No one answers the phone anymore. Back in the Clinton days, pollsters could collect voter opinions from about one in three calls. Today it is fewer than one in 11. Blame disappearing landlines — only 60 percent of U.S. households had one in 2013 — as well as cell-phone caller ID. Yet even as response rates plummet and costs of chasing mobile users soar, most data-driven election predictions still rely on phone-poll results. Even Nate Silver’s Five Thirty Eight models, which perfectly predicted 2012 presidential race outcomes in all 50 states and the District of Columbia, hinge largely on phone surveys. So researchers are now hunting for alternative sources of voter data — and finding them in unlikely places.
At YouGov, volunteers answer online surveys in exchange for gift cards. When the New York Times and CBS News announced in July that they would rely on YouGov for the November midterm elections, traditional pollsters were aghast. Opt-in surveys break the cardinal rule of polling: respondents should be a random sample of the population. But the global polling firm claims its demographic profiles rival those of even the best phone polls. Case in point: YouGov’s prediction for Obama’s two-party 2012 vote share was off by only one percentage point. Gallup’s error was nearly three times as high — and predicted the wrong candidate.
Video gamers are hardly a cross section of voting America. So when more than 340,000 Microsoft Xbox users— mostly male millenials — replied to mini surveys before the 2012 election, researchers at Microsoft and Columbia University mined the data with new statistical techniques that were designed to deal with nonrepresentative data. It worked. Not only could the analyses predict Obama’s vote share within 0.6 percentage point, their estimates for voting preferences of demographic subgroups — women older than 64, say — were a hair’s breadth away from exit polls. More mini surveys are coming soon to other Microsoft platforms.
Some forecasting researchers have turned to eavesdropping. Analyzing candidates’ Facebook friends has hinted at election results in New Zealand; Flickr’s geo-tags showed the spread of 2008 U.S. presidential campaigns; and Twitter’s data have predicted elections worldwide. In 2013 a project at Raytheon BBN Technologies analyzing 13 billion tweets in Latin America — funded by the U.S. intelligence community — found the number of candidate mentions could predict the winner’s vote share with errors as low as 0.6 percentage point. Adding in data from YouTube and Google Trends bolstered accuracy even more.
Ask voters who they think will win, and you will get better results than asking whom they are voting for. Why? The former also captures the election zeitgeist — friends’ opinions, pundits’ rants, and more. Betting pools do, too. At online prediction markets, people buy shares in election outcomes; higher market prices reveal a stronger combined belief in the outcome. Although election gambling is illegal for U.S. citizens, a few international futures markets such as the U.K.’s Betfair and bitcoin-driven Predictious are thriving. Analyses of Intrade’s price history before the company was shuttered predicted Obama’s 2008 win with excellent estimates in contested races — and even more accurate than Nate Silver’s.