Posted in Products and Services

Stellar debut for AP VoteCast

, by Lauren Easton

AP VoteCast, our new election survey of the American electorate developed with NORC at the University of Chicago, delivered impressive results in the U.S. midterm elections one week ago.

“The initial assessment of our survey results shows conclusively that we have developed a replacement for the exit poll that solves its most vexing challenge: heavy Democratic bias,” said Deputy Managing Editor David Scott, who oversees AP’s polling operations.

Scott continued:

The critical moment for an election survey is 5 p.m. ET on Election Day, when decisions about the evening’s editorial coverage are made. At that moment, AP VoteCast correctly forecast Republicans would hold the Senate with about the same number of seats as they had going into the midterms, and that Democrats would take the House.

In races for Senate and governor, AP VoteCast correctly projected the winner in 92 percent of races at 5 p.m. In the others, AP VoteCast had two as a tossup, with a projected difference between the candidates of less than one percentage point; three races remain too close to call a week after Election Day; and one incorrect winner was projected.

At 5 p.m. on Election Day, the average error of an AP VoteCast projection in a race for Senate or governor was 1.7 percent in favor of the Democratic candidate. In past elections at that time, the error in favor of the Democratic candidate reported by the in-person exit poll that AP VoteCast replaced was often measured in double digits.

Election workers sort a new batch of mail-in ballots collected the day before Election Day, in Renton, Washington, Nov. 5, 2018. (AP Photo/Elaine Thompson)

In 2016, AP did not use the exit poll results to call races. The Democratic bias was too severe and pronounced.

“In 2018, we sent AP VoteCast data to our decision team without any such reservation,” Scott said.

The response to AP VoteCast also exceeded expectations by 13 percent, as the survey completed nearly 140,000 interviews. That includes 40,692 random sample interviews that form the backbone of the survey — or 45 percent more than we hoped to complete. 

“We will continue to work hard with NORC over the next year to improve our methodology, but we start from a great place,” Scott said. “We’re eager to get into the weeds in our data analysis in the days and weeks ahead, and will publish that analysis publicly and release our data for academic review. Most crucially, we will be transparent about our completed interviews and how our calibration model yielded our final results.”

Read more about AP VoteCast.