New Jersey pundits called our opponent a shoe-in. Early polling from veteran lawmaker Linda Greenstein’s campaign showed her with a 10-point lead. But microtargeting of voters gave our candidate, Assemblywoman Bonnie Watson Coleman, a resounding 15-point primary win.
We did it by focusing on a narrow communications and voter contact program governed by the targeting. We’ve seen how national and statewide races benefit from individual voter modeling. We show here that microtargeting, like President Obama’s campaign put in practice in 2012, is now affordable in smaller-budget and down-ballot races.
Setting the Stage
Long-term Congressman Rush Holt of central New Jersey’s 12th District announced in February he would not seek reelection, setting the stage for a four-way Democratic primary on June 3. Holt had given no indication before his announcement that he was even contemplating retiring from Congress. The result was a short three-and-a-half-month campaign, which normally benefits candidates with good name recognition and those who have recently spent communication dollars in prior elections. In this case, both of our opponents fit that bill.
The 12th District is mostly suburban and is bracketed by two urban centers, Trenton and Plainfield. The demographic makeup by citizens of voting age is 65 percent white, 19 percent African American, 8 percent Latino, and 8 percent Asian. It’s highly educated, with more voters having completed some post-graduate work than any other education category. It’s home to Princeton University and Einstein’s Alley, New Jersey’s research corridor.
In addition to Greenstein, a state senator, Assemblyman Upendra Chivukula and Princeton University physics professor Andrew Zwicker, who had never before run for office, also jumped into the primary race. Each of the legislators had their own natural geographic constituency and were well known in the portion of the district they represented.
Greenstein, who’s also an attorney, has represented New Jersey’s 14th Legislative District since 2000, first as a member of the General Assembly and then in the state Senate. She had just won reelection in November 2013 in one of the most expensive races in the state, giving her high name recognition.
Watson Coleman, who has represented the state’s 15th Assembly District since 1998, rose in the Assembly to become the first African-American woman to serve as Majority Leader, and had a firm footing in the Trenton area. If she wins in November, Watson Coleman will be the first African-American woman to represent New Jersey in Congress.
Being an African-American candidate in a majority white district is a challenge in its own right, so putting together the right team was crucial. Watson Coleman realized early that to win she needed to raise money and run a very focused and disciplined campaign.
The team was led by long-time Watson Coleman strategist James Gee. The campaign manager was Adam Steinberger, who had just run Sen. Robert Menendez’s field program. The team also included Holt pollster Fred Yang of Garin Hart Yang Research Group. Longtime New Jersey communication operative Sean Darcy did the direct mail and earned media, John Rowley of Tennessee-based Fletcher Rowley provided the TV and online media. My firm, FiftyOne Percent, did voter models and analytics.
Politicos Weigh In
Watson Coleman began the race 10 points down in Greenstein’s internal poll, which was released publicly. As a result, the political “talking class” gave our candidate little chance of winning. Greenstein is a tenacious campaigner and initially had the support of many of the political leaders in the district. There are more registered Democrats in the Middlesex County portion of the district where she lives, and millions of dollars had been spent in her recent state races. PolitickerNJ, New Jersey’s premier political news website, ran analysis after analysis showing the invincibility of Senator Greenstein.
Still, our modeling of the district and our polling clearly showed a path to victory as long as we stayed focused and disciplined with our message and communications universe. This gave our communication team, headed by Darcy, a former staffer of Gov. Jon Corzine (D), the ammunition it needed to chip away at the media’s perceived view of the race. It also helped to bring early support from the Progressive Change Campaign Committee and Laborers International Union, critical early endorsers of the campaign, and to drive the all-important fundraising effort as more and more people came to believe that Watson Coleman could win.
Understanding the challenges in this election, the Watson Coleman campaign invested early not only in a poll but also in a voter model to catalogue those who would support her, those who could be persuaded to support her, and those who were likely never to support her in the primary.
A microtargeting voter model gathers large amounts of data about each person to gauge the likelihood of their behavior. For each voter, that data includes not only their own demographic and voter history, but also things like estimates of income, education level, and particularly consumer behavior. It’s the same type of information that Google and others are using each time you see an ad online that you know has been tailored to you and your interests.
Back in 2012, a combination of campaign contacts, internal polling, past records, and consumer data allowed Obama’s GOTV team to create the best door-to-door interactions with voters when they canvassed. Those interactions went back into the campaign’s database after individual door-to-door conversations and then their microtargeting models were refined based on that information.
To build the support model in our case, we used not only our internal polling but also additional phone support surveys, historic voting behavior, and IDs the campaign has made by phone and door-to-door canvassing. All of this is used to estimate the individual likelihood that each particular voter will support our candidate or not.
As a new campaign tool, a microtargeting model can be difficult for some candidates to wrap their heads around because it’s difficult to see. This is not broad-based “targeting” such as women in their 30s or some other demographic that many people say they do. The key is that microtargeting is specific to each person, because each person has different interests and behaviors.
