Election forecasting foibles

With former Vice President Joe Biden leading or tied with President Donald Trump in most swing states’ polling averages, oddsmakers say he’s the favorite to win on Election Day. But Trump reminded voters on Saturday that polling averages leading up to the 2016 election had Hillary Clinton ahead in Michigan, Pennsylvania, and Wisconsin, all of which he won.

“The Polls are Fake just like much of the reported news. I won it all against Crooked Hillary!” Trump tweeted Saturday, linking to a polling article from Breitbart News. The last election’s result leads to a natural question: Can Americans trust the polls this time around?

“I want to remind everybody that the polls in 2016 weren’t really that bad,” said David Canon, chairman of the department of political science at the University of Wisconsin.

Although Democrat Hillary Clinton lost the election, she won the popular vote by about 2 percentage points, which was only one point off from RealClearPolitics’ polling average on Election Day—well within the margin of error for most surveys.

“[It] wasn’t necessarily much farther off than any previous presidential election,” said Scott Lasley, a former elector. He heads the political science department at Western Kentucky University.

Pollsters have missed the mark by much more in the past. When President Harry Truman took on Republican Thomas Dewey in 1948, various outlets gave Dewey a 5- to 15-point lead over the incumbent, according to The New York Times. Truman won by more than 4 percent, triumphantly holding up the newspaper headline “Dewey Defeats Truman” in a famous photo. Polls also missed Presidents Dwight D. Eisenhower’s and Ronald Reagan’s landslide victories in 1952 and 1980.

For the most part, “polls are accurate within a certain range,” Canon said. But they were wrong by a significant margin in the Rust Belt area of the Upper Midwest and Northeastern United States in 2016. Canon identified a couple of clear mistakes pollsters made that year.

First, firms did not conduct enough polls in Michigan, Pennsylvania, and Wisconsin because they assumed those locations were part of Clinton’s “blue wall”—safe Democratic states. Polls are coming in from those areas with much more frequency this time around.

And second, Canon said, pollsters did not weight their polls for education. Firms usually use weighting to correct inaccuracies in sampling. If a sample of 1,000 likely voters in a state includes a smaller proportion of African Americans than the overall state population, for example, the surveyors will weight African American responses more heavily to correct for the underrepresentation.

In previous elections, education level did not correlate with voting preferences. But in 2016, white voters without a college degree swung hard toward Trump, and that shift didn’t show up in the polls. Pollsters have corrected for that error in 2020, Canon said.

He believes those changes, along with fewer undecided voters and weaker third-party candidates, indicate Biden’s apparent lead is closer to the truth than Clinton’s was in 2016, barring a swift turnaround in the last week of campaigning.

Does that mean Americans can trust the polls? It depends. Lasley said voters need to change their expectations.

“We need to understand that polls do a good job of giving you a sense of how people are feeling,” he said. “They are not designed to predict winners and losers in close races.”