PoTEMU: Policy and Treatment Evaluation under Model Uncertainty

政策和治疗评估的标准量化方法的有效性至关重要地依赖于建模假设，而这些假设在实践中往往难以证明是合理的。当这些假设失败时，标准方法就会崩溃，推理可能会产生误导。相反，一种稳健的因果推断方法将适应与关键假设的偏差，并提供考虑到经验实践固有的模型不确定性的建议。为了制定这样一种方法，本建议将(I)为模型不确定性下的因果分析引入一个新的统计框架，(Ii)开发相应的因果效应估计和推理工具，以及(Iii)在实证实践中扩大因果推断的范围。这一建议用一种新的分布回归方法对已建立的因果推断的控制变量方法进行了重新表述。这种独特的组合允许在不同个人之间区分政策和待遇影响，同时在模型假设有效时提供对这些分布影响的最准确衡量，或者在某些模型假设不成立时提供准确的近似值。因此，这种方法适合于在模型不确定的情况下进行分配策略和治疗评估。这项建议寻求开发的方法将与从经济学到卫生科学的实证研究中普遍使用非实验设计的广泛学科相关。在这些环境中，随机实验的理想条件通常不被满足，即研究中的个体不太可能在所有相关方面都相同，除了由模型直接控制的那些方面。例如，这些理想条件通常只在用观察数据评估新药的有效性时才大致成立。我的建议将适用于这种情况，并提供现实世界实施所需的工具。

The validity of standard quantitative methods for policy and treatment evaluation crucially relies on modelling assumptions that are often hard to justify in practice. When these assumptions fail, standard methods break down and inference is potentially misleading. A robust method for causal inference would instead accommodate deviations from key assumptions and deliver recommendations that account for model uncertainty inherent to empirical practice. To formulate such a method, this proposal will (i) introduce a novel statistical framework for causal analysis under model uncertainty, (ii) develop corresponding estimation and inference tools for causal effects, and (iii) extend the scope of causal inference in empirical practice. This proposal reformulates the established Control Variables method for causal inference in terms of a novel distributional regression method. This unique combination allows for the differentiation of policy and treatment impacts across individuals, while simultaneously delivering either the most accurate measure of these distributional impacts when model assumptions are valid, or an accurate approximation when some ofthe model assumptions do not hold. This approach is thus tailored to perform distributional policy and treatment evaluation under model uncertainty. The methods that this proposal seeks to develop will be relevant to a wide range of disciplines where nonexperimental designs are commonly used in empirical research, from economics to health sciences. In these settings, the ideal conditions of a randomised experiment are generally not satisfied, i.e., individuals in thestudy are unlikely to be identical in all relevant aspects other than those directly controlled for by the model. For example, these ideal conditions often only hold approximately when evaluating the effectiveness of a new drug with observational data. My proposal will apply in this case and provide the tools needed for real-world implementation.