[jeremyjh]

That isn't the only thing that happened. Probably the 1 in 3 odds were too low given the data available, because the polling demographics were not adjusted for education. If you randomly sample 1000 people to represent several millions, you also collect demographic information to ensure that you properly weight the responses based on how skewed that demographic is in your sample compared to the total voting population. In 2016 they weren't correcting for education, which turned out to be a huge hidden variable. This is explained quite well by 538 themselves: https://fivethirtyeight.com/videos/polling-101-what-happened...