The national polls were right. On average, national polls found Hillary Clinton up by 3 points leading up to Election Day. She appears to have won the popular vote by about 1 point.
National polls have averaged a 2-point differential from the final result over the past 12 presidential races, so the 2016 polling was about as accurate as usual.
Based largely on eight national polls I conducted in partnership with Fox News over the fall, I thought Hillary Clinton would win the national vote by 2 to 4 points – enough that an Electoral College upset would have been out of reach.
But when I revisited our pre-election polling, the story of a potentially much closer race was right in front of me. Clinton’s support had been flat and her edge ranged from just 1 to 3 points, with the exception of right after the release of the Access Hollywood video when Trump’s support briefly slipped 4 points.
So why was there such widespread surprise at a Trump victory when Clinton’s edge in polling was within the margin of error all fall?
The main reason for the great surprise was that a lower probability outcome prevailed. Donald Trump drew the inside straight many said he would need in order to prevail. Before Trump’s election, 3 of 44 presidents lost the popular vote but won in the Electoral College. The odds of drawing an inside straight are 4 in 47.
But what happened that led to this lower probability outcome?
Late deciding voters broke disproportionately for Trump.
Undecided voters and those supporting third-party candidates in our polling tended to view both Clinton and Trump unfavorably, while also thinking Clinton had the judgment and qualifications and Trump did not. I thought these voters were unlikely to break disproportionately for Trump at the end, but they did. According to exit polls, voters who were unfavorable towards both candidates voted for Trump by 20 points (49% to 29%).
Nationally, voters who decided in the last week broke for Trump 47% to 42%. But the late break was even more decisive in Wisconsin (59% to 30%) and Pennsylvania (54% to 37%), where Trump achieved his biggest upsets. As it turned out, late-deciding voters were more interested in “draining the swamp” than in traditional qualities associated with competence.
There was a lower and harder ceiling on Clinton’s support than many thought.
Clinton never trailed in any of the eight polls we conducted over the fall, and I assumed this meant she had a higher ceiling of support than Trump. In hindsight, I should have questioned this assumption when Clinton’s support stayed stuck at 45% after the Access Hollywood video release. When voters didn’t move towards her after the release of a video that would have been fatal to any past candidate, why would they on Election Day?
There was no Rust Belt firewall.
Either there were some significant polling misses in Michigan, Pennsylvania and Wisconsin, or there was a huge late break for Trump. But either way, polling in these states contributed to a belief that Clinton had a Rust Belt firewall that would save her in the Electoral College, even if she lost the popular vote…
It is now clear that polling (both nationally and especially in the upper Midwest) was close enough that we should have spent more time exploring and articulating alternate outcomes rather than arguing for the data and logic that supported the higher probability outcome. But at the same time, we should recognize that when a lower probability outcome becomes the result, this isn’t necessarily a failure of polling or modeling. It might just be a lower probability outcome.
Inside straights happen, and Donald Trump just drew one.