Young Mie Kim has a theory about voters. Each one is basically a collection of data points that determine how that person will vote.
“In the data age, voters are defined based on an algorithm, based on a construction and reconstruction of data,” she said.
Two months after an election where political campaigns were blamed or credited for relying on voter data to an unprecedented degree, Kim, a professor and researcher at the University of Wisconsin-Madison’s School of Journalism and Mass Communication, is studying how campaigns used that data in Wisconsin and across the country.
Much like how businesses market products, political campaigns are increasingly focusing on sophisticated data-driven calculations, used to persuade voters and get them to the polls. Those predictions inform campaign strategy and often dictate how a candidate interacts with the electorate.
“Campaigns view micro-targeting as this holy grail (and) I think people need to be aware of it,” she said.
Micro-targeting is the practice of narrowing down large pools of voters to smaller ones that campaigns believe are most likely to be persuaded to support their candidate. The approach, along with sophisticated mathematical models based on large sets of voter data, enable campaigns to more precisely engage individual voters on specific issues that matter most to them.
It’s an aspect of campaigning that has grown over the last 15 years, with campaigns pouring more money into collecting and making sense of voter data each election cycle. According to Kim, campaign finance reports show that paid online ads by both parties in the last election cycle hit $1.1 billion, a nearly 5,000 percent increase from the $22.25 million spent on digital ads in 2008. The spending increase shows campaigns are relying more on technology and data to target voters online, she said.
While campaigns take big sets of data, like all the voters in a state, and pare it down to draw out individuals they want to engage, Kim uses the opposite approach in her research, examining grassroots data, voter-by-voter, to draw big-picture correlations.
In her project, Digital Ad Tracking and Analysis (DATA), she examines ads lobbed at individual voters on websites and social media and tries to understand the algorithm that directed the campaign to target a particular voter with a specific message tailored to him.
“We constructed a model that predicts ad features by user demographics. This sort of reverse engineering technique enables us to unpack algorithms,” Kim said.
Understanding the secret algorithms campaigns use to target voters is a crucial, yet missing, piece in examining the dynamic, Kim said.
Using data to target voters has been standard practice in political campaigns nationwide since at least 2000, but how campaigns acquire data and use it to target voters has continued to evolve with each election cycle.
“It’s definitely grown in every election I’ve been involved in, which is 2000 to this last one,” said Ken Strasma, founder of HaystaqDNA, a Democratic data firm based in Washington, D.C. Strasma works out of a home office in Middleton and has advised several presidential campaigns including Sen. Bernie Sanders last year, John Kerry in 2004 and President Barack Obama’s campaign in 2008.
“I don’t see it stopping,” he said.
HOW IT WORKS
In Wisconsin and nationwide, Republican and Democratic campaigns and state parties amass information, building and refining their databases each year. Both state parties have in-house data analysts. The state parties also contract with data firms who supply them with software platforms and large data sets.
Neither state party made their analysts available to discuss their internal data operations following interview requests by the Cap Times.
Publicly available voter registration rolls serve as the foundation of a party’s data reservoir. That is supplemented and altered each election cycle with information purchased from data companies that compile magazine subscriptions, purchasing history and other data that marketing firms use to target advertisements for products.
There are additional vendors that tailor data and help campaigns strategize to find out not just how a voter is likely to feel about a particular candidate, but which issues are most important to them. Then, the campaign will strategize on the best way to get the voter to show up at the polls on Election Day and vote for the candidate.
Republican Sen. Ron Johnson’s campaign attributed its win over Democrat Russ Feingold in the fall election, in part, to a strategy informed by internal predictions from data on Wisconsin voters.
The Johnson campaign and state party contracted with i360, a D.C.-based company funded in part by Freedom Partners, a conservative and libertarian leaning group founded by conservative billionaires Charles and David Koch.
Campaigns can download voter data from i360, add to it and manipulate it to create their own predictions for specific races. Republican campaigns and advocacy groups nationwide use i360, including Americans for Prosperity and the Republican Governors Association. Gov. Scott Walker’s gubernatorial and recall campaigns also contracted with the firm, along with Johnson and Republican members of Wisconsin’s Congressional delegation.
Data sets from Walker’s campaigns have been distributed to other Republican campaigns in the state. Those campaigns can, in turn, manipulate them to suit their purposes. The same process is used on the Democratic side, which has its own data firms and baseline data sets for campaigns to use.
Political data firms create searchable databases with a profile linked to each voter in the state. The database calculates a percentage “score” for each voter indicating the likelihood that she would vote for the candidate in question. A query for Jane Doe, for example, might show that she is 80 percent likely to vote for Johnson based on an algorithm that considers her voting history, magazine subscriptions and other data the campaign or party has gathered on her over the years.
“The more data we can get on people, the better, the more powerful our database becomes,” said Michael Palmer, president of i360.
Campaigns are looking to target subsets of voters they feel can be persuaded to vote their way by blitzing them with ads and reaching out in person or over the phone.
