Nate Silver's fivethirtyeight.com blog has an interesting graph by Andrew Gelman.  I note that a regression line would clearly have a slope greater than one (draw the best fit straight line through the data points, it will be steeper than the slope of the 1:1 line shown).  This means that high and low predictions were over-conservative - high predictions should have been higher, and low predictions should have been lower.  Was this a feature of only this election, or is it a common factor in polling?  Do pollsters need to adjust their parameters if greater accuracy is desired at the high and low ends, or was there something intrinsically uneven about this election?
 






Silver also has an interesting review of the time course of "gay marriage" issues in which he finds a national trend of about 2% vote shift per year in favor of no marriage bans.