How U.S. polls got it wrong in 2016 — and why they're more likely to get it right this time
Polling firms that failed to see Trump coming have changed their approach, and so have many voters
Donald Trump's victory in the 2016 U.S. presidential election defied long odds and the widespread assumption that Hillary Clinton had the contest in the bag. That experience led many to suspect that — no matter how big Joe Biden's lead is this time — it's inevitable that Trump will prove the polls wrong again.
That's a possibility, of course. But a lot has changed since 2016.
The CBC's Presidential Poll Tracker gave Clinton a 3.4-point lead in national polls over Trump on election day. She was projected to win North Carolina and Florida by narrow margins, with Michigan, Pennsylvania and Wisconsin looking safer for the Democratic nominee.
Instead, Trump won all of these states and secured more than the 270 electoral college votes needed to win the White House. He won Michigan, Pennsylvania and Wisconsin by less than a percentage point, but these states proved decisive.
With forecasters estimating Clinton's odds of victory at anywhere from 71 per cent to 99 per cent, the result came as a shock. The assumption that the polls were wrong — massively wrong — has become part of the lore of the 2016 election.
Actually, the polls weren't far off the final result at all. Clinton did win the popular vote, besting Trump by 2.1 points. So the total error in national polls was quite small — in line with, and even little better than, past performances in national U.S. polling since 1972.
State-level polling was worse. The polls overestimated Democratic support by about four points in Michigan and Pennsylvania, about five points in North Carolina and six points in Wisconsin. But the overall performance of state-level polls was also in line with the last few decades.
The problem wasn't that the polls had missed the mark completely. The problem was where they missed the mark — which turned out to be in the few states that ended up greatly affecting the outcome of the race. If the Democrats had instead been under-estimated by a few points — as was the case in the 2012 election won by Barack Obama — Clinton's margin would have only been padded and there would have been no polling controversy.
Instead, there is probably more doubt right now about Biden's 11-point lead over Trump than there should be. The polls aren't telling the same story they did in 2016 — and they aren't telling that story in the same way.
Counting the voters who were missed in 2016
The polling error that did occur in 2016 wasn't a fluke. Something did go wrong — and many pollsters have taken steps to avoid the same thing happening again.
The politics and data journalism website FiveThirtyEight surveyed a series of American pollsters to find out what changes they had made to their methodologies since 2016. The most common change has been that pollsters are now weighting their samples by education.
The fact that many pollsters did not do this in 2016 was cited as one of the main sources of the error in a post-election report by the American Association for Public Opinion Research (AAPOR). According to the report, only half of national pollsters weighted by education in 2016.
Only a third or fewer of the firms polling the Midwestern states where Trump scored his biggest upsets weighted their results by education in 2016.
Pollsters weight their samples by different factors, such as age, gender, region or household income, to ensure that various demographics are properly represented.
Failing to weight by education boosted Clinton's numbers in some key swing states because pollsters weren't getting enough responses from voters without college degrees. These voters broke for Trump by a margin of seven points nationwide and by 37 points among white non-college graduates alone, according to exit polls.
Now that more pollsters are weighting for education, the chance of making the same error in 2020 should be reduced. But it's an open question whether these voters will be as decisive this time.
According to a recent poll by Marist, Biden is beating Trump among non-college graduates by a margin of three points. In the last Marist poll conducted before election day in 2016, Clinton was losing the non-college graduate vote by 11 points.
Biden's lead is bigger, broader than Clinton's
According to the CBC's Presidential Poll Tracker, an aggregation of all U.S. polls, Biden is leading Trump by a margin of 10.8 points among decided voters. At this stage of the 2016 campaign, Clinton was leading by only 4.8 points.
In some states, Biden is doing far better than Clinton was this far out from election day. He's leading by wider margins in Florida, North Carolina and Pennsylvania. He's ahead of Trump in Arizona and Georgia — states where Clinton was trailing by about three points at this stage of the campaign. In Texas, Trump is ahead by only 0.9 points. In 2016, Clinton was behind in Texas by nearly 12 points.
The Midwestern states are looking redder than they did at this stage in 2016. On average, Trump is doing 1.8 points better now than he was at this point in the campaign four years ago in Iowa, Michigan, Ohio and Wisconsin.
By that measure, Trump would win Michigan and Wisconsin if the polls are off to the same degree they were in 2016. That would also swing North Carolina over to his column and perhaps Pennsylvania.
But 2016's error wouldn't be enough to overcome Biden's lead in Arizona, Florida or Georgia. Without those states, Trump can't win the White House.
And the changes pollsters have made since 2016 make it difficult to compare polls in this campaign to polls in the previous campaign. If pollsters were still failing to include enough non-college graduates in their samples today, Biden's apparent polling lead in the Midwest might be even greater.
Biden is better liked than Clinton was
It's important to note that Biden and Clinton are also very different candidates.
At the end of the 2016 campaign, the RealClearPolitics average put Clinton's favourability at just 42 per cent, with 54 per cent of Americans holding an unfavourable view of her. Trump's numbers were worse (38 per cent favourable, 59 per cent unfavourable) but he performed very well on election night among those who disliked both Clinton and Trump.
According to exit polls, those voters represented 18 per cent of the national electorate and they broke for Trump by a margin of 17 points. In Wisconsin, Pennsylvania and Michigan, those voters backed Trump over Clinton by an average margin of 28 points.
As of Friday, RealClearPolitics puts Biden's favourables at 51 per cent and his unfavourables at 44 per cent. Among those Americans who dislike both Biden and Trump, one recent poll suggests Biden is beating Trump by a margin of seven-to-one.
A two-horse race with few undecideds
Another difference between 2016 and 2020 is that very few American voters are sitting on the fence this time.
At this point of the last campaign, FiveThirtyEight's national polling average suggested that 15.4 per cent of voters were either undecided or leaning toward a third party candidate. That number was still 12.5 per cent on election day — the highest such percentage since 1996.
As of Friday, it was just 5.7 per cent. That's another strike against Trump's hopes of pulling off an upset. Exit polls suggested Trump won the voters who made up their minds in the last week of the 2016 election campaign by three points nationwide, and by an average of 19 points in Wisconsin, Michigan and Pennsylvania.
That helped him a lot last time. This time, there aren't enough undecided voters still available to swing this election for Trump.
Worry about the voting, not the polls
In the end, there are better reasons to have doubts about the outcome of this election than the performance of the polls in 2016. One reason is the pandemic. The other is the man in the White House.
WATCH | First-time voter says he has 'hard decision' to make:
COVID-19 has fundamentally shifted how Americans are voting in this election. Millions more will be voting by mail. These ballots historically have had a higher rate of rejection than those cast in person. Trump has suggested — without evidence — that mail-in voting will be rife with fraud and has hinted that he would not participate in a transition of power if he deems the election to have been unfair.
In many states, restrictions and regulations will make it complicated to vote by mail and difficult to vote in person. The United States has a long, sad history of attempts to prevent some Americans — particularly Blacks and Hispanics — from casting ballots. The pandemic can make such attempts easier.
The president and his surrogates have worked to de-legitimize the voting process and have mused about court challenges after the election — something Trump himself has used as justification for appointing a new Supreme Court justice while his Republicans still have control of the Senate. After a year of protests and rioting across the United States, the threat of more violence resulting from a contested election is very real.
In light of all this, whether the polls will be off again by a few percentage points probably should be the least of anybody's concerns.