The
Edison-Mitofsky evaluation report in early 2005 reported that the discrepancies in precincts that used touchscreens, optical scanner, and punch cards all were similar; the discrepancies in lever-machine precincts were higher, and the discrepancies in the 40 hand-count precincts were lower. (See page 40.)
Some folks have interpreted that as evidence that only the hand counts were accurate -- but as the report pointed out, hand counts were used almost exclusively in rural areas, and the discrepancies were lower in rural areas regardless of the equipment they used. Controlling for size of place, the hand count discrepancies are statistically indistinguishable. (The lever machine discrepancies probably have to do with place, not equipment. In New York, John Kerry did astonishingly well in the exit poll, but his vote count was in line with -- slightly better than -- pre-election polls.)
The analysis turns out the same way using state-level data (see
here). Moreover, if one looks in states like New Hampshire or North Carolina that used a mix of hand counts and other methods (or Ohio, which used several methods although not hand counts), there's no evidence of a technology effect. I've written up some of those results, but never bothered to post them because no one's attention span seemed to be that long.
Here is an interesting analysis.
All that is with respect to the 2004 general; I've seen no reason to think that the results would be qualitatively different in 2006 or the 2008 primaries.