What this boils down to is that the differences in voting pattern between the two groups of counties this year are greater than they were in 2000. I've attached a graph showing both years. As you can see, firstly, in 2000, the percent change for both parties was negative, whereas it was positive in 2004. Secondly, in both years, the democrats did better in counties that used "E-touch" in 2004 and Republicans did better in counties that used "Op-scan" in 2004. So this is a potential confound. However the stats say that this effect was significantly bigger in 2004.
Unfortunately that stats can't tell you the reasons for the difference, they can only tell you that when you analyse results by machine used in 2004, the voting pattern was significantly different from that observed in 2000.
I've looked at things like changes in voter registration in the two groups of counties between the two elections, and for some reason the two sets of counties do appear to differ significantly (possibly simply by chance) in their patterns of party affiliation. This does not rule out machine tampering, but it does say that there are other potential explanations for the observed pattern. It is a sad fact of statistics that no amount of clever maths can attribute causality to an effect unless you randomly allocate the manipulated variable.
Elizabeth Liddle