At The American Prospect, Jamelle Bouie flags an important post by Matt A. Barreto of Latino Decisions, “Why Pollsters Missed the Latino Vote – 2012 edition,” which notes:
Let’s examine how these faulty Latino numbers create problems with the overall national estimates. After all, Latinos are estimated to comprise 10% off all voters this year. If Latinos are only leaning to Obama 48-42, that +6 edge among 10% of the electorate only contributes a net 0.6 advantage to Obama (4.8 for Obama to 4.2 for Romney). However, if instead Obama is leading 70.3 to 21.9 that +48.4 edge contributes a net 4.8 advantage to Obama (7.0 to 2.2), hence the national polls may be missing as much as 4 full points in Obama’s national numbers. […] If these mistakes are being made nationally where Latinos comprise an estimated 10% of all voters, they are even worse in statewide polls in Nevada, Florida, Colorado and Arizona where Latinos comprise an even larger share of all voters. In Florida Latinos are estimated at 17% of all voters. If you are badly mis-calculating the candidate preference among 17% of the electorate (that’s 1 out of every 6 voters), then the entire statewide estimates are wrong.
As Bouie concludes, “The polls might be underestimating Obama’s level of support, due to small and unrepresentative samples of Latino and African American voters. If that’s the case, then Obama’s position is much stronger than it looks.” Barreto doesn’t speculate on the motives behind the Latino undersampling. But it would be quite a stretch to credit these pollsters with honest ignorance about the Latino demographic. The more likely case is that they are distorting their samples to produce a more competitive race — or worse.