PITTSBURGH -- The crowds attending Joe Biden’s post-debate “whistle stop tour” in Ohio and Pennsylvania looked like the kind of voters Hillary Clinton attracted in this state when she campaigned here four years ago: predominantly white and middle-aged.
Every stop in Pittsburgh, Latrobe, Greensburg and Johnstown replicated what Clinton drew here and in Philadelphia in the weeks leading up to the general election in 2016; most of the time, those demographics only became younger or more racially diverse when a concert was included with the speeches.
While Biden consistently leads President Trump in the polls in Pennsylvania -- the RealClearPolitics polling average shows him up 6.3 percentage points -- Clinton had a similar advantage a month from Election Day four years ago. In 2016, FiveThirtyEight’s Election Day model estimated her chances of winning the Keystone State at 77%.
Understandably, this track record gives Trump supporters solace, even as Biden’s backers are confident the race is heading their way. The Biden camp also believes their voter demographics are going to look much different than Clinton’s – and even different from those showing up at Biden’s rallies.
Yet academics and strategists on both sides of the aisle are unsure if a repeat of 2016 is in store. The underlying questions are whether pollsters are biased and if we have entered an era, at least with Donald Trump at the top of the ticket, when the norms of data collection are upended by a lack of understanding who is going to vote. In pollsters’ language, the problem is simple: Are they sampling the right voters?
Youngstown State University political science professor Paul Sracic is among those wondering if the models are accurate. Four years ago, he was convinced by the polling data that Trump would lose. Yet he couldn’t shake the anecdotal evidence he saw in the Steel Valley region and the nagging feeling that something different was happening in the famed Democratic “Blue Wall” of Michigan, Ohio, Wisconsin, and Pennsylvania. His instincts proved to be more accurate than the polls.
If the 2020 election tightens, the same states are the key to reaching the magic number of 270 electoral votes. As happened last time, when a Democratic candidate starts the race with a national lead, a winning Republican candidate must come close to running the table in the swing states.
Although Trump looked certain to win Ohio and had a narrow lead in Florida four years ago, hardly any polls showed him ahead or even close in Michigan and Wisconsin. Certainly Pennsylvania tightened in the final days, but even there Clinton led Trump by nearly two percentage points in the RCP average.
“I knew he was going to do well with working-class voters, but the polls weren’t showing that they would make a major difference in enough of the swing states,” Sracic told me recently. “As it turned out, the state polls were wrong because they were not weighted by education, and therefore undercounted non-college voters.”
In the weeks ahead of the 2016, I spent hours either on the phone or visiting Sracic in person at YSU trying to understand and dissect what I saw anecdotally in Pennsylvania in reporting and looking at the trends in Pennsylvania before that election night. Both of us kept coming back to the fact that while the state had given the Democratic nominee its electoral vote since 1992, every four years the returns had shown Pennsylvania becoming .04% percent more Republican.
Sracic would argue that the numbers just aren’t there; I’d counter that the historical trend was unmistakable. Four years later we are still struggling with that same conundrum, although this time he is less skeptical of a possible Trump win.
This remains a tricky part of polling in 2020. Surveys that show Biden attaining numbers that Barack Obama didn’t earn in 2012 depend on modeling that predicts 33% of this year’s electorate is going to be 39 years old or younger. That under-40 cohort poll does heavily favor Biden, but does this methodology accurately reflect the percentage of young people who really will show up on Election Day?
“There is some guesswork here, since the election has not yet occurred. People who are going to vote in 2020 may not have even registered yet in some states,” explains Sracic. “The most difficult voters to capture in your model as a pollster are brand-new voters -- since so much of the guidance that a pollster relies on comes from past elections.”
Meanwhile, other data points indicate that Pennsylvania is trending more conservative ahead of next month’s election. Republicans have added nearly 198,000 more registered voters since 2017 and a staggering 276,648 people bought a gun for the first time this year in the Keystone State, including a significant percentage of women and African Americans. Are these cautionary flags for Democrats that pollsters ignore at their peril?
“One of the things to look at in the polls that a lot of people aren’t considering is the number of undecided among African American voters,” said Sracic, who wonders aloud whether they could be this year’s version of the elusive (and, some say, mythical) “shy Trump voter.”
Another possible source of hidden Trump support, he says, are first-time voters (but not those who just turned 18), whom pollsters typically don’t capture because they are looking for someone with a history of showing up on Election Day.
In 2016, the uniqueness and energy behind Trump’s campaign attracted voters who were never inspired by Mitt Romney or Barack Obama and who were fairly non-political. The brash outer borough New Yorker was attractive to them precisely because he was an anathema to the political establishment. Think “Drain the Swamp.”
Traditionally, the Democratic Party spends money on voter registration efforts and its candidates benefit from that investment. But four years ago, undetected enthusiasm for Trump in counties such as Luzerne, Erie, Washington and Westmoreland – two of which went for Barack Obama in 2008 and 2012 -- meant that Democrats’ new-voter drives may have had the effect of helping the Republican ticket. Moreover, Trump himself seems to have been a draw: Despite little in the way of organized GOP registration or get-out-the-vote efforts, turnout in those counties increased in 2016.
This movement of infrequent voters, new voters and shy voters was an important but undetected sliver of the president’s win in a very close election. Can the polls be missing such voters again? One clue that should worry Democrats: In several Pennsylvania counties there has been a significant number of legacy Democratic voters who have flipped their registration from Democrat to GOP. Is it possible it’s Groundhog Day again in the state?
“The 2016 state polls were not the only example of polls underestimating the final vote count in an election,” said Professor Sracic. “It is fairly common for polls to be off as much five points or more, and that is why polling aggregators often stress that they are providing the probability of win, not the certainty. Aggregating or averaging polls is an attempt to counter polling problems by assuming that they won’t be replicated by different pollsters using different methodologies.
“This, however, may lead to one source of error known as ‘herding’ -- you may not want to be the one polling company finding numbers that are very out of line with what other firms are publishing,” he added. “We don’t have a lot of evidence to prove herding, but many experts seem to assume that it exists.”
“The real problem in 2016 was a failure to weight for education,” Sracic also told me. “Pollsters just don’t call up a thousand people and report how many said they would vote for each candidate. Instead, and the real art in polling, is in weighting the results to match the demographics of the group you are analyzing."
If you call 100 people and 60 are females while 40 are males, you are going to weight the male response a bit higher if you expect a 50-50 turnout on Election Day. In 2016, the weighting problem was obvious, but only in hindsight: White voters without college educations broke for Trump. The X-factor in 2020 might be the coronavirus pandemic: How it will impact voter turnout is anybody’s guess. That uncertainty is exacerbated by the low survey response rate that has plagued polling since the advent of cellphones.
“I think that the biggest problem pollsters might have this year would involve non-response bias,” Sracic said. “We don’t know exactly how many people refused to answer their phone when a pollster calls or refused to answer questions if they did take the call. Some estimates, however, are higher than 90%.”
Even those who do respond may not pass a “likely voter screen,” since the most basic way to screen for likely voters is to determine whether they have voted in the past. If they haven’t, these Trump voters could be out there in the hills and hollers of rural Pennsylvania, not “shy” at all – but hiding in plain sight and even answering their cellphones. But they’re possibly still being undercounted in a state where just a tiny slice of new voters could determine the outcome of the 2020 presidential election.