
Abstract
Background: The Bayesian approach gives new opportunities such as aggregating different types of data, combining direct and indirect comparison or assessing clinical significance. On the other hand, the variety of Bayesian models can beconfusing, and implementation difficulties can cause unwillingness to apply it.
Objectives: The aim is to systemize the knowledge of application of Bayesian statistics in the area of meta-analysis, and to compare it with traditional statistical methods. We want to identify the situations in which use of the Bayesian approach is really worthwhile.
Methods: Initially the systematic reviews were conducted using existing statistical methods for meta-analyses. Special attention was given to different Bayesian models which were then implemented in the WinBUGS environment and examined on different data sets.
Results: In the case of regular meta-analysis of dichotomous data, applying basic Bayesian models leads us, in fact, to similar results of estimation as the Mantel-Haenszel or DerSimonian-Laird method. The real advantage of the Bayesian approach is noticed if we expect something more than typical meta-analysis, especially if we have to deal with the following problems: 1. Assessing the clinical significance-for instance, assessing the chance that relative risk is greater than 1.25 (or any other level of significance). 2. Combining data from different type of studies including extra information (e.g. results of non-randomized studies) to meta-analysis, keeping a moderate 'level of conviction’ to this extra data. 3.Combining direct and indirect evidence (Mixed Treatment Comparison).
Conclusions: Bayesian statistics give us technical opportunities to improve meta-analysis, especially in the area of aggregating multi-type data. On the other hand, there is no significant difference between the results obtained by the Bayesian and traditional approach in the case of simple meta-analysis of regular data. Moreover, if WinBUGS codes are prepared, conducting the calculations is not as difficult as one may think.
Objectives: The aim is to systemize the knowledge of application of Bayesian statistics in the area of meta-analysis, and to compare it with traditional statistical methods. We want to identify the situations in which use of the Bayesian approach is really worthwhile.
Methods: Initially the systematic reviews were conducted using existing statistical methods for meta-analyses. Special attention was given to different Bayesian models which were then implemented in the WinBUGS environment and examined on different data sets.
Results: In the case of regular meta-analysis of dichotomous data, applying basic Bayesian models leads us, in fact, to similar results of estimation as the Mantel-Haenszel or DerSimonian-Laird method. The real advantage of the Bayesian approach is noticed if we expect something more than typical meta-analysis, especially if we have to deal with the following problems: 1. Assessing the clinical significance-for instance, assessing the chance that relative risk is greater than 1.25 (or any other level of significance). 2. Combining data from different type of studies including extra information (e.g. results of non-randomized studies) to meta-analysis, keeping a moderate 'level of conviction’ to this extra data. 3.Combining direct and indirect evidence (Mixed Treatment Comparison).
Conclusions: Bayesian statistics give us technical opportunities to improve meta-analysis, especially in the area of aggregating multi-type data. On the other hand, there is no significant difference between the results obtained by the Bayesian and traditional approach in the case of simple meta-analysis of regular data. Moreover, if WinBUGS codes are prepared, conducting the calculations is not as difficult as one may think.