So how do you get a non-respondent to predict his non-response?

Mark Blumenthal has a very interesting post about exit polling; he reports on a recent Fox News/Opinion Dynamics poll which suggests that Democrats are more likely to participate in exit polls than Republicans.  If this is a valid finding, it might go a long ways towards explaining the big discrepancy between the exit polls and outcome of the 2004 presidential election.  It might also suggest why Karl Rove is so calm before the midterm election, when all the "right people" are predicting a Democrat landslide.

Too bad most of the commenters on Blumenthal’s site are political partisans; I’d like to see some discussion on how well a survey can detect reasons for non-response, and how this information could be used for a more accurate model-based estimate.   Looks like I’m going to have to dive into Sarndal’s book to get hip on the topic.