Skip to main content
You are the owner of this article.
You have permission to edit this article.
Edit
Polling gets it wrong again in Wisconsin. What happened this time?
alert featured topical
ELECTION 2020 | PARSING THE ELECTORATE

Polling gets it wrong again in Wisconsin. What happened this time?

{{featured_button_text}}
Voting at Elver Park

Sandra Dickman fills out her ballot in a tent at Madison's Elver Park. The pandemic prompted election officials to look for novel polling places where social distancing could be enforced.

For political pollsters, the ghosts of elections past are back to haunt them.

Following the widely covered polling miss in 2016, when many state level polls, including in Wisconsin, failed to closely show President Donald Trump’s support, pollsters this year had reason to believe things would be different.

Trump was an incumbent president this election cycle, creating a more stable political environment than four years ago, the thinking went, and many pollsters tweaked their methods, such as adjusting for education if they hadn’t already, and finding ways to make educated estimates about the vote choices of voters who didn’t want to reveal who they were voting for.

Being one of the most crucial swing states of the election this year, Wisconsin also enjoyed a substantial increase in the number of political polls conducted, raising hopes that accuracy could be found in sheer volume.

But to pollsters’ frustration, they got it wrong again, although polls of some states were closer than others. In Wisconsin, FiveThirtyEight’s combined average of the state’s political polls going into Election Day showed Biden leading Trump by 8.4 percentage points, 52.1% to 43.7%. But once the state finished its final unofficial count early Wednesday, Biden ended up winning the state by a narrow 0.63%, 49.57% to 48.94%, 7.77 percentage points off from the polling average.

One ABC News/Washington Post poll from late October gave Biden a 17-point lead over Trump, a result wildly different from the actual Election Day result. The last Marquette University Law School poll before Election Day was much closer, but still gave Biden a five percentage point lead over Trump, 48% to 43% among likely voters with a margin of error of plus or minus 4.4 percentage points.

The last UW-Madison Elections Research Center/Wisconsin State Journal poll placed Biden’s support at 53% of likely voters, compared with 44% for Trump, a nine point margin in Biden’s favor.

That’s a bigger miss than in 2016, even though the prediction of a Biden win was correct.

In 2016, the polling average headed into Election Day gave Democratic presidential nominee Hillary Clinton a 5.3 percentage point advantage over Trump, 49.6% to 44.3%, an average that ended up being 4.53 percentage points away from the actual 2016 margin of 0.77 percentage points, with Trump getting 47.22% and Clinton getting 46.45% in the recount.

Pollsters every election cycle caution people that their polls are only a snapshot in time on the campaign season rollercoaster. Each candidate’s estimated support in the poll is subject to the poll’s margin of error, which often swings four to five points in either direction.

Plus, there is typically a brief span of time after the last poll is conducted but before Election Day, a time when many candidates make their final rounds, and last-minute developments can occur. Such developments can cause last-minute swings in public opinion, like it did to an extent in 2016.

But with polls for a second time underestimating Trump’s level of support in the state, pollsters say the statistical realities of political polling don’t tell the full story. Something’s off.

“The polls were clearly off again and in a surprising fashion,” said UW-Madison political science professor Barry Burden. “Many of us expected the polls to be more accurate this year than they were four years ago.”

Wrong once more

Burden said pollsters had a lot of reason to believe polls would be better this year: More pollsters were weighting for education, there were fewer undecided voters, and there were fewer third-party voters and strong feelings on both sides of the political spectrum about Trump.

But none of that ended up mattering. For a second time, polls, at least in Wisconsin, substantially underestimated Trump’s support but were closer in estimating the Democratic candidate’s support. Burden said the fixes pollsters made post-2016, chiefly weighting for education, were overstated. Weighting for education may not have been the fix, after all.

Support Local Journalism

Your membership makes our reporting possible.
{{featured_button_text}}

For now, pollsters are scratching their heads and left theorizing about why exactly they missed another shot at accurately predicting Trump’s vote share.

“There’s something mysterious about all of this that we don’t have our hands around just yet,” Burden said.

