Robustness - A New Methodology for Causal Inference

经验社会科学家感兴趣的是检验从理论中得出的关于现实世界的假设，并为政策和决策提供证据基础。要做到这一点，他们需要从事所谓的因果推理，这包括确定因果关系(而不仅仅是两个事件之间的相关性)，理解驱动因果关系的机制，并根据对有限数量的案例的观察或分析，对更广泛的人群得出关于因果关系普遍存在的结论。社会科学家有大量的策略或技术来进行因果推理：他们可以对单个案例或几个案例进行假设的反事实分析。他们可以随着时间的推移分析选定的个别病例，或者横向或随时间比较几个选定的病例。他们可以使用基于观测数据的回归分析来分析大量的案例。最后，他们可以研究治疗分配是随机的选定病例的样本，或者他们可以进行随机样本和治疗的实验。在这些技术中，基于观察数据的回归分析最受欢迎，主要有两个原因。首先，如果基于从所有相关案例中随机抽取的样本，则回归分析产生的结果可以在所研究的样本之外进行推广。其次，回归分析的用途非常广泛，可以根据人口普查、调查或行政记录，在任何存在观测数据的地方部署回归分析。然而，回归分析只有在一个人的估计模型被正确指定的情况下才能产生可靠的推断。然而，目前进行的回归分析并不能提供可靠的推断，因为分析师不能知道他们的模型被正确指定的程度。我们提议的新方法将允许分析师得出更可靠的推论，方法是使他们能够测试他们的推论在面对模型规范的合理变化时是成立的，还是依赖于某些模型规范选择。对模型规范的合理更改具有健壮性的推断可以获得更大的置信度，也就是说，这样的推断更加可靠。总而言之，该项目的目标是将基于观测数据的回归分析--经验社会科学的主要技术--建立在更坚实的方法论基础上。

Empirical social scientists are interested in testing hypotheses about the real world derived from theories and to provide an evidence base for policy and decision-making. To do so they need to engage in what is known as causal inference, which consists of identifying a causal effect (rather than a mere correlation between two events), understanding the mechanism driving the causal effect, and drawing conclusions over the general existence of a causal relation based on observation or analysis of a limited number of cases to a wider population.Social scientists have a large number of strategies or techniques for making causal inferences at their disposal: They can undertake a hypothetical counterfactual analysis for a single case or a few cases. They can analyze a selected individual case over time or compare a few selected cases cross-sectionally or over time. They can use regression analysis based on observational data to analyze a large number of cases. Finally, they can study a sample of selected cases in which treatment assignment is randomized, or they can conduct experiments that randomize both sample and treatment.Of these techniques, regression analysis based on observational data is the most popular because of essentially two reasons. First, if based on a randomly drawn sample from the overall set of relevant cases, then regression analysis produces results that can be generalized beyond the sample studied. Second, regression analysis is extremely versatile and can be deployed wherever observational data exist based on censuses, surveys or administrative records. However, regression analysis only results in reliable inferences if one's estimation model is correctly specified.Regression analysis as currently undertaken does not provide reliable inferences, however, because analysts cannot know the extent to which their models are correctly specified. Our proposed new methodology will allow analysts to come to more reliable inferences by enabling them to test whether their inferences uphold in the face of plausible changes to model specification or are dependent on certain model specification choices. Inferences that are robust to plausible changes to model specification can command a much larger confidence, that is, such inferences are much more reliable. In sum, then, the aim of the project is to place regression analysis based on observational data - the workhorse technique of empirical social science - on a sounder methodological foundation.