How to De-Ice the Windshield of a Political Campaign
Understanding, Targeting, and Influencing Voters
If you don't know where you are going, you might wind up someplace else.
- Yogi Berra
Which takes more guts: starting a company, or running for office?
If starting a company is like jumping off a cliff and assembling a plane on the way down, running a political campaign is like flying a plane through a storm while trying to scrape the ice off the windshield.
Political campaigns start with an icy windshield and an unclear flight path.
Skilled practitioners can de-ice the windshield with a few key data points: an understanding of the voters and their concerns, a data-driven approach to targeting or modeling those voters, and as-good-as-possible data on what works to convince voters.
What don’t they have? Anything that looks like the stories told about Cambridge Analytica.
The Cambridge Analytica Myth
In early 2017, the media was obsessed with a simple question: how on earth had Donald Trump won the 2016 election?
Journalists across the world soon seized on a provocative, engaging answer: that Cambridge Analytica could predict the personality of every single adult in the United States of America. Facebook data, the story went, was leaking everywhere; clever data scientists affiliated with the Trump campaign had scooped up all of your Facebook likes and follows, now knew you better than your spouse did, and could precisely manipulate you and millions of American voters.
Not long after those stories broke, a high-level political candidate told me a very different story. This person had recently run as a Democratic candidate, had raised tens of millions of dollars, and executed what was by most accounts a sophisticated campaign. And yet, the candidate said, the campaign had been “flying blind” day-to-day, lacking the data that could help provide a path to victory.
The first thesis – that a political campaign understands the personalities of every voter and is precisely pulling strings to manipulate them – was absurdly misleading. Cambridge Analytica was (almost) all hype.
At the same time, most political candidates aren’t quite flying blind — but they need to be savvy to scrape the ice off their windshield.
Understanding Voter Sentiment
Campaigns usually rely on polling to understand voter sentiment. Polling gives campaigns a more specific sense of an individual race, answering questions like:
Does the electorate know who our candidate is? Do they know the opponent?
Do they view the candidates favorably? Are there particular issues on which they favor one candidate or the other?
What concerns are most pressing for voters right now?
How do voters feel about a few specific hot-button topics?
How do voters respond to different descriptions of the candidate and her opponent?
Are there local leaders – political or otherwise – who are popular? Would it be helpful or unhelpful to tie our candidate closely to the well-known mayor?
Polling insights will significantly inform a campaign’s broad messaging and approach. Candidates’ websites will usually feature issues that polls show are important; candidates’ bios incorporate the elements that resonate with the electorate.
One longtime political consultant told me about working with a candidate who had previously run a flower shop. The candidate wanted to talk about all of the awards he’d won; polling showed that voters cared very little about his awards, but were excited to support someone who understood the challenges of building a small business in their community.
Polling insights also help campaigns target their communications more precisely. If a candidate is less liked or less well known with a specific demographic group – voters under 30, middle-aged women – the campaign may spend more time or money to reach out to voters in that category.
This relatively simple targeting resembles neither the overhyped claims of Cambridge Analytica, nor the more sophisticated algorithms of companies like Facebook and Google. It is nonetheless quite helpful to resource-constrained campaigns.
In 2016, I volunteered for two city council races, neither of whom ran a poll. Both candidates were effectively flying blind: they had no quantitative sense of which issues really mattered for voters (they had plenty of anecdotes) or whether the way they were framing issues was effective.
In 2017, in part to address this challenge, I started a polling company called Change Research. Change Research has since worked with hundreds of similarly small campaigns (those with a budget of $100,000 or so) to deliver a quantitative sense of what voters cared about.
Time and again, we’ve found that “what voters care about” can be quite different from prevailing wisdom in and around campaigns. While insiders political activists in Virginia were most focused on a potential oil pipeline, voters cared more about “boring” stuff like traffic and the cost of healthcare.
In many ways, polling in California is even more interesting and more valuable than polling in purple states. When issues are nationalized, voters’ views often line up closely with their party. In San Francisco, however, different issues and opinions break through. Voters are concerned about education, homelessness, crime, and the cost of housing in ways that don’t closely adhere to national politics. De-icing the windshield may be especially valuable in our warm, sunny state.
Modeling Voters
Using available data – from the voter file and other commercially available sources – campaigns can get access to predictive models built from large surveys, with scores for every voter on the voter file. These scores are valuable to campaigns, but they’re neither as accurate nor as creepy as the stories told about Cambridge Analytica.
A voter named David might have both definitive data from the voter file:
29 years old
Male
David lives at 37 Main Street in a precinct that voted 71% Democratic in the 2020 election
David voted in the general election 2020 and 2018, didn’t vote in 2016, 2017, or 2019
David voted in the Democratic primary election in 2020
David is registered as an independent
And, with access to probabilistic models, a campaign might predict the following about David:
60% chance that David is Latino
70% chance that David has a college degree
50% chance that David is Catholic
80% chance that David voted for Biden in 2020
40% chance that David will vote in the general election in 2022
10% chance that David expresses climate change as a top concern
50% chance that David expresses housing costs as a top concern
20% chance that David expresses healthcare as a top concern
Because models say David has a 40% chance of voting in 2022 and most likely voted for Biden in 2020, David would almost certainly be the target of turnout efforts by a Democratic campaign. David would likely vote Democratic but is far from guaranteed to vote. The campaign might choose to play up their candidate’s actions on housing, as the model says this is likely a key issue for David.
By contrast, imagine that the model said that Jane, a 60-year-old woman also registered as an independent, was 50% likely to have voted for Biden in 2020 and is 90% likely to vote in 2022. Our campaign would likely work to convince Jane to vote for the Democratic candidate and against the Republican.
