The example code you give does not give
any instances with lower confidence limits
below zero, which I assume is what you mean
by negative confidence intervals. But clearly
the CIs do differ.
Setting aside the complications of -svy-, which
is naturally a big set-aside:
There is no carved on stone, handed down from
on high, method of getting "CORRECT" [your word] binomial
confidence intervals. This is why -ci- (pure
and simple) offers a variety of ways of doing it,
and what you get over a range of real situations is
interestingly scary. Sometimes methods agree nicely;
other times they don't. Also, sometimes a confidence level
means about that much coverage, but often not.
The manual entry for [R] ci gives one entry into
the literature. The paper by Brown and friends in
Statistical Science 2002 is relatively friendly,
and likely to be web-accessible to you.
Regardless of that, as proportions approach 0
(or 1, really the same problem, modulo some measurement
convention), then on any reasonable view the problem becomes
increasingly asymmetric, and thus not best to be thought
in terms of
estimate +/- some multiple of standard error,
which, whatever they may say in introductory treatments,
is at best a crude approximation to what is going on.
A much better scale to work on is logit.
Thus negative confidence limits
are in essence a clear sign that you are using
an inappropriate method, and/or that a one-sided
interval would be more appropriate.
Nick
n.j.cox@durham.ac.uk
Jason Ferris
>
> I have been running survey proportions and observing results with
> negative confidence intervals (which doesn't make sense). When I use
> survey tab (with column percent, se and ci) I get the same point
> estimates and standard errors but different 95% confidence
> intervals. I
> assume this is an issue with the proportion calculations
> using "Binomial
> Wald" for confidence intervals.
>
> I checked the survey manual and have not been able to find why:
>
> Paste the following command to see my dilemma:
> webuse nhanes2b, clear
> svy: proportion race
> svy: tab race, ci se
>
>
> The results show the same point estimates and standard errors (with
> rounding) but different CI's. As mentioned, for my data, I get some
> negative CI's for svy: proportions commands but not for the svy:
> tabulate commands. My ultimate concern is being able to automatically
> extract the CORRECT estimates to excel (from using matrix
> e(b) and e(V)
> - and calculating 95% CI from square-root of e(V) *1.96).
>
> I am using the latest version of Stata 9.2, on Windows XP.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/