Directed Acyclic Graphs: A Reader's Guide to Causal Adjustment

A regression model does not know which variables cause which. It only knows which values move together. A directed acyclic graph, usually called a DAG, forces researchers to state the causal story that tells the model what to adjust for; the diagram cannot prove that story, but it makes hidden assumptions inspectable and prevents the common mistake of treating every available variable as a helpful control.

Read the question before the arrows#

Every DAG is built for a particular causal question. Consider a study asking whether a new treatment improves recovery. The target might be the total effect of assigning the treatment, including all pathways through adherence and follow-up care, or it might be a direct effect that excludes one pathway. Those are different estimands, so they can require different adjustment sets.

The time order also matters. Baseline severity can affect treatment choice and recovery. A complication occurring after treatment cannot cause the earlier treatment choice, even if the dataset lists both variables in adjacent columns. Start with the intervention or factor, the outcome, the time zero, and the effect the authors claim to estimate.

The basic grammar#

A DAG contains nodes connected by one-way arrows. "Directed" means each arrow points from a presumed cause toward a consequence. "Acyclic" means following arrows forward can never return to the starting node. The rule encodes time and logical order: within the causal question, a variable cannot eventually cause itself.

A path is any route linking two nodes, even if some arrows point against the direction of travel. A causal path follows arrows from the factor to the outcome. A noncausal path creates an association for another reason. Adjustment aims to block relevant noncausal paths while preserving the causal pathway the study wants to estimate.

The arrows are not coefficients. They do not state effect size, functional form, or certainty. Missing arrows are also claims: omitting an arrow means the researcher assumes no direct causal relationship relevant to the analysis. That is why a readable DAG should come with an explanation of how its structure was chosen, so that you can argue with it.

Forks: the confounding pattern#

A fork occurs when one variable causes both the factor being studied and the outcome. Disease severity may influence which treatment a patient receives and the chance of recovery. The route treatment <- severity -> recovery is a backdoor path. It can make treatments given to sicker people appear less effective even if treatment helps.

Conditioning on severity, through restriction, stratification, matching, weighting, or regression, can block that path. The goal is not to adjust for severity because it predicts the outcome. The goal is to adjust because its causal position opens a noncausal route between treatment and outcome.

This distinction explains why automatic variable selection is not a substitute for causal reasoning. A weak predictor may be a necessary confounder. A strong predictor may be a collider that should remain untouched. Statistical association alone cannot assign the role.

Chains: the mediator question#

A chain runs factor -> mediator -> outcome. For example, a treatment may lower blood pressure, which then reduces stroke risk. Blood pressure is on a causal pathway, and if the question concerns the treatment's total effect, adjusting for post-treatment blood pressure removes some of the effect the study is trying to measure.

Sometimes researchers intentionally ask how much effect remains outside that pathway; then mediation analysis may be appropriate, but it requires assumptions beyond ordinary confounder adjustment, including careful treatment of causes shared by the mediator and outcome. Simply adding the mediator to a regression and calling the remaining coefficient a direct effect can be misleading.

Timing helps you spot the problem. Variables measured after an intervention deserve special scrutiny. They may be mediators, consequences of early outcome changes, or selection variables. A paper should name why any post-intervention adjustment belongs in the model.

Colliders: when adjustment opens bias#

A collider is caused by two variables: factor -> collider <- outcome, or by causes of those variables. The path is naturally closed. Conditioning on the collider opens it, creating an association between its causes.

Hospital-based studies offer an intuitive example. Suppose both a chronic condition and an acute infection increase the chance of admission. Among all people, those causes may be unrelated. Among admitted patients, however, a person without one reason for admission is more likely to have the other, and restricting the analysis to admitted patients conditions on a common effect and can make the two causes appear negatively related.

Selection into a study, completion of follow-up, availability of a laboratory test, and use of a referral center can all act as colliders; regression adjustment is not the only way to condition. Restricting a sample, matching on a variable, or analyzing only complete records can do the same thing.

Descendants can inherit the problem#

Conditioning on a direct child of a collider can also open a biased path. If admission affects whether a specialist test is ordered, restricting analysis to people with that test may reproduce selection bias even though "admission" is not in the model. DAGs make these longer routes visible.

Likewise, adjusting for a proxy may only partially block confounding if it measures the common cause poorly, and a box labeled socioeconomic position may hide several mechanisms that the available variable captures incompletely. A graph clarifies the intended role, but measurement quality still determines whether adjustment succeeds.

A minimum sufficient adjustment set#

The backdoor criterion provides a systematic aim: find a set of pre-intervention variables that blocks every open noncausal path from the factor to the outcome without blocking the desired causal path or opening a collider path. More than one valid set may exist.

Smaller is not always statistically superior, but "adjust for everything" is not causally safer. Extra variables can reduce precision, introduce measurement error, or open bias. Authors should state the adjustment set implied by the graph and explain departures from it. If a key confounder was not measured, the DAG does not solve that problem; it reveals it.

What a strong DAG report should show#

A strong report answers several questions for you:

1. Is the causal question and time zero explicit?
2. Were arrows based on prior knowledge rather than chosen to fit observed associations?
3. Are all plausible common causes included, even if not measured?
4. Which variables form the proposed adjustment set, and why?
5. Were mediators, colliders, and selection mechanisms kept distinct?
6. Did authors test reasonable alternative diagrams or discuss how conclusions change under them?

Unmeasured variables still belong in the diagram. Omitting them because the dataset lacks them confuses a data map with a causal map. Showing them documents residual confounding and can motivate quantitative bias analysis.

Disagreement is useful when it is explicit#

Two teams may draw different arrows for the same question. That does not make DAGs subjective decoration. It turns a vague disagreement into testable consequences, and if one graph treats a variable as a confounder and another treats it as a mediator, the resulting estimates answer different questions. Reporting both may show how much the conclusion depends on the contested assumption.

A DAG is therefore a pre-analysis contract. It tells you what the investigators believe must be true for the estimate to have its causal interpretation, and the diagram earns trust when it clarifies the argument and its vulnerabilities, not when it is visually complicated.