Not Everyone Has Given Up on Polling

(Bloomberg Opinion) -- In the ongoing quest to render human behavior more comprehensible and predictable, some researchers are harnessing the upcoming election as a sort of natural laboratory for trying out wildly new kinds of prediction methods. Just asking enough people what they plan to do seems like it should work pretty well — but the U.S. 2016 election put the weakness of straightforward polling on public display.

As a new twist, one research group has asked people not just how they plan to vote but also how they think their friends will vote. There are a couple of potential advantages to this, said Santa Fe Institute social scientist Mirta Galesic, who collaborated on the project. First, it allowed them to harvest information from potential voters who wouldn’t normally be polled — people who don’t have land lines or don’t answer them. And, secondly, it taps into the nosiness that is part of human nature. People can be very good at knowing things about their friends’ health or marital problems, she said.

People also are less inhibited talking about other people’s behavior than revealing their own. And asking about friends gives a useful piece of information for political campaigns: Those potential voters whose social circles are predominantly on the other side are more likely to change their minds.

One nice thing about elections as experimental tools is that you get a definite answer against which to test your forecast, said Galesic. If researchers can eventually show it works, she said, social circle polling might prove useful for gathering data on public opinion, science literacy and health-related behaviors. She’s currently studying its effectiveness in predicting who will get a flu shot.

The idea has already been tested in two previous elections — the 2016 U.S. election and the 2017 French one, the results of which the researchers published earlier this year in Nature Human Behavior.

In France, the new method did fine, but so did traditional polls. There was a lot more room for improvement in the 2016 U.S. election. Traditional forecasters couldn’t predict the outcome in critical swing states — especially Michigan, Wisconsin and Florida. Social circle polling got four out of the five major swing states right.

Galesic points out that the election results in 2016 were still within the stated margin of error of the much-disparaged 2016 polls. A subsequent survey revealed that about 25 percent of Americans totally misunderstood the forecasts: polls indicated 80 percent odds that Hillary Clinton would win the presidency, but many people interpreted that as a prediction that she would win 80 percent of the popular vote. That would indeed make a Trump victory seem almost impossible.

Even people who understood probability forecasting might hope for something better, however. In studying the upcoming midterm election, Galesic and her colleagues are also test-driving a concept called a Bayesian truth serum. The idea, which was floated in 2004 by MIT management professor Drazen Prelec, is to follow survey questions with questions about how likely people think others will answer the same way. This can’t detect individual lies, but very broadly, the people whose personal answer is at odds with their estimate of the majority are more likely to be telling the truth.

That’s the theory anyway. The researchers aren’t planning to publish anything in the scientific literature until after the 2018 election result is out, though Galesic did post their forecast on her Psychology Today blog. In the Congressional Generic Ballot, a combination of Bayesian truth serum and social circle polling gives Democrats a 7 percentage point lead, which is lower than most polls, but about the same as what poll-cruncher Nate Silver has come up with using a weighted aggregate of many polls.

On individual states, social circle polling and Bayesian truth serum differed from Silver, but then, the goal is not to compete with him. Most people in the election forecasting game are using different methods to analyze lots of traditional polling data, while these researchers are testing the predictive power of a very different kind of data. If it works it could be very exciting for social science and for political campaigning. If it works.

This column does not necessarily reflect the opinion of the editorial board or Bloomberg LP and its owners.

Faye Flam is a Bloomberg Opinion columnist. She has written for the Economist, the New York Times, the Washington Post, Psychology Today, Science and other publications. She has a degree in geophysics from the California Institute of Technology.

©2018 Bloomberg L.P.