Political polls may have taken a beating in the last presidential election, but we shouldn’t count them out quite yet.
After President Donald Trump, who was predicted to lose the election by a wide margin, emerged victorious from the 2016 presidential race, stories about polls were thrown into the “fake news” shredder. Although the fault may have lain with how we interpret them, polls lost a significant amount of hard-earned trust in the eyes of the public.
If we drill down to the data, polls remain a potent, though not infallible, tool for determining how the democratic process might play out. In a paper published Thursday in Science, researchers examined more than 600 international elections across seven decades and found that polls were reliable predictors of election outcomes—90 percent of the time. They then built a model based on polling data and backed up their tough talk with several real-world tests, finding that they were able to successfully predict more than 80 percent of 36 ongoing elections.
Their model relied on other factors besides polling, of course, such as the level of democracy, whether or not an incumbent ran (incumbents have a slight edge) and foreign relations. When they took polling data out of the model, however, they saw a substantial drop in predictive accuracy. Even in places where polling data was poor, they say that it still affected the reliability of their model, indicating that polls generally picked the right candidate.
Though polls may be important, some refinements seem to be in order. The researchers found that they could improve their predictions by accounting for biases in the polls—a tendency to pick one candidate over the other based on differences in the type of populations sampled, among other things. They also created an updated model by “smoothing out” weaknesses in polling data by accounting for regional variations, whether the incumbent candidate is running and economic factors.
This paper is part of a special section in Science looking at the power of prediction. Other essays cover the accuracy of predictions of violence and scientific breakthroughs, the role of artificial intelligence in predicting the best policies and ways to make predictive science better.
Pollsters, and the statistical machinations that underlie their predictions, may have been mildly embarrassed in November, but the presence of polls, and predictive science in general, isn’t going anywhere—we predict. The authors point out that statistical models never claim the ability to peer into the future, they only rank various outcomes as more or less probable. Even someone with a 16 percent chance of victory, as Trump had in the researchers’ model, will win sometimes. Add to that the fact that most polls tracked the outcome of the general election fairly well, and it’s safe to say that polls will remain a important tool for political scientists.
The authors sum it up best themselves in the last line of their paper:
“We predict that reports of the death of quantitative electoral forecasts are greatly exaggerated.”