Interesting commentary about polls over at electoral-vote.com

[skipos]

Since there is no other major news today (although I will keep an eye on FL-16, where more news is leaking out every day), this is probably a good time to discuss polling methodology. Probably the least understood but nevertheless most overrated term in polling is "margin of error." If Smith polls 53% and Jones polls 46% with a margin of error of 3%, what this means is that there is a 95% chance that Smith is favored by between 50% and 56% of the voters and Jones is favored by between 43% and 49% of the voters. Thus Smith is (barely) ahead by more than the margin of error.

But does that really mean much? Unfortunately, no. Margin of error has to do with how much error is introduced in the polling results by the fact that only 600-1000 people are polled. With bad luck the pollster might happen to get too many Democrats or too many Republicans in the small sample (although they try to correct for that). If the pollster asked every registered (or possibly, likely) voter in the state, the margin of error would be 0, but the results might still be highly inaccurate.

In general, statisticians distinguish between precision and accuracy. Precision has to do with having a small sampling error; accuracy has to do with actually determing the true value you are trying to measure. The classic example is to ask a million people the height of the Emperor of Japan. You will get a very precise (i.e., reproducible) answer down to a fraction of a millimeter, but it won't be very accurate because most people don't know the height of the Emperor of Japan.

Another classic example is the 1936 Literary Digest poll that predicted Republican Alf Landon would crush Democrat Franklin D. Roosevelt. They sent 10 million mock ballots to addresses taken from telephone books and got back 2.4 million replies, an immense sample, with a margin of error less than 1/10 of a percent. But it turned out that anybody rich enough to have a telephone at the bottom of the Depression was probably a Republican. The sample was anything but random.

more at:
http://electoral-vote.com/

[Lasher]

1. What are internals?

I've seen the term used and I think I know what that means but I'm not sure.

[rock]

2. You don't want precision to exceed the accuracy

The old farmer said, "that river is 10,000,006 years old."
I looked perplexed, "How do you know?"
"Six years ago a geologist told me it was 10,000,000 years old."