For one thing, the surveys were conducted online. Studies have shown people are more willing to reveal potentially unpopular attitudes on a computer than in questioning by a live interviewer.
The polls also used a technique aimed at measuring what psychologists call "affect misattribution." This involved showing faces of people of different races quickly on a screen before displaying a neutral image that people were asked to rate as pleasant or unpleasant. Studies have shown that people consciously or unconsciously transfer their feelings about the photograph to the object they are rating.
The researchers compared the subjects' ages, party identification, perceptions of Obama and Romney, and other factors to their racial attitudes. This allowed them to create mathematical formulas predicting the likelihood that people would vote for either Obama or Romney, based on their different characteristics and attitudes. The models allowed them to estimate how much impact each factor has on each candidate's support, controlling for other factors.
By using their formulas, the researchers were able to conclude that race was a factor in how people vote, independent of their other political views and their demographic characteristics. They then used the formulas to predict how much support Obama is losing because of his race. While the model was exhaustive, it is hard to know if it included every variable that could have an impact in this election.
The study on attitudes toward African-Americans included interviews with 1,071 adults between Aug. 30 and Sept. 11. It was conducted by GfK Custom Research under the supervision of AP's polling unit. A 2011 survey assessed attitudes toward Hispanics and was conducted April 8-23, 2011, by GfK Custom Research, known at the time as Knowledge Networks, among 1,053 adults.
The interviews were conducted online. The original samples were drawn from a panel of respondents GfK recruited via random sampling of landline telephone households with listed and unlisted numbers. The company provides Web access to panel recruits who don't already have it.
Results were weighted, or adjusted, to reflect the adult population by factors such as age, sex, race, region and education.
In the 2012 survey, no more than one time in 20 should chance variations in the sample cause results to vary more than plus or minus 3.8 percentage points from the answers that would have been obtained if all adults in the U.S. were surveyed. For the 2011 study, it is 3.6 points.
There are other potentially greater sources of variability in surveys, including the wording and order of questions.
The questions and results for this poll are available at http://surveys.ap.org . A closer look at the results of the 2008 study using the same methodology is available here at http://poq.oxfordjournals.org/content/73/5/943.full .