Although the polls failed to predict Ted Cruz's Iowa win on Tuesday night, it isn't because they were dead wrong.
They just ended too early.
Consider this: SurveyMonkey data from Iowa showed a seismic shift among GOP caucus goers in the final week -- with Donald J. Trump losing six percentage points and Texas Sen. Ted Cruz gaining six in the last six days.
Beyond the very real challenges that polling faces -- from single-digit response rates and the fast growing number of "cell phone only" households -- the Iowa misfire traces to issues that have challenged polls there for decades: large late shifts in voter preferences and the difficulty in predicting exactly who will trek out to caucus meetings on a cold winter night.
The surprise on Tuesday night was two-fold. First, Texas Sen. Ted Cruz defeated Donald Trump after at least 10 polls conducted over the final week of the campaign showed Trump leading. The final HuffPost Pollster average had Trump nearly seven percentage points ahead of Cruz, while the much heralded Des Moines Register/Bloomberg poll had given Trump a five percentage point lead. Cruz won by a better than a three-point margin (27.6 to 24.3 percent).
Second, Florida Sen. Marco Rubio, who had been polling in the mid teens - he registered just over 17 percent on Pollster's final aggregate - finished with just over 23 percent of the caucus vote, nearly overtaking Trump for second place.
The explanation for the Iowa polling miss was largely foreshadowed in an analysis by FiveThirtyEight's Harry Enten published on Caucus eve. Enten's review of past Iowa polling revealed "two types of candidates who tend to outperform their polls." Candidates who gain "late momentum," such as John Kerry in 2004 and Rick Santorum in 2012, and candidates who do well with evangelical conservatives. According to Enten, the final Des Moines Register polls "in 1988, 1996 and 2012 all missed the candidate favored by Christian conservatives by at least 8.5 percentage points."
Data collected by our firm, SurveyMonkey, helps confirm that both issues were in play for polling on the Republican caucuses this year.
In our data, there was a dramatic difference between the first (Jan. 21-26) and second (Jan. 27-Feb.1) six-day stretches. From the first six days to the second Trump lost six percentage points, while Cruz gained six and Rubio five -- looking at GOP voters expressing at least some intention of showing up Monday evening.
Momentum was running against Trump. Collectively, New Jersey Gov. Chris Christie, former Florida Gov. Jeb Bush, Ohio Gov. John Kasich, and Kentucky Sen. Rand Paul lost nine percentage points over this time period.
We're in the midst of collecting more than two million interviews about the 2016 election, as a continuation of the work we did for the 2014 midterms and the UK parliamentary elections in 2015. We're packaging some of this work in partnership with NBC News.
We collected the Iowa data as part of this larger effort. Since we're still refining likely voters models -- aren't we all? -- and we have not previously released any caucus or primary data, we shared the broad, 12-day results only with subscribers, showing plausible Trump leads of between four and nine points. We then segmented the data Monday evening, and the momentum leaped off the screen. Even in the end, our final six-days of data had Trump at 27 percent, Cruz at 25 percent; both close to their finally tallies, but the media-genic order still wrong.
Something else happened as well when dividing the time period: the evangelical share of GOP caucus-goers in the first six days was 39 percent; it jumped to 55 percent in the second six days (closer to the 64 percent evangelical share of Republican caucus reported by the networks' entrance poll). There was a lot shifting on the ground in Iowa.
One factor there was Cruz's campaign. In our data, fully 71 percent of Cruz supporters said they received mail from the campaign, easily outpacing the 56 percent of Trump's supporters who said the same. Cruz backers were also 17-points more likely than Trump's to say they'd received an email from a campaign, and 11-points more apt to say they'd gotten a phone call.
In addition to the five-point bump up we see for Rubio in the autopsy of our numbers, other sources of data confirm that a late trend favoring Rubio: The public polls had shown Rubio rising slightly over the final days of the campaign. The network entrance poll found that Rubio ran first, and Trump third, among voters who made their decision within the last few days before the Caucuses. And there are hints that other data showed a Trump decline, including Google search data in Iowa and, perhaps, the analytical models produced internally for the Rubio campaign.
This isn't a claim of any sort of hindsight victory -- it's a call to stop some of the breathlessness around the the sky is falling storyline about polls.