“The power of this data isn’t in what we have on any one person, it’s what we can classify a group of people as being,” Palmer said.“We can infer and make predictions about how other people might vote and might think. It’s really helpful for campaigns and other organizations to focus their limited resources to persuade and turn out the people they need.”
Johnson campaign manager Betsy Ankney gave a lot of credit to the campaign’s data operation for his win.
“We relied heavily on our data operation and trusted in that at the end of the day,” Ankney said. “We used the information that we got from them for everything that we did, from our targeting on digital and optimizing our TV buys. They gave us a road map on where Ron needed to spend his time.”
The Democrats’ answer to i360 is NGP-VAN, also based in Washington, D.C. It offers similar foundational databases of voter information for campaigns to download and manipulate. Firms including Strasma’s HaystaqDNA use that data to help campaigns target voters with specific messages and determine the best times of day to reach voters.
“We help answer the who, what, when, where questions,” said Strasma.
For the Sanders campaign, HayStaqDNA ran about 50 issue-based data models, along with Sanders-specific models in order to determine groups of voters most likely to be persuaded to agree with Sanders and what issues they cared about. Those lists were provided to the campaign’s digital advertising firm, Revolution Messaging, based in Washington, D.C.
Revolution Messaging then matched the lists of individuals to lists of online device IDs.
How does that match process work?
Any device that connects to the internet — laptop, desktop computer, tablet, smartphone — has its own ID. Digital advertising firms are able to match a person’s identity to the device she is using any time that person provides information about herself on a website from that device. This would include registering for a news website, an online account at a retail store or a discount website like Groupon, Strasma said.
“Most people have signed up for at least something,” he said. “From that moment on, you know that at least in one point in time this particular device was used to sign in for a particular person’s name and that is how the device ID-to-name matches.”
From the matches, advertisers pool similar types of voters into groups as small as 20 people and “serve” them ads on their internet devices, Strasma said. The method allows campaigns to target, follow and adapt to the search habits of specific voters instead of simply posting ads on websites and hope the right people see them.
Strasma calls this kind of advertising technology and strategy the “next generation” and said it broke ground as a major campaign tool in 2016.
Another common campaign technique, separate from the device-to-voter ID matching, is search-based targeting.
Campaigns use this technique to send ads to voters based on how they use the internet. For example, if a voter is trying to do general research about the candidates in an election, they will likely be blitzed with ads from both political parties.
But if a voter is doing research online using specific search terms that reveal his political position, like “Which candidate will defend the Second Amendment?” or “Which candidate supports common sense gun control?” his chances of seeing some types of political ads decrease.
“These are people who have already made up their minds in one way or another, so it is less valuable to get them a persuasion message, though it could be valuable for them to receive a get-out-the-vote message,” Strasma said.
HayStaqDNA also developed a strategy for the Sanders campaign to recruit volunteers and donors. It used computer models to predict which people would be likely to make small donations. Then they would ask for a bigger donation, then to make a phone call for the campaign, then to volunteer at a campaign event.
“We did a lot of tests and modeling on how to ask, how frequently to ask,” Strasma said. Those predictions helped the campaign more efficiently use its resources to raise money.
“Fundraising is probably the number one pleasant surprise for me,” he said. “(Sanders) started as a fairly obscure protest candidate and ended up shattering fundraising records.”
Sanders’ campaign raised $229 million before dropping out of the race in July, with 99 percent of the contributions coming from individual donors, according to the campaign finance website Open Secrets.
WHAT HAPPENED WITH THE DEMOCRATS' DATA IN 2016?
Having sophisticated tools to make predictions, with comprehensive voter and issue databases, did not translate to November wins for Democrats. Analysts on both sides of the aisle have acknowledged they were surprised at the margin of Donald Trump’s wins across the country, but particularly in Wisconsin and Michigan.
“I think everyone in public polling and in micro-targeting underestimated Trump support, so that was something going on across the board,” Strasma said.
Both Democrats and Republicans acknowledge the pitfall of overreliance on data and predictive modeling. A model is only as accurate as the data upon which it is based. Data that is incomplete or misleading is a mistake that might not be realized until it’s too late.
“Sometimes it can be the case where people go a little bit far with reliance on the data and forget sometimes about how the underlying information underneath the data may not be as strong and as accurate as people might think it might be,” said Joel Rivlin, the data analytics director at Pivot, a Democratic data firm based in Washington D.C.
Data should be one of several drivers in a campaign, which is most successful with a blend of strategies, analysts say.
“In politics, as in anything, there’s no silver bullet,” said Palmer, of i360.
Palmer said his company aims to keep an edge by continuing to experiment, evaluate its strategy and develop new and better products. Palmer said that election results nationwide and especially in Wisconsin with the Johnson campaign show they were effectively able to translate their data to results.
As companies and campaigns spend more money developing new tools to target and persuade voters, a balance remains paramount, Rivlin said.
“I think that’s a real problem. As we invest more and more and think that we are more precise in our measurements than we are,” Rivlin said. “(Data and predictive modeling) is really useful at narrowing your universe and the people you need to speak to so that you can exclude a good number of people who aren’t worth speaking to, but whether you can say confidently that you can narrow it to only the people that you need to speak to is often a stretch.”
