Methods for Synthesizing Regression Results

智力价值：在这个项目中，研究人员将检查现有的用于表示荟萃分析回归分析结果的指数，并将开发新的方法来合成回归分析结果。目前，由于很难比较不同模型的结果，元分析研究经常忽略回归结果。调查人员将考虑这些研究指数的适当标准误差，这些标准误差是大多数研究结果组合的基础。此外，他们将确定现有指数是否可以用来代表结构方程模型和分层线性模型(HLMS)等相关分析的结果，并将在需要时提出新的指数。在研究中将考虑几种特殊情况。特别是，将考虑旨在表示组差异的结果(即，虚拟变量的斜率)，包括回归模型包含多个虚拟变量以表示具有多个水平的因素的情况(例如，两个处理组和一个共同的控制组)。由于元分析人员也经常希望总结不同设计的研究(例如，具有回归研究的HLM研究)，建议的研究将概述这种交叉设计组合是合理的条件。更广泛的影响：这项研究将对数学和科学教育以及在回归模型和其他形式的线性模型常见的许多其他领域(例如经济学、政治学)的研究综合产生广泛影响。这项研究提供的方法将使评价者能够将回归结果与更简单的指标(例如相关性、标准化平均数差异)一起纳入Meta分析。将特别注意综合工作，在这些工作中，亚群体比较(如在服务不足群体内部和跨群体差异)特别重要。

Intellectual Merit: In this project, the investigators will examine existing indices for representing the results of regression analyses for meta-analysis, and will develop new methods for synthesizing results from regression analyses. Currently, meta-analytic studies often omit regression results due to the difficulty in comparing results across different models. The investigators will consider the appropriate standard errors for these study indices, which are fundamental to most combinations of study results. In addition, they will determine whether existing indices can be used to represent results of related analyses such as structural equation models and hierarchical linear models (HLMs), and will propose new indices if needed. Several special cases will be considered in the research. In particular, results meant to represent group differences (i.e., slopes for dummy variables) will be considered, including cases where regression models contain several dummy variables to represent factors with multiple levels (e.g., two treatment groups and a common control group). Because meta-analysts also often wish to summarize across studies with different designs (e.g., HLM studies with regression studies), the proposed research will outline the conditions under which such cross-design combinations are reasonable. Broader Impact: This study will have broad impacts on syntheses of research in mathematics and science education, and in many other areas (e.g., economics, political science) where regression models and other forms of the linear model are common. The methods provided by this research will enable reviewers to include regression results in meta-analyses together with simpler indices (e.g., correlations, standardized mean differences). Special attention will be paid to synthesis work where subpopulation comparisons (such as within and across underserved group differences) are particularly important.