The Bradford Hill Criteria

From association to causation

Austin Bradford Hill proposed nine viewpoints in 1965 to help judge whether an observed statistical association is likely to be causal. The criteria are strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, and analogy. They are aids to reasoned judgment rather than a checklist, and temporality is the only near-necessary condition. They remain foundational in epidemiology and observational causal reasoning.

What Is the Framework and Why Does It Matter?

The Bradford Hill criteria originate from Sir Austin Bradford Hill's 1965 presidential address to the Royal Society of Medicine, titled The Environment and Disease: Association or Causation? The central question is whether an observed statistical association between two variables is causal or merely a product of chance, bias, or confounding. In medicine and public health, where randomised controlled trials are not always feasible or ethical, this question is critical. The criteria give researchers a structured, transparent framework for moving from association to reasoned causal judgment using observational data.

The Nine Viewpoints in Order

1. Strength: A stronger association is less likely to be due to confounding or bias. 2. Consistency: The association should be replicated across different studies, populations, and methods. 3. Specificity: An association limited to a specific disease or site strengthens the causal claim. 4. Temporality: The cause must precede the effect; this is the only viewpoint Hill treated as near-necessary. 5. Biological gradient: A dose-response relationship adds credibility. 6. Plausibility: The association should be coherent with known biological mechanisms, though Hill acknowledged this depends on current knowledge. 7. Coherence: The causal interpretation should not conflict with the natural history and biology of the disease. 8. Experiment: Evidence that removing or reducing the exposure changes the outcome strongly supports causation. 9. Analogy: If a similar agent is known to cause a similar effect, the bar for accepting the new association may be lower.

How the Criteria Are Applied in Practice

Researchers apply the criteria systematically to the available body of evidence, documenting where each viewpoint supports or weakens the causal hypothesis. No single criterion is sufficient or necessary on its own — except temporality — and judgment depends on the overall weight of evidence. The classic application is the causal case against cigarette smoking and lung cancer, where strong effect size, cross-study consistency, a clear dose-response pattern, and correct temporal ordering together built a compelling causal argument. Systematic reviews and meta-analyses also use the criteria as a judgment scaffold when synthesising observational evidence.

Common Pitfalls and Limitations

The most common pitfall is treating the nine viewpoints as a checklist to be scored, which Hill explicitly warned against. A second error is assuming that an unmet criterion rules out causation; some genuine causal relationships are not specific or even particularly strong. The plausibility criterion is bounded by current biological knowledge, meaning unknown mechanisms can obscure real causation. Importantly, the criteria do not prove causation — they structure and discipline the causal argument. Analytical power increases when they are combined with formal causal frameworks such as directed acyclic graph approaches, which provide more rigorous tools for identifying and adjusting for confounding.

Key terms

Causation: A relationship in which a change in one variable produces a change in another.
Temporality: The requirement that the cause precedes the effect; the only near-necessary criterion.
Dose-Response Relationship: A pattern where increasing exposure leads to increasing effect, supporting causation.
Confounding: A third variable associated with both exposure and outcome, creating a spurious association.
Observational Study: A study design where the researcher does not intervene; data are collected under natural conditions.