US election polls look good for Joe Biden. But can they be trusted?

 

US election polls look good for Joe Biden. But can they be trusted?

 

After Donald Trump’s surprise victory in 2016, some Democrats are reluctant to believe the numbers

 

Tom McCarthy

 @TeeMcSee  Email

Mon 28 Sep 2020 08.30 BSTLast modified on Mon 28 Sep 2020 08.31 BST

https://www.theguardian.com/us-news/2020/sep/28/us-election-2020-are-polls-accurate

 

The numbers are looking pretty good for Democrats in Arizona. The party’s Senate candidate is up more than five points in polling averages, and Joe Biden leads Donald Trump by more than three points in the same averages. If those numbers hold, the state could hand Democrats the Senate and Biden the White House in one fell swoop.

 

But a key Democratic organizer in the state can’t say whether he thinks the numbers will hold – because he does not believe the numbers exist in the first place.

 

“I would say the polls are a mirage,” said Larry Bodine, president of the Democrats of Greater Tucson group. “After 2016, I decided from then on, I was just not going to rely on what the polls had to say, and instead rely on what my fellow Democratic volunteers encounter out in the field.”

 

Inside Democratic party offices from coast to coast, and under a good number of roofs where anti-Trump voters dwell, polling results that show promise for Biden and down-ticket Democrats are being handled with a similar mix of arms-length trepidation and not-today-Satan refusal. Feeling they were misled by polling to believe that Hillary Clinton was a shoo-in in 2016, only to be ambushed by Trump’s win, many progressives in 2020 vow that they’re done with the numbers game.

 

“I’m really active in Democratic circles and pretty much nobody talks about the polls,” said Bodine. “I believe that all the positive polls do is give a false sense of security.”

 

That attitude seems to have little downside for political organizers. But the question of the reliability of polling has broader implications for campaigns, for public policy – and ultimately for daily American life, on issues ranging from racial discrimination in interactions with police to skepticism about a potential coronavirus vaccine.

 

Ultimately, the health of polling is bound up in the health of the democracy, analysts say. Asking people what they think is, among other things, an expression of faith that what the American people think matters – a notion that can seem even more worthwhile amid Trump’s demand to “get rid of the ballots” in November.

 

With those stakes hanging overhead, and under intense public scrutiny on the eve of a watershed election, pollsters across the country have made adjustments to address their mistakes of 2016 and are working hard to capture an accurate snapshot of 2020.

 

The picture is not simple. While some key state polls were off in 2016, the national polls in aggregate were right on target, showing Clinton three points ahead at the end; she won the popular vote by two points but lost in the electoral college.

 

The mistakes last time, according to a full buffet of postmortem analyses, included: pollsters did not have an eye on educational attainment as a potential fault line in the electorate; they were foiled by an unusual wave of undecided voters breaking for Trump at the last minute; there was too little polling in key swing states to really know what was going on; conclusions extrapolated from that paucity of data were broadcast with far too much certainty; and there might have been some “shy” Trump voters who didn’t want to say they were supporting him.

 

The results remain a political shock. In the final days of the 2016 election, the “average” of the scant polling in Wisconsin had Clinton ahead 6.5 points. In Michigan, Clinton’s average “lead” was 3.6 points, while in Pennsylvania it was what looks in retrospect like an extremely tenuous 2.1 points. Yet Trump won all three states and with them the White House. The immediate criticism of the Clinton campaign was that it had failed to visit the upper Midwest, taking the voters of Michigan and Wisconsin for granted, lulled by the siren song of reassuring polls.

 

But have the polls improved since then?

 

Changes since last time

The well-known and widely followed Franklin & Marshall College poll based in Lancaster, Pennsylvania, had Clinton ahead by double digits in the state in its last poll before the 2016 election. Trump won the state by a razor margin of fewer than 50,000 votes, or less than a percentage point.

 

The director of the poll, G Terry Madonna, said an unusual wave of late-deciding voters mostly breaking in the same direction – toward Trump – created the polling blind spot.

 

“In some cases, including ours, we were out of the field, meaning we completed the interviews, 10 days before the election,” Madonna said. “What we found in exit polls was that in the last 10 days 20-some per cent of voters made up their mind or they changed their mind and then went for Trump far more than for Clinton.”

