Election polls come in a lot of flavors. Some sample registered voters; some try to screen for only “likely” voters. Some use humans to make the calls; others use robots. But most have one thing in common: They ask people whom they plan to vote for. And it’s those answers that are the headline result — the numbers that get the most attention from the politicians and the press.
A new paper by Microsoft’s David Rothschild and University of Michigan’s Justin Wolfers suggests that reading by those numbers is a mistake.* They compared state-level polls of voter intentions (“whom are you voting for?”) to polls of voter expectations (“whom do you think is going to win?”) to see which gives a more accurate prediction of an election’s outcome. Their surprise conclusion was that voter expectations polls are more useful than voter intention polls in predicting those results.
Now, they’re not saying that the percentage of people who think a candidate will win is the percentage who end up voting for that candidate. In the wipe-out elections that were studied, huge majorities expected a victory by the winning candidate. The researchers found, for instance, that 91 percent of voters in 1964 expected Lyndon Johnson to beat Barry Goldwater. LBJ did, but with 61.3 percent of the vote, not 91 percent.
So the numbers need to be weighted. But once they are, the error in predictions is much smaller than the error of conventional polls, at least for states. For 80.9 percent of the time, voter expectations polls correctly guessed the winner of the state in the elections charted below, while voter intention polls did so only 69.3 percent of the time:
Compare the correlation for conventional intentions polls:
And that for expectations polls:
Polls that ask which candidate voters support explain only 57.1 percent of the variation in final results. But polls that ask which candidate they expect to win explain 75.7 percent. That’s pretty good. An average of national polls in the last week explains a staggering 96 percent, and Emory University’s Alan Abramowitz has a model using GDP and presidential approval data with an even better 97 percent correlation. But those only predict national vote share, not the state vote shares necessary for Electoral College calculation. For noisy state polls, 75.7 percent is very respectable.
Rothschild and Wolfers conclude that the ideal model would give 90 percent weight to expectations and only 10 percent weight to intentions. That’s in contrast to the models of Princeton’s Sam Wang, the New York Times’s Nate Silver and Emory University’s Drew Linzer, which rely on intention data. Wolfers is working on a precise estimate, but at first glance the model is good news for President Obama, who benefits from a big lead in expectations polls. In the most recent Gallup poll respondents expect Obama to win, 54 to 34 percent. That would imply a significant lead for the president.
Update: Wolfers has released his projection. He predicts that Obama will get 52.5% of the two-party vote. By comparison, Nate Silver projects he’ll get 50.5%.