The US election and big data: an EUI perspective
Election forecasts based on polls and on past voting patterns have limitations, as we saw again in the US Presidential election in November 2020. Despite massive number-crunching and adjustments made since the elections of 2016 and 2018, the analysts for the website FiveThirtyEight.com underestimated support for Donald Trump in some key states. This might be lightly dismissed as unavoidable trial and error in the social sciences, except that people who perpetuate conspiracy theories about a fraudulent election are only too happy to point to erroneous polls as support for their claims.
Both polling evidence and the record number of early, absentee ballots make it clear that most voters had made up their mind which candidate they would support long before election day. This means that turnout was key to the result.
The inscrutable side of voter turnout
But turnout is very hard to predict based on past experience or on polls. As we explain in an October article in American Economic Review, this is because both the voters and party campaign managers are strategic actors. Every election plays out differently: the leaders of the parties contesting the elections try to adjust and redirect their campaign efforts and expenditures based on what they think the other side is doing and saying to get its vote out. People are also strategic in dealing with the social pressures that have to do with performing one’s civic duty, expressing loyalty to the party’s cause, or simply making good on a promise you gave to a canvasser.
The analytic model we introduce focuses on the crucial role of turnout and efforts to influence it, rather than on voters’ choice of a candidate per se. It contends that people in democracies weigh two kinds of ‘costs’ related to voting. The first, turnout costs, are those related to the effort to show up at the polls (or cast your vote through other available means). The other, which the authors suggest should be considered ‘monitoring costs’, are those related to the social pressures to vote. People want to avoid opprobrium or being shut out of their ‘community’ – online or flesh-and-blood – for not doing their part. The tag line for this peer pressure might be “Friends don’t let friends not vote”.
For even casual observers, it is clear that turnout costs are likely to differ in each election. Think of a blizzard on election day, long waits when you have pressing work or family matters, or – this year – the fear of catching COVID-19 at the polling station.
But the point a casual observer might miss is that voters make strategic decisions in response to the peer pressure, and that campaign leaders and grassroots also strategise about where and how to generate or apply such pressures. These aspects of monitoring costs are what we elaborate in our analysis.
Monitoring costs and election outcomes
The ease with which people can avoid peer pressure to vote, can hide their vote, or can invent excuses not to vote, gives one party an advantage over another. When social control is difficult (high monitoring costs), people who are only weakly committed might not turn out to vote; this means that parties with larger numbers of regular supporters lose their advantage of sheer numbers.
We suggested in our article, written well before November’s election day, that the smaller party, the Republicans, might do better than expected for this reason. The widespread mail-in/advanced voting options, touted by election officials and civic groups as an easy and safer way to cast a ballot during the COVID-19 pandemic, adds yet another twist to the strategic calculations. As we put it, “measures designed to increase turnout by lowering participation costs may actually have the perverse effect of decreasing turnout because they also raise monitoring costs”.
The model we describe also takes account of the effects of a hotly contested versus a less contested election. The high stakes perceived by both sides in the 2020 election – the choice of candidate being made out to be the end of American life and values as we know them – means that this time around, it is fairly certain that turnout was influenced more by the stakes than by any monitoring effects.
Upgrade your model, man
So, what are FiveThirtyEight and other election analysts doing wrong? Turnout and elections are intrinsically unpredictable. They understand this well enough to give odds rather than certainties of victory. There is a lot of talk about the quality of their data, but that is not and has not been the problem. Just as in the 1960s the mechanical models used by economists to forecast the impact of policies systematically failed until rational expectations were incorporated into those models, so the odds given by FiveThirtyEight will have little meaning until their mathematical models take account of the equilibrium nature of the strategic interaction between the two parties.
David Levine is a Professor of Microeconomics and Joint Chair in Economics – Robert Schuman Centre for Advanced Studies/Department of Economics at the European University Institute.
Andrea Mattozzi is a Professor of Microeconomics and Head of the Economics Department at the European University Institute.
Their article “Voter Turnout with Peer Punishment” in the American Economic Review is available online.