I’m Here To Remind You That Trump Can Still Win

It’s tempting to write this story in the form of narrative fiction: “On a frigid early December morning in Washington, the Supreme Court ruled 5-4 that disputed mail ballots in Pennsylvania—” You know, that kind of thing. But given the stakes in this election, I think it’s important to be prosaic and sober-minded instead.

[Live Updates: We’re Tracking The Vote And Voting Problems]

So let’s state a few basic facts: The reasons that President Trump’s chances in our forecast are about 10 percent and not zero:

• As in 2016, Trump could potentially benefit from the Electoral College. Projected margins in the tipping-point states are considerably tighter than the margins in the national popular vote.
• More specifically, Joe Biden’s lead in Pennsylvania — the most likely tipping-point state, according to our forecast — is solid but not spectacular: about 5 points in our polling average.
• Without Pennsylvania, Biden does have some paths to victory, but there’s no one alternative state he can feel especially secure about.
• While a lot of theories about why Trump can win (e.g., those about “shy” Trump voters) are probably wrong, systematic polling errors do occur, and it’s hard to predict them ahead of time or to anticipate the reasons in advance.
• There is some chance that Trump could “win” illegitimately. To a large extent, these scenarios are beyond the scope of our forecast.
• There’s also some chance of a recount (about 4 percent) or an Electoral College tie (around 0.5 percent), according to our forecast.

Before we proceed further, a short philosophical note. I hate it when people use phrases — to be fair, we often use phrases like these ourselves! — such as “Nate Silver is giving Biden a 90 percent chance” or “FiveThirtyEight still gives Trump a 10 percent chance.” We aren’t giving anybody anything. Instead, as former FiveThirtyEight politics host Jody Avirgan puts it, what we’re doing is “mapping uncertainty.” In other words, if Biden leads by about 9 points in national polls, 8 points in Wisconsin, 5 points in Pennsylvania, 2 points in Florida, etc., how does that translate into a probability of victory? That’s what our model is trying to figure out.

And indeed — although nobody needs any reminders of this after 2016 — Trump can win. All the election models are bullish on Biden, but they are united in that a Trump win is still plausible despite his seemingly steep deficit in polls.

A huge part of why our model and others’ think Trump can still win is the Electoral College. Trump has only a 3 percent chance of winning the popular vote in our model. Other models put his chances at less than 1 percent. It’s very likely that Democrats will win the popular vote for the seventh time in the last eight elections.

But while a roughly 8-point deficit in the popular vote is hard to overcome — as of this writing, at 7:30 p.m. ET on Sunday, our model forecasts Biden to win the popular vote by 7.8 percentage points — a 5-point gap is a lot easier to close. And that’s our current forecast in Pennsylvania: Biden wins by 4.7 points. Note the roughly 3-point gap between the popular vote and the outcome in Pennsylvania, the most likely tipping-point state. That’s similar to 2016, when Hillary Clinton won the popular vote by around 2 points but lost the tipping-point state, Wisconsin, by just a little under 1 point.

Or if you want a more sophisticated version of this, we can look at how often Biden is projected to win the Electoral College from various potential margins in the popular vote:¹

If Biden wins the popular vote by 2 to 3 percentage points, the Electoral College is roughly a toss-up. But if Biden wins the popular vote by less than 2 points, Trump is a fairly heavy favorite to win the election. Even popular vote margins of up to 6 points are not entirely safe for Biden if his votes are distributed in exactly the wrong way. So you can see why an 8- or 9-point lead in the popular vote shouldn’t make Biden feel that secure; despite being a landslide margin, it’s also only a few points removed from the inflection point where the Electoral College starts to become competitive.

Biden’s position would simply be a whole lot safer if one of two things were true: If either the polling in Pennsylvania were like that in Wisconsin and Michigan, where he has a larger lead … or if another state such as Florida were also polling more like Pennsylvania to give Biden a clear Plan B. But neither of those things are true. The gap between Pennsylvania and Wisconsin and Michigan has actually grown in the waning days of the election. And no state has emerged out of the pack of Arizona, North Carolina, Florida and Georgia to be Biden’s clear Plan B (Biden is forecasted to win each state by between 1 and 3 points). Arizona is probably Biden’s best bet in this group, but winning it would also require him to win either Nebraska’s 2nd Congressional District or Maine’s 2nd Congressional District to break a 269-269 Electoral College tie; he’s favored in both districts but they aren’t sure things.

