Week 2: Intro to DAGs Flashcards
What are three key reasons we use DAGs for?
Know whether our estimated effect is close to the truth, i.e., whether it is unbiased.
Know whether our effect of interest can even be estimated with the measured variables we have in our data.
Be transparent/explicit about how we perceive the causal relationships, i.e., our assumptions.
DAGs are an acronym for what? What do these words mean?
Directed: Arrows point from cause to effect, indicate direction of causality.
Acyclic: no variable cycles back to cause itself.
Graph: a picture showing relationships between variables.
What three variables go into a DAG?
Mediators; confounders; factors that influence selection into study.
A common cause in a DAG is what?
A confounder.
A common effect or collider in a DAG is what?
Covariate that is descendent of two other covariates.
Opening a backdoor path means what?
Confounding - a biased connection between A and Y that does not follow arrows.
Collider bias is said to occur when we do what?
Condition on common effects.
Conditioning refers to what four mechanisms? What do we do to denote it in a DAG?
Using restriction, stratification, adjustment, or matching to measure effect size within levels of conditioned variable.
Draw a box around a variable that is being conditioned on.
If we condition on a common cause, what are the implications?
Remove the non-causal association (control for confounding), close/block the backdoor path.
If we condition on a common effect (e.g, collider), what are the implications?
False association between A and Y. Not conditioning on a collider blocks the flow
of association between A and Y.
What do we control for and not control for in a DAG? Why?
Control for a variable that is a common cause – this “blocks” non-causal path.
Don’t control for a variable that is a collider on the path (i.e., with two arrows pointing to it) as this blocks the non-causal path.
What is selection bias in the context of DAGs?
Collider bias where the common effect is conditioned on via restricting to a level of that strata.
Consider a simple DAG where A causes B, A causes C, and B and C are both causes of
D. What does conditioning on D do?
Induces an association between B and C even if they are independent in the population. Conditioning on the collider D “opens” a backdoor path between B and C that allows for spurious associations to appear.
DAGs involve these four terms: parent, ancestor, child, descendant. What do these mean?
Parent: direct cause of variable.
Ancestor: direct cause (parent) or indirect cause (grandparent) of variable.
Child: direct effect of variable.
Descendant: direct effect (child) or indirect effect (grandchild) of variable.
