The goal of this project is to study the statistical properties of a variety of tests that have been proposed for detecting the presence of publication bias prior to performing a meta-analysis. Publication bias occurs when only a subset of the completed studies on the topic of interest are published, and when the decisions to publish individual studies ar influenced by the results of the studies. This can lead to two observable patterns in the data: an induced correlation between the observed effects and the sample sizes; a """"""""bunching"""""""" of the observed effects in regions where selective publication is strongest, e.g., when p<0.05. Each of these patterns provides leverage for the construction of a statistical test. In the former case we use a rank correlation test of effect size versus sample size (or study variance). In the latter case we impute the selection function using weighted distribution theory, and construct a test based on the observed pattern in the data. In this project we will study the operating characteristics of these tests using simulations, with a view to developing practical strategies for meta- analysts.
|Begg, C B; Mazumdar, M (1994) Operating characteristics of a rank correlation test for publication bias. Biometrics 50:1088-101|