sigma = sqrt(p * (1-p) / N)

But what they’re really reporting is 1/sqrt(N). Whether that quantity should be referred to as a “margin of error” is mostly semantics. It’s the quantity that’s reported, so it’s the quantity I based my calculations on.

]]>The margin of error depends not only on the sample size, but also the recorded result: Margin of error

]]>I figured out where the 1.644 was coming from. It’s equal to

sqrt(2) * erfinv(2*0.95 – 1) = 1.64485

Plug in values other than 0.95 to get other confidence levels.

]]>http://www.econtalk.org/archives/2008/07/rivers_on_polli.html

fascinating on why these margins of error systematically underestimate the true uncertainty. Of course I don’t know anything about the particular poll you are pointing too, but if what this speaker says is true, sloppy statistics are pretty pervasive. In particular the part about not adjusting error calculations for sample re-weighting.

]]>