Directed Acyclic Graph (DAG) Framework for Causal Inference

What it is

— The directed acyclic graph (DAG) framework is a non-parametric language for stating, sharing, and auditing the causal assumptions behind a study. A DAG is a graph whose nodes are variables (exposure, outcome, measured and unmeasured common causes, mediators, selection indicators) and whose directed edges encode assumed cause→effect relationships; "acyclic" means no variable can cause itself through a feedback loop, and arrows respect time order. The framework was articulated for epidemiology by Greenland, Pearl, and Robins (1999), building on Pearl's structural-causal-model theory. It is not a reporting checklist maintained by EQUATOR, Cochrane, ISPOR, or a regulator — there is no single steward; the methodological community maintains it through textbooks (Hernán & Robins, Causal Inference: What If) and the DAGitty software/web tool (Textor et al.), which mechanizes the graph-theoretic algorithms. Its purpose is narrow and powerful: make the identification step of a causal analysis explicit and falsifiable-by-debate, so that the chosen covariate-adjustment set is a derived consequence of stated assumptions rather than an opaque modeling choice.

When to use

— Draw a DAG whenever the estimand is causal (comparative effectiveness/safety, treatment effects, policy effects) and the design is non-interventional or a target-trial emulation in claims, EHR, registry, or linked data — i.e., whenever confounding control is the central threat to validity. The DAG belongs at the protocol/design stage, before data are analyzed, and the resulting adjustment set should be pre-specified. It is the right tool when (a) you must justify which covariates to adjust for and, equally, which to leave alone; (b) reviewers (FDA/EMA, an HTA committee, a journal) will ask why your model is sufficient to remove confounding; (c) you need to expose unmeasured confounding honestly by drawing U-nodes; or (d) you must defend against subtle structural biases (collider stratification, M-bias, overadjustment, selection bias from informative censoring or differential enrollment). Use it alongside, not instead of, the reporting and protocol guidelines that actually govern the manuscript (STROBE/RECORD-PE, HARPER, STaRT-RWE) — those tools increasingly require a DAG or equivalent causal diagram as a design artifact.

What it requires

— Proper use of the framework — as opposed to drawing a decorative bubble chart — demands: (1) Encode every common cause of exposure and outcome, including variables you cannot measure; unmeasured confounders are drawn explicitly as U-nodes — omitting them because they are unmeasured defeats the entire purpose. (2) Respect time-ordering: no arrow points backward in time; baseline covariates precede exposure, post-baseline variables are flagged because adjusting for them can open bias paths. (3) Subject-matter justification for every edge present and every edge absent — a missing arrow is the strong assumption (no direct effect), and it must be defensible. (4) Apply the back-door criterion (mechanized by DAGitty) to derive a minimal sufficient adjustment set that blocks all confounding paths without opening new ones. (5) Classify each node as confounder, mediator, or collider relative to the exposure–outcome contrast, because the correct action differs: adjust confounders, do not adjust mediators when the estimand is the total effect, never condition on colliders or their descendants. (6) Pre-specify and version the DAG as a study artifact (DAGitty exports a machine-readable model and the testable implications), so the adjustment set is auditable and the assumptions are stated before results are seen.

When NOT to use — limitations and common misapplications

— A DAG encodes assumptions; it does not certify they are correct, and a completed graph never makes an observational study causal. Concrete failure modes: - DAG-as-theater: drawing the graph after choosing an adjustment set to retro-justify it. The DAG must precede and constrain the analysis, not decorate it. - Treating the DAG as data-testable: the core structural assumptions (which arrows exist) are not falsifiable from observed associations. DAGitty's testable implications check conditional independencies implied by the graph; they cannot confirm the graph is right. - Conditioning on a collider (or its descendant, or a selection variable) — this induces association where none existed (M-bias, selection bias), often worse than the confounding it was meant to fix. - Adjusting for a mediator when the estimand is the total effect — overadjustment that biases the estimate toward the null and can introduce collider bias at the mediator. - Omitting unmeasured confounders from the graph because they are inconvenient; the U-nodes are precisely what communicate residual confounding to a reviewer. - Feedback loops / time-varying treatment-confounder feedback: a standard DAG cannot represent cycles. For time-varying exposures with confounders affected by prior treatment, use time-expanded DAGs and g-methods, or single-world intervention graphs (SWIGs); a single cross-sectional DAG will mislead. - Substituting the DAG for subject-matter expertise or for the reporting checklist the venue actually requires (using a DAG where RECORD-PE/HARPER content is mandated, or vice versa).

How it maps to this catalog

— The DAG framework sits upstream of nearly every causal-inference concept here and tells the analyst what to do; the concepts tell them how: - target-trial-emulation — the emulated trial's eligibility, assignment, and time-zero structure is the design a DAG makes explicit; draw the DAG to justify the emulation's adjustment set. - active-comparator-new-user — the design choice that removes confounding by indication and immortal time by structure; the DAG shows why the active comparator blocks the back-door path that a non-user comparator leaves open. - high-dimensional-propensity-score-hdps-rwe and propensity-score-methods-psm-iptw — implement adjustment for the confounder set the DAG identifies; hdPS operationalizes proxies for the U-nodes you drew but cannot measure directly. - estimands-ate-att-intercurrent-events-rwe and estimand-analysis-traceability-rwe — the estimand (total vs. direct effect) determines whether a node is a mediator to leave alone or a confounder to adjust; the DAG enforces that distinction. - diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe — measurement-error and misclassification nodes belong on the DAG; phenotype validation quantifies the arrows into and out of the measured-outcome node. - attrition-and-loss-to-follow-up-rwe, database-feasibility-attrition-funnel-rwe, and selection-bias-sensitivity-analysis-rwe — censoring and enrollment are selection nodes; conditioning on them (or on their causes) is the collider/selection-bias failure mode the DAG is designed to catch. - immortal-time-bias-handling — a time-ordering violation the DAG exposes when follow-up starts before the exposure decision. - empirical-calibration-negative-controls-rwe, negative-control-outcomes-rwe, negative-control-exposures-rwe, and e-value-sensitivity-analysis — the quantitative bias analyses that probe the residual-confounding (U-node) and unblocked-path assumptions the DAG could not eliminate. - claims-analysis — the applied substrate: in claims/EHR/registry RWE, draw the DAG before extracting any covariate, derive the minimal sufficient adjustment set, then build only those baseline covariates (measured up to time zero) into the propensity model. Applied note: in administrative data the most consequential nodes are usually unmeasured (disease severity, frailty, lab values, over-the-counter use) — draw them as U-nodes, document which measured proxies stand in for them, and reserve a negative-control outcome and an E-value to bound the confounding the graph admits it cannot close.