Causal DAGs make the assumptions behind an analysis explicit, so that others can see and challenge them.
They exist because the same data can support different conclusions under different causal stories.
Drawing the graph forces a decision about what causes what before any numbers are run.
Each variable is a node, and each arrow means a direct cause.
A missing arrow is a strong claim, since it says one variable does not directly affect the other.
A path is any route between 2 variables along arrows, whichever way the arrows point.
A backdoor path runs from the treatment back through a common cause to the outcome, and it carries confounding.
Adjusting for a variable on a backdoor path blocks it.
The backdoor criterion picks out a set of variables whose adjustment blocks every such path without opening new ones.
A chain, such as treatment to blood pressure to recovery, carries part of the real effect and should not be blocked when the total effect is wanted.
Variables where 2 arrows meet head to head need special care, because adjusting for them creates links.
Software can read a DAG and list which sets of variables are valid to adjust for.
Different experts may draw different graphs, and comparing results under each shows how much the conclusion depends on assumptions.
In practice, a DAG on one slide often settles a long argument about which variables to control for.
The usual misconception is that a DAG is learned from data, when it mostly comes from knowledge of the subject.