DAGs: Directed Acyclic Graphs for Drawing Assumptions and Guiding Causal Inference

In pediatric hospital medicine, we often rely on observational data to conduct hospital-based research studies. Observational studies may establish a correlation between an exposure and outcome, but for this correlation to be considered causation, researchers must incorporate causal inference methods to control for confounding and reduce study bias. This methodology article reviews the concept of a directed acyclic graph (DAG), which is a causal inference tool that visually depicts the assumed relationships among study variables to guide study design and analytic plans. First, we outline the introductory steps to guide researchers toward building a basic DAG for their research question. Next, we explore the different types of study variables, including mediators, effect modifiers, confounders, and colliders, and clarify the causal assumptions linking them to the exposure and outcome. Finally, we share recommendations for how investigators can incorporate findings from their DAG into their study design and analysis plan. Although DAGs are only as strong as their creator's understanding of the system of interest and its underlying context, we hope to equip readers with the tools necessary to build powerful DAGs that communicate their study assumptions and guide causal inference in research.