Moreover, because polling errors are somewhat correlated from state to state, if Biden loses Pennsylvania, he would no longer be a favorite in states such as Florida and Georgia — where he’s narrowly ahead now — because it would be a sign that Trump had outperformed his polls again. Of course, this only goes so far: In 2016, Trump massively outperformed his polls in the Midwest, but there wasn’t much of a polling error in Arizona. Still, losing Pennsylvania would take Biden from favorite to underdog.

You’ll notice that I’ve mostly been focusing here on the whats and not the hows or the whys. In other words, I’m describing what combinations of states could plausibly produce winning maps for Trump given the possibility of a polling error, but not describing why such a polling error might occur.

To some extent, that’s on purpose. If pollsters knew what the source of a polling error might be, they’d presumably try to fix it. Many pollsters are weighting by education now, something many didn’t do in 2016, and that was a big source of error that year. Another big source of error in 2016 was the large number of undecided voters, who broke toward Trump in the Midwest. To some extent, that one isn’t on the pollsters, since polls aren’t really supposed to try to predict the vote of people who say they’re undecided. Nonetheless, that’s much less of an issue this year, because there are far fewer undecided voters.

There are, however, some new potential sources of error this year. One of them is the huge growth in people who are voting early or by mail. It’s not only that more people are using these methods, but also that — unlike in the past, when they were relatively bipartisan — Democrats are far more likely to vote by mail than Republicans. Republicans are much more likely to vote in person on Election Day, conversely, while early in-person voting falls somewhere in between.

One issue for pollsters here could be the rate of ballot spoilage. Because of processing delays, some mail ballots won’t be received by states’ deadlines. And voters might not complete the instructions correctly, as we’ve already seen in Pennsylvania with the issue of “naked ballots”, that is, voters forgetting to enclose their ballot in its extra, secrecy envelope.

So imagine, for instance, that in a certain state, the vote is divided evenly at 50-50 in a poll between Biden and Trump. But two-thirds of Biden voters are voting by mail, whereas two-thirds of Trump’s supporters are voting in person, and the rate of mail ballot spoilage is 3 percent. That would be enough for Trump to win 50.8 to 49.2, meaning that you had a polling error of 1 or 2 points.

However, there are several mitigating factors here. First, 3 percent is probably on the high side for mail ballot rejection rates; other estimates hover at closer to 1 percent, although the number could be higher this year with so many first-time mail voters. Second, some ballots are also spoiled during in-person voting because of errors with voting technologies (think about hanging chads, for instance). Third, there is considerable evidence that Democrats mailed in their ballots early, which could put them less at risk of spoilage. According to the United States Elections Project, the party registration on mail ballots received so far favors Democrats by 24 percentage points. But for mail ballots requested but not yet received — those that might trickle in late — the partisan gap is just 11 points in Democrats’ favor. Polls find that Democrats are also more likely to drop off their ballots at drop boxes, which reduces the risk a ballot is rejected.

In addition, there’s something to be said for the idea that it’s worthwhile to lock in a vote. If someone has already voted, they’re 100 percent likely to vote (and 98 or 99 percent likely to have their vote counted, depending on the rate of ballot spoilage). What about someone who says they’re planning to vote on Election Day but hasn’t done so yet? They’re certainly not 100 percent likely to vote. Something could come up on Election Day — they get stuck late at work, they blow out a tire, they feel sick, they don’t bother because they think their candidate is losing. Indeed, even some of the people that pollsters deem to be the most likely voters don’t wind up voting. If 2 percent of mail voters have their votes rejected, but 5 percent of “likely” Election Day voters don’t wind up voting, then polls could underestimate Democrats.

Wait, wasn’t this supposed to be a post about how Trump could beat his polls? Well, the point is just that mail voting creates additional uncertainty this year, and it’s easy to imagine how that could help out Trump or Biden.

Another potential source of anxiety for pollsters is the Hispanic vote. Polls show Trump having made significant gains relative to 2016 with Hispanic voters — and to a lesser extent with Black voters, especially Black men. This is not enough to offset gains that Biden has made with white voters, however, including white voters both with and without a college degree.

