For those charged with analyzing polling and turnout data to predict the outcome of the 2016 U.S. election, the question now is, “How did we get this so wrong?”
Following the election of Donald Trump on Tuesday, one thing is clear: the models for polling voters and predicting election outcomes will need to be reworked. Many of the leading predictive models completely inverted themselves over the course of election night as actual results emerged.
“There’s going to be a massive, massive postmortem,” says Larry Huynh, partner with digital strategy firm Trilogy Interactive. “There were many conversations about the future of the polling industry after the results of the mid-term election in 2014. Now with the results of the 2016 election, there will be a huge indictment of the polling industry.”
Where the voting models went wrong
Political polling was once a much simpler exercise: reach out to a large enough sample of U.S. citizens, and you were likely to extrapolate a reasonably accurate picture of the electorate. Not so today. Polling companies must craft and run turnout models to estimate the numbers of registered voters that will actually cast ballots. It became clear on Wednesday that most of those public turnout models were mistaken, says Huynh, particularly in states like Michigan, Wisconsin, and Pennsylvania where both parties assumed that Hillary Clinton would win.
The question is whether there were problems with the models themselves or the data fed into them. “People will be looking at whether this was a turnout model issue or whether people just weren’t honest with pollsters,” says Huynh. “No turnout model can account for respondents not being forthcoming.”
As that most important analytical input—how people voted—becomes available, political organizations, analysts and journalists can assess what actually happened with the electorate and—perhaps—why.
The first new data are exit poll results. “Looking at those helps us frame the election—who voted and how they voted,” says Alex Lundry, co-founder and chief data scientist at media analytics firm Deep Root Analytics. Journalists and other media professionals will use that data to craft—and in this unusual election, re-craft—their election narratives.
However, that immediate, crunched-on-the-fly data can be misleading. “Exit polling data has to be taken with a grain of salt because there are few controls regarding who gets interviewed or decides to answer questions,” Huynh says. “There tends to be a lot of lazy analysis and overgeneralization right after the election.”
Decoding the results for future campaigns
The more nuanced analyses will take place in the coming months as secretaries of state make their voter files available. “We have a wealth of information about these registered voters,” says Lundry, who has served as political data scientist for presidential candidates, national organizations and Fortune 50 companies. Data scientists can pore over those updated voter files and the data they have appended to them, such as a voter’s consumer preferences, personal interests and detailed demographic information. “They’ll get a better sense of the turnout and see if African American and Latino turnout was over-predicted or rural turnout was under-predicted,” Huynh says.
Political organizations will also use this data to assess their marketing and advertising tactics. “The data sources we have for understanding what we’re trying to accomplish with online, mobile, and television advertising is incredibly granular. Campaigns can map their ad exposure to micro-targeted segments,” says Andrew Lipsman, vice president of marketing and insights at ComScore. “There’s a huge opportunity to go back and understand what worked.”
Political organizations can measure the effectiveness of the controlled experiments they conducted throughout their campaigns. Given the new types of digital communication and targeted advertising used this election cycle, that in-depth analysis will provide critical guidance going forward. Advertising intelligence will be most beneficial to down-ballot candidates who rely on media buying for name recognition. At the presidential level, the focus will be the effectiveness of ground operations. “You finally have that outcome variable so you can compare what you tried to do with the result,” Lipsman says. “It’s like any other marketing efforts: what percent of people did you get to go to the polls and to what degree did they convert.”
That analysis will help determine how campaigns may be run in the future. Analysts will be looking at how Trump, clearly outspent and out-organized, managed to connect to the electorate. “The Trump campaign was strong and was able to tap into certain aspects of voter sentiment by delivering a consistent theme,” Huynh says. “That was clearly enough for Trump. The question is whether that will be transferrable to other candidates or whether this was an anomaly.”
The data may not provide a clear direction on how to influence voters in a rapidly evolving media environment. Voters today are “living in this media bubble where they self-select information that reinforces their existing belief, which is completely changing the way campaigns will be handled going forward,” says Lipsman. “That’s what we need a better understanding of for the future. But getting good data around that is excruciatingly difficult. You’re trying to understand not just what someone was exposed to but how it made them feel.”
Analyzing the accumulated data may inform not only future campaign tactics, but overall strategy. “Don’t underestimate the power of understanding who the electorate was and what they wanted,” Lundry says. That clarity will also be used to refine the predictive analytics models used by campaign two and four years from now.