Reply #4: not a straw man, actually -- bear with me

[OnTheOtherHand]

It's true, as we've discussed, that a single line doesn't describe the shape of the data.

But Ron's comments above miss something important. Regression lines are actually much more heavily influenced by points that are far from the center of gravity than by points near the center. Therefore, if a bunch of high-Bush precincts have too-high WPEs, and a bunch of high-Kerry precincts have too-low WPEs -- which I think is the point of Baiman's part of the USCV paper -- the regression line really ought to tilt positive.

When I looked at O'Dell's simulation, that's what actually happened -- the regression line had significant positive slope. The same thing was true of the Dopp simulation in the AAPOR version of the paper, although I don't know what it does right now -- I haven't figured out the new modeling of vote shift yet.

I'm pretty sure O'Dell agrees with me about this. Coincidentally, I was talking with him a little while ago, and he offered to post here.

Now, here's the thing. I'm pretty sure that Baiman's response rate arguments are undermined by the actual data. And I don't know about you, but when I look at the scatterplots, I don't see fraud concentrated in Bush strongholds, which was one theory. I do believe the Bush-stronghold theory would entail a positive slope (at least if the fraud were nearly enough to swing the election). But there are lots of other theories that wouldn't entail a positive slope. I've spent most of three weeks now (mostly not on DU) arguing against one pretty weak theory, but it doesn't mean that I'm against all the other theories.

If you see people insisting that folks like me and Febble assume or insist on a linear model, that really is a straw man. I don't think that any line or curve we fit through those points is going to tell the whole story.