But are these changes real? Both white voters and Hispanic voters without a college degree can be hard groups to reach on the phone. It can also be hard to get a representative sample — if, for example, you don’t get enough Cuban American voters in Florida, or if you aren’t reaching enough Hispanic Americans who primarily speak Spanish. Hispanic voters and white voters without a college degree can also work somewhat at cross-purposes to one another when you’re weighting a poll, because while white voters without a college degree are more Republican than whites with a college degree, the opposite is true for Hispanics.

Overall, I don’t particularly think there is any reason to distrust the polls here. If anything, polls have tended to underestimate Democratic support in recent elections in states such as Nevada that had a large number of Hispanic voters. Still, suppose that Trump’s growth in Hispanic support is real, while Biden’s gains among white voters without a college degree are not, for whatever reason. That could lead to a rough night for Biden: The lack of white non-college support could cost him Pennsylvania, while a mediocre performance among Hispanics could keep Arizona and Florida in Trump’s column. Maybe Biden would eke out a win in Georgia or North Carolina, but that’s a much narrower path then he’d planned on.

And what about those “shy” Trump voters? There’s no particularly good evidence that Trump voters are likely to conceal their intentions to pollsters. Nor — if we want to expand the sample size a bit — is there any reason to believe that nationalist or right-wing parties tend to beat their polls in other countries.

Conversely, there is quite a bit of evidence that most of Biden’s polling gains relative to Clinton come from vote-switchers, rather than from an expectation of higher Democratic turnout. If a respondent tells a pollster that they voted for Trump in 2016 but will be voting for Biden this year — and there aren’t a lot of those people, but a few make a big difference — it’s hard to consider them a “shy” Trump voter.

Still, the theory isn’t completely crazy. Social desirability bias — not wanting to provide an answer you think the person on the other side of the line won’t like — has been a problem in some other polling contexts. The point is that even if you mostly aren’t worried about “shy” Trump voters — or think it’s equally likely that there are “shy” Biden voters! — that’s different than being 100 percent sure that the theory isn’t true. And if we’re trying to account for how 10 percent chances happen, we have to accept that sometimes it’s because our assumptions are wrong.

Finally, there are the factors our model doesn’t try to account for, such as the many, many things we’re tracking on our election administration blog: attempts to disqualify various groups of ballots, voter intimidation, polling-place irregularities, and so forth. Plus, there’s a good chance that Trump will try to declare a premature victory.

It isn’t terribly easy to sort out the heat from the light here. But it may sometimes be worth putting magnitudes on things. Problems at one polling place are not going to have nearly as much of an impact as a Republican attempt to throw out 120,000 ballots cast in Texas’s Harris County, for instance. Yet, as much of an affront to democracy as that would be, even that would still only amount to around 1 percent of the vote in one state. Let’s up the ante: What about Republican legislatures trying to send alternative slates of electors to the Electoral College? Now, that could have a really big impact, although it’s not clear how likely it is.

It’s also worth recognizing that there is another side to this, too. There is — long overdue in my view — far more attention paid to voter suppression and voter disenfranchisement than there used to be. (We’ve certainly made a big effort to put far more resources into those stories at FiveThirtyEight.)

But … is it actually harder now to vote than it has been in the past? It depends on the state, but in most states, the answer is no. According to the Brennan Center’s annual reports, recent years have seen more efforts to expand voting rights than to restrict them. And the COVID-19 pandemic has brought about an expansion of voting options, some temporary and some permanent. If it’s still too hard to vote for disadvantaged groups — but it’s easier than it used to be — that could lead to a net increase in turnout for Democrats relative to past elections. The FiveThirtyEight model does try to account for changes to voting laws in each state; that’s part of the reason why it’s relatively bearish on Biden in Texas, for example, which has some of the strictest laws in the country.

So when I say that there are certain things outside the scope of the model — well, the truth is a little bit messier than that. One reason that we make relatively conservative assumptions, such as by using fat-tailed distributions, is to account for “unknown unknowns.”

Here’s what it seems safe to say, though. In an election that is very close, a 6-3 conservative majority on the Supreme Court is likely to side with Trump. Our model shows a 4 percent chance of an election that winds up with one or more decisive states within 0.5 percentage points, close enough to trigger a recount. If you want to round up Trump’s odds slightly by assuming he wins the lion’s share of those 4 percent of cases, plus most of the 0.5 percent of the time that the election ends up in an Electoral College tie, I wouldn’t strenuously object to that. Mostly, though, I’d just be worried about the meltdown that could occur if a recount or a tie comes up. The odds are against it, but the stakes are awfully high.