Pew did a study of initial non-respondents back in about 1999, where they spent a year tracking down the people who they were initially unable to contact or who iniitally refused to answer in a survey. Within a year they were able to get the vast majority of the inital refusers/"not-homers" to answer the survey, and found there wasn't a dramatic difference on the big variables like party preference, ideology, etc. Somewhat more likely to be distrustful of insitutions, and significantly less likely to vote. Overall, it indicated that we were not getting significantly skewed results due to non-responses, especially in terms of predicting election outcomes.

With non-response rates contuing to climb, and with non-response rates correlating increasingly with demographic variables (e.g., under 30s who don't have telephones), there's still room for concern.

A much greater worry, one that has been clearly documented to ocurr, is people lying about their voting intention in particular cases where they feel embarassed about that choice. The classic case was underreporting of intentions to vote for David Duke, but there was also substantial underreporting of intentions to vote for Guliani over Dinkins among liberal whites in NY (the evidence indicates that many answered "don't know" instead when polled). You'd probably see the same if say, Houston suburbanites decided to vote Democratic in 2008, and were being interviewed by someone with a Texas accent.

You are welcome to be as paranoid as you like about survey houses cooking the books. I'm fully aware of conservative blogs complaining about polls having "too many Democrats", but from I what know professionals who depend on having good polling data to analyze don't share those fears (nor for that matter do conservative bloggers who understand polls well).