HUDSON | Will data tracking make polls obsolete?

Miller Hudson | Aug 17, 2020

Political campaigns are changing in ways that have rapidly outpaced the capacity of pollsters and pundits to accurately comprehend how voters perceive candidates. The communication between campaigns and voters, once largely confined to TV ads and direct mail, is being replaced with ephemeral messages dispatched across the Internet to social media accounts. Even when you receive one of these dispatches, it’s unlikely to be the same appeal dispatched to your neighbor’s computer screen. Algorithms are mining your digital profile to capture those public policy issues of paramount concern to you and then craft a personalized campaign pitch highlighting your preferences. More than half of political campaign budgets today are spent on these platforms.

Two years ago, I spent a week in Eagle County a few months before the 2018 election. Each time I logged onto the Internet through our condo’s Wi-Fi connection, it opened to an ad for State Sen. Kerry Donovan’s re-election campaign. With a single click I could then access her campaign website. As a visitor, I couldn’t vote in her Senate election. The notice, of course, had nothing to do with who I was — it had everything to do with where I was. The Donovan campaign had erected a “digital fence” around the Senate district where all Internet users were greeted by the senator and their addresses captured when they proceeded to her website. Once I returned to Denver, I continued to receive campaign updates and fundraising solicitations. I still receive the occasional legislative update.

This relationship with the senator has been entirely virtual — invisible to pollsters, but not to Google analytics. Polling, despite the best efforts of Nate Silver to slice and dice and weight and average results, is growing increasingly suspect as a measure of public opinion. Sample sizes have had to be enlarged to improve reliability, while the mechanics of reaching a representative sample of voters are slow and expensive. Enter “social listening.” This euphemism may sound a bit like eavesdropping on political arguments at your favorite bar. Historically, reporters and pundits have referenced such snooping as, “anecdotal reports.” While watering holes are closed to inebriated inquiry these days, Google is soberly tracking our Internet searches, bundling the results to protect individual privacy and then selling that data.

In Colorado, Skylar White and his associates at unumAI now offer a scan of website visits as a substitute, or proxy, for polling — at a third its cost and a fraction of the time needed to turnaround reports. The 2020 election cycle will serve as something of a Beta test for his software. unumAI is providing advice to several campaigns this year, but not actively recruiting candidates. The premise behind its analysis is that voters’ on-line search behavior offers a more accurate glimpse of their opinions than responses provided in a poll. This requires a “leap of faith” many campaigns are not yet ready to take — not to mention, I suspect, nasty grumbling from the political consultants and pollsters they employ. If White is right, however, and his forecasts prove superior to polling predictions during this election cycle, pollsters will not face market disruption but, potentially, swift extinction. Political campaigns require and reward reliable results.

There are several potential pitfalls in replacing polls with data analytics. Who gets left out? Does curiosity translate into support? Skylar explains that 88% of Colorado voters access the Internet through Google, in the low 90s among Democrats and the low 80s for Republicans – providing near universal saturation in urban centers with scantier searches from rural areas, reflecting the state’s digital divide. Consequently, unumAI failed to see Lauren Boebert coming for Scott Tipton in the 3rd CD’s Republican primary, but neither did anyone else.

In the U. S. Senate race White’s results differ substantially from reported polls. Before COVID-19 arrived in Colorado, unumAI found Cory Gardner and John Hickenlooper running neck-and-neck. Only then did they see Hick’s campaign jump to the 10-12 point lead that was being reported in polls. Differing from a recently reported tightening in the Senate race to 5-6 points, Internet searches show the former governor holding on to his ten-point advantage. Trump appears down by 15.

Time will determine accuracy, but one statistical advantage to the unumAI approach is that it relies on a huge dataset appraising literally millions of searches. Pollsters, on the other hand, must extrapolate from a statistically significant sample (between 600 and 1,200 likely voters) and then project from those results to the entire electorate – thus their 3-5% margins of error. unumAI (they really need to rethink that name) has ambitions for 2022. They hope to market a Bloomberg-style terminal to clients where they can run their own analyses using the company’s proprietary software. Marketers do this kind of thing every day, so White’s proof-of-concept trial may well succeed.

See the original article on