As I come upon my first 100 days in office, I’ve been reflecting on the time we spent campaigning. As this article explains, my campaign was reliant on statistical modeling, which provided the evidence we used to pinpoint the voters most likely to vote for me. We allowed the evidence to guide us – and this is the same approach I have been using to legislate.
Campaigning Like a Physicist
By Emily Conover
When plasma physicist Andrew Zwicker ran for a seat in the New Jersey General Assembly last November, he approached the challenge like a physicist, employing a powerful trick from the scientist's toolkit: a statistical model. Using data about constituents in his district, he worked with Sherrie Preische, a physicist who now works on election modeling, to pinpoint the voters most likely to help tip the balance in his favor, and reach out to them.
Zwicker was facing an uphill battle, running as a Democrat in New Jersey's 16th district, which had a 40-year streak of electing Republicans. But the method apparently worked: Zwicker pulled off an upset victory — by a hair. He won by just 78 votes, defeating the incumbent, Donna Simon, 16,308 to 16,230.
In 2012, the Obama campaign successfully employed a similar strategy on a larger scale. The technique, known as micro-targeting, aims a candidate's messages at narrow slices of the voting population, even tailoring communication to individual voters. Until recently, such modeling has been out of reach to all but the biggest, wealthiest campaigns, Preische says. "My approach has been to try to make it available to smaller campaigns, and particularly close [races] where it can make a real difference."
The model Preische developed for Zwicker assigned registered voters a "support" score representing the probability they would support Zwicker, and a "turnout" score representing the probability they would show up at the polls on election day.
To most effectively reach out to the 150,000 registered voters in the district, the campaign tailored its communication efforts according to these scores. Voters with a high support score and a low turnout score might be contacted by a volunteer encouraging them to vote, while voters that the model predicted were likely to vote, but were on the fence about whom to vote for, might merit an in-person visit from Zwicker himself.
"Every piece of mail we sent out, every phone call we made, and every door we knocked on was guided by the model, and we had to be incredibly disciplined to do that," Zwicker says.
To create such models, Preische uses public information, like a voter's registration as a Republican, Democrat, or independent; whether the voter cast a ballot in a given election; political donations; the results of elections by precinct; and census data. She combines this with consumer data that companies gather about individuals, which are available for a price. Such data range from the type of car you drive to the shows you watch on television. "There’s thousands of pieces of information," Preische says.
To create the model for Zwicker's campaign, Preische combined these methods with data from a phone survey, in which voters responded to questions relevant to Zwicker's candidacy — for example, how they felt about a scientist running for office, or whether they supported Zwicker.
Presiche then used machine-learning algorithms to look for correlations between the public data and the survey responses, and used the results to build a model that could make predictions for voters who hadn't been surveyed.
On election night, the outcome was so close that the Associated Press initially called the election for his opponent, and Zwicker conceded. But it soon became clear that the final results would have to wait for every last ballot to be counted, which took two weeks. In fact, Zwicker says, he was at the 2015 APS Division of Plasma Physics Meeting in Savannah, Ga. when he got a phone call informing him that he won.
Zwicker says that the model alone wasn't enough to tip the balance in the election: "We outworked them — that’s the only reason why I won," he claims. "Altogether, we knocked on 22,000 doors; we made 78,000 phone calls; and we sent out many pieces of mail. So the model just guided it."
Zwicker campaigned with the promise that he would use evidence to make decisions. When he knocked on doors of voters with this line, Zwicker says, "They would laugh sort of cynically, like 'good luck with that.'" But, Zwicker says, many voters, although they didn't necessarily promise to vote for him, said they were intrigued. "That happened over and over and over again."
Now, in his first months on the job, Zwicker is sticking with the strategy. "If I don’t agree with your position, show me the evidence, and I have to do what I promised to do — what any good scientist would do. I cannot stick to only ideology," Zwicker says. "People respond to that, because they say to me, 'We may not agree but at least I understand where your position comes from.'"
Zwicker is enjoying his new role as an assemblyman. "I feel incredibly honored to have this opportunity," he says. He is splitting his time between the assembly and his job at Princeton, where he is head of science education at the Princeton Plasma Physics Laboratory (PPPL). Zwicker says his fellow assembly members have responded well to his scientific approach, and he is serving on three committees: telecommunications and utilities, environment, and judiciary.
Preische notes that in previous micro-targeting campaigns, her model has proven to be "amazingly accurate," when its predictions are compared to election outcomes. Preische notes that as a physicist she predicted the behavior of electrons instead of people. "People are less predictable," she says, but "people are more predictable than they think they are."
Preische got involved in politics when she left a career in plasma physics to work on another physicist's election campaign: former U.S. Representative Rush Holt (now the CEO of AAAS). She now is a partner in the political consulting firm FiftyOne Percent, where she works on election campaigns and advocacy campaigns for organizations. She previously employed a micro-targeting strategy for U.S. Representative Bonnie Watson Coleman, in a heavily contested 2014 New Jersey Democratic primary race for Holt's former seat, which was left open after he announced he would not seek reelection. Zwicker also ran in this race, unsuccessfully.
Preische knew Zwicker from PPPL when she was a graduate student and he was a postdoc. Holt himself was previously the assistant director of PPPL. Sources were unable to confirm whether there's something in the water at PPPL that motivates physicists to go into politics.
As for Zwicker, he says Holt was part of his inspiration to get into politics, but he also grew up in a politically active family, and his parents volunteered on local campaigns. "My mom, at 83, still reads the paper every day and wants to talk about politics all the time," he says. Zwicker was also frustrated at the politicization of science, and felt that as a scientist he could help steer his state toward economic prosperity. "We know historically that investment in science comes back multiple-fold in terms of economic activity," he says.
"New Jersey has real economic challenges — our economy has not recovered as fast as our neighboring states. And we have this rich, rich history of scientific innovation." Zwicker believes science is the way to turn his state's economy around. "Who better than myself as a scientist and as a science educator to try to do that?"