I basically agree with this: ruling out unobserved variables is an unusual way to use causal graphical models.
Also, taking the set of variables that are allowed to be in the graph to be the set of variables defined on a given sample space makes the notion of “intervention” more difficult to parse (what happens to F:=(X,Y) after you intervene on X?), though it might be possible with cyclic causal relationships.
So basically, “causal variables” in acyclic graphical models are neither a subset nor a superset of observed random variables.
I basically agree with this: ruling out unobserved variables is an unusual way to use causal graphical models.
Also, taking the set of variables that are allowed to be in the graph to be the set of variables defined on a given sample space makes the notion of “intervention” more difficult to parse (what happens to F:=(X,Y) after you intervene on X?), though it might be possible with cyclic causal relationships.
So basically, “causal variables” in acyclic graphical models are neither a subset nor a superset of observed random variables.