I have just repeated my original analysis of mid-size counties counties, this time first of all omitting NW panhandle counties, then omitting all N Florida counties. The machine effect remains robust ((p<.01) despite loss of statistical power.

I also repeated it after converting the raw percentages (which may have a binomial distribution) to values with a Gaussian distribution. Again, the effect of machine remains.

Of course this does not rule out demographic effects, but it does say that the apparent machine effect is not confounded by something special happening in N Florida.

I found data on senate versus presidential votes here http://electionexamination.blogspot.com/

Using the same very conservative group of mid-size central and southern Florida counties (no N Florida counties), Machine is a significant predictor of Bush proportion of votes over and above the prediction made by Republican senator votes. Opscan nets significantly more votes (p<.05) for Bush than would be predicted on the basis of votes for republican senators.

So if it is a demographic effect, it is not just a historical party registration factor - it is also evident in current party voting for senate.

Elizabeth Liddle