Why election models didn't predict Trump's victory — A primer on how polls and election models work

Also, to be clear, in order to compute his prediction, Wang did assume that non-sampling errors were somewhat correlated, just not nearly enough. As I say in the post, he is a very smart guy, so it's not as if he didn't know the things I explain.

I agree with you that the probabilities of Clinton winning individual states are correlated, but I'm not sure this makes what I wrote false, although you're probably right that it's a bit misleading. The fact that the probabilities of Clinton winning individual states are correlated is only relevant to calculate the probabilities for each possible outcome in the electoral college. It means that, as I explain later in my post, you have to take into account the fact that non-sampling polling errors in different states are correlated in order to calculate the probabilities for each possible outcome in the electoral college. One of the sources of non-sampling error that I describe in my post is measurement error, which if you read my post carefully I define in such a way that if someone doesn't vote for the candidate they claimed they would vote for when they participated to a survey for whatever reason (e. g. because they heard a news story that made Clinton or Trump look bad), it counts as measurement error. I agree that it's probably an unusual definition of this concept, which is typically construed more narrowly. But I defined measurement error in that unusually broad way precisely because I didn't want to introduce the complication that, even if someone who tells a pollster n days before the election that he's going to vote for X and would really vote for X if the election took place on the day he participated to that survey, he might not vote for X on election day. (Wang takes that, among other things, into account in order to calculate his prediction, but I was only describing the way in which he calculates a snapshot of where the race stands at any given time, since I think it's where the most interesting mistakes were made. I may be wrong about that, but judging by what he said after the election, I think Wang would agree with me on that.) Now, if the probabilities you calculated for each possible outcome in the electoral college are correct, then you can just use the aggregation method I describe above the passage you quoted in my post. What is misleading in my post is that I say the assumption for that method to be reliable is that the probabilities of Clinton winning individual states are correct (instead of the probabilities for each possible outcome in the electoral college), because it suggests that we can assume they are probabilistically independent (although I never said that and the rest of my post makes clear that I wasn't making that assumption), which of course they are not. Do you agree with that or do you think that there is a more serious problem here?

I think this is a misunderstanding of what "oversampling" means in polling. See e.g. this.

I think models that rely on fundamentals are worthless. I don't have time to explain why in details, though perhaps I will post something on that at some point, but if you want to know the gist of my argument, it's that models of that kind are massively underdetermined by the evidence.

A cursory glance through Fivethirtyeight's collected poll data shows a survey with over 84,000 voters (CCES/YouGov) giving Clinton a +4 percentage point lead, with 538 adjusting that to +2. Google and SurveyMonkey routinely had surveys of 20,000+ individuals, with one SurveyMonkey one having 70,000 with Clinton +5 (+4 adjusted). There was no clear reason to prefer your poll (whichever that one was) over these. https://projects.fivethirtyeight.com/2016-election-forecast/national-polls/

And it should go without saying that Clinton did end up at +2 nationally.

I'm sure pollsters sometimes "cheat" by constructing biased samples, but this can happen even if you're honest because, as I explain in my post, polling is really difficult to do. To my mind, the problem had more to do with commentators who were making mistaken inferences based on the polls, than with the polls themselves, although evidently some of them got things badly wrong.