Modeling turnout was thought to be easier than modeling support, since how citizens vote is a private concern but whether or not they do is a matter of public records. In reviewing the attributes of 2008 voters, Obama’s analysts confirmed that the most useful predictor of turnout was past vote history. For years, campaigns would sort voters by these criteria—assuming that someone who had participated in two of the past four elections was more likely to turn out than someone who had turned out for only one—and the Obama campaign used this as the basis for a basic typology. Those who had voted in the 2010 elections were classified as “midterm voters”; those who had voted in 2008 but not in 2010 were called “sporadic voters.”
But splitting the electorate into only three or four categories didn’t leave much opportunity to assign priority to individual voters, or room to draw granular distinctions among them. More perilously for Obama, a pure reliance on vote history didn’t help to predict the behavior of young voters, because the campaign simply did not have enough of a track record to assess. The president will rely for his re-election on young people, who overwhelmingly supported him in 2008 but tend to be fickle voters. Which kids had a thin voting history because they only recently became eligible to cast a ballot and which ones because they weren’t the type of people who could be counted on to vote?
Obama’s analysts built statistical models to pull out other factors that distinguished voters from nonvoters. Socioeconomic factors like income and housing type played a role; those who lived in multi-tenant dwellings, for instance, were less likely to vote. But within those households Obama’s analysts found a twist. A voter living with other people who had a demonstrated history of voting was predicted as more likely to turn out herself.
The Obama campaign’s algorithms ran the numbers and predicted the likelihood that every voter in the country would cast a ballot, assigning each a turnout score. Obama’s analysts knew how good their support score was because they polled a new group of voters to validate it: 87 percent of the time it would accurately predict an individual’s preference. But it would be impossible to confirm their algorithm’s turnout predictions until after the election. But they did their best to assess its accuracy, by calling voters and asking them how likely they are to vote. Analysts know that people are poor predictors of their future behavior, but they got answers that confirmed that their rankings were at least sensible. Among the 10 percent of voters seen as most likely to vote, 95 percent said when contacted that they definitely would. Based on 2008 figures, Obama’s analysts assumed that 40 percent of those who told callers that they would “definitely not” vote ultimately would. “Possibly,” the analytics department advised in a September memo, this was “due in part to our GOTV efforts among these voters.”
In late October, or as early as September in early voting states like Ohio, campaigns shift from registering new voters and persuading wavering ones to harvesting votes from people they already count as on their side. Such get-out-the-vote operations include pre-election reminders, providing information on polling-place locations and absentee-ballot protocols, and, increasingly, psychological nudges informed by the behavioral sciences—all delivered over personalized channels like mail, phones, or in-person visits. Indiscriminately subjecting voters who might support your opponent to such motivators is considered an unsustainable risk. As a result, field organizers typically move a voter into a so-called GOTV universe only if he or she has told a caller or canvasser that they are a supporter, or statistical models predict they are likely to go your way.
Campaigns tend to focus their mobilization efforts on voters who have been assigned high support and mid-range turnout scores. Those with turnout scores outside a span of, say, 30 to 80 are not worth the effort: Those above it are self-motivated enough to vote already, and those beneath it unlikely to do so under any circumstances. Democrats approach the question of prioritizing voters for turnout in much the same way as Republicans. Obama, however, goes a step farther.
Since 2008, Democrats have administered randomized-control experiments to test the impact of GOTV contact on voters with different score combinations, with the goal of quantifying where those contacts are most likely to produce a net vote. The most fruitful terrain turned out to surround voters with turnout scores centered around 45. Delivering a GOTV contact to a voter with a 100 support score and a 45 turnout score increased the likelihood of netting a Democratic vote by 4.5 percent; delivering a GOTV contact to a voter with a 75 support score and a 45 turnout score increased the likelihood of netting a vote by 2.7 percent.
Obama’s analytics department synthesized all of this research into a new GOTV score that combines predictions about one’s likelihood of voting and supporting Obama. It, too, ranks voters from zero to 100, but this one doesn’t assess voters’ characteristics so much as prioritize them based on their susceptibility to the campaign’s efforts to modify their behavior. When canvassers like Darley-Emerson get a list of names, it has been edited according to the one criterion that matters: how likely her visit is to generate a new vote towards the president’s re-election—whether the canvasser remembers to ask who the voter is supporting or not.