Efficient nonparametric estimation of heterogeneous treatment effects in causal inference

项目总结 在几乎所有关于比较有效性的研究中，研究人员试图估计一种不同类型的药物是如何发挥作用的。 平均而言，干预会改变结果。也就是说，接受治疗的受试者的结果是否更好 比未经治疗的受试者平均吗？虽然平均治疗效果(ATE)是一个有用的总和- 根据治疗效果的不同，不同患者的治疗效果可能有所不同。这道菜就是 一种治疗效果的低维摘要，因为它总结了 使用单一量化措施进行处理，并忽略可能的效果异质性。许多人在- 研究人员试图通过估计异质性治疗来超越低维总结 效果(HTE)。最常见的估计HTE的方法依赖于简单的统计 方法：研究方法。具体地说，回归模型被广泛使用，但由于线性关系，回归模型可能会有偏差 泛函形式，尤其是当HTE是非线性的或基于复杂的 患者分组。目前，人们对开发更灵活的方法很感兴趣 用于HTE的估算。在这个项目中，我们将使用双鲁棒机器学习 (DRML)框架，为各种HTE开发改进的方法。DRML框架 是半参数理论、机器学习(ML)方法和双重健壮性的组合 估计者。DRML框架的主要优势是，它允许使用 ML估计方法，同时保留了参数模型的效率。

PROJECT SUMMARY In nearly all studies of comparative effectiveness, the investigators seek to estimate how an in- tervention changes outcomes on average. That is, are outcomes for treated subjects better on average than for untreated subjects? While the average treatment effect (ATE) is a useful sum- mary of the treatment effect, the treatment effect may vary from patient to patient. The ATE is a low-dimensional summary of a treatment effect, since it summarizes the overall effect of the treatment using a single quantitative measure and ignores possible effect heterogeneity. Many in- vestigators seek to go beyond low-dimensional summaries by estimating heterogenous treatment effects (HTEs). The most common approach to the estimation of HTEs relies on simple statistical methods. Specifically, regression models are widely used but may be biased due to the linear functional form especially where HTEs that are nonlinear or based on complex combinations of patient subgroups. Currently, there is considerable interest in developing more flexible methods for the estimation of the HTEs. In this project, we will use the the doubly robust machine learning (DRML) framework to develop improved methods for a variety of HTEs. The DRML framework is a combination of semiparametric theory, machine learning (ML) methods, and doubly robust estimators. The key advantage of the DRML framework is that it allows one to reduce bias using ML estimation methods, while retaining the efficiency of parametric models.