Sample Sizes and Accuracy.

Just to keep this Lancet thing going a little longer (and to remind myself that I **cked up) Virginia Postrel has something interesting. The largest survey done in the US (Except of course for the election itself which by definition is correct....if your sample size is the entire group that you are trying to estimate the behaviour of your certainty of your result is, well, pretty certain.) was the 7-11 coffee cup choice. Very close to the actual result, much much better than the exit polls for example. 51.08 percent for Bush, 48.92 percent for Kerry. As the company itself noted, their sample size was way way larger than anyone else's so we would expect it to be closer to the real result. (There wasn't anyone else sampling a million people a day.)