I.e., suppose that 3 of the questions were meant to test some theory that the researchers have, and the 21 other questions were thrown in for the reasons stated above.
Now suppose that the answers to the 3 seriously-chosen questions weren't predictive of long life, but 5 of the other 21 questions were predictive of long life.
It sounds like you're saying that somehow that 5-question correlation is bogus, but the 3-intentional-question correlation would be legitimate. But the only difference I see between the two groups is subjective: whether or not the researchers anticipated the questions being meaningful. I don't see how/why that would be relevant to making inferences about the wider population.