One more time into the question about the VA-Gov polls and the election results.
I came up with a possible (and speculative) explanation based on the condition of two candidates who were not especially popular among their own parties, along with a sufficiently well-publicized third candidate. Perhaps Democrats trusted the polls, believing that the seven-point lead Terry McAuliffe held was real, and therefore believed it was safe to vote for the Libertarian; meanwhile, Republicans didn't trust the polls, believed it was close, and thus switched at the last minute back to the Republican.
It occurs to me there doesn't have to be a difference in partisan bias to get the results I speculated on.
Suppose partisans on both sides read the polls with a 5 point bias in their own direction. Then, in a 7 point race with a third party candidate, the winning side party is going to perceive a 12 point landslide and vote their true preference, while the losing side party will perceive a 2 point barnburner and reluctantly return home. After all, a heavily-polled election with a 7 point spread will surely have some 12 point and some 2 point polls, and in fact this one did.
It further occurs to me that if true, this should be a pattern we see in other elections. The main conditions, I would think, would have to be a polling lead in the 5-10 point range...that's small enough that the losing side can plausibly believe that it's very close, but large enough for the winning side to feel safe -- again, if both sides interpret the polling with about the same bias.
All of which gets to that 1992 Perot vote. The research on that was pretty clear: Perot took equally away from both candidates. However, if this (speculated! hypothesized!) effect was operating, then it would have cost Perot Republican votes while leaving Democrats parked with him at the end. If that was the case, then the "real" Perot vote (that is, those who would have voted for him all else equal) would have been more Republican the observed Perot vote.
Note, of course, that it still wouldn't be the case that Perot cost Republicans the 1992 election; no third party candidate, and all those votes come back home, half to Clinton and half to Bush.
At any rate: I suspect this should be testable. The basic idea should be that an effect kicks in when there's a 5-10 point lead and a plausible third party candidate...say, one getting better than 5% of pre-election polls. Under those circumstances, the final result should be more favorable to the losing candidate than the polls predicted. It's also possible, if detailed polls are out there, to compare the third party supporters before election day and in the final vote.
It should also be an effect, if real, that may stronger when the party-aligned press has a larger share of the overall press, since presumably that would produce stronger cues for interpreting the polls with a bias.
Not sure if anyone wants to bother doing the work on this (I don't), but I think there's a fairly good chance that it's real -- both for VA-Gov and for Perot.