 

Madonna said this year the poll would stay in the field longer – and he sees fewer undecided voters this time.

 

“When you have 85-90% of Republicans saying they approve of the job Trump’s doing, and Democrats are in single digits – people are locked in in this race,” Madonna said. “There’s a relatively small number of undecided voters. And it may turn out that what they do might make a difference, but it’s probably more important for the campaigns to get out their base of voters.”

 

For readers who have decided not to ignore the polls: Franklin & Marshall released a poll on Thursday morning that showed Biden up by 6 points in the must-win state, where averages have Biden ahead by 4.1 points.

 

Other state-level pollsters, including the Muhlenberg College Institute of Public Opinion in Bethlehem, Pennsylvania, have expanded their methodologies since the final days of 2016.

 

Christopher Borick, the poll director, said that this year the poll was not only looking to the most familiar categories for insights on voter behavior – gender, age, region, party and race – but had also added one more category: educational attainment.

 

“We had never included educational attainment as a variable in our weighting formula, largely because it never mattered,” Borick said. “When you went back historically over time, if you had weighted, it didn’t do much – it was a wash. But now we’re seeing more of a divide, where the upper-level attainment are voting one way and the lower are voting the other way. And that really started to emerge in this decade and blossom in 2016, to the point where I think it’s silly to ignore that.

 

“And so we are now weighting with education attainment, and it does slightly move the polls. So our last poll had Joe Biden up four, with an educational weight built in. If we had not done that, like we did not in 2016, his lead would have been six points.”

 

Communicating the limits of polling

Of 453 pollsters ranked by the FiveThirtyEight data analysis web site, the Muhlenberg College Institute of Public Opinion survey is one of only six to be awarded the top rating of A+. The poll gets the top rating based on its reliable track record, minimal observable bias, methodological rigor and the fact that it does the expensive, difficult, time-consuming kind of polling, meaning live telephone interviews, including calling cellphones.

 

Muhlenberg’s last poll of the presidential field in Pennsylvania before the 2016 election had Clinton with a narrow, single-digit lead. The actual result – Trump won the state by less than 1% – was within the poll’s margin of error. Statistically speaking, the poll was not wrong.

 

 

But Borick points out that when the gap between a poll number and an election result falls across the line separating winner from loser, it’s impossible to tell anyone the poll was not wrong.

 

“If you looked at 2012, the polls were just about as off with Barack Obama against Mitt Romney, and they understated Obama’s performance that year,” Borick said. “But no one really cared at the end, if Obama won by one point or four points or five points, because it was all on the same side of the ledger. The error was going in the direction of the person that won it anyway.”

 

“If you cross over to the other side, by even one vote, two votes, half a per cent, whatever – it changes the whole outcome, but the math isn’t all that different.”

 

This election cycle could prove unusually challenging for pollsters because of a significant climb in the number of voters casting ballots early and via mail, said Nate Silver, founder of FiveThirtyEight. Silver forecast all 50 state results correctly in the 2012 presidential election and was one of the few polling analysts to articulate clearly on the eve of the 2016 election that Trump had a path to victory.

 

“I do think the transition to high rates of mail voting is one of the bigger potential sources of polling error, especially with the mail vote likely to be disproportionately Democratic, although it’s hard to know in which direction the error might occur,” Silver tweeted this week.

 

Borick said it was down to pollsters, especially academic pollsters, to explain what their results mean to a public that does not always think about election outcomes in terms of percentage likelihood and margin-of-error.

 

“I think pollsters, people that do public opinion research, want to be able to give citizens a sense of where the broader public is on issues, on races, and to do that accurately. And to also educate about the limits of the very things we do,” Borick said.

 

“We sample. We take small groups to make inferences about big groups. And that inherently has error involved in it. And trying to communicate the reality of what we do is a big role for us.”

 

Bodine, the Arizona organizer, says those lessons are not for him.

 

“Democrats should not take anything for granted, and my advice for them is to call up their local Democratic party and get active,” he said. “Stop yelling at the TV, stop complaining about what they see on the news and get out there and do something about it.”