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

Mayta
May 7
2 min read

🎯 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

Path Type	Example	Open/Closed?	Adjust?	Why?
Causal Path	X → M → Y	✅ Open	❌	Transmits treatment effect
Backdoor Confounder	X ← C → Y	✅ Open	✅	Opens false link unless blocked
Collider Path	X → Z ← Y	❌ Closed	❌	Adjusting opens bias

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

Path	Type	Adjust?	Why?
EarlyFluids → Hypotension	Direct causal	❌	This is your effect of interest
EarlyFluids → Volume → Hypo	Mediated causal	❌	Leave open for total effect
EarlyFluids ← Severity → Hypo	Confounding	✅	Must close this bias path
EarlyFluids → LOS ← Hypo	Collider	❌	Closed by default; adjusting creates false path

✅ 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.