[R] survey package: proportion estimate confidence intervals using svymean

[Thomas Lumley]

On Tue, 12 Feb 2008, Peter Holck wrote:

> Using the survey package I find it is convenient and easy to get estimated
> proportions using svymean, and their corresponding estimated standard
> errors.  But is there any elegant/simple way to calculate corresponding
> confidence intervals for those proportions?  Of course +/- 1.96 s.e. is a
> reasonable approximation for a 95% CI, but (incorrectly) assumes symmetrical
> distribution for a proportion.

'Incorrectly' is a bit strong, I think.

> About all I can think of is to transform onto some scale like the logit
> scale, use the delta method to estimate standard errors, and transform back
> for appropriate confidence intervals.  Or perhaps use the machinery of the
> survey package svyglm function with logit link (family=quasibinomial) to do
> the calculations for me.  Still, seems rather cumbersome.  Is there a better
> solution?
>

I think the most straightforward solution would be svyglm().

You could also use svycontrast() to do the delta-method calculations for you.

        -thomas

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle