Data-adaptive causal inference methods for effect modification

“因果推论”是对因果关系的研究。例如，我们可能有兴趣了解一种治疗是否对减轻一种疾病的症状有影响。从统计学上讲，我们可以估计治疗对一组特定个体的平均影响。但可能是有些人比其他人更容易受到治疗的影响。因此，我们可能也有兴趣评估群体中个体的哪些特征改变了个体的影响。这些知识可以帮助我们确定哪些人将从治疗中受益最多，哪些人不会获得好处，甚至可能受到治疗的伤害，从而指导决策。我的研究计划专注于开发统计方法来识别个人特征，以帮助预测个人的影响。我使用的数据来自研究，在这些研究中，治疗不是随机分配给个人的(称为“观察性研究”)。这导致了一个重大的挑战--个人“选择”的治疗方法与他们的一般状态之间往往存在关系。因此，我把重点放在能够强有力地纠正这些关系的统计方法上，以便估计影响。我研究的另一个重点是开发能够“自动”发现改变个体效果的特征的方法。从本质上讲，这包括让算法处理数据，并确定哪些特征与不同大小的效果相关。我的具体贡献涉及为不同数据结构和类型开发方法。例如，当个体随着时间的推移改变他们的治疗方法时，或者当特征可能以复杂的方式影响效果时，特定的挑战就会出现。因为它们通常对任何个体影响的调查都有用，所以我开发的方法可以用于学术界、政府和工业界的健康、经济、政策和工程研究。

"Causal inference" is the study of cause and effect. For instance, we may be interested in knowing whether a treatment has an impact on lessening the symptoms of a disease. Statistically speaking, we can estimate the impact of treatments on average in a given group of individuals. But it may be that some individuals are more greatly impacted by a treatment than others. Therefore, we may also be interested in evaluating which characteristics of the individuals in the group modify the effects for the individual. This knowledge can guide decision making by helping us identify which individuals will most benefit from a treatment, and which individuals will not obtain a benefit and may even be harmed by a treatment. My research program is focused on the development of statistical methods for identifying individual characteristics that can help predict effects for the individual. The data that I use come from studies where treatments were not randomly assigned to individuals (called "observational studies"). This leads to a major challenge - there are often relationships between treatments that individuals "choose" and their general state. Therefore, I focus on statistical approaches that can robustly correct for these relationships in order to estimate effects. Another focus in my research is the development of methods that can "automatically" discover the characteristics that modify effects in individuals. Essentially, this involves letting an algorithm crunch the data and determine which characteristics are related to different sized effects. My specific contributions involve the development of methods for different data structures and types. For example, specific challenges arise when individuals change their treatments over time or when characteristics can influence effects in complex ways. Because they are generally useful for any investigation of individual effects, the methods I develop can be used in health, economics, policy and engineering research across academia, government, and industry.