Today the New York Times started live-polling registered voters here in Virginia’s 7th U.S. Congressional district. Historically, our district is reliably red, but some forecasters are now calling it a toss-up between Republican incumbent Dave Brat and Democratic nominee Abigail Spanberger. As responses stand now, Brat has an edge, but it’s within the still-large margin of error.
What is live-polling? The Times has put together a variety of real-time visualizations of polling responses as they come in, supplemented by explanations of how the poll works and assessments of its predictive power. This type of transparency helps to demonstrate the social-science principles that motivate good poll design.
Like meteorologists, pollsters get a lot of grief when they predict future events incorrectly. But despite our complaining, we still check the weather forecasts every morning because there’s no better source to help us decide whether to grab our umbrellas on the way out the door. Polls are the same: They’re the best information we can get now, in September 2018, about conditions in 2019 that might affect our decision-making.
Every comparison of poll responses to a political result is more data that can improve predictive power. That’s what pollsters are after following Election 2016, when most polls underestimated the likelihood of a Trump win. (I’ll never forget watching the Times’ victory-probability dial that night.) The Times feature is refreshingly transparent about its methodological limitations and how to think about them.
For example, check out the incredibly low response rate for 18- to 29-year-olds compared to other age demographics. As of this posting, only 10 out of 1,274 people called in that group have been interviewed (not shocking). That rate is only halfway to the pollsters’ goal for that group, but they’ll be satisfied if they can get just 8 percent to respond. By contrast, they’re hoping to reach 32 percent of the 30- to 64-year-olds they call. So they were always expecting to have to weight their youngest respondents more heavily in quantifying their prediction. The uncertainty of that thumb-on-the-scale is part of what determines the margin of error (an unhelpful 7 percent at the current sample size).
For what it’s worth, fivethirtyeight.com gives Brat a 5/8 chance of victory on November 6.
Update, Sept. 12, 2018: The poll is still underway, with over 20,000 calls made and just shy of 500 responses. My favorite intermediate result thus far is one of those funny overlaps you sometimes see when you compare different response groups: As of now, an even 50 percent agree that the Russia investigation is a “politically motivated witch hunt.” Yet 52 percent support Robert Mueller’s investigation “into the Trump campaign’s ties with Russia.” I’m curious about that two percent who support it despite believing that it’s politically motivated.