Sensitivity analysis applied to epidemiology

说明： 拟议的出版物将收集和综合量化影响非随机化流行病学研究的系统误差的方法。在非随机化流行病学研究中，传统的可信区间作为总体不确定性的度量表现不佳，因为它们只考虑随机误差，然而这些研究中的大多数误差来自系统来源。流行病学研究产生效果估计(如风险比)，其结果对卫生干预的各个方面都有影响。伴随这些效应估计的误差被定义为它们与实际效应的差异，可分为随机误差和系统误差。随着研究规模的增加，随机误差趋于零，但系统误差不变。精度量化效果估计中的随机误差量，通常用一个可信区间来表示。有效性衡量的是系统误差的大小，但很少被量化。可以通过灵敏度分析对与效果估计相关的系统误差进行定量评估。 拟议出版物的主要受众包括非随机化流行病学研究的所有作者、分析师和消费者。该出版物将指导计划进行敏感性分析的读者，并将说明在进行分析和汇总结果以供出版时可应用于数据集的敏感性分析方法。出版物将有一个文本部分和一个软件部分。文本将作为一本常规书籍出版，并以电子方式在国家生物技术信息中心的在线书架上发布。该软件将在北京邮电大学的互联网上免费提供。软件文件将是Microsoft Excel文件和SAS宏代码，用于说明这些方法，并允许读者使用其数据实施敏感性分析方法。这本出版物将增强调查人员评估效果估计中的系统误差的能力。

DESCRIPTION: The proposed publication will collect and synthesize methods for quantifying systematic errors that affect non-randomized epidemiologic research. Conventional confidence intervals perform poorly as measures of total uncertainty in non-randomized epidemiologic research because they account for only random error, yet the majority of the error in these studies arises from systematic sources. Epidemiologic studies yield effect estimates (such as the risk ratio) and their results have an impact on all aspects of health interventions. The error accompanying these effect estimates is defined as their difference from the true effects, and can be divided into random error and systematic error. The random error approaches zero as the study size increases but systematic error does not. Precision quantifies the amount of random error in an effect estimate, and is usually represented by a confidence interval. Validity measures the amount of systematic error, but is seldom quantified. A quantitative assessment of systematic error associated with an effect estimate can be made by sensitivity analysis. The primary audience of the proposed publication comprises all authors, analysts, and consumers of non-randomized epidemiologic research. The publication will guide readers who are planning for sensitivity analyses and will illustrate methods of sensitivity analysis that can be applied to data sets as they conduct analyses and summarize results for publication. The publication will have a text component and a software component. The text will be published as a conventional book and electronically at the National Center for Biotechnology Information on-line bookshelf. The software will be freely available on the BUSPH internet. The software files will be Microsoft Excel files and SAS macrocode that illustrate the methods and allow the reader to implement sensitivity analysis methods with their data. This publication will enhance the ability of investigators to assess systematic error in effect estimates.