- FiveThirtyEight founder and editor-in-chief Nate Silver has a message: while the polls in 2020 did underestimate Republican support, they weren’t horrendously wrong in the grand scheme of things.
- Biden led polls at the national level by 8.4 percentage points on Election Day in FiveThirtyEight’s average, but is set to win the national popular vote by around four percentage points instead.
- “The simple fact is that polls miss on average by about three points so for three-and-a-half or four points, that’s pretty normal,” Silver said on the latest episode of the FiveThirtyEight politics podcast.
- Silver also wrote on FiveThirtyEight’s website that polls performed pretty well given low response rates, the possibility that voters could change their minds, and the difficulties of surveying in a pandemic.
- Visit Business Insider’s homepage for more stories.
FiveThirtyEight founder and editor-in-chief Nate Silver, one of the best-known election polling analysts and modelers, says that 2020 polls were not horrifically wrong, and we don’t need to discard them altogether.
In 2020, the polls and election forecasts like FiveThirtyEight’s predicted the correct winner both nationally and in key swing states, unlike many polls conducted of the 2016 presidential election.
But polls underestimated Trump’s vote share and Biden’s margin of victory nationwide and in several key swing states, in addition to largely underestimating Republican support in US Senate and House races, spurring initial post-mortems into what went wrong.
As of November 13 with most of the vote counted, Biden leads Trump by four percentage points, 51% to 47%, indicating that while the national polls correctly captured Biden’s share of the vote, they underestimated Trump’s share, leading to a polling error of around four points.
But when we zoom out and look at the broader picture, Silver says, the polls are (mostly) alright.
“They were okay,” Silver said of the polls on the latest episode of the FiveThirtyEight Politics podcast. “If you wind up with a three or four-point polling error for the presidency in national polls and in key swing states, the simple fact is that polls miss on average by about three points so for three-and-a-half or four points, that’s pretty normal.”
After 2016 and now in 2020 too, Silver has worked to dispel the myth that 2016’s errors were uniquely bad and that polling is getting worse over time, pointing out that historically, national presidential polls since 1972 missed the final result by 2.3 percentage points on average.
So while 2020’s polling error was slightly larger than the average over the past half-century, national polls were also four to five points off in the 2012, 2000, and 1996 presidential elections, according to FiveThirtyEight’s analysis.
“As someone who analyzes the polls and models the uncertainty, this was very normal. It’s very common. We’re not even on the edge of the distribution,” Silver said in the podcast. “I think part of what pollsters do is as part of a communications strategy, they sometimes want to make it seem like their instruments are more exact than they really are in practice. In practice, there are many things that make polling difficult and so you hope to get close and you usually do get close, including this year, for the most part.”
At the state level, polls most severely missed the margin between Biden and Trump and underestimated Trump’s vote share in Maine’s 2nd Congressional District, which accounts for one electoral vote, Wisconsin, Iowa, and Florida, according to FiveThirtyEight.
Polls much more accurately estimated the tight race between Biden and Trump in Georgia, and slightly underestimated Biden’s margin of victory in Colorado and Nebraska’s 2nd District.
Election analysts like Silver have put forth some preliminary theories as to why many of the polls underestimated Trump’s support nationally and particularly in states with lots of non-college-educated white voters, as they did in 2016.
In the wake of 2016, pollsters adjusted their samples to more properly account for the difference in vote preferences between college-educated and non-college-educated white voters, a method known as weighting the sample for education levels.
Pollsters survey random samples of the population and weight those to be representative of the population writ large. But that method is only effective at correctly predicting results if if the people who answer the polls are representative of the ones who don’t answer. And this year, some analysts argue, that may not have fully been the case.
And the results so far have led election data analysts including The New York Times’ Nate Cohn and Democratic data guru David Shor to hypothesize that the polls may have suffered from non-response bias overall and within subgroups, with Democrats and the voters who are most politically engaged more likely to pick up the phone and take a poll than the less engaged, lower-propensity turnout voters that swung to Trump.
But as Silverargued in a piece published on FiveThirtyEight’s website, it’s impressive that polls turned out to be as accurate as they were in 2020, given the persistently low response rates for public opinion surveys, the possibility that voters could change their minds after taking the poll, and the difficulties with developing samples and weights that accurately represent the population — not to mention the challenges caused by polling in the COVID-19 pandemic.
“There are a whole bunch of new problems that crop up that you didn’t necessarily anticipate ahead of time, some of which may be permanent, some of which may be temporary, some of which may relate to COVID, or Trump, or mail voting, or whatever else,” Silver said in his podcast. “But in the grand scheme of this election, the polling, A), was not that bad, and B), was like, the 15th most important story.”
Friends, this isn’t the time to be complacent. If you are ready to fight for the soul of this nation, you can start by donating to elect Joe Biden and Kamala Harris by clicking the button below.
Thank you so much for supporting Joe Biden’s Presidential campaign.