During the Second World War, engineers examined aircraft returning from missions and recorded where each one had been hit. The pattern was clear: many holes along the wings and the body, far fewer around the engines.
The obvious recommendation was to add armour where the holes were. That is where the aircraft were being hit, so that is where the protection was needed.
A mathematician working with the group asked one question that changed the recommendation completely. Which aircraft, he asked, are in this data?
Only the ones that came back. An aircraft hit in the engines had a poor chance of returning at all, so it never reached the yard where the counting was done. The near-empty areas on the diagram were not the safe places. They were the places where a hit meant the aircraft was lost.
The armour went on the engines.
The point was not that the engineers had counted badly. Their counting was exact. What they had not asked was which aircraft the yard could ever have received.
The error the engineers nearly made has a name now — it is what happens whenever a group is studied after something has already filtered it. Successful companies, finished buildings, patients who recovered enough to attend a follow-up: each of these is a sample chosen by survival, and the cases that matter most are the ones that are absent.