Reply #2: I think E/M were sloppy-wrong here

[OnTheOtherHand]

I'm not sure that anyone noticed this issue until a few weeks ago (Bruce O'Dell noticed that the alphas in his regressions seemed too high), but it probably takes a bigger alpha to yield a 6.5-point mean WPE.

It's sort of a nasty calculation to estimate, because WPE is a mess. But I just roughed out that for a precinct that is really 51.5% Bush to have a -6.5 WPE, it takes about a 7-point difference in completion rates; if it were really 80% Bush or 20% Bush, maybe an 11-point difference.

The actual numbers are going to be all over the map, but if you stare hard at page 12 of the O'Dell paper, you'll see that it implies an average alpha of somewhere around 1.158 (because the average ln alpha will be around 0.1467). So that could be interpreted as a bit under an 8-point gap, say, about 56.9% to 49.1% (if we are trying to hit about 53% overall), which seems to fit with what I roughed out.

At this point, I don't think that this matters much because (1) I don't have any strong prior assumptions about how bad the response bias could be, and (2) the vote shift models I've seen create a slope that doesn't show up in the E/M scatterplot. Some folks find it crazy to think a poll could be that far off, but that scatterplot is a mess, so I figure the poll was a mess, too. But we need more out of E/M, and we need to keep thinking about what kinds of wholesale fraud would be consistent with the data we have.