Polls face challenges in predicting outcomes
We are coming around the last turn and getting close to the sprint to the finish, with a few incredibly important debate hurdles to navigate. A little more than four years ago, Donald Trump won a majority of Electoral College votes and was elected president. This was a surprise – apparently to Trump, and to most of the rest of us. Most reputable surveys had indicated a victory for Hillary Clinton.
Any single survey has a real chance of generalizing from an unrepresentative sample. Pooling all recent surveys reduces the chances of a random outlier. But that doesn’t address two major challenges: failure to capture last-minute trends and failure to accurately project who would actually turn out to vote.
Last-minute trends
The final Realclearpolitics national poll average before the 2016 election showed Clinton with a 3-point lead. The actual margin was 2.1% This was hardly a great miss. The national numbers, of course, don’t predict outcomes in a federal election.
Presidential elections were intentionally set up not to be national contests, but to preserve the role of the states in the federal system. The outcome in 2016 boiled down to contests in a few key states. The final average margin for Clinton in Pennsylvania was 2.1%. Trump prevailed by 0.72%. In Michigan the final average was 3.2%.
Trump won by 0.23%. In both cases the polls were off by about 3%.
Final polls in key states missed by about 3%. Why? Some of it was about late trends, which seemed to be in Trump’s direction. Another was a failure to account for turnout, especially in battleground states. Depressed African American turnout in Michigan clearly cost Clinton. The same pattern emerged in Pennsylvania and Wisconsin.
It’s not accurate to make much of supposed poll failure in 2016. There was clearly a late Trump trend. There is too much statistical error in any poll to expect perfect accuracy. Most state polls ended days before the election.
Who will actually vote?
The main failure of survey research in forecasting elections is tied to making predictions from likely voters. So how does one decide if a respondent is likely to vote? It is impossible to have certainty about whether one will vote. It’s pointless to ask directly. People know that they are expected to vote and will almost universally indicate their intent to do so.
Researchers use responses to a variety of questions to classify a respondent as a probable voter. They ask how much they care who wins, how closely they follow the election and, if available, the person’s voter history. Estimating who will actually vote is the secret sauce of election polling.
Surveys’ value lies in isolating the votes of key groups. Trump has lost the support of some older voters and college-educated women who were key to his 2016 victory. He appears to have gained a bit among Hispanics.
COVID complication
The COVID-19 pandemic further complicates making sense of who will actually vote, and when. Many may neglect to send in paper ballots. Many will plan to vote on election day and then stay away. Turnout should be significant this year despite the virus, but modeling it will be difficult. A growing challenge is the spreading out of the time of the vote, especially with mail-in voting.
This year the national and state polls are again indicating a Trump loss, and by a greater margin. There are many weeks to election day, but mail in voting has begun. Joe Biden’s lead should be well over 3% and stable during the days preceding the formal election day, especially in the battleground states, if he is to win.
The election process is under attack as never before. The result may not be clear on election night. Nothing is served by continuing to question the legitimacy of the voting process. We are all served by a clear outcome and an end to the chaos.
William Lyons worked as a professor of political science at the University of Tennessee and served for more than 16 years in a number of policy-related roles for Knoxville Mayors Bill Haslam, Daniel Brown, Madeline Rogero and Indya Kincannon.