Directed Acyclic Graphs: How to Read a Causal Map Before You Trust the Adjustment

A directed acyclic graph, or DAG, is a simple diagram in which arrows point from causes to effects, drawn from knowledge before any data are analyzed. Its purpose is to make your assumptions about the causal structure explicit, so you can decide by clear rules which variables must be controlled to estimate an effect and which ones must be left alone. The counterintuitive payoff is that adjusting for the wrong variable, such as a common effect of two others, can manufacture bias where none existed, and a DAG is how you see that trap coming.

What a DAG actually is

A directed acyclic graph is a diagram of what you believe causes what. Each variable is a point, and an arrow runs from a cause to its effect. Directed means the arrows have a direction, and acyclic means you can never follow the arrows in a loop back to where you started, because a thing cannot cause itself.

Crucially, a DAG is drawn from knowledge and assumptions before the analysis, not fitted to the data. It is a way of writing down your theory of the causal structure so that its consequences can be checked.

Three roles a variable can play

Between a treatment and an outcome, a third variable can sit in one of three positions, and the position decides how you should handle it. A confounder is a common cause of both treatment and outcome, such as age influencing both whether someone gets a drug and whether they recover. A mediator sits on the path from treatment to outcome and carries part of the effect. A collider is a common effect, a variable that both the treatment and the outcome influence.

The same variable can be a confounder in one question and a collider in another, which is why you cannot decide how to treat it without a map.

The backdoor path and why blocking it matters

A backdoor path is a route from treatment to outcome that runs against the arrows, through a common cause. It is how a confounder creates a false association. The job of adjustment is to block these backdoor paths, and a DAG shows you exactly which variables you must control to close them.

Control too few and confounding leaks through. This is the familiar reason studies adjust for factors like age, sex, and severity of illness before comparing treatments.

The collider trap: adjusting can create bias

Here is the part that surprises people. Adjusting for a collider, a common effect of two variables, does not remove bias. It creates it. When you control for or select on a common effect, you open a path that was closed and make two otherwise unrelated causes appear associated.

A classic example is a study restricted to hospitalized patients, where selecting on admission can make two independent risk factors look linked, an artifact known as collider or selection bias. A DAG is often the only way to notice that a well intentioned adjustment is itself the source of the problem.

What a DAG cannot do for you

A DAG is only as good as the assumptions inside it. It does not prove that your arrows are correct, and two thoughtful experts can draw different graphs for the same question. What it does is make the disagreement visible and specific, so it can be argued about and tested.

It also cannot tell you the size of an effect. It tells you which variables to adjust for, and the data still have to supply the number.

How to read one in a paper

When a study includes a DAG, trace the paths from treatment to outcome and check that the variables the authors adjusted for are the confounders, not the mediators or colliders. Ask whether any adjusted variable could be a common effect rather than a common cause.

If a paper reports a long list of adjustments with no causal reasoning behind them, a DAG is what is missing, and the reader is left unable to tell whether the adjustments helped or hurt.