Drawing Directed Acyclic Graphs for Causal Inference

Causal inference with observational studies is challenging because the identification and estimation of a causal effect rests on the untestable unconfoundedness assumption. The assumption requires researchers to determine a set of observed covariates which then successfully removes all the confounding bias. Causal directed acyclic graphs (DAGs) provide (automatized) guidance for selecting an adjustment set of covariates. However, before one can select the covariates a valid causal DAG about the presumed data-generating process of the data in hand must be drawn. But published guidance on how to draw a valid causal DAG is at best still incomplete. This article introduces and discusses a stepwise procedure for drawing DAGs for causal inference.