The cohort that was followed

A study that followed twelve thousand people for thirty years produced a finding that no trial could have produced and that no trial will ever confirm.

The design is the point. A randomised trial can establish cause, because the assignment is controlled, but it cannot run for thirty years on an ordinary population, and it cannot assign people to a diet or an occupation they did not choose. A cohort study observes, which means it cannot control what else differs between groups, but it can follow a lifetime. The two designs answer different questions, and the common summary that trials are strong and cohorts are weak describes a ranking that does not exist.

What a good cohort study does instead of randomising is make its confounders explicit. It measures income, education, smoking, existing illness and a dozen other things at the start, adjusts for them, and reports how the result changes as each adjustment is applied. A finding that survives adjustment for everything measured is worth attention; one that disappears when income is included was a finding about income.

The remaining danger is the confounder nobody measured, and there is no statistical remedy. The partial remedies are comparisons: a finding that holds in populations with different confounding structures, or that shows a dose-response relationship, or that appears before the supposed cause could plausibly have been reversed, is harder to explain away.

The study's own authors described their result as the best available answer to a question that cannot be tested directly. That sentence is unusual in a press release and it is the correct description of most of what is known about how people live and how long they live.