Causal inference in network settings

许多统计学问题旨在回答病因学问题：当考虑到可能的偏倚因素，如非随机抽样（选择偏倚）、测量误差、缺失数据或混淆时，暴露Z是否会导致结果Y的变化？因果推理提供了一种形式化假设的方法，这些假设是从分析中得出因果结论所必需的。虽然因果方法通常用于简单的设置，如点或甚至纵向暴露，但在这个令人兴奋的统计学分支中有许多新的领域需要探索。拟议的研究将处理几个新环境中的因果关系，在这些环境中，传统的回归技术是不合适的。我将重点关注与因果推理相关的三个主要目标：(1)在环境数据的背景下扩展因果方法；(2)开发用于受访者驱动样本的因果估计器；(3)从存储在不能共享个人层面数据的分布式站点网络中的非实验数据中估计顺序定制决策规则。这三个目标都处理网络中因果推理的不同方面，根据网络结构的类型产生不同的挑战。在环境数据中，数据往往是空间分布的，违背了通常假设的“无干扰”。在被调查者驱动的抽样中，网络结构的产生是由于抽样设计，参与者招募他们的联系人，因此抽样单位不是独立的，对于某些暴露，可能会再次出现干扰。第三个目标解决了一个非常不同的问题：虽然给定站点内的数据可能是相关的，但分析人员无法获得单独的度量，并且只有汇总或摘要信息可用于了解量身定制的决策策略。这项工作的影响将是改变研究人员进行旨在实现重要研究目标的分析的方式，从了解环境暴露对经济结果的影响到更好地使用来自多个站点的数据，而不会在共享完整数据时出现隐私泄露的风险。这项工作还将为相关性和复杂网络结构对可以因果回答的问题类型的影响以及用于回答这些问题的估计器提供重要的见解。

Many statistical problems aim to answer etiological questions: does exposure Z cause changes in an outcome Y, when accounting for possible biasing factors such as non-random sampling (selection bias), measurement error, missing data, or confounding? Causal inference offers a means of formalizing the assumptions that are required to draw causal conclusions from an analysis. While causal methods are routinely used for simple settings such as point or even longitudinal exposures, there are many new frontiers to explore in this exciting branch of statistics. The proposed research will tackle causality in several new settings, where traditional regression techniques are not appropriate. I will focus on three main objectives related to causal inference: (1) to extend causal methods in the context of environmental data; (2) to develop causal estimators for use in respondent-driven samples; and (3) to estimate sequential tailored decision-rules from non-experimental data that are stored across a distributed network of sites who cannot share individual-level data. These three objectives all tackle different aspect of causal inference in a network, with different challenges arising depending on the type of network structure. In environmental data, data are often spatially distributed, and the commonly-made assumption of "no interference" is violated. In respondent-driven sampling, the network structure arises due to a sampling design in which participants recruit their contacts and hence the sample units are not independent and, for some exposures, interference may again arise. The third aim tackles a very different problem: while data within a given site may be correlated, individual measures are unavailable to the analyst, and only pooled or summary information is available to learn about tailored decision strategies. The impact of this work will be to change the manner in which researchers carry out analyses designed to achieve important research goals, ranging from understanding the impact of environmental exposures on economic outcomes to better using data from multiple sites without the risk of privacy breaches that occur when full data are shared. The work will also provide important insights into the impact of correlation and complex network structures on the types of questions that can be answered causally, and the estimators used to answer them.