Austin Bradford Hill's nine viewpoints separated causal claims from lucky correlation

By 1965 the evidence linking smoking to lung cancer was strong enough that opponents had switched tactics: few disputed it; they disputed what it meant. Correlation is not causation was unanswerable by experiment — you cannot randomize people into decades of smoking. Occupational medicine faced the same trap in miniature: workers exposed to dusts and fumes, claims contested by employers and insurers. The Royal Society of Medicine's new Section on Occupational Medicine asked Austin Bradford Hill, the statistician behind the 1948 streptomycin trial, to give its first President's Address.

Hill's answer was deliberately modest. He offered nine 'viewpoints' — strength, consistency, specificity, temporality, biological gradient, plausibility, coherence, experiment, analogy — and insisted none was necessary and the list carried no rules. Before any of them, he said, ask whether the association is chance or confounding; the viewpoints matter only once it is real. The genius is in what the list does rather than what it contains: it turns an argument with no referee into one with shared terms, so a critic must attack specific viewpoints rather than wave the problem away.

It works because it is honest about epistemology. Causal inference from observational data is a judgment, and Hill made the judgment inspectable. Each viewpoint names a distinct failure mode — reverse causation fails temporality, cherry-picked populations fail consistency, a real effect with no dose-response looks different from a confounded one — so evidence accumulates by adding viewpoints instead of re-litigating settled ones. The framework needs no math beyond counting studies, which is why it traveled from dust exposure to drug safety to toxic-tort courtrooms within a generation.

Nine questions, no formula: judgment made inspectable

Randomizing people into carcinogens is unethical, so observational data needed its own defensible standard of proof rather than a borrowed experimental one.

Regulators, courts and insurers needed a procedure they could examine and contest — a statistician's private intuition settles nothing in a public argument.

None of the nine is decisive by design, which blocks both gullible acceptance and reflexive denial: a weakness in one viewpoint cannot dispose of the evidence.

The viewpoints need no data format and no math beyond counting, so they transfer intact from dust exposure to drug safety to litigation.

When experiments are impossible, don't fake certainty with a scoring rule. List the considerations and where the evidence points on each. Arguable frameworks outlive formulas nobody believes.

The address became known as the Bradford Hill criteria and is the most-cited framework for causal inference in epidemiology; a 2015 review of its modern use already drew 457 citations. The viewpoints shape how agencies weigh carcinogenicity and appear routinely in product-liability litigation, where plaintiffs must establish general causation. Hill never called them criteria and warned against checklist use — a caution both critics and users credit for the framework durability. Molecular epidemiology keeps reinterpreting the viewpoints for biomarker and genomic data rather than discarding it.