Patterson's numbers

Check out the strange data under 'Sample Error' at the website of Patterson, who conducts the Westpoll. [Last checked 11 Aug  2007.]

They reckon that a sample of 400 "produces results that are accurate to within about +/- 2% depending on the actual survey estimate". Just 2% - can this be so?

No. Their table chooses to illustrate various error margins given a result of 82% - an odd number, you would think; when was the last time an opinion poll showed 82% support for any party? They also use three confidence intervals, one of which is 70% (most pollsters use 95%)

Westpoll is trying to pull a swifty. For very high or very low "results" (say approaching 0 or 100%), the error margin becomes small. It is biggest around 50% - which is, of course, the ballpark of most political polls.

Patterson uses 82, with 70% confidence, because it gives an impression of low margin. (Even for that, two percent is too low by my calculations.)

For the record, a sample of 400 which gives a result around 40-60% for a party is within +/- 5% of the "true" number, with a 95% confidence interval.