Bayesian Methods for Meta-Analysis in the Presence of Publication Bias

在社会科学和行为科学中，积极结果和消极结果发表率的差异被称为发表偏倚（publication bias），越来越受到关注。本研究项目将开发一种新的元分析方法，明确考虑到有偏见的出版过程的可能性。荟萃分析在解释学术文献中提出的主张方面发挥了重要作用。然而，学术期刊，尤其是社会科学和行为科学的期刊，似乎更倾向于假设某种效应存在的手稿，而不是无显著结果的手稿。这阻碍了经典的元分析方法，因为一组有偏见的经验结果的总和也会有偏见。新方法将允许更好地汇总已发表的结果，并将提供对各种实验操作和治疗效果的更准确的看法。将开发和发布为各种情况实现此方法的软件。该研究项目将开发一种新的元分析方法，称为“统计缓解”，将行为模型与最先进的统计方法相结合。该方法将基于贝叶斯模型平均技术，其中使用一组合理的选择模型计算效应大小估计，并对这些选择模型进行平均。利用这种方法，可以在测量误差噪声中隔离出真实效果的信号。研究者将在各种情况下测试该方法，将新方法与存在发表偏倚的现有推断方法进行比较，并进行模拟以评估该方法的效率。通过单一的方法进行荟萃分析，研究人员将能够解释发表偏倚的可能性，确认或不确认null和非null假设，并进行效应大小估计。

The differential rate of publishing between positive and negative results, which has been called publication bias, is of increasing concern in the social and behavioral sciences. This research project will develop a new approach to meta-analysis that explicitly takes into account the possibility of a biased publication process. Meta-analysis has been instrumental in interpreting the claims made in the academic literature. However, academic journals, especially in the social and behavioral sciences, seem to strongly prefer manuscripts that posit the existence of an effect rather than non-significant outcomes. This hinders classical meta-analysis methods because the aggregate of a biased set of empirical results will be biased as well. The new approach will allow for better aggregation of published results and will provide a more accurate view of the effect of various experimental manipulations and treatments. Software will be developed and published that implements this approach for a variety of situations.This research project will develop a new approach to meta-analysis called "statistical mitigation" that combines behavioral models with state-of-the-art statistical methods. The approach will be based on a Bayesian model averaging technique in which effect size estimates are computed using a set of plausible selection models and averaging across these selection models. With this approach, it will be possible to isolate the signal of true effects within the noise of measurement error. The investigator will test the method under various circumstances, compare the new approach to existing methods for inference in the presence of publication bias, and perform simulations to assess the efficiency of the method. With a single approach to meta-analysis, researchers will be able to account for the possibility of publication bias, confirm or disconfirm null and non-null hypotheses, and do effect size estimation.