If the polls are to be believed, Democratic nominee Joe Biden is the favorite to win the 2020 election against his Republican opponent, President Donald Trump. At the time of this writing, FiveThirtyEight.com, which aggregates and analyzes polls, gives him a 76 percent chance of winning; all eleven of the most recent polls listed at RealClearPolitics predict a Biden victory with an average spread of almost six points.
This article first appeared in Salon.
And yet — as anyone who followed the 2016 election remembers — pollsters heavily favored Clinton to beat Trump when she was the Democratic nominee. Clinton supporters who felt confident of her victory on election night 2016 have a right to feel once bitten, twice shy about trusting polls again. Should we all feel similarly suspicious of polling firms in the lead-up to the 2020 election?
To answer that question starts with understanding the science of polling — and how it works (or fails, in the case of the 2016 presidential election).
"It's sampling theory, which is just the idea that if you interview a subset of the population [and] if the sample is drawn randomly from that population, then the distribution that you get of responses should reflect the characteristics of the entire population within a known margin of error," Dr. Alan Abramowitz, a professor of political science and polling expert at Emory University, told Salon. "That's the science and that can be used not just for human populations, but in lots of other fields like quality control, where you might have an assembly line and you draw a sample of the products is to test for defects. It's used in a lot of different fields."
As Abramowitz explained, there are good reasons to question the accuracy of polls.
"The problem with public opinion polling nowadays is that because response rates are so low to telephone polls, you start using internet polls [and] you have to use a lot of special techniques to try to come up with a sample on the internet that matches the characteristics of the population," Abramowitz pointed out. "It's very, very hard to really say how accurate the polls are."
Abramowitz emphasized that post-2016 polling had been quite good. "They don't always get it right, but in the 2018 midterm elections, if you went down and looked at the polling average for all the races — like governor, senator — and where they were at polls for House districts, by and large they were pretty accurate."
Dr. Christopher Wlezien, a professor of government at the University of Texas – Austin, had similar opinions about contemporary polling practices. Wlezien said the history of American polling began in 1916 when Literary Digest asked readers to inform their editors by mail about how they were going to vote — a non-scientific approach, as Literary Digest readers did not represent a random sample of the population. More scientific polling started being used in 1936, after Gallup correctly predicted that President Franklin Roosevelt, a Democrat, would defeat Republican challenger Alf Landon using scientific methods. (Literary Digest infamously predicted a Landon victory.)
At the same time, Wlezien explained, the challenges of conducting accurate polls as the main methods for conducting them have changed. Mail-in polls have been replaced with door-to-door polls, telephone polls and online polls over time, with each practice bringing with it certain sets of advantages and disadvantages. The overall goal, always, is to make sure that the sample size is large enough and diverse enough to accurately represent the larger population, so that the data which pollsters provide to the public is statistically likely to correlate with actual public sentiment.
So did a systematic error emerge when Clinton was viewed as the likely winner over Trump in 2016?
"The American Association for Public Opinion Research conducted an analysis of the 2016 polls," Wlezien explained. "And some of it was published as a big report and there's a smaller side that was published as a journal article. I happened to be on that committee on that task force. We have some familiarity with what we did and basically what the analysis shows is that the national polls were pretty good. In fact they were probably better than they'd been in the past."
He added, that while many of the state polls were good in 2016, "a lot of the state polls were not, particularly the ones in the decisive states. And we learned a lot in that report about what the polling organizations did wrong in their polling and it has a lot to do with their underweighting non-college types. And the ones that got that right were a lot more accurate than those who didn't, which was quite a number got that wrong. And even to the extent some were getting it right, a lot of pulling organizations pulled out of some states early, so we couldn't really gather any of the late movement, which was toward Trump."
He noted, hopefully, that "my understanding is that pulling organizations are aware of what went wrong in 2016 and the states, and some states at least, they've corrected that."
In the end, the key to properly understanding polls is to realize that they are not meant to be prophetic.
"The problem is, of course, that polls are not predictors," Dr. Allan Lichtman, a historian at American University who studies the science behind elections, told Salon. "I'll repeat that: Polls are not predictors. They are abused as predictors. The easiest thing in the world is to write a story about polls. You don't even have to get out of bed in the morning, just read the polls and write the story. And of course, we also have compilers of polls like Nate Silver [the statistician who founded FiveThirtyEight.com], whose forecasts are no better than the polls themselves. So that was the real problem with 2016, was to use the polls as predictors when they are not predictors."
Lichtman emphasized that state polls can be unreliable, and "tend to have a large margins of error." He also noted that "the national poll was just a point or two off of the plurality that Hillary Clinton compiled" in 2016, meaning that they were not far off the mark.
Lichtman also reviewed his own system of predicting presidential elections, one that has successfully anticipated the results of every presidential election since 1984. It uses a series of true-or-false statements that anticipate whether the incumbent party's presidential candidate will be elected in a given year. If six or more of the statements are false, the incumbent candidate will lose; if fewer than six are false, he or she will win. His system has caused him to make predictions that were rejected by the expert consensus at the time: He anticipated that Republican candidate George H. W. Bush would defeat Democratic candidate Michael Dukakis in 1988, even when Dukakis was ahead in the polls by 17 points, and was one of the few pundits to publicly predict that Trump would beat Clinton in 2016.
"I'm sitting here in my study and I have over my shoulder a note written on the Washington Post interview where I predicted Trump's wins. And the note says, 'Professor, congrats! Good call!' in a big Sharpie letter signed 'Donald J. Trump.'"
Yet while Trump was appreciative of Lichtman's prediction in 2016, the professor noted that "he didn't understand the deeper meaning of the keys, which is that governing, not campaigning, counts. And when you are the incumbent president, you're going to be judged by your record. And rather than dealing substantively with these challenges, [Trump] reverted to his 2016 playbook when he was a challenger and the result was a disaster for Trump's reelection prospects."
In 2020, Lichtman predicts that Biden will beat Trump, citing the poor economy, widespread social unrest, his scandals, his lack of major foreign policy successes, his losses in the 2018 midterm elections and the fact that he is not popular (or "charismatic") beyond his own base. Lichtman's only caveat was that, because Trump has slowed down the post office and has engaged in voter suppression, the election may be "stealable."