Confounding
When a third factor explains the link you thought you'd found between two things.
Confounding happens when a third factor is associated with both the exposure being studied and the outcome, creating an apparent relationship between them that is not causal. It is the central difficulty of observational research.
The classic illustration: coffee drinking looks associated with lung cancer, because coffee drinkers were historically more likely to smoke. Coffee is not the cause; smoking is the confounder, and it produces a real association with no causal link.
Researchers handle it by adjusting statistically for the factors they measured, by restricting the study population, by matching participants, or by stratifying the analysis. All of those help and none of them fully solves it.
The residual problem is what was not measured, or was measured badly. Adjusting for self reported activity, for example, corrects for what people said rather than what they did, so a great deal of the confounding remains inside the adjustment.
There is also a healthy user effect running through most lifestyle research. People who take a supplement, attend screening or exercise regularly differ from those who do not in income, education, other health behaviours and access to care, in ways no questionnaire fully captures.
Randomisation is the only design that handles unmeasured confounders, because chance distributes them evenly, which is why a randomised trial outranks a cohort study for the same claim. Mendelian randomisation offers a partial substitute where randomisation is impossible.
In practice, when a headline reports an association, the first question is what else is true of the people who did the thing, and the second is whether a trial has ever tested it.