Easy fixes don’t provide an explanation. Burden said even if you took all the undecided voters in the polls and handed them to Trump, it still wouldn’t provide a spot-on result. Marquette Poll director Charles Franklin tried a version of this in his last 2020 poll of likely voters, where 48% supported Biden, 43% supported Trump and 7% said they are undecided or declined to say how they would vote.

Franklin tried estimating how these undecided respondents would vote by “allocating” vote assignments to Biden if the respondent had a favorable view of Biden and not a favorable view of Trump; or to Trump if the respondent had a favorable view of Trump and not a favorable view of Biden. If the respondent was favorable to both candidates or unfavorable to both, the respondent would remain unallocated and remain undecided.

With this experiment, Trump’s share increased by two percentage points, to 45%, but so did Biden’s, to 50%. The experiment got fairly close to Biden’s actual 49.57% support, but it still underestimated Trump’s 48.94% support once the votes were tallied.

What’s the problem?

That leaves pollsters to think something else is going on, especially because Franklin’s analysis shows national polls overstated Biden’s lead by 5.8 percentage points this year, far above 2016, when national polls overestimated Clinton’s lead by only one point.

One of the leading theories is that there must be a small percentage of Trump supporters who are systematically declining to take part in surveys. It’s different from the “shy Trump voter” effect, another theory that some Trump voters who do participate in surveys choose not to say who they’re voting for.

The effects could be compounded by the fact that Trump has been so critical of polls, which could implicitly or explicitly discourage his own supporters from participating in the polls.

Franklin thinks the likelihood of there being people who just systematically don’t participate in polls is more likely than the “shy Trump voter” theory because when he breaks down who isn’t disclosing their vote, those people split evenly in their favorability of each of the two candidates.

“If it were shyness, and it was asymmetrically shy Trump voters, then I ought to see a bigger rise in Trump support when I look at people who like him and don’t like Biden,” Franklin said. “What we seem to see more here is these are folks that are simply reluctant to tell a stranger on the telephone who they voted for, especially if they’ve already voted.”

Burden said this effect could potentially be explained by a psychological theory of people who aren’t “community oriented,” meaning they don’t trust government institutions, media, universities, large companies and other authorities.

“It’s not being a shy Trump voter, it’s that they just won’t engage with outsiders who they don’t trust,” Burden said. These populations, he said, tend to be older, whiter and more rural: the demographic groups that Trump attracts.

If it’s the case that some Trump voters just aren’t agreeing to be a part of surveys, that could present a big challenge for pollsters, who need to be able to reach these types of people to learn more about them. But identifying that demographic group won’t be as easy as classifying people based on age, race or gender.

“It’s something attitudinal. It’s a set of beliefs or opinions that maybe are not correlated with other things like education that we can look at,” Burden said. “We have to get them in the survey in the first place to ask those questions.”

Whatever’s happening, it’s not isolated to a particular method of polling, but is seen across both internet and live-caller polls. It also seems to be a phenomenon specific to Trump voters. During the 2018 Wisconsin gubernatorial election, for example, the Marquette poll ended up with a result pretty close to the final election result. The poll showed a 47% to 47% tie. Evers went on to win the election by just over a percentage point: 49.5% to Walker’s 48.4%.

Another theory Franklin said hasn’t been tested out yet is whether or not there was a greater than expected surge in in-person Election Day voting, something the Trump campaign had been counting on to boost its support.

Burden said he expects pollsters in the coming months and years to undertake yet another investigation into why the polling was off in order to preserve its value.

“I think polling still has tremendous value because it provides a look at the electorate,” Burden said. “It’s the only way to measure the attitudes and perspectives of a representative group of residents. There’s really no other time in our social life where everyone’s voice is equal.”

Concerned about COVID-19?

* I understand and agree that registration on or use of this site constitutes agreement to its user agreement and privacy policy.

Related to this story

Get up-to-the-minute news sent straight to your device.

Topics

News Alerts

Badger Sports

Breaking News

Crime

Politics