Causal Inference with DAGs in Pediatric Clinical Research

🎯 Learning Objectives

By the end of this guide, you’ll be able to:

✅ Classify variables into confounders, mediators, colliders, and effect modifiers ✅ Detect bias paths using DAGs ✅ Use DAGitty to build and test your DAG ✅ Apply causal logic to real-world clinical studies

---

🩺 Clinical Question (New Example)

Does initiating antipyretic treatment within 30 minutes reduce the risk of seizure recurrence in children presenting with febrile seizure?

DEPTh Type: Therapeutic (intervention → outcome)

---

📌 1. DAG = Directed Acyclic Graph

What’s a DAG?

• Directed: Arrows show assumed causality

• Acyclic: No loops allowed—no feedback

• Graph: A visual model of assumed causal structure

🧠 DAGs map your thinking, not just your variables.

---

🧠 2. Classifying Variable Roles

Role	Meaning	Adjust?	Example
Exposure (X)	The treatment or risk factor	—	Antipyretic within 30 minutes
Outcome (Y)	The endpoint	—	Seizure recurrence within 6 hours
Confounder	Common cause of X and Y	✅	Initial fever severity
Mediator	Lies on causal path X → Y	❌*	Rate of temperature decline
Collider	Caused by both X and Y	❌	ED length of stay
Effect Modifier	Alters strength of X → Y	➕	Underlying epilepsy
Precursor	Precedes X or Y, not necessarily a cause	⚠️ Care	Prior hospitalization

*Adjust for mediator only if estimating direct effect.

---

🧪 3. Drawing the DAG

dagitty

dag {

FeverSeverity -> EarlyAntipyretic

FeverSeverity -> SeizureRecurrence

EarlyAntipyretic -> TempDecline

TempDecline -> SeizureRecurrence

Epilepsy -> SeizureRecurrence

Epilepsy -> EarlyAntipyretic

ED_Crowding -> EarlyAntipyretic

}

✅ Adjust for: FeverSeverity, Epilepsy

❌ Don’t adjust for: TempDecline (mediator), ED_Crowding (precursor), LengthOfStay (collider)

🔗 Try it: Open DAG in DAGitty

---

🔁 4. Understanding Path Types

Path Type	Example	Open?	Adjust?	Why?
Causal	X → Y	✅	❌	This is the path you want to estimate
Backdoor (confounder)	X ← C → Y	✅	✅	Biases effect estimate
Mediated causal path	X → M → Y	✅	❌*	Keep open if estimating total effect
Collider path	X → Z ← Y	❌	❌	Adjusting opens a spurious path

🧠 Collider adjustment = new bias creation, not removal.

---

🧱 5. Backdoor Control

🎯 Backdoor paths = non-causal routes that simulate effects.

To close them:

• Identify confounders on these paths

• Adjust using regression, matching, and stratification

• Validate with DAGitty’s “Adjustment Sets”

In our case:

• EarlyAntipyretic ← FeverSeverity → SeizureRecurrence

• EarlyAntipyretic ← Epilepsy → SeizureRecurrence ✅ Adjusting for FeverSeverity and Epilepsy blocks both paths.
  

---

💥 6. Collider Trap in Pediatric Context

New Scenario:

Let’s add:

dagitty

LengthOfStay -> EarlyAntipyretic

LengthOfStay -> SeizureRecurrence But wait — that's wrong. LengthOfStay is affected by both:

dagitty

dag {

EarlyAntipyretic -> LengthOfStay

SeizureRecurrence -> LengthOfStay

}

Here, LengthOfStay is a collider. ❌ Adjusting for it opens a backdoor: EarlyAntipyretic → LOS ← SeizureRecurrence

---

🎭 7. Effect Modification vs. Interaction

Concept	What It Means	Example
Effect Modifier	X → Y effect varies by Z	Antipyretic works better in children without epilepsy
Interaction	X1 and X2 together influence Y differently	High fever + fast antipyretic = exponential benefit

Model it with: X + Z + X*Z Visualize it by: Stratified DAGs or subgroup KM curves

---

🧩 8. Clinical Reasoning First: Then Model

Use this workflow:

Clinical Question → Variable Roles → DAG → DAGitty Adjustment → Regression Model

❌ Don’t start with a dataset ✅ Start with causal structure

---

📋 9. DAG Construction Checklist

1. Define exposure (X) and outcome (Y)

2. Identify variable roles:

   • Confounders

   • Mediators

   • Colliders

   • Precursors

3. Draw DAG (DAGitty or pen + paper)

4. Run Adjustment Sets in DAGitty

5. Build a regression model with justified covariates
   

---

✅ Final Takeaways

• DAGs ≠ illustrations — they are logic blueprints for your research.

• Adjust only confounders; avoid colliders unless modeling effect mediation.

• DAGitty helps verify if your adjustment set is both valid and minimal.

• Model causal paths, not just correlations.

• DEPTh logic ensures alignment with clinical decision-making.