2016 Elections

How the Polls Could Be Wrong

Third parties, shy voters, ground organizations and technical details all could have an impact.

How many will turn out?

Photographer: Ty Wright/Getty Images

If the polls are correct, this election is likely over. In HuffPollster’s estimate, Hillary Clinton is winning by 8 percentage points right now, based on two-candidate surveys. That edge looks close to insurmountable with slightly more than three weeks remaining and with voters in many states already casting their ballots. Poll-based forecasters (some of whom also take into consideration “fundamentals” such as the economy) estimate Clinton’s chances of winning at 86 percent to 98 percent.

This is when partisans start speculating that the polls are getting it wrong. It’s not entirely impossible, although the numbers can go in either direction. While the election might be more of a blowout than it appears right now, it could be a much closer contest, too. Here are some variables.

Third-party craziness. Polls show third parties doing unusually well in this cycle, with Gary Johnson at 7 percent nationally. In normal years, third-party candidates underperform their polling numbers on Election Day, especially the early numbers. This leaves pollsters (and forecasts based on surveys) with a dilemma. If they include Johnson (and the Green Party’s Jill Stein and independent conservative candidate Evan McMullin) in their survey, they probably overestimate votes for third-party candidates. Asking only about Clinton and Donald Trump will underestimate third-party voting.

We know that Clinton’s estimated lead is somewhat lower (about 6 percent) in polls that include Johnson, for example. But we can’t know how many people who think now that they’ll vote for the third party will do so, and whether Clinton or Trump will be the preferred second choice for those who opt for a major-party candidate after all.

Shy voters. Trump’s supporters argue he will benefit from “shy” voters who are hesitant to admit they will cast their ballots for him. That’s plausible. But so is the possibility that some people (perhaps in red states) might be reluctant to admit supporting Clinton.

Normally, we can mostly dismiss the chances that people are lying to pollsters. But the candidates this time are unusual: Trump for many reasons, Clinton mainly because she’s the first woman to be a major-party nominee. So it’s more believable than usual that a small percentage of the electorate isn’t being honest -- perhaps because they aren’t being honest with themselves.

Organization. By all accounts, the Clinton campaign and the Democratic Party have a large advantage in field operations over the Trump campaign and the Republican Party. This should mean that, all else being equal, a Clinton supporter will be a bit more likely to cast a ballot than a Trump supporter.

Lots of voters intend to vote but never actually do; this is one reason polling is difficult to begin with. It’s possible that effective get-out-the-vote drives work not by changing voter intent, but by changing the likelihood that voters will follow through on their intentions. Pollsters won’t have any trouble identifying changing intentions. But they have a harder time measuring the likelihood that someone will act on those intentions (that is, actually vote). Maybe this unknowable helps Trump. Or maybe it gives Clinton an advantage not measured in polls, especially in swing states, where campaign organization is most intense.

Again, voting has already begun, and more and more people are casting ballots each day. Therefore, the trick for pollsters is less and less about predicting what people will do on Nov. 8, and more and more about figuring out what people have already done. 

Technical stuff. Many people find it hard to believe that the opinions of a few thousand Americans interviewed in polls can predict the actions of millions of voters. Yet the math of these surveys works.

We’ve seen plenty of examples of polls successfully predicting election results. This year the polling was fairly good even in the primaries, though they are harder to measure than general elections. 1  But even experienced professionals can make mistakes, and not everyone publishing survey results is a top-rated pro (see Nate Cohn’s item at the Upshot on how experts can differ even after the calls are all made).

Polling averages (the basis for all the poll-based forecaster estimates) are good in part because the decisions pollsters make tend to cancel each other out. 2  But if there are unpredictable changes in the electorate -- say, a large demographic group suddenly becomes more likely to vote than usual -- then pollsters may miss them.

Overall. None of the potential errors in any of these categories would be enough by itself to mean, for example, that the apparent 8-percentage-point lead for Clinton is actually a tie. It’s possible that a “shy voters” effect could help Trump while an organization effect helps Clinton, canceling each other out. A major polling error would not only require most of these outcomes to be real, but for them all to be mistakes in the same direction. 3

In other words, while all of these effects are plausible, the most likely result is that the polls will wind up being correct. 

Why Polling May Be Worse Than Ever

This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

  1. Why? Because turnout is much lower, and candidate choice is a lot less firm, and therefore both are less predictable by voters themselves. The polls were far from perfect in the primaries, with a particularly large miss in Michigan for the Democrats, but overall they were fairly accurate.

  2. That is, one pollster might choose a model which turns out to help Clinton, while another opts for a model which helps Trump. Generally, public pollsters try hard to get it right. They rarely deliberately try to boost one candidate, since their reputations depend on accuracy. Its just that there are reasonable choices of which procedures to use and how, and they will produce different results. 

  3. Furthermore, each of these effects should be independent of the others, meaning that even if one of the errors turns out to exist, theres no reason any of the others does -- or, if it does exist, that it would be in the same direction. 

To contact the author of this story:
Jonathan Bernstein at jbernstein62@bloomberg.net

To contact the editor responsible for this story:
Katy Roberts at kroberts29@bloomberg.net

Before it's here, it's on the Bloomberg Terminal.