Pollsters in the United States have declared the 2024 presidential election a statistical dead heat. After months of polling, they do not provide even an inkling of whether Kamala Harris or Donald Trump will be declared the next President on Tuesday. Those of us watching from afar are forced to make guesses about the election based on “vibes” and “gut feeling”.
Also Read: Decoding Kamala Harris’s India equation: What will her win in US election mean for New Delhi
Given India’s regular misadventures with political polling, it is reasonable to ask whether pollsters in the US have missed a systematic trend towards one of the candidates. After all, it is clear that pollsters underestimated Donald Trump’s support among key demographics in both the 2016 and 2020 elections, although most major pollsters claim to have made adjustments to prevent these sort of mistakes. It seems that even the American pollsters are afraid to be wrong. Josh Clinton, a political scientist at Vanderbilt University, has suggested that cautious pollsters are “herding” themselves around a dead heat because they are afraid to go out on a limb among all of the statistical noise. He has also shown that perfectly defensible assumptions by pollsters could shift the results up to 3 percentage points in either candidate’s favour — more than enough to push the election to the Democrats or the Republicans.
How should we interpret the data that we see, and what are sources of biases in election prediction?
Also Read: Melania Trump leaves NYC with son Barron for next ‘big move’ ahead of the US elections
Let us begin by laying out a few differences in the US from what we are used to seeing in India. First, the quality of polling in the US is significantly higher, with many of the best pollsters using sophisticated techniques for sampling and predicting “likely voters”. Second, what we see today in the US are “pre-election” polls; in India, the pre-election polls are mostly useless and cannot be released just before an election due to the Model Code of Conduct. Third, much of the reporting on election polling we see today are passed through “aggregators” using statistical techniques such as Nate Silvers’s FiveThirtyEight, which account for the quality of pollster, the bias it displays towards Republicans or Democrats, and the quality of the underlying sample. This certainly yields higher quality inferences than focusing on a single pollster who either dances or cries on television.
The quality of prediction from pre-election polls are dependent on two factors. First, the quality of the sample (its size, its representativeness, etc.) tells us the extent to which the survey accurately captures the political preferences of each key demographic group. Second, the accuracy in projection of who is likely to vote (through a likely voter model), or more precisely, the relative turnout rate among different demographic groups, will affect estimates. The final prediction is basically the predicted vote of each key demographic group, weighted by how much it is expected to turn out to the polls.
Let us begin with question of sample quality. Like in India, the proliferation of polling agencies and political party consultancies has meant a certain fatigue with answering the survey, not to mention a greater share of people who value their privacy. In the US, in-person surveys are seen as intrusive, and the best surveys are still conducted over the phone — but response rates have dropped over time. The New York Times/Siena poll, widely believed to be among the best , which makes multiple calls to a voter for a survey reports only about a 2% response rate (the fact that the predictions in the US are so much closer than those in India is a testament to sophistication of their statistical machinery as compared to India). There are online polls and SMS polls as well out there, but these are still viewed as lower quality. With such low response rates, there is always a fear that the most reticent responders will have voter preferences than those who answer the phone.
The second challenge is predicting a “likely voter” from the set of registered voters. The 2020 US election saw a 66% turnout, the highest recorded turnout since 1900. But that still means that one-third of registered voters did not show up to the polls. Those that don’t come out tend to have certain demographic characteristics — they are younger and poorer. One of the stories we hear is that turnout models have had a hard time predicting turnouts in many parts of battlegrounds states where the election is close, such as Michigan and Pennsylvania. Still, in a polarised electorate, even those who rarely vote may be agitated enough to come to the polls. This is perhaps why the closing arguments for each campaign have focused on “triggering” core voters, such as almost comical fighting about who called whom “garbage” and why.
My own view is that the US has a good handle on the relative level of support among different key demographic groups (and there is a view that as voting day approaches, demographic groups such as women and Black voters shift in predictable ways). What is harder to understand is patterns of turnout. It’s anybody’s guess as to how much and whose voters in critical counties in states such as Michigan and Pennsylvania will come to the polls. In an election that looks so close, that may make all the difference.
(Neelanjan Sircar is Associate Professor in the School of Arts and Sciences at Ahmedabad University)