While it might be too early to know just how far off political polls were in relation to the outcome of Nov. 3 races, it is not too early to say that they were wrong. Again.
The News interviewed Gregory Huber, Alexander Coppock and Jasjeet Sekhon — three Yale political science professors who conduct research related to polling — to learn more about the industry as a whole, what their thoughts are on what went wrong in the most recent election and what the future of polling might look like — if there is one at all.
“I do think polling is in this weird middle ground currently, in the sense that the standard way, the scientific justification of random sampling, because people don’t respond to you that much, that’s fallen apart,” Sekhon, a professor of political science and statistics and data science, told the News.
Coppock, an assistant professor of political science, works to understand political persuasion and the reproducibility of social science research, especially in the context of elections. While Coppock said he can’t definitively explain why the polls were wrong, he has numerous papers on what he considers to be a main misconception: the “Shy Trump Supporter” hypothesis.
This hypothesis claims that Trump voters are too embarrassed to tell pollsters whom they voted for in fear that the pollsters might think that the voter is racist or sexist.
“The main thing I want to communicate … is that some people might think that that might be a reason why the polls are wrong, but I don’t think that’s true,” Coppock said.
Proponents of this hypothesis, according to Coppock, point to higher support for Trump in online polls versus those that involve directly interacting with a pollster. However, Coppock adds, that can also be attributed to different types of people taking online or telephone polls.
He points to list experiments — in which people can reveal their true beliefs without directly sharing them with the person administering the survey — as evidence against the hypothesis. In the experiment, voters are divided into two groups. Both groups are asked about their opinions on a specific policy, but one group is also asked about whether they plan to vote for a specific candidate. Respondents add up the number of policies they support, and since pollsters do not know which specific policies a respondent supports, the assumption is that answers will be more honest. Subtracting the number of those who were only asked about policy decisions from everyone asked directly about a candidate will, therefore, give researchers a more truthful estimate of Trump voters. And that estimate is similar to the result of a direct poll, which, according to Coppock, refutes the idea that Trump voters are lying on polls.
“The real explanation here is that so few people respond to polls that you have to do all this wizardry on the backend to reweight the data in order to get an estimate of support for politicians, and the wizardry is just a little wrong,” he said.
Huber, who chairs the Department of Political Science, studies American politics and the attitudes of Americans surrounding different issues. For him, one of the challenges of polling stems from the advent of caller identification and the shift to cell phones. With it, fewer people answer their phones for polls, which makes representative sampling difficult. Poll analysts have used weighting in an attempt to make relatively small data sets more representative.
But Huber said the problem is that weighting assumes the people who do answer the phone can stand in for the rest of their demographic, which is not necessarily the case.
Huber noted, for example, that more wealth once indicated that a voter was more likely to be Republican, but that relationship is “less clear” now. He added that there might be more of an inclination for Democratic voters to talk to pollsters.
“If that’s true, that’s a really hard problem,” Huber said. “[We] seem to be missing people systemically,” he added.
The solution? Huber told the News that the key to more representative data is more funding to reach more people and hire more pollsters. But he also said that money might only be a “partial” answer and that more expensive surveying efforts have not completely solved the fundamental problem of sometimes missing entire groups of people.
He also stressed that accurate polling has “always been a problem,” although that problem has been exacerbated in the last 20 to 30 years. Huber cited as an example a famous 1936 poll by the Literary Digest that incorrectly predicted that Alfred Landon would overwhelmingly win that year’s presidential election — instead, Franklin Delano Roosevelt won 46 states.
But he also remarked that political polling is held to a “hard standard” compared to other types of polling, like predicting the outcomes of sports games. The New York Times, for example, previously had a column forecasting the exact score of NFL football games. According to Huber, the column never once correctly predicted the results of a game, even after running for several years.
Sekhon studies machine learning and statistics and has previously worked on polling and survey design. Similarly to Huber, he noted the assumption that those who answer polling surveys have similar political beliefs to those who don’t answer surveys may not be true.
“That’s not like a statistical assumption, that’s just like a behavioral assumption that there’s no reason that has to be true,” Sekhon said. “It would just be convenient for us if it were.”
For Sekhon, it’s not surprising that the polls are wrong, but rather that it took this long for everyone to recognize them as flawed. He credits the Trump era — having a Republican candidate be more anti-trade than the Democratic candidate, shaking up the political alignment of policies associated with certain parties — with bringing those problems to the forefront of discussion.
Sekhon said that the problems with polling are more than an interesting phenomenon — they have real-life consequences.
Sekhon cited the run-up to the 2016 election, in which 2016 media coverage of Hillary Clinton was more negative than that of Donald Trump, as an example.
“Partly [the negative coverage] could be an overcorrection. … [Reporters] want to be nonpartisan and they correct for it,” Sekhon said. “But I think also part of it is I think they thought based on the polls, based on their own prejudices, ‘Oh, Hillary’s gonna win. So we can do this.’ So these things have consequences when polls are wrong.”
As for how to move forward, none of the three professors had definitive answers. But they all agreed that something needs to change.
“The approach that we’ve done for the last two cycles is not good enough,” Coppock said.
Some popular political polls include FiveThirtyEight, The New York Times and Gallup.
Madison Hahamy | firstname.lastname@example.org