Tuesday, January 14, 2020

A Point about Polling

If you had a truly, truly random sample of US voters, how many would you have to poll to get as good an estimate as a you get from a pretty good sample of 2.3 million voters?

401.

If you knew your sample was truly random, you could do very well just polling 401 voters. Even a tiny deviation from true randomness makes a sample of millions less accurate than a perfectly random sample of 401. All the effort pollsters put into trying to contact thousands and thousands of people is necessary because getting a truly random sample is all but impossible.

2 comments:

G. Verloren said...

This seems absurd to me.

A truly random sample could, by sheer randomness, be wildly skewed in favor of one group or another. It's not common to flip a coin and get heads 9 out of 10 times, but it does happen - and if you base your predictions on such an uncommon chance sampling, they're going to give you wildly skewed results.

Sure - on average, a purely random sampling could be accurately representative. But there will be a substantial number of times when your sampling isn't average, and unless you somehow could KNOW that and take a new sample to work off instead, you're setting yourself up for mistakes.

Shadow said...

I thought 401 was high. I've seen many with fewer respondents.

And it's getting more difficult. People don't answer their phone any more unless they recognize the number. It's forcing pollsters to move their polls to the Net. Should be interesting.