Common DAG Patterns and Pitfalls
You'll learn to recognize frequently occurring DAG structures and typical mistakes that lead to biased causal estimates.
The Three Fundamental Patterns
After building DAGs and learning about confounders, mediators, and colliders, you need to recognize these patterns in the wild. Three structures appear constantly:
Chain (Mediation)
A → M → Y
Here, M is a mediator. A affects Y through M. Pitfall: Conditioning on M blocks the path from A to Y, hiding the causal effect you want to measure. Only control for M if you specifically want to isolate the direct effect of A.
Fork (Confounding)
A ← C → Y
C is a confounder creating a backdoor path. Must control for C to isolate A's effect on Y. This is the classic "third variable problem." Miss this, and you'll mistake correlation for causation.
Inverted Fork (Collision)
A → C ← Y
C is a collider. Pitfall: Conditioning on C opens a path between A and Y where none existed, creating spurious correlation (selection bias). Never control for colliders unless they're also on a causal path you need to block.
Common Mistakes
- Over-controlling: Adding every available variable to your model without checking the DAG. You might condition on mediators or colliders.
- Under-controlling: Ignoring confounders because they seem unimportant or weren't measured.
- Mistaking correlation patterns: Seeing A and Y correlate and assuming A→Y, when really both are caused by unmeasured C.
Think of it like plumbing: controlling for a variable either blocks or opens "pipes" of association. You need to know which pipes carry real causal flow versus confounding "leaks."
Key Takeaway: Chains, forks, and inverted forks are the building blocks of causal inference—master recognizing them to avoid conditioning on the wrong variables and introducing bias.