Advancing Instrumental Variable Methods in Comparative Effectiveness Research

描述（由申请人提供）：比较有效性研究（CER）主要涉及评估替代治疗对结果的因果影响。为此，观察性数据库是一个很有希望的关于患者治疗和结果的信息来源。然而，观察性数据分析容易出现选择偏倚或指征混淆，这是由于在接受不同治疗的患者中观察到的和未观察到的危险因素的水平存在差异，这使得对治疗因果效应的推断复杂化。工具变量（IV）方法是最有效的方法之一，可以解决这些挑战，并帮助估计这些数据的因果关系，但这些方法在CER中未得到充分利用。本应用程序的目标是通过克服采用这些强大的CER方法的三个重要障碍，增加工具变量方法的适当利用。适当使用IV方法进行CER取决于选择好的仪器和适当的估计。一个好的工具必须1)在治疗选择中引起实质性的变化（即“强”），但2)除非通过治疗选择（即“排除限制”），否则不会影响结果。虽然已经研究了使用弱仪器的后果，但违反排除限制的后果尚不清楚。即使在传统的均质治疗效果假设下，也正在开发几种新的静脉注射方法。了解哪种方法适合于任何特定的应用程序仍然具有挑战性。默认情况是使用两阶段最小二乘，但CER常见的许多情况需要替代方法，例如近距离匹配或两阶段残差包含。本应用程序旨在通过应用工具变量分析来解决这些挑战，目的是为CER的应用从业者提供适当的指导。如果治疗效果是异质的，静脉注射分析的结果可能被推广到错误的亚群，因为这些效果依赖于分析者对静脉注射的选择，并且难以为临床和政策目的解释。我们还将开发新的静脉注射方法，以解决治疗效果的异质性，并产生可解释的CER结果。由于缺乏易于实现的软件或统计代码，许多当前的CER应用并没有充分利用最近IV方法的进步，导致IV分析科学转化为实践的延迟。因此，我们将开发相关的统计代码，帮助从业者使用常用的统计软件包来实现这些方法，并通过前列腺癌和心血管疾病的实证例子来说明这些方法。

DESCRIPTION (provided by applicant): Comparative effectiveness research (CER) primarily involves estimation of causal effects of alternative treatments on outcomes. To this end, observational databases are a promising source of information on patient-level treatments and outcomes. However, observational data analyses are prone to selection biases or confounding by indication, which arise due to the differences in levels of observed and unobserved risk factors across patients receiving different treatments, and which complicate inference on causal effects of treatments. Instrumental variable (IV) methods are one of the most powerful methods that can address these challenges and help estimate causal effects from such data, yet these methods are underutilized for CER. The goal of this application is to increase the appropriate utilization of instrumental variables methods by overcoming three important barriers to adoption of these powerful methods for CER. Appropriate use of IV methods for CER hinges on selecting good instruments and appropriate estimation. A good instrument must 1) induce substantial variation in treatment choices (i.e. be "strong") but 2) not affect outcomes except through treatment choices (i.e. the "exclusion restriction"). While the consequences of using weak instruments have been investigated, the consequences of violating the exclusion restriction are not well understood. Even under the traditional assumption of a homogenous treatment effect, several new IV approaches are being developed. Knowing which method is appropriate for any particular application remains challenging. The default has been to use two-stage least squares, but many situations common to CER require alternative approaches such as near-far matching or two-stage residual inclusion. This application aims to address these challenges with applying instrumental variables analysis with a goal of providing applied practitioners of CER with appropriate guidance. Results of IV analyses may be generalized to the wrong subpopulations if treatment effects are heterogeneous as these effects become dependent on the analyst's choice of IV(s) and are difficult to interpret for clinical and policy purposes. We will also develop novel IV approaches that address treatment effect heterogeneity and generate interpretable results for CER. Many current applications of CER do not take full advantage of recent IV methodological advances, due to unavailability of readily implementable software or statistical code, resulting in delays in the translation of the science of IV analysis to practice. Therefore, we will develop relevant statistical code to help practitioners implement these methods using common statistical software packages and illustrate the methods through empirical examples in prostate cancer and cardiovascular disease. PUBLIC HEALTH RELEVANCE: Comparative effectiveness research (CER) primarily involves estimation of causal effects of alternative treatments on outcomes. To this end, observational databases are a promising source of information on patient-level treatments and outcomes. However, observational data analyses are prone to selection biases or confounding by indication, which arise due to the differences in levels of observed and unobserved risk factors across patients receiving different treatments, and which complicate inference on causal effects of treatments. Instrumental variable (IV) methods are one of the most powerful methods that can address these challenges and help estimate causal effects from such data, yet these methods are underutilized for CER. The goal of this application is to increase the appropriate utilization of IV methods by overcoming three important barriers to adoption of these powerful methods for CER.