Studies on Model Inference of Spatial Causal Effects

在一个日益城市化的世界里，我们的大部分生活都是在高度结构化的空间中进行的，并与其他人的生活交织在一起。因此，我们感兴趣的大多数社会现象正变得越来越难以理解和干预。在此背景下，近年来，迫切需要建立解决城市和政策问题的能力，重新引起了对空间因果效应的研究兴趣。面对这些挑战，调整基于模型的推理以适应当前因果关系的概念比以往任何时候都更加重要。在经典的因果框架中，“空间”的缺失给书面叙述带来了不公平的空间复杂性负担。如果我们的模型没什么可说的，为什么还要用它们呢？提出的研究从一个乐观的注意开始，即值得将空间编码到因果模型中。这种乐观主义将被仔细评估并付诸实践。作为激励问题，我们问：空间在因果过程中的作用是什么？空间信息在因果模型中应该出现在什么地方，以什么形式出现？从概念上讲，我们质疑空间如何作为因果过程的背景和组成部分呈现的模糊性。这与一些紧迫的挑战有关，例如未观察到的混淆作为模型错误规范的来源，以及由于空间回归中的共线性而产生的偏差。为了应对这些挑战，该研究将具有去偏见机器学习（ML）框架的双鲁棒性（DR）扩展到空间设置。对于新的方法结构，我们要问：它在存在空间结构化数据时如何执行？有什么替代方案可以使它“空间化”？它在空间环境中是否像在空间环境中一样“健壮”？本课程旨在回答这些问题。

In an increasingly urbanised world, much of our lives are performed in highly structured space and intertwined with those of others. As such, most of the social phenomenon we are interested in are becoming more difficult to understand and intervene on. Against this background, the urgency in building the capacity to address urban and policy problems has renewed the research interest in spatial causal effects in recent years. With these challenges, tuning model-based inference to current conceptions of causality is more relevant than ever. The absence of 'space' in classic cause-effect frameworks has put an unfair share of the burden of conveying spatial complexities on written narratives. If our models have so little to say, why use them at all? The proposed study starts from an optimistic note that it is worth encoding space into cause-effect modelling. This optimism will be carefully evaluated and put to practice.As motivating questions, we ask: What is the role of space in causal process? Where should spatial information appear in cause-effect models and in what form? Conceptually, we question the ambiguities with how space presents as both context to and component of causal processes. This is related to some pressing challenges, such as unobserved confoundedness as a source of model misspecification, and bias due to collinearity in spatial regressions. In response to these challenges, the study will extend the doubly robust (DR) with debiased machine learning (ML) framework to spatial settings. With the new methodological construct, we ask: How does it perform in the presence of spatially structured data? What are the alternatives to make it 'spatial'? Is it as 'robust' in spatial settings as aspatial ones? This studentship aims to answer these questions.