The Victory Lab (27 page)

She started plotting with Mark Steitz, a former DNC official who had drifted out of politics for a decade before resurfacing as the philosopher-king of Podhorzer’s early geek lunches. Steitz was a garrulous former economist who had followed his boss, Senator Gary Hart, into presidential politics, and had grown frustrated with the idiocies of most political consultants. (Only a handful of them earned Steitz’s highest praise: to be called “a serious human.”) He usually took the long view of the political developments his peers preferred to assess in news cycles, and would often say things like “the Clinton campaign in 1992 was the full flower of what might be called the neoclassical synthesis of polling, media and field” or, when referring to the state of public opinion polls, “I think what you’ll find out is there is more of the doctrine of eternal recurrence rather than Kuhnian scientific revolution against it. You know what I mean?”

Steitz had spent his self-imposed exile from politics in the corporate world. He had overseen communications for the 2004 Athens Olympic Committee and marketing for the Body Shop, the lefty British cosmetics retailer. (“They didn’t believe in marketing per se,” says Steitz. “So it was an interesting place to work.”) When Steitz reinserted himself in the Washington campaign world during the 2004 election season, he found a circle of Democratic operatives eager to develop a culture of learning he had usually found absent from strategy sessions. They shared a spirit of entrepreneurship, inspired by recent events to conjure something altogether new rather than fight over market share in existing campaign budgets. “Losing is a great tonic for internecine bullshit,” says Steitz. “We had been beaten badly enough and repetitively enough that lots of people were willing to sit down with one another who previously would not have.”

THOSE DEFENDING THE UNPOPULAR VIEW
that there is actually not enough money in politics frequently take refuge in the fact that Procter & Gamble each year spends more money advertising soap than Americans do on the quadrennial marketing pageants that choose their presidents. Political operatives have often gazed covetously at their analogues in the corporate suite, with their big budgets and multiyear market research studies, and imagined that on the other side someone stood on the verge of a major breakthrough in understanding human behavior. But the envy runs both ways. “The political business and the corporate business are like movie stars and rock stars,” says Alex Lundry, who works with Gage at TargetPoint Consulting. “Everybody wants to be doing what the other side is doing. Every movie star wants to be a rock star, and vice versa.”

After a decade in each sphere, Steitz maintained a more realistic view of the two sectors. “I had higher hopes for what the consumer world would do. I kept assuming that the commercial world had everything solved. When you actually got neck-deep, you realized that there are many things they knew that we didn’t know, but there are many challenges that were very similar,” says Steitz. “Neither is as far ahead of the other as they would hope.”

Still, the commercial sphere had produced one piece of data that mesmerized Steitz, and it came from credit-card companies. It was a single predictive variable that allowed its analysts to compare at a glimpse tens of millions of people in behavioral terms. Through the first half of the twentieth century, lenders—from banks issuing mortgages to retailers with store accounts—would hire underwriters to look at individual borrowing histories, compiled by financial institutions and merchant associations, before deciding to issue credit. This was a subjective process, and an arduous one.
In 1956, engineer Bill Fair and mathematician Earl Isaac began developing computer programs that could automate this analysis, and two years later started rating would-be borrowers for the St. Louis–based American Investment Company. By using these scores, the bank was
able to cut delinquencies by one-quarter, or, depending on the adventurousness of loan
officers, increase its lending volume by as much without spiking delinquencies. Fair, Isaac & Company credit scores were the common point of reference enabling a postwar lending boom that included the introduction of mass-market credit cards in the 1960s. Now everyone who wanted to lend money—the small-town credit-union officer, independent landlord, store with a house charge account, or corporate risk officer at Bank of America—could decide whether an individual was worthy of credit. A process that once often took weeks collapsed to minutes.

Fair, Isaac funneled consumer information from a variety of sources (there are now three major credit reporting agencies in the United States) into a shared numerical language.
In 1989, the company introduced its FICO score, which assessed every individual on a 300–850-point scale of universal risk. A few years later, Fannie Mae and Freddie Mac started to require FICO scores for home sales, and they quickly became part of a standard mortgage application. As they matured, credit scores went from being merely a hunt for red flags in someone’s past to a prediction of their future actions based on patterns in others’ behavior.

In April 2005, Quinn and Steitz started a firm to market a credit score for voting. They called their new company Copernicus Analytics, after the sixteenth-century Polish astronomer. “
Copernicus took individuals out of the center of the physical universe; we are trying to put them back at the center of the political world,” Steitz announced at the time. They looked to the credit industry for an early hire. Ben Yuhas had studied math as an undergraduate and then earned an advanced degree in electrical and computer engineering. He went to work at Bellcore, the research institute opened by the regional phone companies alongside the legendary Bell Labs in northern New Jersey, and in 1995 joined AT&T as it beefed up antifraud efforts around its recently introduced no-fee credit card. Yuhas quickly became a creature of Wilmington, Delaware, America’s credit-card company town, changing affiliations as banks consolidated but never having to move.

As a mathematician, he loved the fact that everything the companies did was very neatly measurable, and he looked admiringly at Capital One.
An offshoot of Virginia’s Signet Bank,
Capital One was the first credit-card company to take analytics seriously. Two consultants, Richard Fairbank and Nigel Morris, had approached fifteen banks in the late 1980s offering “information-based market strategy,” and Signet was the only one to say yes. Fairbank and Morris used the large database that Signet had built of its transactions to look for patterns between customer behaviors, their credit scores, and the revenue they generated for Signet. When they saw that the most lucrative customers for Signet were ones who quickly borrowed large sums and slowly but responsibly paid down the balances, Fairbank and Morris proposed the bank introduce a balance-transfer card. (Signet spun off Capital One as a publicly traded company in 1994.)

