A microaggregation framework for reproducible research with observational data: addressing biases while protecting personal identities

项目总结/摘要 当前提案的主要目标是促进努力实现透明和 循证医学的可复制知识库。医疗保健的财富 可以更好地利用电子健康记录中已有的数据， 治疗选择和随机对照试验的结果。这项建议 解决了两个主要障碍。第一个挑战是从以下方面获得高质量的证据： 观察数据中存在的偏见和混杂因素，特别是与时间数据。 第二，患者隐私和其他问题阻止了源数据的披露， 阻碍了可重复的研究--目前有大量的医学文献， 研究结果指导临床实践，但不能独立审查。我们将解决这些问题 通过创新的方法，本地控制，既纠正偏见， 能够披露特定问题的微观汇总数据，以重现研究结果 不披露个人信息。地方控制背后的关键思想是形成许多 同质患者群，在其中可以在统计学上比较替代治疗 校正测量偏差和混杂因素，类似于随机区组设计。我们 方法提供了一个统一的框架，使开放的，高质量的，比较 有效性研究结合新的特征选择方法，基于分数 析因实验设计，生存分析的进展，包括竞争风险。我们 我将创建一个公共R包，其中包含一系列非参数偏差校正方法 以及在横断面、病例对照和生存分析设置中的统计披露控制。 这项研究的成功还将使一种新的模式成为可能，我们称之为“打包数据共享”， 促进针对具体问题公开、有选择地发布专有数据源- 同时保护患者隐私、专有利益和公共利益。我们 研究将有助于实现循证医学的目标， 全球知识库，基于数百年来数千个比较有效性问题， 数百万病人的健康记录此应用程序通过协助执行以下任务来支持NLM使命： 通过传播和交流医学和相关科学， 医学和健康进步的重要信息。具体目标是：（1） 开发和评估基于生存率的局部控制方法，用于偏倚校正治疗 时间-事件观察数据的比较;和（2）制定和评估当地控制- 基于微聚集的可重复研究。

PROJECT SUMMARY/ABSTRACT The primary objective of the current proposal is to foster efforts towards transparent and reproducible knowledge repositories for evidence-based medicine. The wealth of healthcare data already available in electronic health records could be better utilized to help guide treatment choices and compliment findings from randomized controlled trials. This proposal addresses two major obstacles. The first is the challenge of deriving high-quality evidence from observational data in the presence of biases and confounders, particularly with temporal data. The second is that patient privacy and other concerns prevent disclosure of source data, which hinders reproducible research -- currently there is a vast body of medical literature whose findings guide clinical practice, yet cannot be independently scrutinized. We will address these challenges through an innovative methodology, local control, which both corrects biases and enables disclosure of question-specific microaggregated data to reproduce research findings without disclosure of individual information. The key idea behind local control is to form many homogeneous patient clusters within which one can compare alternate treatments, statistically correcting for measured biases and confounders, analogous to a randomized block design. Our methodology provides a unified framework for enabling open, high quality, comparative effectiveness research by combining novel feature selection approaches, based on fractional factorial experimental design, with advances in survival analysis, including competing risks. We will create a public R package containing a family of methods for nonparametric bias correction and statistical disclosure control in cross-sectional, case-control, and survival analysis settings. Success of this research will also enable a novel model, we term “parcelled data sharing” to facilitate open selective release of proprietary data sources for specific questions -- simultaneously protecting patient privacy, proprietary interests, and the public good. Our research will contribute to the goal of evidence-based medicine being supported by national and global knowledge bases on thousands of comparative effectiveness questions from 100’s of millions of patients’ health records. This application supports the NLM mission by assisting in the advancement of medical and related sciences through the dissemination and exchange of important information to the progress of medicine and health. The specific aims are to (1) Develop and evaluate a survival-based local control methodology for bias-corrected treatment comparisons in time-to-event observational data; and (2) Develop and evaluate local control- based microaggregation for reproducible research.