Mind Poison

Dick Morris has just announced to the world that he's a major idiot. His argument goes something like this: "Exit polls are never wrong, and they were wrong last night, so clearly the lying liberal media tampered with the exit polls." Mm hm.

I mentioned this before, but anyone who understands statistics a little bit should understand why this is nonsense (you'd think a political advisor would understand that much about statistics). I linked yesterday to MysteryPollster's explanation of exit polling, but if you didn't read it, you should. They try to sample 100 people at about 1500 precincts across the country throughout the day. In a tight race, with sample sizes that small, it would not be unlikely that you see results that suggest the opposite of what the actual distribution is.

If you've got say 5 million people voting, and their underlying distribution is that 51 percent support Bush, when you sample 100 of them, there's no reason you shouldn't expect to see that 51 percent of your sample support Kerry. Taking this sample, your maximum likelihood estimator would lead you to believe that 51 percent of the 5 million people support Kerry.

Now, they don't forecast a state with 100 voters. If the precincts are distributed across the states equally (probably not), then say there's 30 per state, or 3000 people. That's a much better-looking sample size. But keep in mind that these are not ideal sampling conditions; MysteryPollster brought up several problems that make exit polls not statistically ideal. Furthermore, that 3000 people isn't 3000 randomly selected, it's 3000 from 30 randomly selected precincts. Some of the precincts will be heavily democratic and some heavily republican, and some neither. By the luck of the draw, your 30 random precinct draws in that state will probably sample one or the other too much (keep in mind that their "corrective" weighting is based on elections in years past, and voter attitudes and demographics, could have changed greatly in that time, to say nothing of redistricting). Voting patterns might also be different throughout the day, and also, reading Drudge and seeing a close race that your guy is losing might get your ass out to the polls at the last second.

In the end, just about every early exit poll number I can find was within the margin of error (supposedly 4%), and well within it at that. Kerry supposedly had 51% in Ohio and he had 49%, supposedly 52% in Michigan and he had 51%, 53% in Pennsylvania when he only had 51%, 50% in Iowa when he only had 49%, 51% in Wisconsin (which was correct), 52% in Minnesota when he had 51%, and so forth. The farthest from the exit polls were New Hampshire and Florida, and I imagine if you go back through the polling data in previous elections, there are plenty of exit polls in very close states that were not correct within the margin of error. Only it wasn't a big deal because they were in states that didn't matter, or they predicted the same as the actual result anyways.

So the mid-day exit poll numbers showed a small lead in the wrong direction in the closest races. Calm down, Dick.