The Victory Lab (46 page)

During the summer, Obama’s campaign also found fresh reason to believe that the Bradley effect might be more than a historical artifact. Mike Podhorzer, then the deputy political director of the AFL, called a colleague at the Obama campaign to relay something that had caught his attention. Obama was consistently performing better on polls the AFL conducted with live callers than those, called robopolls, where respondents were prompted by a recording to submit their multiple-choice answers by pressing touch keys. As a result, Podhorzer began tracking the race of the callers at the phone banks who conducted the AFL’s polls, and the race of each respondent, looking for patterns between the two. There was one, and it attained the level of statistical significance: Obama repeatedly did better in
polls conducted by black callers than those of other races. The AFL’s finding reverberated around Obama’s headquarters.

Strasma thought that holding racial prejudice was an individual behavior, just like voting or shifting one’s preference, and that he should be able to draw up an algorithm to measure its likelihood for every voter in his file. People had all sorts of reasons for preferring John McCain over Barack Obama; maybe they were concerned about Obama’s lack of experience or his health-care proposals or always trusted Republicans to oversee foreign policy during wartime. Strasma did not build his models to account for these reasons, merely to predict whether the sentiment was so strong that Obama stood little chance of changing the voter’s mind and, if not, what issues could unstick it. He had added a question about gun control to the microtargeting surveys not because it was a big theme of Obama’s campaign—in fact, the candidate was unusually quiet on the issue for a Democrat—but because it could help to flag a certain type of likely shifter. Strasma built a model to find voters who were liberal on guns and abortion; even if they were undecided or showed high McCain-support scores, he thought, Obama’s campaign should never give up on winning their vote. “When this person pays attention, unless they’re nuts they’re going to pick Obama,” he says.

But racial prejudice appeared to be a different type of block on a voter’s part than any constellation of policy issues. It seemed the type of attitude that was irreconcilable, which no amount of new information about McCain’s positions or Obama’s background could budge. Obama needed to find these people early so that he could give up on them, and not—like Bradley or Wilder or Dinkins—emerge shocked on election day when they quietly cast ballots for his opponent.

During past campaigns, Strasma had built models to pick out individuals with potentially sensitive personal characteristics that were not a matter of public record. When he had tried to identify military veterans for Kerry, in some places—like Iowa, with its county-level tax benefits—it was
easy and straightforward. In others, it was roundabout: Strasma had the campaign buy subscriber lists for military-themed magazines, or records of those who had purchased commemorative license plates connected to military service, and fed those into the algorithms. Each of those was an imperfect net for catching veterans, but Strasma was not too worried. If he was catching nonveterans, they looked demographically a lot like veterans and still cared about military affairs. If they got a brochure describing Kerry’s military service and attacking Bush’s handling of veterans’ hospitals, what was the harm?

Yet Strasma also had experience with cases where that kind of false-positive modeling had significant risk. For years, like many of those who targeted voters for Democrats, he had been trying to find a way to profile gays and lesbians so campaigns could speak to them directly on issues of specific interest. The common method was some version of looking on the voter file for two adults of the same sex but different last names living at the same address with birth dates within a decade of each other, a mix of variables that usually sorted out parent-child or sibling combinations. It could not account, however, for people of close age living together platonically, leading to a legendary (and possibly apocryphal) instance in which a campaign ended up targeting some of its gay-themed mail to pairs of roommates at a military academy. Then in 2000, the Census Bureau began to offer an option on its long-form survey for respondents to identify themselves as part of same-sex couples, which were counted in block-group-level profiles. Strasma added the intensity of gay couples in one’s neighborhood to sharpen the algorithms predicting if an individual was gay, but also used it to reduce the damage of a false positive. Heterosexuals who lived in significantly gay neighborhoods, he assumed, were less likely to take umbrage at receiving literature addressing gay issues.

Now, as he tried to quantify racism for the Obama campaign, Strasma was in essence looking for a different kind of false positive. He started by focusing on people whom the algorithms positively identified as someone who should be a likely Obama supporter, on the basis of their demographic
and political qualities, but answered ID calls by saying they were backing McCain. To build a model that could predict which voters might be biased against his candidate, Strasma had to identify a variable, or a combination of variables, that explained that gap. At one meeting, one member of the targeting staff with little political background suggested just adding a battery of new questions, including “Are you racist?” and “Does racism affect your vote choice?” Strasma shook his head. (That’s one of the problems with hiring campaign staff directly from computer science or statistics backgrounds, he thought.) Already pollsters were wary of even inquiring about respondents’ race in polls; when they did it was often the last question, after asking about income, so that if the respondent hung up the whole survey wouldn’t be ruined. Strasma knew that any question that successfully uncovered a voter’s racist sentiment would have to be more sly. He needed to find a publicly available data point that would be an efficient proxy for asking people one of the most indelicate questions in American life.

Strasma tried adding questions about a few different policy issues considered racially-tinged, like affirmative action, but they didn’t yield useful patterns in the modeling. So he started paying close attention to the focus groups that David Binder, a San Francisco–based opinion researcher, conducted all over the country. Each night Binder or one of his deputies would moderate a session, which typically lasted an hour, with undecided voters in a different battleground state. Some days, Binder would read statements from the candidates to hear what voters thought of them. Other times he would show the latest ads, or a series of mock anti-Obama ads that the campaign’s media consultants would produce to test voters’ reactions and audition countermessages. Sometimes Binder would just ask open-ended questions to guide a conversation.

