I read this Freeman paper. I am not impressed.

The main problem is that Freeman does not acknowledge that the numbers he uses are not actual counts, they are massaged by a computer program. That computer program takes the real numbers, compares them to the last election, throws in assumptions about what changes are important, and then makes the other changes disappear. If the assumptions are good, the poll will be accurate. If the assumptions are bad then the polls will be way off.

The strongest assumption is that people who voted in past elections tend to vote again, and vice versa. The polls will miss any big change in turnout, unless they are specifically calibrated to look for that. In fact they will throw out the first-time voters as an anomaly.

Here's how this works. For instance, the poll Freeman used had 2000 "respondents" in Ohio. What, do we guess these were taken at 40 polling places, with 50 respondents each? OK, Ohio has about 10,000 polling places, so the 0.2 percent chosen for polling must be very carefully selected. In my area, there is a real patchwork of outcomes from neighboring polling places.

OK, let's say that in one "typical" polling place 13 of the 50 respondents are White Evangelicals. The computer knows that an average of only 10 of the 50 respondents at that polling place have been White Evangelicals in all past elections. The computer "knows" this is just a statistical fluctuation and weights those 13 votes less. (These numbers are chosen to exactly correspond with my idea of the increase in White Evangelical vote that soundly gave Bush the election.)

This is not a glitch, this is exactly how polls are able to get such precise results from polling so few people, 2000 from all of Ohio.

And btw, Freeman doesn't mention the discrepancy in the non-contested states. If the exit polls were just as incorrect in North Dakota as in Ohio, it is hard to think it is a nationwide conspiracy. This is just one example how Freeman's article is not just wrong, it is irresponsible.