How to Think Causally in Clinical Research — The Counterfactual + DAG Blueprint
- Mayta
- 2 days ago
- 3 min read
“Causation isn’t just about what happened—it’s about what would have happened instead.”
🎯 The Clinical Dilemma
Your patient, Mr. G, age 50, has stage IV lung cancer. He receives an herbal treatment and survives.
Question: Did the herb save him?
Or would he have survived anyway?
That’s the counterfactual question. And it’s the gold standard for causal inference in medicine.
🔄 1. Association ≠ Causation
Let’s separate noisy thinking from causal clarity.
Concept | Meaning |
Association | Knowing X helps predict Y |
Causation | Changing X will change Y |
Example:
“Patients who take Vitamin D have better COVID outcomes.”That’s association.
“Vitamin D improves COVID outcomes.”That’s a causal claim—and it needs counterfactual proof.
🧬 2. The Counterfactual Logic
Key Concept:
A causal effect = Observed outcome minus Counterfactual outcome
But we never observe both for the same person. You only see one timeline: treated or untreated.
So, how do we estimate the missing outcome?
🧪 3. Approximating the Counterfactual: Your Toolbox
We create “what-if” worlds using:
Tool | Function |
RCTs | Create two comparable worlds via randomization |
Matching | Pair treated patients with similar untreated ones |
Stratification | Compare within subgroups (e.g., ECOG = 2) |
Regression | Adjust for confounders statistically |
Propensity Scores | Balance groups based on treatment likelihood |
DAGs | Show where bias might sneak in (more below) |
🔁 4. The DAG: Your Bias-Detecting Blueprint
What is a DAG?
A Directed Acyclic Graph is a visual map of your assumptions:
Arrows = causal directions
Nodes = variables
Helps decide what to adjust for
✍️ Example DAG: Does the Herbal Drug Help?
Age → Herbal Treatment → Survival ↘-------------------------↗
Age affects both treatment and survival → confounder
Must adjust to isolate the drug’s effect.
🛑 Don't Adjust for These:
Role | Rule | Example |
Confounder | Adjust ✅ | Age, ECOG |
Mediator | Don’t adjust if estimating total effect ❌ | Side effects |
Collider | Never adjust ❌ | ER admission caused by both severity and treatment |
🔍 Secret Insight: Adjusting for colliders introduces false associations.DAGs protect you from this silent sabotage.
🎲 5. Mr. G’s Real-World Counterfactual
You want to answer:
Would Mr. G have survived without the herbal treatment?
Since we can’t observe that:
Define precise inclusion/exclusion: age, cancer stage, ECOG.
Find matched patients who didn’t receive the drug.
Adjust for known confounders (e.g., smoking).
If survival differs → credible evidence for causality.
🔄 6. Time Matters: Temporality
Causal claims demand:
Cause precedes effect
No reverse logic
If treatment happens after improvement, it can’t be the cause.
⚔️ 7. Confounding, Indication Bias & Comparability
Real-world mess:
Doctors prescribe based on prognosis = confounding by indication
Sickest avoid treatment = confounding by contraindication
Self-selection = selection bias
How to beat this:
Design with comparability in mind
Use DAGs to plan adjustments
Use propensity scores, matching, or inverse probability weighting
📘 Occurrence Equation for Causal Studies
🧠 Summary Table
Element | Function | Caution |
Counterfactual | The unseen alternative | Estimate, don’t assume |
DAG | Bias map | Build before analyzing |
Confounder | Adjust ✅ | Blocks bias |
Mediator | Adjust ❌ if estimating the total effect | |
Collider | Never adjust ❌ | Causes a spurious association |
RCT | Gold standard | Randomizes counterfactual |
✅ Final Takeaways
Causal inference = imagining alternate realities and estimating them scientifically.
Counterfactual thinking is not optional—it’s the foundation.
DAGs are essential for identifying what’s safe (and unsafe) to adjust.
Comparability is key, and it can be achieved through design or statistics.
Always clarify the variable roles: confounder, mediator, and collider.
Comments