You mostly can rely on the polls, but that doesn’t mean Trump can’t win.
by Matthew Yglesias
The polls are pretty clear at this point — Joe Biden is in position to win the presidential election.
He’s up nationally, he’s up in Pennsylvania, he’s up in Florida, and he’s even up in North Carolina. He’s definitely up in Michigan and Wisconsin. If the states go the way the polls say, he’s going to win the election, and it won’t even be particularly close.
But many people remember reading popular poll aggregation sites four years ago and the confident predictions that Hillary Clinton would win: She had an 85 percent chance of winning, according to the New York Times, for example. Clinton, however, not only lost the key battlegrounds of Florida and Pennsylvania but even lost Wisconsin, where public polling had her so far ahead she didn’t bother to campaign in person.
Is it really different this time?
The answer is mostly yes. The extreme confidence in Clinton’s 2016 victory was despite a modest lead in the national polls. Forecasts in 2016 differed a lot in their assumptions. But there were generally two factors at play, on top of the state polling misses: a flawed way of thinking about how state-level races relate to one another, and a misperception of the electorate in key states in the Electoral College.
Today’s forecasts built in more GOP-friendly assessments of state dynamics. The main reason forecasters still think Biden has a very good chance of winning? His polling lead is just genuinely large. As of October 28, his chance of victory in the FiveThirtyEight forecast model (88 percent) is higher than Clinton’s was on Election Day in 2016 (72 percent).
Uncertainty remains primarily because even though Biden’s lead is big and has been remarkably stable, things could change and it could shrink — particularly in states key to winning the Electoral College. And if it does shrink, we’d see that a lot of things have not changed since 2016. It continues to be unclear if pollsters can more precisely gauge public opinion in the key Midwestern swing states, and the Electoral College has a large bias toward Republicans. Still, even as those factors remain constant, there have been some key changes over the past four years.
The modeling has improved
There are two distinct steps involved in poll-based election forecasting. First, they look at polls to try to assess the state of public opinion; second, they build a model, using the polls and sometimes other data, that goes from those polls to a prediction.
Common sense says a 5 percentage point lead is better than a 2-point lead and a lead in October is better than a lead in July. But to generate a precise forecast, you need to formalize those intuitions.
If you look back at the 2016 forecasts, some models were super-bullish on Hillary Clinton — but not all. On Election Day, she had an 85 percent chance of winning, according to the New York Times, and a 98 percent chance, according to the Huffington Post. Nate Silver and FiveThirtyEight, however, gave her a more restrained 72 percent chance. The main reason was a disagreement about modeling, not about what the polls said.
The Huffington Post looked at the election as 50 separate state races where deviation from the current polling could happen in each place, but would happen independently. So if Trump had a 35 percent chance of winning Pennsylvania, a 40 percent chance of winning North Carolina, a 40 percent chance of winning Florida, and a 45 percent chance of winning Georgia, etc., then the model assumed that added up to something like a 2 percent chance of Trump carrying enough battleground states.
The FiveThirtyEight model — to oversimplify — built in data from previous elections and also assumed that if the polls were off, they just might be off everywhere in the same direction. So in the moderately unlikely universe where Trump narrowly wins Pennsylvania, he is likely winning North Carolina, Georgia, and Florida.
Now, of course, you don’t want to model state election results as completely correlated with one another. Biden could probably boost his numbers in Wyoming by a point or two by running unanswered TV ads there, but this would have no implications for the national election. And effective outreach to Mexican American voters could help in Arizona, Colorado, New Mexico, and Texas without moving the needle in Maine or Wisconsin.
But most everyone agrees now that Silver had the better of this argument. The current models all assume polling errors and swings in public opinion will be correlated. When the Economist says that Biden has a 93 percent chance of winning the Electoral College, it is making a methodologically sounder claim than the New York Times or Huffington Post did four years ago.
Silver, whose methods have not changed in this regard, appears significantly more confident in Biden’s chances than he was in Clinton’s.
All of which is to say that apples to apples, Biden has a bigger lead.
Biden has a larger, more stable polling lead
RealClearPolitics does a simple, naive polling average with no fancy math or house effects. It says Biden has an 8.6-point national polling lead, which is a bit bigger than Clinton’s lead was at the height of her 2016 convention bump. But Biden is not in the midst of a convention bump, and there are reasons to believe his lead will be more permanent:
- Biden’s lead is bigger than Clinton’s lead was at any point in the 2016 campaign.
- Biden’s big lead comes just two weeks before Election Day, so there is limited time for things to change.
- While her lead was at its peak, Clinton was only polling in the low 40s with lots of undecideds and third party voters. Biden is above 50 percent in the RCP average.
