Statistical significance

[davywavy]

It's been 15 years since I did any stats work and I'm damned if I can remember how to do it now.

If I have an audience size of about 22,000 people, how many would I need to survey to get results significant to first 5% and then 1%?
Any thoughts on how I can work this out?

Comments

[drplokta]

It depends on your null hypothesis and how different the survey results are from it. You can't pick your desired significance level and then choose a sample size to achieve it, you can only pick a sample size, conduct your survey, and then see what significance level you've achieved.

[davywavy.livejournal.com]

Nadgers, I was afraid I remembered it was something like that.
I'll aim for several hundred and then see what that produces.

[hiromasaki.livejournal.com]

Well, what you can do is take your best-guess estimate of what you feel the results should be, add in a margin of error, and work backwards to find out what minimum number you should shoot for is.

[davywavy.livejournal.com]

Yeah, that's whay I'm doing. I'm coming out as needing about 400 responses for a 5% margin of error, which sounds like a fair starting point.

[omentide.livejournal.com]

You can do confidence intervals which aren't quite the same thing...

I have a link to a website that does them for you, but it's at work. Try google.

Cheers for the tip.

[davywavy.livejournal.com]

Google has provided such a thing. Looks like I'm aiming for a minimum of 400+ people.

[davywavy.livejournal.com]

Yes, I've been looking at confidence intervals instead which are probably more relevant to me at this stage. Like I say, it's been 15 years!

[(Anonymous)]

If you're about this weekend, I'll haul 6-sigma for dummies round for you.

[davywavy.livejournal.com]

I've heard of that in Dilbert. Do I want to know more?

[(Anonymous)]

No. It's yet another faddish doomed-to-inevitable-failure compilation of buzzwords designed to trick companies out of lots of money.

[davywavy.livejournal.com]

I got a 3rd in it then, so it's never been my forte.

Stats through Beer

[applez.livejournal.com]

Drink 5 pints of beer with a standard deviation of 1 pint - and all statistical questions are resolved. ;-)

Go on, you know that's exactly what you were doing 15 years ago. Don't pretend otherwise. ;-p