Disclaimer. Don't rely on these old notes in lieu of reading the literature, but they can jog your memory. As a grad student long ago, my peers and I collaborated to write and exchange summaries of political science research. I posted them to a wiki-style website. "Wikisum" is now dead but archived here. I cannot vouch for these notes' accuracy, nor can I even say who wrote them. If you have more recent summaries to add to this collection, send them my way I guess. Sorry for the ads; they cover the costs of keeping this online.
Berinsky. 1999. The two faces of public opinion. AJPS 43 (4).
People hesitate to give surveyors socially undesirable answers. Thus, surveys underestimated opposition to forced integration of the schools, and they overestimated the black candidate's support in the 1989 NYC mayoral race: those with racist views simply claimed that they had no opinion. Thus, social desirability leads to a selection bias: Those with socially unattractive opinions abstan from certain questions, claiming they have no opinion.
Berinsky demonstrates an empirical technique for estimating the degree of bias introduced by this problem. Basically, he uses a bunch of variables to predict two things: (1) How likely are you to support integration (the "outcome" equation), and (2) how likely are you to abstain from the question (the "selection" equation). By noting the variables that predict abstention, he can identify the types of people who are more likely to abstain, and then weight the responses of those who did respond.
Here's the intuition, though it's not exactly what he did. See Table 1. In the left column, he has used probit to predict the probability of supporting school integration, using a few dozen variables. Obviously, the left column includes only those who actually answered the question; this with socially undesireable answers have claimed not to have an opinion. Based on the coefficients for these demographic and partisan indicators, he calculates the predicted responses for 'all' respondents--including those who claimed not to have an opinion. Thus, he has an estimate of what the general population--including the "abstainers"--really thinks. (Again, that's 'not' what he did, it's just the intuition. What he did is actually rather persuasive.)
He demonstrates that the gap between black and white respondents would have been even larger if everybody had responded, among other things (see Table 1; compare left to right column; right column is "corrected"). Moreover, the effect of having a black interviewer would have been weaker if everybody had responded--because none of the abstainers lied to a black interviewer (they abstained instead), but some of the non-abstainers did lie. On the aggregate, these selection effects were significant: If everybody had responded, support for integration would have fallen from 49.4 to 35.9 percent.
The less persuasive part is his application to the NYC 1989 mayoral race. Pre-election polls showed Dinkins, the black candidate, enjoying around a 15% lead over Guiliani. On election day, however, he won by only a point or two. Many of those who did not want to support a black candidate (but didn't want to admit it) abstained in the surveys, but voted for Guiliani. Using the same method as described above, he finds that his correction to the surveys matches the findings from the ballot booths.
Research on similar subjects