Via Patrick Ruffini. To put the sample size here into context, note that most national polls typically survey between 800 and 1,500 likely voters. The Cooperative Election Study released today by Harvard is upwards of 50 times that number.

The CES is conducted every two years before and after the national election, which raises the question of how it did forecasting Trump’s 2016 upset. All told, pretty well. It had Clinton ahead by four; she went on to win the popular vote by two. More importantly, the numbers showed a lot of uncertainty in the final days of the race and, with one exception, recognized that swing states were so tight that Clinton couldn’t properly be said to be favored there.

Nationally, 43 percent of survey respondents chose Hillary Clinton and 39 percent chose Donald Trump for President. The remaining 18 percent either were undecided or chose Libertarian Gary Johnson, Green Party candidate Jill Stein, or another candidate.

The survey was also designed to have representative samples in each state. In Alaska, Florida, Iowa, Nevada, North Carolina, Ohio, and Pennsylvania, Trump and Clinton are separated by 3 points or less. Statistically speaking, we cannot be confident of either candidate’s lead in these states.

In the remaining battleground states, Clinton appears to have the edge. Trump holds leads in Arizona and Georgia. Clinton holds leads in Colorado, Michigan, New Hampshire, New Mexico, and Virginia.

The one state they missed on was Michigan. They had Clinton ahead there and Trump came back to win by less than half a percentage point. The outcome of the 2016 race was especially hard to predict, though, for two reasons. One, as noted above, was the uncertainty inherent in having two unpopular low-polling major-party candidates in the mix with several high-profile third-party opponents. The other was the upheaval generated in the closing days by James Comey’s letter announcing the reopening of the Emailgate probe, which Nate Silver believes ended Clinton’s chances by convincing late deciders to opt for Trump.

This year the race is much simpler. The third-party candidates are marginal and there’s been no earth-shaking last-second surprise despite the best efforts of Team Trump to get Huntergate off the ground. CES has Biden ahead by eight, and their data matches the surprising trend in other polling this year in which Trump is overperforming among minorities relative to 2016 whereas Biden is overperforming with white voters. Ruffini has the key numbers:

Swing to and from Trump from 2016 Post-Election CCES

White Non-College: Trump +27 to Trump +19 (-8)
White College: Clinton +3 to Biden +22 (-19) (!!!)

Black: Clinton +80 to Biden +77 (+3)
Hispanic: Clinton +35 to Biden +24 (+11) (!!!)

— Patrick Ruffini (@PatrickRuffini) October 30, 2020

Trump really has made inroads with Hispanics this year, but Biden’s made inroads with whites — of both education groups — and whites are a much bigger group. And as Ruffini notes, Biden is killing it with people who either stayed home four years ago or voted third-party. He’s up 56/27 among that group, a better than two-to-one margin.

Some readers are looking at that and thinking, “If Trump won despite being behind in the polls in 2016, he can win despite being behind in the polls now.” Absolutely, especially if there’s a last-second shift in Pennsylvania. But bear a few things in mind. First, Silver estimated awhile back that Trump has a realistic shot of winning the electoral college again if he loses the popular vote by no more than five points. If he trails by more than that, the odds get very long. CES has him trailing by eight. If we assume they’ve underestimated him today to the same degree they did four years ago, lowballing him by two points in his final margin against Hillary, then he’s trailing Biden by six. Still long odds.

His comeback in 2016 was like a team rallying from two runs down in the bottom of the ninth to win. This year, it’s more like rallying when you’re five runs down. You might see the first scenario happen a few times over the course of a season. You might see the second scenario happen a few times over the course of your life.

Moreover, there’s no sign of a rally right now the way there was four years ago. Here’s what the final two weeks of national polling looked like back then:

You can see Trump start to make his move right after the Comey letter appeared. He went from 4.6 points down on October 29 that year to 1.3 points down by November 3, which is just about as far from Election Day that year as we are from Election Day this year right now. He had momentum. The bases were loaded when he came to bat.

Here’s what the last two weeks of national polling look like this year:

Two out, nobody on, five runs down. Various other pieces have made that point over the last few days, stressing that it would take a much bigger polling error to explain a Trump victory this year than it did four years ago. And not just a bigger error but a less excusable one considering that pollsters know more about Trump’s supporters now than they did in 2016. They’ve adjusted their methods to make sure those voters are accounted for in samples so they shouldn’t be missing the extent of his support to the degree they did then. On top of all that, notes Derek Thompson, Trump has a massive liability weighing on him in how he’s handled the pandemic. And voters are about to issue a verdict on it at a moment when cases are reaching an all-time high nationally.

It’s not a simple “if he won then, he can win now” calculation, in other words. It’s the difference between having to hit a routine three-pointer at the buzzer to win in 2016 versus having to hit a shot from beyond halfcourt to win today.

There’s one other factor he has to contend with, the surge in early voting. No one knows which way the ballots already cast are leaning, but traditionally Democrats vote early while Republicans vote on Election Day. And there’s an enormous partisan gap this year across many polls on mail-in voting, with Democrats far, far more likely to vote by mail than Republicans are. Given all that, the massive number of early votes cast thus far suggests Trump and the GOP will have a daunting hole to climb out of on Tuesday:

Here’s the big hurdle for Trump. With 80 million votes — a bit over half our projection for total turnout — already cast, it is very, very unlikely that he could swing enough of the remaining voters to win the election fairly. Polls will have to have been 2-3x as wrong as in ‘16.

— G. Elliott Morris (@gelliottmorris) October 30, 2020

You can squint and get around this w/ the idea that what's baked was overwhelmingly just cannibalized D votes anyway, but man that math gets tough. https://t.co/qRvIfcQTd8

— Liam Donovan (@LPDonovan) October 30, 2020

Morris is the election modeler for The Economist magazine. Donovan has worked in elections here in the U.S. They’re not random yahoos. No one’s writing Trump off given the relative tightness in Pennsylvania at the moment, but it looks all but certain that the national polling as of Tuesday morning will suggest that a Trump electoral college win is close to impossible. That wasn’t the case in 2016. Stay tuned.