Polling in the Age of Trump: What's the point?
On November 8th 2016, the day of the last US presidential election, TIME Magazine wrote that “most agree that Hillary Clinton will win, but no one agrees by how much.” The Huffington Post boasted that they had run 10 million simulations of their own election polling data to conclude that “Clinton has a 98.0 percent chance of becoming president.” Finally, The Economist ran the headline “Hillary Clinton has got this. Probably. Very Probably.”
Fast forward four years to September 2020, less than two months out from the next presidential election between former Vice-President Joe Biden and President Donald Trump. The Economist runs the headline “Right now, our model thinks Joe Biden is very likely to beat Donald Trump in the electoral college” and has placed Biden’s odds of election at an overwhelming 87%. Surely, I can’t be the only one feeling a sense of déjà vu. How were polling companies and the media so wrong in 2016, and should we trust them again in 2020?
Why were polls so wrong in 2016?
US presidential candidates compete for control of 538 delegates, divided amongst the states, roughly, by population. The first candidate to win more than 270 of these delegates wins the election. However, as each state awards all of their allocated delegates to that state’s victor (a system known as first-past-the-post voting), a small number of ‘swing states’ effectively decide the outcome of the election. These states, such as Arizona and Florida, are not overwhelmingly Democratic nor Republican and tend to alternate between parties in recent presidential elections.
Even the slightest majority in any battleground state can have enormous ramifications for the wider race to the White House. In 2000, the Democratic nominee Al Gore lost Florida, and therefore the election, by 1,754 votes. In 2016, Clinton lost the election in the key swing states of Wisconsin, Michigan and Pennsylvania by a combined 79,316 votes, or 0.057 percent of total voters, despite winning the nation-wide popular vote by almost three million ballots. The point being that a very small number of votes in a small number of states often dictates the outcome of the presidential election.
Late Swing Vote Preference Towards Trump
A review by the American Associate of Public Opinion Research (AAPOR) concluded that in 2016 polling companies encountered a record high number of undecided voters in these key swing states and these voters ultimately decided to support Trump in the final days of the election. Given both Trump and Clinton’s poor favourability ratings, voters, unhappy with their options, may have waited until the final stages of the election before deciding. Further, as these voters were less likely to be politically anchored, they tended to be more influenced by campaign events than voters who decided early.
Likely Voter Screens Favoured Clinton
Late swing voters are largely out of the control of pollsters. Their existence can be noted in the data but their impact on the election result is inevitably hard to gauge. Similarly, pollster’s methodology for “screening-in” likely voters and weighting samples by factors such as education is more difficult than it appears. Ultimately, this was the second reason the polling in 2016 was so wrong.
Weighting for education is critical across all US polls. Over 45 percent of respondents to a typical poll will have a bachelor’s degree or higher, while only 28 percent of US adults have a degree. It is necessary that pollsters adjust their findings to account for the large number of less educated respondents. This was especially needed in the case in 2016 as Clinton fared well among well-educated voters, often up to 25 points better than Trump in pre-election polls. Crucially, this was a new phenomenon in US election history. For example President Obama held only a four-point lead over Mitt Romney among college graduates in 2012, lowering the necessity to weight by education. The AAPOR review found that less than half of states’ polls were weighted by education, leading to an incredibly skewed perception of the election’s likely outcome.
This issue is unlikely to resolve itself in 2020. As The New York Times points out, most state polls contact only registered voters, and there isn’t an authoritative and up-to-date measure of the educational composition of registered voters. The publication also points out that even amongst those state polls that did weight for education, these polls were still often inaccurate, leading to hypotheses such as that Trump voters were more likely to be politically disengaged and decline polling requests.
Should we trust the polls in 2020?
In many ways, the circumstances of the 2016 election were highly unusual and are unlikely to be repeated. For example, both candidates were historically unpopular, and Clinton won the popular vote but not the election, only the fifth such occurrence in American history.
Analysis of 2020 polling data by the University of Sydney’s, United States Studies Centre, concluded that if the statistical range of poll errors seen in 2016 reoccurs in 2020, current polling implies Trump has a roughly one in three chance of re-election. However, that finding is complicated by two facts. First, studies show that there are less than half the undecided voters in 2020 when compared to 2016, pointing towards more accurate polling data. Second, there is unprecedented uncertainty driven by COVID-19 as to who will vote and how they will vote, leading to unpredictable polls that no methodology can effectively remedy. Trump has raised concern as to the security of mail in ballots, a crucial point given that a recent poll indicated Democrats are far more likely to rely on mail-in voting compared to Republicans (72% to 22%).
There appears to be consensus amongst research bodies that polling data, while effective at predicting large national based voting trends, is highly susceptible to error on the state and local level, especially in the politically volatile swing states. Many of the late factors contributing to a candidate’s success cannot be guarded against or effectively incorporated into the polling data no matter the improvements to methodology.
Perhaps then, polls should be looked at as a gauge of sentiment rather than as a tool to predict the winner of the election. This is not the approach taken by the majority of American (and indeed international) media, whose desire for sensational headlines anticipating the future perpetuates the idea that polling data presents an entirely accurate picture of the electorate, even when our own experience as recently as 2016, tells us that this cannot be true.
Declan Curtin is the North American Regional Correspondent for YDS. He is currently studying the Juris Doctor at Melbourne University and holds a Bachelor of Arts degree majoring in international politics and history. He has a keen interest in public law and its role in international relations.