Is 538’s Bearish Clinton Prediction Likely to be the Most Accurate?

The NYT’s Upshot has a useful grahpical summary of the major poll aggregators:

They all have Clinton between 85% and 99+%, except for the most famous one of all, which has Clinton at 65%.

Fundamentally, I think there should be a presumption, although not an unrebuttable one, that the outlier is wrong. Silver’s model has a good track record but not better than Drew Linzer’s or Sam Wang’s.

There are two additional reasons why I think 35% is a very substantial overestimation of Trump’s chances of winning. If I understand correctly, the assumption of the 538 model is that 1)there is an unusual degree of uncertainty and 2)this uncertainty is equally likely to favor Trump as Clinton. Accepting the first premise arguendo, I don’t really buy the second. Trump’s unusual candidacy, to the extent that it can be expected to differ from the typical Republican coalition, should be more likely to attract middle-class white voters without college educations and less likely to attract Hispanic voters. The problem for Trump is that given historical turnout patterns, this presents much more upside for Clinton than Trump. And, in fact, in early voting the Latino turnout has been up substantially, including in Florida (where a loss would leave Trump drawing dead) and in the firewall states of Nevada and Colorado. Between this and Clinton’s far superior organization, I think the polls are much more likely to be overestimating Trump’s results than underestimating them.

There’s another reason I think the 538 model is significantly overestimating Trump’s chances — the state-by-state predictions, which the Upshot also helpfully compiles. Right now, 538 has Trump with a roughly 5% chance of winning Connecticut, a 10% chance of winning Delaware, Oregon and Rhode Island, and a 20% chance of winning Minnesota and New Mexico. Do any of these estimates seem remotely plausible to you? I mean, I’m sure Erik is going from Woonsocket to Narragansett rallying support for Stein to send a message to BIG VAXX and BIG WIRELESS, but I don’t think it will be enough. The 538 model is showing Trump with a real chance in states in which Republican candidates have been completely noncompetitive for a long time, and that no other model shows him with any chance of winning. To me, that’s dispositive.

I’m not saying don’t panic, exactly. If we more realistically assume that Trump has something like a 10% chance of winning, that’s still pretty terrifying. But I am confident that the 538 model is an outlier because it’s substantially overestimating Trump’s likelihood of winning.

…since this post is getting a little Twitter attention, I should say that I don’t think Silver et al. are trying to attract traffic or otherwise acting in bad faith. They have a good track record and deserve to be taken seriously. I don’t even think the assumptions of their model are unreasonable in the abstract. But ultimately the burden of proof is on the outlier, and looking at the specific facts of this election as best as I can I think the chalk is right here.