Haber Noah A, Wood Mollie E, Wieten Sarah, Breskin Alexander
Meta Research Innovation Center at Stanford University, Palo Alto, CA.
Department of Epidemiology, Harvard TH Chan School of Public Health, Boston MA.
Ann Epidemiol. 2022 Apr;68:64-71. doi: 10.1016/j.annepidem.2022.01.001. Epub 2022 Feb 3.
Directed acyclic graphs (DAGs) are frequently used in epidemiology as a method to encode causal inference assumptions. We propose the DAGWOOD framework to bring many of those encoded assumptions to the forefront. DAGWOOD combines a root DAG (the DAG in the proposed analysis) and a set of branch DAGs (alternative hidden assumptions to the root DAG). All branch DAGs share a common ruleset, and must 1) change the root DAG, 2) be a valid DAG, and either 3a) change the minimally sufficient adjustment set or 3b) change the number of frontdoor paths. Branch DAGs comprise a list of assumptions which must be justified as negligible. We define two types of branch DAGs: exclusion branch DAGs add a single- or bidirectional pathway between two nodes in the root DAG (e.g., direct pathways and colliders), while misdirection branch DAGs represent alternative pathways that could be drawn between objects (e.g., creating a collider by reversing the direction of causation for a controlled confounder). The DAGWOOD framework 1) organizes causal model assumptions, 2) reinforces best DAG practices, 3) provides a framework for evaluation of causal models, and 4) can be used for generating causal models.
有向无环图（DAG）在流行病学中经常被用作一种编码因果推断假设的方法。我们提出了DAGWOOD框架，将许多这些编码假设置于突出位置。DAGWOOD结合了一个根DAG（所提出分析中的DAG）和一组分支DAG（根DAG的替代隐藏假设）。所有分支DAG都共享一个通用规则集，并且必须1）改变根DAG，2）是一个有效的DAG，以及要么3a）改变最小充分调整集，要么3b）改变前门路径的数量。分支DAG由一系列必须被证明可忽略不计的假设组成。我们定义了两种类型的分支DAG：排除分支DAG在根DAG中的两个节点之间添加单向或双向路径（例如，直接路径和对撞机），而错误导向分支DAG表示可以在对象之间绘制的替代路径（例如，通过反转受控混杂因素的因果方向来创建对撞机）。DAGWOOD框架1）组织因果模型假设，2）强化最佳DAG实践，3）提供因果模型评估框架，以及4）可用于生成因果模型。
