Novel Designs and Methods to Remove Hidden Confounding Bias in Health Sciences

摘要 在因果推论文献中，一个旨在减轻由于未测量的混淆引起的偏差的主要方法是这样的-- 称为辅助变量(IV)设计，其依赖于识别影响治疗过程的变量， (Ii)除通过治疗外，对结果没有直接影响，以及(Iii)独立于任何不可测量的 混血儿。四种方法在社会科学和卫生科学中得到了很好的发展和广泛的应用，尽管 如果违反所需的任何假设(I)-(Iii)，则IV推论可能不可靠。这项提案旨在发展 (A)对违反(I)-(Iii)中任何一项的行为具有强大抵抗力的新的IV方法；。(B)可用于检测 并有时非参数地解释未测量的混杂偏差；(C)新的括号方法 比较中断时间序列研究中因果效应的推断。建议的方法将用于 解决当前在三个主要实质性公共卫生领域提出的科学问题：(1)了解 空气污染；(2)量化阿尔茨海默病和相关疾病的可改变危险因素的因果影响； (3)揭示随机化的一揽子干预措施大幅减少碳排放的机制 最近在非洲博茨瓦纳进行的一项治疗作为预防的主要分组随机试验中的艾滋病毒发病率。我们的建议 将提供迄今为止最好的分析方法，以解决这些高影响力公众中令人困惑的问题 在健康科学的观察性研究中，更广泛地说，在健康应用方面。

Abstract A major approach in causal inference literature aimed at mitigating bias due to unmeasured confounding is the so- called instrumental variable (IV) design which relies on identifying a variable which (i) influences the treatment process, (ii) has no direct effect on the outcome other than through the treatment, and (iii) is independent of any unmeasured confounder. IV methods are very well developed and widely used in social and health science, although validity of IV inferences may not be reliable if any of required assumptions (i)-(iii) is violated. This proposal aims to develop (a) new IV methods robust to violation of any of (i)-(iii); (b) New negative control methods that can be used to detect and sometimes to nonparametrically account for unmeasured confounding bias; (c) New bracketing methods for partial inference about causal effects in comparative interrupted time series studies. The proposed methods will be used to address current scientific queries in three major substantive public health areas:(1) to understand the health effects of air pollution; (2) to quantify the causal effects of modifiable risk factors for Alzheimer's disease and related disorders; (3) To uncover the mechanism by which a randomized package of interventions produced a substantial reduction of HIV incidence in a recent major cluster randomized trial of treatment as prevention in Botswana, Africa. Our proposal will provide the best available analytical methods to date to resolve confounding concerns in these high impact public health applications and more broadly in observational studies in the health sciences.