| Summary: | Meta-analyses usually combine published studies, omitting those that for some reason have not been published. If the reason for not publishing is other than random, the problem of publication bias arises. Research into publication bias suggests that it is the ‘interest level’, or statistical significance of findings, not study rigour or quality, that determines which research gets published and subsequently publicly available. When the results of the scientific literature as a whole are considered, such publication practices distort the true picture, which may exaggerate clinical effects resulting in potentially erroneous clinical decision-making. Therefore, meta-analyses (as well as other more complex evidence synthesis models) based on the published literature should be seen as ‘at risk’ of publication bias, which has the potential to bias conclusions and thus adversely affect decision-making. Many methods exist for detecting publication bias, but this alone is not sufficient if results from meta-analyses are going to be used within a decision-making framework. What is required in the view of this thesis is a reliable way to adjust pooled estimates for publication bias. This thesis explores different novel and existing approaches to publication bias adjustment, including frequentist and Bayesian approaches with the aim to identifying those with the most desirable statistical properties. Special attention is given to regression-based methods commonly used to test for the presence of publication bias (and other ‘small-study effects’). The regression-based approach is seen to produce very encouraging results in a case study for which gold standard data exists. The incorporation of external information about the direction and strength of the bias is also explored in the hope of improving the methods’ performance. Ultimately, the routine estimation of the bias-adjusted effect is recommended as it improves the overall results compared to standard meta-analysis.
|
|---|
