I think this is a case where doing the automated analysis on a large dataset is misleading, because the automated analysis is based on an automated evaluation of how good a move is. Another way of saying "don't blunder" in this context is "choose a move that is not much worse than the best move". That is hardly more useful advice than "choose the best move". The advice "don't blunder" only becomes useful when you can also give advice about how to recognize blunders: "check whether your pieces are hanging", "check whether your opponent has mate in 1", "check for forks". Probably many of the blunders in that dataset are simple things like this, but others are long forcing lines or counterintuitive sacrifices that are difficult to recognize both for you and for your opponent. The computer doesn't distinguish, but it's much easier to improve by focusing on the former than the latter. (obviously to continue to improve you have to do both, but "not hanging pieces" is a lot less work).