The size of Biden’s lead is clearly good news for him. The significance of the stability of the lead is something modelers disagree about. One way of looking at it is that views of the incumbent are pretty locked in, as are views of the former vice president — and consequently, otherwiseearth-shattering events like the Covid-19 pandemic and the mass protests following the police killing of George Floyd don’t move the polls very much.
Under those circumstances, it just looks incredibly unlikely that anything more dramatic is going to happen over the next week — which is one reason the Economist is so bullish on Biden.
One might note, though, that past performance is no guarantee of future results. We know that, historically, public opinion sometimes does exhibit sharp sudden moves. It’s unlikely that there will be a huge, last-minute swing toward Trump, but why rule anything out?
All this assumes that the polls are accurate.
Most pollsters have fixed a mistake
National polling averages said Clinton was up by a bit more than 3 points on the eve of Election Day, and she won the popular vote by about 2 points, a very modest one-point error that nobody would remember as a big deal had she actually become president.
But those same polling averages showed Clinton up 6 points in Wisconsin. That’s a fairly large error, a bit outside the normal variance you’d expect. One major culprit, in retrospect, is that pollsters’ samples turn out to include way more college graduates than are present in the actual electorate. Since college graduates’ voting behavior differs systematically from non-grads, that biases the polls — especially in the Midwest — toward Democrats.
These days, the better pollsters address this by “weighting” their sample to reflect the actual education attainment of the population. Not every pollster weights, and among those who do, there is some variance in exactly how they do it.
Education weighting is also not a panacea. National polling in 2018 was accurate, but national polling was accurate in 2016 as well. On the state level, the 2018 polls underrated the GOP in Florida and the Midwest and overrated it in California and the Southwest.
State polling was just as in 2018 as it was in 2016, while in both years national polling was about right pic.twitter.com/BOqqPuTsV7— (((David Shor))) (@davidshor) March 1, 2020
The issue, pollsters tell me, is that educational attainment matters politically because it’s a proxy for underlying differences in personality. Poll samples are biased toward college graduates because non-graduates are more likely to have low social trust, which, among other things, makes them less likely to speak to pollsters. And in recent cycles, low-trust voters have skewed toward Republicans. But there are low-trust college graduates and high-trust non-graduates. So while education is a convenient proxy, relying on it does not fully eliminate bias in the polls.
The continued issues with Midwestern polling got obscured to some extent in 2018 by the fact that Democrats won the key races. But Debbie Stabenow and Gretchen Whitmer both underperformed their polls by about 2 percentage points. It was just a strong enough year for Democrats nationally that it didn’t matter.
The Electoral College is still tilted for Trump
As long as Biden is up by 8 to 10 points in the national polling, the fact that he’s “only” up by 5 or 6 points in Pennsylvania doesn’t seem so significant. But that 6-point lead in Pennsylvania represents a huge Electoral College disadvantage.
If instead Biden were up 6 nationally and 2 in Pennsylvania, forecasters would be saying that Biden is a very narrow favorite. Two-point polling errors are pretty common, though 6-point misses aren’t totally unheard of.
This huge Electoral College bias is why coverage of the 2020 race can end up giving you whiplash — either the polling looks like a landslide for Biden or else it’s a toss-up, with nothing in between. That’s a reflection of the underlying reality of a situation in which Biden probably needs to win the national popular vote by a large margin to carry the Electoral College.
Back in 2012, by contrast, the electoral map was tilted modestly in Obama’s favor, so it was possible for him to have something like a safe but narrow 3-point lead in the national polls.
Due to the polling problems discussed earlier, many observers believed that Clinton retained Obama’s Electoral College advantage during the 2016 race. Because Michigan, Wisconsin, and Pennsylvania had voted for Al Gore and John Kerry even as they lost nationally, they were seen as a “blue wall” that could safeguard Clinton’s election, as long as she was able to secure wins in Virginia, Nevada, and Colorado, seen as swing states.
In retrospect, of course, that was not true. Clinton won Virginia, Nevada, and Colorado while losing the “blue wall” states. A modest 4- to 5-point national polling lead for Clinton never should have been seen as secure.
It was this misperception, more than anything else, that drove the misguided complacency about the campaign. A candidate up 3 points in preelection national polling averages is in fact very likely to win the national popular vote, which Clinton did. But a candidate up 3 points and facing a 2-point Electoral College bias is at serious risk of losing the election — which she did.
The good news for Biden is his national polling lead really is a lot bigger than that. The bad news is that the questions about the reliability of state polling have not vanished and the Electoral College bias does not appear to have diminished.
Biden’s national lead is currently simply too big to be plausibly overcome by state-specific polling error or Electoral College bias. But if it shrinks by only a few points — due to debates or news events or whatever else — the likely outcome could tip from potential landslide to a squeaker in the blink of an eye.