I think it’s probably to do with the standard error for the population drawn from not actually being the same thing as the standard error for the population as a whole, due to the conditions that would link them being violated. In particular, I don’t think they’re likely to be allowing for model error at all.One thing about the '+/- error'. If these statements of possible error really were true, then nearly all the polls would fall within twice the error margins of all the others. Since they don't, the statements of possible error margin by the pollsters are demonstrably false.