Bradford Hill Criteria for Causation
What you'll learn: How to systematically evaluate whether a correlation you've found might represent a true cause-and-effect relationship.
Finding a correlation is exciting—but how do you know if X actually causes Y, or if it's just coincidence, confounding, or reverse causality? British epidemiologist Sir Austin Bradford Hill developed nine criteria to help answer this question. Think of them as a checklist that strengthens (or weakens) your case for causation.
The Nine Criteria
-
Strength: Strong correlations are more likely causal. If smoking increases lung cancer risk 20-fold, that's harder to explain away than a weak association.
-
Consistency: Has the relationship been found repeatedly, across different studies, populations, and settings? Replication matters.
-
Specificity: Does the exposure lead to a very specific outcome (rather than many unrelated effects)?
-
Temporality: The cause must come before the effect—this is the only non-negotiable criterion. You can't reverse time.
-
Biological Gradient: Is there a dose-response relationship? More exposure → stronger effect suggests causation.
-
Plausibility: Does a reasonable mechanism explain how X could cause Y, given current scientific knowledge?
-
Coherence: Does the causal interpretation align with existing theory and evidence?
-
Experiment: Can you manipulate the cause and observe the effect change? Experimental evidence is powerful.
-
Analogy: Are there similar cause-effect relationships already established (e.g., one drug causes birth defects, so a chemically similar drug might too)?
How to Apply Them
No single criterion proves causation—they work together like pieces of evidence in a trial. The more criteria satisfied, the stronger your causal case. Temporality is essential; others add weight.
Key Takeaway: Bradford Hill criteria provide a systematic framework for moving beyond "correlation ≠ causation" to actually assessing whether causation is likely, by examining strength, consistency, temporality, and six other dimensions of evidence.