DAGs vs. Causal Diagrams: What to Use for Clinical Causal Inference
- Mayta
- May 7
- 2 min read
🎯 TL;DR Summary Table
Feature | DAG (Directed Acyclic Graph) | Causal Diagram (General) |
✅ Arrows only | Yes – all edges are directional | Often, but may include non-directional paths |
🔁 No cycles | Yes – strictly acyclic | Can have feedback loops or dynamic components |
🧠 Logic model | Purely structural assumptions | Broader – includes feedback, timing, processes |
📐 Graph theory-based | Yes – rooted in Pearl’s structural causal model | Sometimes informal or illustrative |
🔬 Research use | Adjustment, backdoor paths, DAGitty analysis | Teaching, conceptual modeling, sometimes DAG |
🎯 Objective | Identify bias paths and valid adjustment sets | Often to explain a general system |
🧪 1. What is a DAG?
✅ Definition
A Directed Acyclic Graph (DAG) is a precise, rule-based model used in causal inference to:
Specify assumptions
Reveal bias paths
Calculate adjustment sets
DAGs only use arrows, and never have feedback loops (acyclic). They are mathematical tools with causal logic and can be tested in tools like DAGitty.
🔬 Example
dagittydag {
ShockSeverity -> EarlyFluids
EarlyFluids -> Hypotension
}
Used to test:
Backdoor paths
Minimal sufficient adjustment sets
Mediator vs confounder logic
🧩 2. What is a Causal Diagram?
🧠 Definition
A causal diagram is a broader concept that includes any visual causal system model.
May include:
Loops (feedback, delay)
Non-directional links
Boxes for stages or processes
Time-sequence arrows
Often used in complex systems modeling, conceptual frameworks, or clinical teaching.
📌 Example (non-DAG causal diagram):
Fever → Cytokine Storm → Multi-organ Failure↑ ↓
Immunosuppression ←—┘
This feedback loop violates DAG rules but is valid in systems thinking.
🔍 3. Why This Distinction Matters in CECS
Use Case | Use a DAG | Use a Causal Diagram |
Designing a regression model | ✅ To identify confounding paths | ❌ Too vague or circular |
Teaching system dynamics | ❌ Too rigid | ✅ Ideal for concept mapping |
Planning a target trial | ✅ DAGs emulate randomization logic | ❌ Can't compute adjustment sets |
Modeling feedback in physiology | ❌ DAGs can't handle cycles | ✅ Systems or influence diagrams |
🔍 Secret Insight: In causal inference, precision trumps beauty. DAGs look simple, but what matters is what they let you prove.
🔄 4. When Causal Diagrams Fail as DAGs
Common violations:
Violation | Example | Fix to Make It a DAG |
Feedback loop | A → B → A | Remove loop or model time-sequenced DAGs |
Unlabeled arrows | A — B (unclear direction) | Specify arrow direction |
Ambiguous timing | Diagnosis ↔ Treatment | Break into nodes: "Before diagnosis", etc. |
Unstructured effects | Vague arrows “influences” | Use structured paths: X → Y or X ← Z → Y |
✅ Final Takeaways
A DAG is a formal, strict subset of causal diagrams, optimized for inference.
Causal diagrams are conceptual tools; DAGs are analytical weapons.
For CECS study design, always build a valid DAG first—use other diagrams for teaching or exploring mechanisms.
Use DAGitty or pen-and-paper to construct DAGs with rules: directional, acyclic, causal paths only.
Comments