Capital One became known for its culture of testing, eventually running three hundred different experiments at once, many of them using mailed credit-card offers that could be easily randomized in the way Gerber and Green had with campaign brochures. There were plenty of dependent variables to track: Capital One analysts could account for responses to a particular promotional offer, or examine the rates at which cardholders paid off their bills or went bad altogether. At the same time, credit scores offered a matrix to compare different types of customers. Did lowering fees appeal more to low-risk or high-risk borrowers, and how did different groups adjust their buying patterns when given new terms? “You had this really clear feedback loop, so you could focus on making the math better,” says Yuhas. The goal was to convert a once-binary decision about risk (should we issue the card or not?) into a dynamic one (what rates and fees should we charge each of them to maximize our return?).

It was Yuhas’s job to apply that math to politics. Ultimately his goal was to locate an algorithm with predictive power for two basic questions: how likely someone was to vote and how likely he or she was to support a certain candidate. An algorithm was, in effect, nothing more than a complex equation in which each variable was given a different weight, with those variables tested in different combinations to see which exerted the most force on the desired outcome (support and turnout) and weighted accordingly.
Yuhas’s challenge was to design a model that accounted for which personal variables would play a consistently predictive role in a particular election. (Those variables could change from election to election; in a statewide race between San Francisco’s mayor and Los Angeles’s, for instance, having an Oakland ZIP code might be a major determinant in predicting a voter’s support. In a race between two Angelenos, it might not matter at all.) Credit agencies had learned that household income and age played different roles in determining creditworthiness; Yuhas had to determine what influence they—and hundreds of other variables to which Copernicus had access—had on voting behavior.

In the spring of 2005, Tim Kaine hired Copernicus to put its new algorithm to work in his campaign to be Virginia’s governor. Democrats saw the state moving gradually in their direction, but there was something else about Virginia that made it an appealing venue for Quinn and Steitz to put their scoring strategy to the test. Virginia is among the states that do not allow voters to register with a party, which means that the most useful predictor of general-election behavior was not available on the voter file. “We were at a real disadvantage knowing who was a Democrat and who wasn’t. We are taking a guess on a lot of people as to whether they’d vote for us,” says Mike Henry, Kaine’s campaign manager. “There were only two ways to get it: call ’em or go and talk to ’em.”

Even the old method of looking at precincts would have been of limited use to Kaine. His strategists thought their candidate, a former Richmond mayor finishing a term as lieutenant governor, was a different kind of Democrat than had run statewide before. Even the map from 2001, when Kaine had first won statewide office as Mark Warner’s running mate, offered little guide for those seeking to put together a coalition for Kaine. Any Democrat had to turn out black voters in cities like Richmond and Norfolk, but the question was where white votes would come from in a state that had voted Republican in every presidential campaign since 1964. Warner lived in the Washington suburbs but had used his conservative stance on guns as well as a NASCAR aesthetic to win over rural voters.
Kaine’s positions on cultural issues—he was a former missionary and civil rights lawyer who opposed capital punishment—made Warner’s map impossible. “No one had ever run a race in Virginia that was against the death penalty and had an F from the NRA and we had to overcome these obstacles,” says Henry. Where Warner had been carried to victory on a rural-urban coalition, Kaine’s advisers looked to the suburbs and exurbs of central and northern Virginia. These included some of the fastest-growing counties in the country, the type of place where Bush had used microtargeting to pick up new votes the year before.

Based on a large-sample poll in the summer of 2005, Copernicus was able to give each voter two different scores derived from Yuhas’s algorithms. Each set to a ten-point scale, one predicted the likelihood that an individual would support Kaine, and the other that the voter would go to the polls in November at all. A person with a 7 support score was more likely to back Kaine than anyone with a 6. Steitz prepared a pair of maps to show Kaine’s advisers how a new statistical method could effectively change the political geography of a state they thought they understood. One highlighted the counties that would get attention under a traditional precinct-targeting strategy, where large numbers of voters lived in areas regularly delivering 65 percent of their votes for Democrats: about ten counties appeared in the most intense color, and another ten in the next darkest shade. Yet a map that colored counties by the numbers of voters whose Kaine-support scores were in the top 20 percent darkened nearly every county in the state. From the void, a thick column of newly targetable counties emerged, stretching from Washington to Richmond.

Many of the Kaine targets in these counties had scored very high in Yuhas’s models for Kaine support, but unusually low for turnout. That was a familiar profile for Democratic organizers: largely minority urban precincts overwhelmingly supported Democrats but lagged behind suburban precincts in turnout. This was much of the reason that the party and its allies had made such a conscious push to invest in GOTV during the 1990s. But some of the other counties that became filled in with Kaine
targets on Steitz’s maps were suburban, generally white and fairly affluent. When Yuhas looked more closely at who the individuals were in these areas, he found a distinctly different political profile than the poor African-Americans the party was used to rousing with another round of phone calls and a flotilla of vans ready to drive them to the polls. These upscale suburbanites, Yuhas found, lived in reliably Democratic precincts and regularly voted in midterm congressional elections but sat out the state elections held in odd-numbered years. “They were definitely our voters,” says Henry. “If we could turn them out we’d get them.” The campaign referred to them internally as Federal Democrats: their lives revolved around Washington, not Richmond, and their local newspaper and television stations often paid more attention to Maryland issues than their own. It would take more than a phone call and a van to push them to the polls. “We needed to show them why it’s important that they vote in a governor’s race,” says Mo Elleithee, Kaine’s communications director. When Copernicus produced a list of 250,000 such voters, Kaine’s team designed a mail campaign about transportation spending—traffic was a perennial issue in Northern Virginia, and one that could only be fixed by local governments—and started sending canvassers to areas they had never gone before. “They almost had a list of the people by name they needed to win the election,” Steitz says.