Reply #7: will 1988 suffice?

[OnTheOtherHand]

That year, in the CBS/New York Times exit poll, the average weight on white Bush voters was about 11% bigger than the average weight on white Dukakis voters, which implies about a 5-point error on the margin. The model error may have been smaller, or larger. In the raw data, there were more Dukakis supporters than Bush supporters -- but the researchers may have deliberately oversampled African-American precincts, so I limited the comparison to whites.

In 1984, there was about a 7% swing on weights (among whites) in the other direction. In 1980 there was about a 3% swing overstating Carter's vote share -- that was pretty good, but probably still biased. Carter voters actually outnumbered Reagan voters by about 2.4 points in the whole sample, although again, I'm guessing that the researchers deliberately oversampled African-American precincts.

I think the error in 1992 is well known by now.

The question isn't whether these were "before electronic machines with no paper trail were used," since such machines have never been used everywhere in the nation. But I daresay that none of these elections was decided on DREs.