Theories often start with a random observation. Inuits, for example, have little heart disease and eat fish. Eureka!

The next stage, usually, is a case-control study in which the diets of those with heart disease are compared with those who are healthy. But there are normally so many other differences between the two groups that the results look a lot more persuasive than they are.

Worse, only the positive findings are published, building up a literature that emphasises beneficial findings and ignores the rest. This is called publication bias.

Only when randomised control trials (RCTs) are held does the emperor turn out to be wearing no clothes. A properly run RCT has two matched groups, followed for as long as possible, one taking pills and one a placebo.

But even RCTs are subject to random error. Combining them in a “meta-analysis” smoothes out the errors, although it does not eliminate publication bias. A single positive or negative study, if it is big, can swing the result.

So the normal life-cycle of a hypothesis is to start with a nice idea, confirm it using a poor evidence base, create a dietary fad sufficient to justify proper trials — then find it was never true to start with.

Times Recommends

• Olympic-size television coming to your town
• Mosley opens new front in battle for privacy
• On-the-spot fine for man smoking in his van