Rise 2020: How Data Science Can Influence Election Turnout
Springboard’s annual Rise 2020 virtual conference continued this week with a very special panel discussion touching upon one of this year’s most talked-about topics: the upcoming presidential election.
If you read the news (or have a Facebook account), you’ve likely noticed a slew of ads in recent months from the presidential campaign that reflects your political persuasion. These ads are no coincidence. Candidates running for elected office tap audience solutions providers like PredictWise, a company that uses data science to increase voter registration and influence election turnout, to help them reach voters through data-driven ad targeting.
Recent Stanford Ph.D. graduate and soon-to-be Cornell professor Nikhil Garg took the virtual stage to discuss his work at PredictWise.
“We don’t focus on the current state of the race: our goal is to change the state of the race through the ads that are delivered,” Garg explained via videoconference.
Using data to target voters
While print ads and TV spots were once the prime advertising mediums for reaching voters, today’s political candidates must vie for a voter’s attention in a fragmented digital landscape.
To do this, they need granular data showing people’s voter registration status, which social media platforms they use, how frequently they consume podcasts and video content, what publications they read, their political beliefs, and the type of messaging most likely to galvanize them to the polls.
Identifying those voters in the first place is no easy feat. No central database currently exists documenting all registered voters and eligible would-be voters. Instead, each state maintains its own voter file of registered voters, but this data isn’t always reliable.
During the coronavirus pandemic, large swaths of people have deserted major cities like New York, Seattle, and Los Angeles for areas with a lower cost of living, resulting in droves of voters who are unregistered in their new home state—and, consequently, harder to reach.
PredictWise has a solution for that. “We want to identify people who no longer live where they’re registered to vote and re-register them,” said Garg, who holds a Ph.D. in electrical engineering from Stanford. To do this, the company partners with a nonpartisan voter registration organization whose goal is to maximize voter turnout. Using anonymized GPS data from voter’s smartphones, they can determine if someone has recently relocated.
“We acquired 180 billion GPS points for 180 million unique people over the course of three months,” said Garg. “So someone’s voter ID might say they’re registered in Florida, but their GPS data says they’ve been spending many, many nights in Pennsylvania.”
Each ad is targeted to a specific audience segment according to voter registration status and a mixture of demographic and psychographic factors. For example, a voter who recently moved might receive a nudge reminding them to register in their new state. Conversely, a registered voter is more likely to see ads alerting them to important election deadlines.
How data can change the outcome of an election
Beyond providing handy reminders to encourage voter turnout, election analytics offer a game-changing value proposition: persuading unlikely voters to support a particular candidate. “For example, a democratic candidate might ask for a list of people in their state who lean Republican but are really progressive on healthcare,” explained Garg.
The candidate would then use a combination of video ads, banner ads, email campaigns, and sponsored social media posts to convince the voter to support their healthcare platform.
Changing a person’s political leanings is exorbitantly difficult, but swaying undecided voters or mobilizing ambivalent voters can transform the outcome of an election. Current polling data puts undecided voters at about 10% of the registered voter population—half of what it was in 2016.
To assemble a full picture of a candidate’s potential voter base, data scientists aggregate data from numerous sources including third-party vendors, voter registration organizations, and state-run databases created and maintained by the secretary of state.
Garg says each voter is distinguished by “digital identifiers” like their mobile identification number (a 10-digit number that a wireless carrier uses to identify a mobile phone) rather than identifiable personal information. “The data we work with is extremely sensitive and we never want to connect this with an actual person,” Garg explained, adding that privacy and security are a number-one priority when handling election data.
Having access to telemetry data from voter’s smartphones allows data scientists to gather very fine data points, such as a list of apps installed on someone’s smartphone along with real-time usage statistics. For example, if a voter spends more time on Instagram than Facebook, a candidate would find it more financially worthwhile to target them with an Instagram banner ad or live video rather than a Facebook ad.
A very messy business
Given the piecemeal nature of this type of data gathering—such as cross-referencing consumer bios from a credit agency with financial data indicating a voter’s income bracket—the data isn’t guaranteed to be accurate. Garg says it’s important to validate the data before acting upon it by asking voters to reconfirm certain information through surveys.
“We have a bunch of polls out in the field,” he said. “We collect behavioral, attitudinal, and personality data tied to their demographics.”
This data is then aggregated with historical data collected over the years, where applicable, and each voter is assigned a score indicating the likelihood they’ll support the political candidate in question.
“Finally, now that we have these scores on all these mobile ad IDs for over 100 million individuals, we can create ad audiences,” said Garg. “If a candidate asks for a particular [voter profile], we can also serve ads to that audience.”
Garg says the vast troves of data they collect has enabled data scientists to make some intriguing inferences about potential voters. For example, people who have the Hobby Lobby app installed on their phone are more likely to be Republican. They can also build highly specific data pipelines, such as ‘voters who frequently post on TikTok and also support a billionaire’s tax.’
For the most part, however, Garg says election analytics serve simply to parse out the data at scale and build audience segments that help candidates target voters most likely to mobilize or convert. Hence why basic techniques like cleaning messy data are really important for a data scientist to do this kind of work, rather than the ability to glean earth-shattering insights about a certain voter cohort.
“If you’re someone who’s starting a career in data science right now, data cleaning and intuition are a thousand times more important than whatever fancy machine learning technique you might use,” said Garg. “Ninety-nine percent of the time you don’t need those techniques.”
In a time of deep political polarization, where social media companies have been accused of promoting “echo chambers” that discourage dissent while purposely riling up people who have opposing views, one may wonder whether political advertising further entrenches these biases.
While some ads do promote incendiary content, Garg says that political advertising, when done right, can be used to educate voters on causes that matter to them—especially when it comes to voters who are not strongly partisan.
“A whole bunch of people have quite diverse views when it comes to these issues,” he said. “They might tend to vote for law & order but are liberal when it comes to healthcare and economic issues. Our work serves to inform people about these issues.”