Each of the focus groups was broadcast via the Internet to a room in Chicago, where members of the campaign’s different departments could wander in and watch. For senior staffers who had given their lives over to Obama’s quest, the focus groups were nightly prime-time entertainment.
Someone would head off to a local supermarket for snacks and a group would settle in for the session, fortified with peanut M&Ms. One night Binder asked, “Do you think your neighbors would be willing to vote for an African-American for president?” Some of the voters answered no, and Strasma watched them closely. Something in that response—perhaps a feeling of being liberated to publicly share an unpopular opinion—convinced him that the people who acknowledged their neighbors’ racism might really be confessing a view of their own.

Strasma added the neighbors question to his survey and saw quickly that it worked. Those who had high Obama-support scores but ended up backing McCain said yes to it, so Strasma made it the core of a new “openness” model: another score, out of 100, that assessed how open a voter would be to casting a ballot for a black candidate. Those voters with low openness scores and mid-range turnout scores could be removed from Obama’s contact universe altogether, for fear that contact from Obama could have a backlash effect and make them more likely to turn out for McCain. Those with high turnout scores were fairly likely to be voting already, so Strasma thought it may still be worthwhile to try to make a pitch to them. “We knew it wasn’t going to be the same motivation message that was working with African-Americans or young people or antiwar voters,” he says. So instead of mail with “Change” themes he thought the campaign could send them something with a very straightforward economic message. “That message made sense on the gut level for those voters,” says Strasma, “and based on what we were able to see it worked better.”

But the openness score was most useful as yet one more variable being added to the hundreds already on Strasma’s computers. When added to the models, a low openness score would pull a voter’s overall support score down to account for the likelihood that his or her willingness to ultimately vote for an African-American had been overstated. It worked in much the same way that the pro-choice and gun control scores did, only in the opposite direction: a way of correcting for the inability of voters to be as honest and self-aware as pollsters like to pretend they are.

The targeting desk felt confident that it had identified a statistical fix for what had been the biggest uncertainty hanging over its inability to accurately predict voters’ preferences. Still it did not fully calm nerves in Chicago. The models had been early to pick up the jolt Sarah Palin delivered to the race, and after the mid-September collapse of Lehman Brothers and the subsequent financial crisis, they saw her influence subside and the race stabilize. The targeting desk would circulate weekly reports to Plouffe and the campaign leadership, with one chart becoming preeminent: a histogram that showed how the campaign’s weekly ID calls matched up against the modeled support scores. By October, the charts were reliably producing the elegant step function that analysts wanted to see—voters were telling callers that they supported Obama at the same rate that the algorithms predicted they would. “I was surprised by how unsurprised I was,” says Strasma. They would have to wait for election day to get the only further confirmation possible.

THE ROOM SQUEALED
when Barack Obama walked into the Greater Columbus Convention Center, where 750 Ohio volunteers had gathered for a daylong training seminar in October. They had expected an immersion in the art of winning votes Obama-style, but not an encounter with the man himself. “
We’ve been designing and we’ve been engineering and we’ve been at the drawing board and we’ve been tinkering, and we’ve been—now it’s time to just take it for a drive,” Obama said, as people mounted chairs for a better look at their candidate. “Let’s see how this baby runs.”

Election day was less than one month away, but the campaign was already producing votes at full thrust. Obama had put a particular emphasis on early voting, which many states had introduced since 2004 to alleviate pressure on their election day operations. Such a service, which often gave people the freedom to cast their ballots either by mail or in person for as long as a month, was almost perfectly catered to the needs of the Obama
electorate. First-time and minority voters who might be overwhelmed or intimidated at a polling place, and those too busy to contend with long lines on election day, could do so at a less harried time. Election officials would produce a daily list of those who had voted early, and Obama’s tacticians pored over those rolling returns. Getting likely supporters to cast an early ballot locked in their votes and allowed them to be removed from target universes for future mail, phone calls, and canvassers, so the campaign could expand its energies elsewhere. Then the fact that a person had voted early could be used as one more data point to refine projections of who would actually turn out.

Two weeks before Obama’s visit to Columbus, a plane carrying an early-vote appeal had floated over the city’s Ohio Stadium during an Ohio State–Minnesota football game—the fruit of one of Moffo’s many gambits—and buses had begun to roll through Cleveland, Canton, Youngstown, and Akron with the same message. When the soul singer John Legend offered to perform on Obama’s behalf, the campaign made sure one of the stops was a midday concert on a modest amphitheater stage in Legend’s hometown of Springfield—because the data team’s projections suggested Springfield, midway between Columbus and Dayton, was lagging larger neighbors in early-vote participation rates. “We never would have put John Legend in Springfield if we hadn’t seen that,” says John Hagner, the Ohio Democratic Party’s voter-file manager.

The confidence Obama demonstrated before his troops was grounded in his campaign’s success at audaciously remaking the political world. It had redrawn states to reflect a preferred geography, redefined individuals and households through algorithms, put political messages where no one ever thought they had a place—but elections were still conducted in the old world. No matter how much Obama believed that his organization had reinvented the machinery of politics, it still had to contend with the existing infrastructure of elections. Votes needed to be cast on paper or behind a curtain, and the institutions that handled those votes could not handle
data as nimbly as the MyBarackObama calling tool or Strasma’s algorithms. So the Obama campaign turned that lag into a tactical advantage.