Pre-election polls are more important this primary election cycle than ever before.
Last week, Mike Bloomberg was given a place on the debate stage in Las Vegas because of his rising support in pre-election polls. Polls provide more than just interesting talking points — they can have direct consequences for candidates' futures.
With the first three Democratic contests of the season behind us, it's worth looking at how well pre-election polls have done so far at predicting candidates' support on Election Day. (We didn't analyze the accuracy of pre-Nevada polling but will do so in a future report.)
So far, we find that the ability of pre-election polls to predict a candidate's support on Election Day isn't exactly reassuring. Large errors are being made for some candidates, most notably Pete Buttigieg and Amy Klobuchar.
In the two weeks leading up to the Iowa caucuses, 14 polls were released and reported publicly. Two weeks before the New Hampshire primary, 31 polls were released, including 19 that were conducted and released after the Iowa caucuses.
There are several ways to analyze the performance of pre-election polls.
Because we are interested in how well pre-election polls are able to measure strength of support for a candidate, we simply compare the election results of the five leading Democratic candidates to their average support in the polls released in the two weeks before the election. For Iowa, we use the results from the initial preference ballot, because it's the vote that most closely resembles the original views of caucusgoers before any reallocation, in which they may choose candidates who weren't their first picks.
To keep things simple, and to focus on their overall performance, we treat all pre-election polls equally and make no attempt to judge whether or why are doing better or worse.
To evaluate how well polls are able to predict a candidate's support, we examine how the percentage of votes each candidate received compared to their average support in the polls. Candidates with positive values did better than the polls predicted, and candidates with negative values did worse.
The results are revealing.
Joe Biden underperformed — he did worse than the polls had suggested he would — in both contests relative to the pre-election polling average. Buttigieg significantly outperformed his polling average in Iowa and New Hampshire, and Klobuchar's surprise third-place finish in New Hampshire wasn't predicted by the polls at all. The performances of Bernie Sanders and Elizabeth Warren, however, were more in line with what might be expected given the polls' margins of sampling error.
Even though there were about twice as many polls in New Hampshire than Iowa, they continued to err in particular ways for certain candidates. Thirty of the 31 polls overpredicted Biden's performance, and all 31 underpredicted Klobuchar's result in New Hampshire. Nearly every poll — 28 out of 31 — underpredicted Buttigieg and overpredicted Warren.
It is tempting to suggest that the performances of the New Hampshire polls were the result of momentum after Iowa.
Looking only at the 19 polls conducted and released between the Iowa caucus and the New Hampshire primary reassuringly reveals that although the polling errors are smaller for every candidate — and although the polls dramatically underpredicted Klobuchar's performance — nearly every poll made similar mistakes by overpredicting the support for Biden and Warren or underpredicting the support for Buttigieg and Klobuchar.
Download the NBC News app for breaking news and politics
Large errors are being made for some candidates — most notably for Buttigieg in Iowa and Klobuchar in New Hampshire — and many of the polls are making similar types of errors for most of the leading candidates. Some of those errors may be the result of voters' changing their minds.
The NBC Exit Polls, for example, revealed that 36 percent of Democratic caucusgoers in Iowa and 51 percent of voters in the New Hampshire primary decided whom they would choose in the last few days before Election Day — last-minute decisions that may be difficult to capture in pre-election polling conducted before decisions were made. We will continue to track whether the polls are better able to predict candidates' support as the campaign continues.