Causal structure · Planning

Causal Directed Acyclic Graph (DAG)

A non-parametric encoding of the assumed causal relationships among exposure, outcome, confounders, mediators, and colliders — the figure that justifies which variables to adjust for.

Download SVG Open code and full worked example

When to use it

At the design stage, before choosing covariates. The DAG makes the backdoor criterion operational: adjust for confounders that open exposure–outcome backdoor paths; do NOT adjust for mediators or colliders, which would induce bias.

How to read it

Arrows are direct causal effects; a path is 'open' unless blocked. A confounder (L) with arrows into both E and Y opens a backdoor path you must close; a mediator (M) on E→Y is part of the effect — conditioning on it removes the indirect effect or opens a collider path.

Worked example

A study of a drug (E) on an outcome (Y) where comorbidity burden (L) drives both treatment choice and the outcome, on-treatment blood pressure (M) mediates part of the effect, and an unrelated risk factor (V) affects Y. The DAG encodes which to adjust for.

Edges: L→E, L→Y (backdoor); E→M→Y (mediated); E→Y (direct); V→Y. No arithmetic — the graph structure IS the content.

Result: The backdoor path E←L→Y is closed by adjusting for L; M is left unadjusted (conditioning on it would block the indirect effect and could open a collider path); V need not be adjusted for confounding but improves precision. The minimal sufficient adjustment set is {L}.

Produced by

DAGs & the backdoor criterion Target trial emulation

Reference: Gatto NM, Wang SV, Murk W, et al. Visualizations throughout pharmacoepidemiology study planning, implementation, and reporting. Pharmacoepidemiol Drug Saf. 2022;31(11):1140-1152.