Degrees of Freedom, Garden of Forking Paths — How to reach p < 0.05 every time
Another Scott Alexander piece, The Pyramid and the Garden ( link ):
Andrew Gelman writes about the garden of forking paths. The idea is: the scientific community accepts a discovery as meaningful if p < 0.05 - that is, if equally extreme data would only occur by coincidence 5% of the time or less. In other words, you need to win a lottery with a one-in-twenty chance if you want to get credit for discovering something absent any real effect to be discovered. But if a scientist forms their hypothesis after seeing their data, they might massage the precise wording of their hypothesis to better fit their data. If there are many different ways to frame the hypothesis, then they have many lottery tickets to choose from and a win is no longer so surprising. Gelman discusses a study claiming to find that women wear red or pink shirts during the most fertile part of their menstrual cycle, which sometimes involves red or pink coloration changes in primates. The study does detect the effect, p < 0.05. But there were a couple of different ways the researchers could have framed the problem. They could have looked at only red shirts. They could have looked at only pink shirts. They chose days 7-14 as most fertile. But they could also have chosen days 6-15 without really being wrong. They could have looked only at the unmarried women most likely to be trying to attract mates. A recent paper listed 34 different degrees of freedom that can be used in this kind of thing. Add up enough of them, and you have more than twenty tickets to the one-chance-in-twenty lottery and success is all but certain.