Collaborative Research: Propensity Scores and Randomization-Based Inference, or Modeling Assignment to Treatment Conditions as Manifestly or Latently Random

合作研究：倾向评分和基于随机化的推理，或将治疗条件的分配建模为显性随机或隐性随机

• 批准号: 0753168
• 负责人: Jacob Bowers
• 金额: --
• 依托单位: University of Illinois at Urbana-Champaign
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2008
• 资助国家: 美国
• 起止时间: 2008-05-01 至 2011-04-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0753168&HistoricalAwards=false
• 关键词: Collaborative; Research; Propensity; Scores; Randomization

The project comprises methodological research to aid the analysis of comparative studies; that is, studies comparing "treated" subjects to controls in order to estimate effects of the treatment. Its starting point is an influential approach associated with Donald Rubin and Paul Rosenbaum, among others. In this approach, if such a study can be assumed latently randomized -- conditional on the covariate, treatment is assigned as if at random -- then the sample is reorganized into suitably similar blocks, and later analysis proceeds as if assignment to treatment had been manifestly at random, at least within these blocks. Techniques with which to reorganize samples in this way, such as matching and propensity scoring, have received much methodological attention in recent years, but whether and when it is inferentially valid to treat the transformed sample as if it had manifestly been randomized remains incompletely understood. In propensity-score matching, for example, extant methodological literature is relatively silent on estimation uncertainty in the propensity score, on imprecision in matching on it, and on how to determine whether these errors may be large enough to invalidate the adjustment. The primary aim of the present project is to use new characterizations of these errors, which are compatible with widely used forms of propensity stratification, to develop diagnostics for the suitability of a given propensity matching or stratification. Secondarily, the project assumes that manifestly randomized studies, along with those being treated as such, are to be analyzed with methods that rely only on properties of randomization, rather than model-based methods. These methods differ from those most commonly used in the social sciences to analyze comparative studies, but they can build on and borrow strength from better-known methods as this project will also make clear. The project's methodological advances will be demonstrated in social science applications, and will be made freely available in high-quality open source software. A motivating premise of the project is that good statistical methods not only better meet the needs of specialists who use them but also enable nonspecialists to more readily and accurately appraise the quantitative evidence those specialists produce. The research therefore concerns itself with a type of statistical adjustment, matching, that is distinguished for its simplicity and appeal to non-technical audiences. Because it and related methods are already quite popular, the methods and methodological extensions developed by the project will be immediately and directly applicable to any number of empirical investigations in the social and behavioral sciences, many of which have public health or public policy implications. These methods' emphasis on diagnostics, and on statistical conclusions based on the same independence assumptions that the diagnostics seek to corroborate, may help to demystify the statistical component of empirical investigations to which they contribute, and to make transparent the evidential strengths and weaknesses of particular studies. The project also seeks to make the methods themselves more broadly available. By generating freely available software, it will enable all who are so inclined to use the techniques in their own analyses. By including undergraduate and graduate students in the research, it contributes to the training of future quantitative social science methodologists. Because it includes a plan for special marketing of this opportunity to underrepresented minority college students, it complements efforts by others to cultivate research expertise among underrepresented groups.

该项目包括方法学研究，以帮助分析比较研究，即比较“治疗”对象与对照对象的研究，以估计治疗效果。 它的出发点是与唐纳德·鲁宾和保罗·罗森鲍姆等人有关的一种有影响力的方法。 在这种方法中，如果可以假设这样的研究是潜在随机的-以协变量为条件，治疗被随机分配-然后将样本重新组织成适当相似的块，随后的分析进行，至少在这些块内，治疗的分配似乎是明显随机的。 以这种方式重新组织样本的技术，如匹配和倾向评分，近年来受到了很多方法学上的关注，但是否以及何时将转换后的样本视为明显随机化的样本仍然不完全清楚。 例如，在倾向分数匹配中，现存的方法学文献对倾向分数中的估计不确定性、匹配的不精确性以及如何确定这些误差是否大到足以使调整无效相对沉默。 本项目的主要目的是使用这些错误，这是兼容的广泛使用的倾向分层的形式的新的表征，开发诊断的适合性，一个给定的倾向匹配或分层。 其次，该项目假设，明显随机化的研究，沿着那些被视为这样的，将被分析的方法，只依赖于随机化的属性，而不是基于模型的方法。 这些方法不同于社会科学中最常用的分析比较研究的方法，但它们可以建立在更知名的方法的基础上，并从中汲取力量，正如本项目也将阐明的那样。 该项目的方法进步将在社会科学应用中得到展示，并将以高质量的开放源码软件免费提供。该项目的一个激励性前提是，好的统计方法不仅能更好地满足使用它们的专家的需求，而且能使非专家更容易、更准确地评估专家提出的定量证据。 因此，本研究关注的是一种统计调整，即匹配，以其简单性和对非技术受众的吸引力而闻名。 因为它和相关的方法已经相当流行，该项目开发的方法和方法扩展将立即直接适用于社会和行为科学中的任何数量的实证调查，其中许多具有公共卫生或公共政策的影响。 这些方法强调诊断，并强调基于诊断所寻求证实的相同独立性假设的统计结论，这可能有助于揭开它们所促进的实证调查的统计部分的神秘面纱，并使特定研究的证据优势和弱点变得透明。 该项目还寻求使这些方法本身更广泛地提供。 通过生成免费提供的软件，它将使所有倾向于在自己的分析中使用这些技术的人都能使用。 通过包括本科生和研究生的研究，它有助于培养未来的定量社会科学方法学家。 因为它包括一个向代表性不足的少数民族大学生特别推销这一机会的计划，它补充了其他人在代表性不足的群体中培养研究专长的努力。