The two Senate races where the actual winner was different from the leader in our polling-based projections were Colorado and Nevada (it’s also probable that we’ll miss Lisa Murkowski’s victory in Alaska once all her write-in votes are — eventually — counted).
The miss in Colorado wasn’t bad — our forecast had Michael Bennet as about a 1-point underdog, and he won by 1 point instead, although there are still a few votes left to count.
In Nevada, however, where most polls showed Sharron Angle ahead and Harry Reid instead won by almost 6 points, the polls were pretty far off the mark. Errors like that occur quite frequently in primaries and in House races, where the polling landscape is tougher. They also occur sometimes in lopsided races, which are more difficult to hit on the nose. It’s fairly unusual, however, to have the consensus of polls off by 7 or 8 points in an extremely competitive Senate or gubernatorial general election.
I riffed a little bit last night on why the public polls might have been wrong in Nevada; I speculated, for instance, that the fact that Mr. Reid is the sort of candidate whom one votes for unenthusiastically might have skewed the turnout models.
There is another theory, however, which was proposed to me last night by Matt Barreto of the polling firm Latino Decisions.
“There is one overarching reason why the polls were wrong in Nevada,” Mr. Barreto wrote in an e-mail to FiveThirtyEight. “The Latino vote.”
His firm, which conducts interviews in both English and Spanish, had found that Latino voters — somewhat against the conventional wisdom — were relatively engaged by this election and for the most part were going to vote Democratic. Mr. Barreto also found that Latino voters who prefer to speak Spanish — about 40 percent of Latino voters in California meet this description, he told me — are particularly likely to vote Democratic. Pollsters who don’t conduct bilingual interviewing at all, or who make it cumbersome for the respondent to take the poll in Spanish, may be missing these voters.
Mr. Barreto had noted to me earlier this year that he felt polls might be overestimating support for Republican candidates in California. Indeed, the polls in that state called the right winners — Democrats Jerry Brown and Barbara Boxer — but underestimated their eventual margin of victories.
Colorado, the other state where most polls picked the wrong winner, is also a state with a fairly heavy Latino population.
As a back-of-the-envelope test of Mr. Barreto’s theory, I compiled results from the eight states with the largest share of Latinos in their population: these are Arizona, California, Colorado, Florida, Nevada, New Mexico, New York and Texas. There were 15 gubernatorial and Senate races last night between these states.
The table below compares the results from FiveThirtyEight’s simple weighted polling average (not the fancier version we use to make our official projections, although the differences are usually slight) against the results from last night.
(Note: in the two races in the chart where a third-party candidate finished in second place — the Colorado gubernatorial race and the Florida Senate race — we count Tom Tancredo as being a Republican and Charlie Crist as being a Democrat, respectively.)
In 10 of the 15 races, the polling average underestimated the Democrat’s margin by at least 2.5 points. There were 5 other races in which the Republican somewhat beat his polls.
Overall, however, the Democrats outperformed their polls by 2.3 points in these 15 races. There’s enough state-to-state variance in the results that we can’t come to any firm conclusions about whether inadequate sampling of Latino voters was the cause. Still, if you look at the presidential polling in 2008, it also underestimated Democrats’ performance in states like Nevada, New Mexico and Colorado, where they won by larger-than-expected margins.
So, we have at least the beginnings of a pattern — and considering how rapidly the Latino population is growing, it’s one that pollsters are going to need to address in states like Nevada, California and Texas if we’re going to be able to take their results at face value.