Some Of 2016’s Biggest Polls Are Works In Progress

Fifty-state polls are the hot new thing this year, but pollsters admit they’re still learning how to use them.
Members of the community came to Univision's Vote For Your America Community Celebration on Nov. 3, 2016, in Norwalk, Califor
Members of the community came to Univision's Vote For Your America Community Celebration on Nov. 3, 2016, in Norwalk, California.

There’s a new kind of poll in the mix this year, an approach known as the 50-state internet poll. It’s notable because it provides an unprecedented trove of data on the status of the presidential election in every single state ― including the deep-red and solid-blue states that no one ever bothers to watch closely.

The outcome of the vote in those places isn’t really in doubt. But thanks to new developments in polling, we actually have an idea of just how big each candidate’s advantage or disadvantage is in the underpolled states. Three pollsters have embarked on 50-state polls this year, surveying thousands of voters in each state and releasing their findings regularly.

The most ambitious 50-state poll comes from SurveyMonkey, which has released daily updates on every state for the last two full weeks of the campaign. UPI/CVoter has released updates for every state for the last couple of weeks as well. And the Ipsos/Reuters poll has been polling most states for a couple of months.

But why do pollsters bother to survey all the states if we know how some will turn out?

“That’s where the world is going,” said Cliff Young, U.S. president of Ipsos Public Affairs. “The world is going toward large N-sizes in polls as well as aggregating. It’s only natural that any single pollster will opt for larger and larger samples.”

“Taking on an election and correctly projecting the outcomes is a way to demonstrate the power of the platform,” said Mark Blumenthal, SurveyMonkey’s head of election polling and a former Huffington Post editor.

The 50-state trackers aren’t giving you completely new data every day. They’re giving you the past several days’ data with a couple hundred more cases added on from yesterday’s polls. (At the same time, they’re dropping older data to ensure things stay current.) Most of these are what we call “rolling” averages, showing the average level of support for each candidate across a span of several days. Some trackers use a data window of as little as three days, while others are looking at several weeks’ worth of polls at any given time.

“There’s some risk involved, but I think that’s where the world is headed,” Young said.

The risk Young refers to is that the concept of surveying enough voters to create separate polls in all 50 states ― rather than just a national poll ― is new. In order to produce estimates for each state, the pollster needs at least a few hundred interviews per state. Telephone polls are too costly a way to obtain that kind of sample size, but online polls are cheaper and able to handle the volume more easily.

The question is how well online polls are able to cover every state. These polls work by recruiting people to a “panel,” meaning they agree to take surveys in exchange for some type of reward, like a small amount of money or points toward prizes. Pollsters try to match the people available in the sample to the known demographics of whatever population they want to represent in the poll. That means the panel needs enough people of all demographic groups to be able to target a representative sample. Panels have to be very large and well-curated to achieve this in every one of the 50 states.

The pollsters doing these 50-state web polls are quick to acknowledge the untested nature of the work.  

“This year we shifted from [telephone surveys] to online for our U.S. polls,” said Yashwant Deshmukh, founder and director of CVoter International. “We just wanted to cover as much as possible in order to learn on what worked and what didn’t at macro and micro levels in our model.”

Young similarly emphasized that Ipsos’ polls are a work in progress. “We’re still learning how to optimize the sample,” he said.

Blumenthal, too, spoke of a learning process. “SurveyMonkey has done lots of work on all aspects of survey research,” he said. “We’re learning from that to make the best estimates we can. Everything we’re doing is about learning lessons to enhance the experience of our customers ― both the tools that they have and the data quality.”

A couple of other online polling firms, YouGov and Morning Consult, have used polls plus a statistical technique called multilevel regression and poststratification, or MRP, to project the election in every state. MRP is a way of using data at the national level to estimate opinion in the states. Instead of relying solely on how people say they will vote in the survey, MRP incorporates information about the respondents and the states they live in that’s known to predict vote choice. Because these methods use other data besides polls, we haven’t included them in the HuffPost Pollster charts.

Young predicts we’ll see a lot more of these 50-state polls in the future. But for now, he said, testing out the method is a way to get a leg up on the competition.

“We’re learning this electoral cycle,” he said. “And next cycle, we’ll have it in good shape.”