Causal vs Non-Causal Paths in DAGs — How to Detect Bias and Adjust Causal Paths in Clinical Research Using DAGs

🎯 Objectives

By the end, you’ll be able to:

✅ Distinguish causal from non-causal paths ✅ Diagnose if a path is open or closed ✅ Know when to adjust to block bias ✅ Simulate path logic using DAGitty

---

🩺 Fresh Clinical Setup (New Example)

Does giving intravenous fluids within 15 minutes reduce the incidence of hypotension in dengue shock patients in the ED?

• Exposure (X): Early IV fluids
• Outcome (Y): Hypotension
• Confounder (C): Shock severity
• Mediator (M): Blood volume recovery
• Collider (Z): ED length of stay
• Effect Modifier: Initial hematocrit level

---

🧬 1. Causal Path = What You Want to Measure

🔗 Causal Path Example:

EarlyFluids → BloodVolumeRecovery → Hypotension

• This path transmits the real effect of treatment.
• Keep it open if estimating the total effect.
• Adjust only if interested in direct effect (e.g., Fluids → Hypotension without mediation via volume recovery).

🧠 Total Effect = Direct + Indirect

---

⚠️ 2. Non-Causal Path = Where Bias Leaks In

These are routes that mimic the association between exposure and outcome but aren’t due to the intervention.

🛑 Non-Causal Path #1: Confounding (Backdoor)

EarlyFluids ← ShockSeverity → Hypotension

• This is a classic backdoor path
• Open by default — must be blocked by adjusting
• If ignored, it creates a false attribution of effect

✅ Action: Adjust for ShockSeverity to close this path

🧨 Non-Causal Path #2: Collider Bias

EarlyFluids → EDLengthOfStay ← Hypotension

• This path is naturally closed
• Adjusting for the collider (ED stay) opens a spurious path
• Creates a collider stratification bias

❌ Never adjust for colliders

🔍 Secret Insight: Colliders are seductive—they feel like good control variables. Don’t fall for it.

---

📊 3. Open vs Closed: The Real Mechanism

DAG Rule of Thumb:

• Open + Non-Causal = BAD (must block)
• Open + Causal = GOOD (don’t block)
• Closed + Non-Causal = GOOD (leave alone)

---

🧮 4. Visualized in DAGitty

dagitty

dag {

ShockSeverity -> EarlyFluids

ShockSeverity -> Hypotension

EarlyFluids -> BloodVolumeRecovery

BloodVolumeRecovery -> Hypotension

EarlyFluids -> EDLengthOfStay

Hypotension -> EDLengthOfStay

}

✅ Adjustment Set: ShockSeverity 🚫 Avoid adjusting: EDLengthOfStay, BloodVolumeRecovery (if total effect)

🔗 Test it in DAGitty

---

📋 5. DAG Path Review Table

---

✅ Key Takeaways

• Causal paths transmit signal — leave open unless isolating direct effect.
• Non-causal paths transmit bias — must be blocked if open.
• Confounding = adjust to close; Collider = avoid adjusting.
• DAGitty confirms whether your chosen adjustment kills bias or kills effect.