We know that voters can behave differently as they move down the ballot to different races. An area that might be reliably Democratic at the federal level may also lean Republican down the ballot at the county level. Likewise, primary voters behave differently in different elections depending on the candidates. Therefore, having a model specific to the race you’re running is critical in close elections.
While support models in general elections are usually dominated by the partisan behaviors of each voter, in a primary, with little space between the candidates on issues, we had to find the many other factors driving voter decisions. Some support factors were well understood by all candidates; particularly that each of the state legislators in the race had his or her own geographic regions of support. But even within these regions, our support model could tell us which particular voters might be more open to our candidate. Each voter was given a support score, representing the likelihood of that voter supporting our candidate.
We also had a model for individual voter turnout in this primary election. Recent Democratic primaries had been uncontested, so we expected higher turnout than in recent primaries for this contest. The challenge was to figure out who the extra voters would be. Partly because primary voters are pretty dedicated voters anyway, our turnout model brought out nuances in people’s voting behavior in primaries. It allowed us to see not only the voters who vote in every primary, but also a second set of voters who would be the easiest to move to the polls, voters who might not vote but whom we needed to pull out to win.
Normal turnout for a Democratic primary in this district is between 15,000 and 27,000 voters. But our model showed that if all sides inspired their voters to come to the polls, we might expect higher turnout—up to 44,000 voters.
Combining support and turnout models showed us exactly which voters we needed to reach out to with a persuasive message, which voters we shouldn’t talk to at all, which extra voters we needed to pull out, and which voters we were confident would come out for Watson Coleman. This allowed the campaign to come up with a win scenario based on real numbers and individual voters. It also allowed our field program, mail, and media buys to focus on the select group of voters who were most important to winning.
Voter Model in Action
The model informed every aspect of the program. While it was an up-front investment, it allowed the campaign to use all its resources—the candidate’s schedule, volunteers’ time on the
phones and on the streets, and paid media—most effectively.
The model identified more than 42 percent of the probable primary voters who we knew would likely be supporting one of our opponents and another set of voters who would likely come out and vote for Watson Coleman if we did nothing. This allowed the campaign to concentrate its resources—money and time—on the select group of primary voters, persuasion and pull targets, who we knew were going to make the difference in this race, with a message tailored to their interests.
For the model to be the most useful, it needs to be more than just an analysis of the numbers and a score assigned to each voter in the database. Having the campaign and political experience to know what the numbers tell us about voter behavior, and how the campaign can best deploy the model, is critical to making this tool most effective.
Although Greenstein and Chivukula had significant TV buys, the model allowed Rowley and Darcy to focus 100 percent of our paid media, TV, and mail in our top persuasion areas. All of our online pre-roll ads were directed to townships and areas that were our top persuasion and GOTV targets. Rowley also integrated the model into the online advertising campaign, targeting specific voters to receive pre-roll video ads.
While we were outspent 7-to-1 on paid media, the model identified the individual voters we needed to reach. We repeatedly communicated with this select group of voters who we knew would make a difference in the election. As the primary vote approached, we needed to expand on our initial support model.
We conducted an IVR (interactive voice response) support survey about a week before the election and integrated the results with our turnout model and the voter IDs the campaign identified through its phone and direct voter contact program. That allowed us to create a GOTV model and refine our categories of voters. The campaign was then able to prioritize its GOTV program and resources for those extra voters we needed to win: those most likely to support Watson Coleman but who needed a little extra incentive and reminder to vote.
Results
Two weeks before the election, a public poll was released from Monmouth University showing Greenstein and Watson Coleman essentially tied at 25 percent and 24 percent respectively, with 11 percent and 6 percent for Chivukula and Zwicker.
Often in these types of elections voters are somewhat hardened in their positions, and the universe of voters who can actually be persuaded is small. It can also be expensive to earn a vote through persuasion. Persuasion is the most expensive communication of a campaign. Limiting that universe to those most likely to be persuaded more than pays for the cost of modeling. That’s why broad TV buys in a race like this can be less effective. In addition to our persuasion universe, our model also identified people who might not be regular primary voters but who would be relatively easy to turn out with a pre-GOTV and GOTV program. We focused a lot of attention on energizing them. The result was that the political establishment was shocked by the large turnout in support of the perceived underdog.
The final tally was 43 percent for Watson Coleman and 28 percent for Greenstein, with Chivukula taking 22 percent and Zwicker 7 percent. And more than 36,000 people voted. The Watson Coleman team had identified and communicated effectively with the right voters.
There’s an adage that repetition is the key to victory. That continues to be true, but with voter analytics and modeling we can now target the individual voters that matter most, giving campaigns more repetitions without much additional cost. A voter model is a wise investment even for down-ballot races. It’s a small fraction of the campaign’s budget, but it makes the overall program more effective and efficient, and it ensures the communications and field operations are targeted to the voters who will make a difference on Election Day.
Dr. Sherrie Preische is a principal at the firm FiftyOne Percent.