Casual Inference with Interference for Evaluating Air Quality Policies

 描述（由申请人提供）：公共卫生干预通常针对健康的上游决定因素（例如，社会或环境因素），以促进人口的健康。尽管这些干预措施是公共卫生政策的基石，但评估其有效性的因果推理方法受到当前对个体水平治疗的临床研究的关注的限制。一个极具争议的例子是一套旨在通过限制美国发电厂有害排放来减少污染相关健康负担的监管政策。与临床环境不同的是，比较这些监管干预措施的有效性受到以下事实的挑战，即污染排放在整个大气中演变，使特定地点的污染和健康结果部分取决于许多发电厂采取的干预措施。一个给定的单位的依赖于监管干预在多个发电厂引起了什么是已知的因果推理文献中的干扰。在一个观察水平上应用干预措施的事实（例如，发电厂）并且感兴趣的结果在另一级别（例如，个体或群体）对数据呈现二分结构。这些特征的组合提出了具有干扰的二分因果推理的挑战。目标1开发了新的贝叶斯方法，用于在观测可以聚类的环境中进行二分部分干扰（例如，通过地理或污染传输模式），使得干扰存在于集群内而不存在于集群之间。目标2发展新的贝叶斯方法与一般的干扰结构。目标3将我们新开发的方法部署到一个前所未有的数据库中，该数据库涉及整个美国的发电厂，排放，环境空气质量和健康结果，以比较减少发电厂排放的监管政策的有效性。目标4将通过开发可重复研究的工具来支持所有其他目标。我们开发和传播的方法，数据和软件将允许系统和严格地评估复杂的公共卫生干预措施的相对有效性，这些干预措施在多个观察单位之间表现出干扰。激励性的例子是空气质量监管政策，但 这些方法将证明适用于评价各种其他类型的复杂公共卫生干预措施。新开发的方法将通过放松在实践中经常违反的关键假设来推进因果推理领域。应用我们的 评估发电厂法规的方法将提供此类政策对健康影响的第一个基于实践的证据，并在评估有争议的空气质量干预措施以支持政策决定的方式方面构成范式转变。

 DESCRIPTION (provided by applicant): Public health interventions routinely target upstream determinants of health (e.g., social or environmental factors) to advance the health of populations. Even though such interventions are corner- stones of public health policy, methods for causal inference to evaluate their effectiveness are limited by a current focus on clinical investigations of individual-level therapies. One highly contentious example is the suite of reg- ulatory policies designed to reduce pollution-related health burden by limiting harmful emissions from US power plants. Unlike in clinical settings, comparing the effectiveness of these regulatory interventions is challenged by the fact that pollution emissions evolve throughout the atmosphere, rendering pollution and health outcomes at a given location determined in part by interventions taken at many power plants. A given unit's dependence on regulatory interventions at multiple power plants gives rise to what is known in the causal inference literature as interference. The fact that interventions are applied at one level of observation (e.g., power plants) and outcomes of interest are measured at another level (e.g., individuals or populations) presents a bipartite structure to the data. The combination of these features presents the challenge of bipartite causal inference with interference. Aim 1 develops new Bayesian methods for bipartite partial interference in settings where observations can be clustered (e.g., by geography or pollution transport patterns) so that interference is present within cluster but not between clusters. Aim 2 develops new Bayesian methods with general interference structures. Aim 3 deploys our newly-developed methods to an unprecedented database on power plants, emissions, ambient air quality, and health outcomes across the entire US to compare the effectiveness of regulatory policies for reducing power plant emissions. Aim 4 will support all other aims with the development of tools for reproducible research. The methods, data, and software we develop and disseminate will allow systematic and rigorous evaluation of the comparative effectiveness of complex public health interventions that exhibit interference among multiple levels of observational unit. The motivating example is air quality regulatory policy, but the methods will prove applicable to the evaluation of a variety of other types of complex public health interventions. The newly-developed methods will advance the field of causal inference through relaxation of key assumptions that are routinely violated in prac- tice. Application of our methods to the evaluation of power plant regulations will provide the first statistically-based evidence of the health impacts of such policies and constitute a paradigm shift in the way controversial air quality interventions are evaluated to support policy decisions.