Read on Twitter
This is a helpful reminder. George W. Bush was also up by 10-11 pts on Gore in August before going on to lose the popular vote. However, in my view, the VERY big diff b/w 2020 and these prior cycles is fundamentals. Both 1988 and 2000 featured parties with popular 2-term 1/n https://twitter.com/ddiamond/status/1287412579854438404
incumbents running for a very difficult 3rd straight term amid a pretty-good economy. In my fundamentals model, that yields an expectation of a very close race (in the end, not a very good prediction for 1988). The basic patterns political scientists have identified 2 explain 2/n
why people wind up voting the way they do yielded a very different conclusion than the polls did. So there was very good reason to doubt that those polling leads would hold up. By contrast, in 2020, the fundamentals suggest that an unpopular incumbent seeking reelection amid 3/n
mass unemployment should be down by anywhere from ~6-7 points (if you treat this as just a very bad recession) to ~10-11 pts (if you treat it as a catastrophic one). And sure enough, that incumbent is down by like 9 pts. My modeling suggests that a comeback big enough for 4/n
Trump to win the popular vote, against what the fundamentals would have us believe, is exceedingly unlikely. Like, a 1-2% chance. By contrast, at this point in 2000, it would have still favored Gore to win the popular vote, and would only have predicted Dukakis to get 51%. 5/n
Now, I'm well aware of the crusade by @NateSilver538 against overfitting fundamentals models. I think our version is pretty parsimonious, supported by the poli-sci consensus, and rigorously cross-validated (model tested against data it wasn't trained on). And its mean 6/n
forecast of Biden +7 or so is very close to what Nate gets using his extremely impressive work extending the time series back to 1880. (We can't do that with our approach, because we treat POTUS approval as a fundamental predictor, and that data doesn't really exist pre-WW2.) 7/n
But we're not really talking about the mean here. We're talking about the variance. The 1948-2016 sample suggests fundamentals predictions this far out have a confidence interval of around 3.2 pts of vote share/6.5 of margin. I believe that a) Nate finds a much weaker 8/n
relationship, with wider confidence intervals, in the dataset going back to 1880 (though again, he's not using POTUS approval) b) he thinks that therefore, fundamentals should only provide a gentle nudge to polling averages, rather than being the anchor of a forecast until 9/n
very late in a campaign c) we're mistaken not to allow for greater uncertainty than the 1948-2016 data would indicate, just because the sample of POTUS elections is so small, and d) we're mistaken not to explicitly code in extra uncertainty in recognition of covid-19, above 10/n
and beyond the extra uncertainty that comes from being in unusual economic conditions. I think all of those concerns are valid, to greater or lesser degrees. I'll be fascinated to see what mathematical steps 538 takes to account for covid-19, given that we've never had a 11/n
POTUS election amid a pandemic before. But I also think that the fact that voters are telling pollsters they plan to vote in roughly the same way that fundamentals would expect is pretty suggestive that this campaign ISN'T breaking the mold of the 1948-2016 precedents. 12/n
