Replies

Maybe I'm stupid, but I thought the idea of statistical sampling went like this: you have a few million marbles in a hopper, and each marble be red or blue. You don't have time to count them all, so you shake and turn and stir the hopper so the marbles inside become randomly distributed, and after you do that you put on a blindfold and draw out a sample of few hundred marbles at random from the randomized hopper.

Then you count the red and blue marbles in your sample, and - if you know some statistical formulas, or have a table of statistical data - you can now estimate the total number of red and blue marbles in the entire hopper on the basis of the count you come up with in your sample.

That's why it's called statistical sampling.

OK. Fine. We know all about that. If there are just a few marbles in your sample, the chance that your estimate will be off is higher. If, on the other hand, you draw out a lot of marbles for your sample, your estimate can still be off, but the probable error of your estimate will be smaller than it would have been if you'd only picked out a few marbles.

But now, we add a wrinkle. The red and blue color on the marbles is only a wrapper on the real marble. Inside the red wrapper you usually find a red marble, and inside the blue wrapper you usually find a blue marble... but not always. Sometimes you open a red wrapper and find a blue marble, and sometimes a red marble hides inside a blue wrapper.

What the polling organizations seem to be doing is this: they draw the random sample, count the red and blue wrappers, and - based on the number of red and blue marbles inside the wrappers in the last trial - they throw out a certain number of red-wrapped marbles until the color ratio of the wrappers is the same as the color ratio of the marbles inside the wrappers found the last trial. After they throw away red-wrapped marbles, they open all the wrappers and count the red and blue marbles that were inside.

Now, as I said, maybe I'm stupid, but this seems like false mathematical reasoning on the part of the polling organizations. If they stuck to basic statistical polling methods, the sampled ratio of people who say they're going to vote for Obama to the sampled ratio of people who say they're going to vote for Romney would work out by itself without any post-sample tweeking.