The danger comes in getting lost in the numbers and missing the bigger picture.
“I think there’s a danger as we move into a data culture, to target a little too precisely and try to thread the needle. If the target changes over time, then you could be missing the mark,” Rivlin said.
It’s a problem some say Hillary Clinton’s campaign encountered, relying heavily on internal predictions to inform its approach to the exclusion of other traditional campaign strategies.
In December, Politico reported on the disconnect between the data operation at Clinton’s campaign headquarters in Brooklyn and campaign operatives on the ground in Michigan, a battleground state like Wisconsin that pollsters and analysts thought would solidly swing Democrat.
Politico reported that it talked to dozens of operatives in Brooklyn and Michigan with many relaying similar stories of how the Clinton campaign steadfastly clung to a data operation that proved to be wrong.
“The Brooklyn command believed that television and limited direct mail and digital efforts were the only way to win over voters, people familiar with the thinking at headquarters said. Guided by polls that showed the Midwestern states safer, the campaign spent, according to one internal estimate, about 3 percent as much in Michigan and Wisconsin as it spent in Florida, Ohio and North Carolina,” according to the Politico report.
Clinton made no campaign stops in Wisconsin after the primary election. And she neglected to visit union shops in Michigan, according to the Politico report.
But some Democrats say it is still too early to accurately pinpoint what exactly went wrong or cast Clinton’s data approach as misguided.
“The amount of money spent on data analytics is definitely rising … and sometimes you can point to instances where it’s not necessarily justified,” Rivlin said. “Every dollar that’s spent on a spreadsheet is a dollar taken away from talking to voters.”
Experts say it is important to be open to innovation on the data operation, but also remain skeptical of it.
THE NEXT FRONTIER
Though the trend in data technology is moving toward more software that is publicly available, competition between liberals and conservatives to develop the next big tool is often tight.
“It is an arms race of sorts,” Strasma said. “My Republican counterparts don’t tell me what’s going on under the hood and vice versa … everyone is keeping the best stuff secret.”
Adding more information to their databases can improve how successful a campaign can be, but what is often more important is how analysts can translate raw data to winning strategies.
“That is probably more valuable and is where a lot of the innovation is happening,” Strasma said. “I do think Democrats do have the better infrastructure technology and a deeper bench of analysts to do the number crunching aspect…one thing the Republicans did very well was psychometric modeling, where they were modeling voter attitudes.”
I360 is looking for more ways to encourage supporters to engage their friends and other people they know to discuss campaign issues. The company is looking into ways it can create social media platforms for people to engage one another, Palmer said.
“What we’ve found is the greatest thing to persuade and get people to do something isn’t TV ads, necessarily, or pieces of mail — they work — but the most effective thing is the human part. People they know reaching out to them and convincing them and sharing their views,” Palmer said. “Campaigns do a great job of knocking on doors, (and) we’re really trying to expand that to tap into more sources for campaigns in the online and social space.”
A UW RESEARCHER QUESTIONS DATA
While some may find micro-targeting manipulative or invasive, analysts say it is merely an extension of what marketers and other companies are already doing with the data that consumers freely offer up when using smartphones or allowing apps to track their behavior.
“The stuff on the cutting edge straddles the edge between cool and creepy,” Strasma said. “I wouldn’t be surprised if we see more regulation coming in the United States as more people realize how much data is out there about them.”
There is still a lot about the political campaign process the public should know, said Young Mie Kim, the UW researcher. She is still poring through ads she collected during the general election to try to understand how voters are targeted. Her findings are due in the spring.
Kim is examining ads received by more than 10,000 voters nationwide during the general election. She collected ads six weeks before Election Day from volunteers who agreed to download an internet browser extension that tracked the political ads they received. The browser extension worked like an ad blocker, but instead of blocking ads, it captured them and sent them to Kim.
Potential voters browse the internet and make choices about what they research and read that is then tracked by campaign data operations and used to blitz them with digital ads tailored to their behavior.
For example, Google ads are generated from feedback to search terms. If voters want to learn more about women’s issues and search for related information, they are more likely to get political ads focused on a candidate’s positions on women’s issues, she said.
“It’s a combination of a voluntary choice, but then involuntary feedback,” she said. “It’s more about persuasion than about get out the vote.”
Kim also looked at ads received by more than 800 voters in Wisconsin during the primary election. In that election, she saw that 50 percent of the people who participated didn’t receive any political advertisements at all. Those who received ads were typically people who cared a lot about a specific issue that maybe was out of line with the position of the political party they most identified with.
“Ironically, people who get the most political information get a narrower view,” she said. “The results of my primary election study indicate that algorithms perhaps propel the information gap between those strategically important to campaigns and those who are not, and sharply divide the public along the lines of campaign tactics.”
Kim is an advocate for making public the algorithms that campaigns use to make predictions. The way voters are targeted has implications for “information equality,” she said.
“No one knows how the algorithm is configured,” she said.
And as campaigns get better at targeting voters, analysts agree it’s a trend that won’t slow anytime soon.