Meta-analysis has become established as a fundamental tool of evidence- based practice, and is relied upon the AHRQ and many other government agencies and professional organizations. Still, key methodological issues remain, including the problem of publication bias, which can threaten the results of any meta-analysis. It is generally recognized that medical studies that achieve statistical significance are more likely to be published than those that do not. Several methods have been developed for detecting and adjusting for publication bias. The funnel plot, a simple graphical tool for detecting publication bias, is frequently used with heterogenous meta-analysis, even though it is only appropriate when all studies come from a single underlying population. Selection models are a promising alternative to methods based on the funnel plot. They model the chance of suppression of studies and us maximum likelihood or Bayesian methods to calculate an adjustment treatment effect. More research is need on this sophisticated class of models. A large-scale comparison of selection models will be undertaken. This will involve developing a selection model that appropriately incorporates baseline risk, and will include the evaluation o f several existing models. Simulated data with a wide range of characteristics of meta-analysis normally found in practice, as well as a large database or real meta- analysis will be used. Information will be extracted from heterogeneous meta-analysis to better understand the relation between heterogeneity and funnel plot asymmetry. The long-range goal is to make available, for nay meta-analysis, the most appropriate tool for dealing publication bias, depending on the characteristics of the data.
