As is well known, a good sample can be more accurate than a poorly
administrated census. In the 2004 U.S election, were Bush
supporters innately more disinclined to talk with exit pollers, or did
the exit pollers get it right, and the machines get it wrong?
Check out an analysis performed by some highly reputable statisticians
at http://uscountvotes.org/ucvAnalysis/US/Exit_Polls_2004_Edison-Mitofsky.pdf