Valid methods for meta-analyses with few studies or small sample sizes - Part II

主要目标是开发有效的方法进行荟萃分析，即使在困难的条件下也能得出值得信赖的结果。挑战是(i)研究之间的异质性，（ii）不平衡的设计，（iii）少量或中等数量的病例，以及（iv）研究较少或(v)以前荟萃分析的汇总结果。虽然传统方法在这一点上经常达到极限，但需要可靠地提供有效结果的方法，例如，使置信区间可靠地达到其名义覆盖概率。为此，将使用、研究或结合现代贝叶斯、自举、重采样和统计学习技术以及鲁棒三明治估计器。在第一个项目阶段的主要焦点是“经典”治疗效果估计器（例如标准化的平均差异）之后，第二个项目阶段将特别关注限制区间估计器（例如比例或相关性）。这些估计量使得证明共同正态分布假设更加困难，特别是在情况（iii）和（iv）中。为此，我们将开发单变量和多变量固定和随机效应模型以及混合效应元回归模型的新方法。此外，将研究回归方法来绘制更复杂的成分干预或解释可能的选择机制（发表偏倚）。为了对它们的充分使用提出建议，将在模拟研究中对所开发的方法进行广泛调查，并示范应用于当前的数据集。此外，基于元分析数据库的分析方法将进行实证评估。这些方法也在免费提供的开源软件中实现，这些软件将以易于理解的方式记录下来。这不仅将增加发布的模拟结果的透明度，而且还为更广泛的应用用户提供了对项目中开发的方法的轻松访问。与第一阶段一样，项目的继续将得到著名的墨卡托研究员的支持。

The main goal is to develop valid methods for meta-analyses that lead to trustworthy results even under difficult conditions. The challenges are (i) heterogeneity between studies, (ii) unbalanced designs, (iii) small or moderate numbers of cases, and (iv) few studies or (v) aggregated results from previous meta-analyses. While conventional methods often reach their limits at this point, methods are needed that reliably deliver valid results so that, for example, confidence intervals reliably reach their nominal coverage probability. For this purpose, modern Bayesian, bootstrap, resampling and statistical learning techniques, as well as robust sandwich estimators will be used, investigated or combined. After the main focus in the first project phase was on "classical" treatment effect estimators (such as standardized mean differences), the second project phase will in particular focus on restricted interval estimators (such as proportions or correlations). These estimators make it even more difficult to justify the common normal distribution assumption especially in cases (iii) and (iv). For this purpose, we will develop new methods for univariate and multivariate fixed- and random-effects models as well as mixed-effects meta-regression models. Furthermore, regression approaches will be investigated to map more complex component interventions or to account for possible selection mechanisms (publication bias). To give recommendations for their adequate use, the developed methods will be extensively investigated in simulation studies and exemplary applied to current data sets. In addition, analysis approaches based on meta-analysis databases will be empirically evaluated. The methods are also implemented in freely available open-source software, which will be documented in an easy understandable way. This will not only increase the transparency of published simulation results, but also provides easy access to the methods developed in the project for a wider audience of applied users . As in the first phase, the continuation of the project will be supported by renowned Mercator Fellows.