Directed Acyclic Graphs for Causal Modeling

Core Idea

A directed acyclic graph (DAG) is a visual representation of causal assumptions about the relationships among variables. DAGs help identify minimal sufficient sets of confounders to adjust for to block backdoor paths (non-causal paths from exposure to outcome). DAGs clarify whether a variable is a confounder, mediator, or collider, preventing unnecessary or harmful adjustment.

Explainer

You already understand confounding intuitively: a third variable that is associated with both the exposure and the outcome can make a non-causal association look causal (or vice versa). The trouble is that deciding what to adjust for in a study — which variables to include in a regression, which to stratify on — has historically been treated as an art guided by subject-matter intuition. Directed acyclic graphs (DAGs) make the causal assumptions explicit and then let formal rules determine the correct adjustment strategy.

A DAG is a graph where nodes represent variables and directed arrows (edges) represent direct causal effects. The "acyclic" constraint means no variable can be its own ancestor — there are no feedback loops, which forces you to think of the causal structure as unfolding over time. When you draw a DAG, you are not describing statistical associations; you are committing to a causal story about the world. The power is that given that story, an algorithm can tell you exactly which variables to condition on to estimate a causal effect without bias.

The three key variable types in a DAG define the logic. A confounder creates a non-causal path between exposure and outcome — it is a common cause of both. You need to block this path, usually by conditioning on the confounder. A mediator lies on the causal path from exposure to outcome (exposure → mediator → outcome). Adjusting for a mediator blocks the very effect you are trying to estimate — so you should *not* adjust for it when you want the total effect. A collider is a variable caused by two other variables (exposure → collider ← outcome). Colliders are the most counterintuitive: you should never condition on a collider, because doing so opens a spurious association between its causes, introducing bias where there was none. This is the "collider bias" or "selection bias" problem that has generated considerable rethinking of observational study design.

The backdoor criterion formalizes when adjustment is sufficient. A set of variables S satisfies the backdoor criterion if (1) no variable in S is a descendant of the exposure and (2) S blocks every "backdoor path" — every non-causal path from exposure to outcome that starts with an arrow pointing *into* the exposure (indicating a common cause). If you can find such a set S, adjusting for S gives you the causal effect. The practical implication is that you can often find *multiple* sufficient adjustment sets, and the DAG helps you choose the smallest or most easily measured one. DAGs do not tell you whether your causal assumptions are correct — that requires domain knowledge and study design — but they make those assumptions transparent and testable in principle, which is a major advance over the implicit and inconsistent practice of "just control for everything."