In California’s areas that are heavily Democratic, many of which feature races between Democrats, predictions of who will vote are likely to be more accurate than predictions of who they will vote for.
However, clever and well-resourced campaigns can (usually with outside help) field large surveys and use them to build predictive models gauging likely support for each candidate. Using a poll of a few thousand respondents, a data scientist can build a model that uses data from the voter file and elsewhere to predict support.
These models allow a campaign to rank order voters based on their likelihood of supporting the candidate and/or turning out to vote. This allows campaigns to more efficiently use resources to turn out voters likely to vote for their candidate and persuade voters who might reasonably be persuaded.
These models are very useful, though not perfect. A campaign that has the expectation “we should knock on this person’s door and tell them about our candidacy as we believe there’s an 80% chance they’ll vote for us if we can get them to vote” is setting itself up to de-ice the windshield. A campaign that is expecting the sort of dystopian predictiveness that Cambridge Analytica claimed is setting itself up for failure.
Moving Voters
It’s easy to get internet users to sign up for a new product; it’s hard to get voters to change their votes.
In previous lives, I developed a/b testing systems for consumer internet companies. Especially in an early, unoptimized product, successful a/b tests can increase conversion massively. Before it was a standard feature, a colleague tested showing friends’ faces in a signup flow and saw a 30% increase in conversion.
By contrast, in a partisan political race, it is incredibly difficult to persuade someone from voting for a Republican to voting for a Democrat — it is usually easier to persuade someone to change their vote in a primary election.
Because polling people is expensive, and getting people to politically change their minds is (in most cases) difficult, it is surprisingly rare to see political campaigns use scientific, experimental techniques to see what moves voters.
Small campaigns will often run polls that ask voters to evaluate how convincing a message is – which is not necessarily indicative of the message actually moving minds. Larger campaigns and organizations more frequently conduct scientific experiments – e.g., breaking respondents into 2-3 buckets and treating each bucket differently – but they are far less exhaustive and rigorous than consumer tech companies.
We used this approach with the Pod Save America team to find effective messaging on education policy: it’s better for Democrats to talk about protecting teachers’ freedom than to talk about denying the ugly parts of American history.
Because rigorous content testing is difficult, it’s likely that most of the political video ads you see have not been empirically tested. And the Facebook ads that are shown most frequently tend to be those that are especially engaging, but not necessarily those that are especially persuasive.
In California, there’s some good news for political practitioners: because so many races are Democrat vs. Democrat, it’s easier to find messages that move voters. In a low name-ID race like AD-21 (San Mateo County), where the two leading candidates (Giselle Hale & Diane Papan) have name recognition below 40%, poll-tested messaging could mobilize Democratic voters to vote for them.
Targeting Voters
When I went into politics, I was struck by the public availability and use of political data, especially after witnessing many politicians bash tech companies for their use of personally identifiable information. In early 2017, I filled out a form with the New York Secretary of State to get access to the voter file; I was quite surprised when a few days later I received a CD in the mail with the voting history of every single voter in the state of New York.
Access to data like that is extremely important: political targeting is driven in large part by the information available on the voter file.
I strongly recommend Hacking the Electorate to anyone who wants to understand how political campaigns target and reach voters.
The book, written after the 2012 election, looks closely at the way that campaigns work in different states. It finds that the structure of each state’s voter file – the publicly accessible record of every voter in the state – has a significant impact on the way that campaigns are run in that state.
In some states, voters register by political party; in most of those states, political campaigns can accurately distinguish between likely supporters of the candidate (who are registered to vote with the campaign’s party) and likely supporters of the opponent. That means they can effectively deploy resources to excite and turnout fellow partisans.
By contrast, some states do not register voters by political party. In these states, it’s harder to gauge whether someone is a likely supporter.
Likewise, some states have data on voters’ race on the voter file; for Black voters in particular (who tend to vote overwhelmingly for Democrats), this is extremely predictive.
In California, the voter file includes the party with which each voter has registered (or No Party Preference if a voter declines to register with a party); it does not include race or ethnicity. It also includes voting history.
That data means that in California, a Democrat running against a Republican might try to turn out voters with the following characteristics:
Voted at least once in the past election, but not in every election
Is registered as a Democrat
Such voters will most likely vote Democratic, have shown that they will sometimes vote, and have also shown that they won’t always vote.
Campaigns with more resources will use more sophisticated models — but in California the simple model is a good start.
Moving Other Needles
The part of a political campaign that is best measured and understood is fundraising. Campaigns can easily gauge the effectiveness of their phone calls, emails, and “the sky is literally going to fall if you don’t rush $25 to our campaign by midnight” Facebook ads on driving donations.
Measuring success on a fundraising campaign is easy: the candidate need only see whether someone gave money. Measuring success on a persuasion campaign is much more challenging: it’s harder to change someone’s vote, and more difficult to register the change.
That means that a campaign’s messages are more likely to be tightly optimized for fundraising than for persuasion – part of the reason that that polarization is increasing in the US.
In Search of the 4K Windshield
Five years ago, I started a polling company with the hope that we would crack the problem of attribution in public opinion: what changed someone’s mind?
I now believe the problem may be impossible to solve — Cambridge Analytica stories aside, it’s extremely difficult to understand why we humans think the way we do.
However, understanding voter sentiment, targeting the right voters, and finding effective messaging are all incredibly important to political campaigns. They collectively serve to de-ice the windshield and prevent the pilot from flying blind, even if they don’t provide the perfect 4K view of the world.