I am unsure why you think the CI can't contain zero. If you get enough bootstrap samples where no success are observed (i.e. 2.5% in a 95% CI), then the bootstrap CI will contain zero.
The analytical solutions implemented by Stata are pretty good (wilson, jeffrey's etc), but when it comes to survey weighted data you need to simulate.
-Tim
________________________________
From: [email protected] on behalf of Jun Xu
Sent: Fri 7/8/2005 4:28 PM
To: [email protected]
Subject: RE: st: RE: Help with confidence intervals generated with nlcom from survey data
I am also puzzled by this problem (sometimes, for example ci derived from
delta method might be out of reasonable range; in this case, below zero).
However, the ci provided by bootstrap will always NOT contain zero (like in
this case and many other cases involving probabilities). Re probablility or
rates ratio, on one hand we should have them between 0 and 1. On the other
hand, the ci provided by bootstrap will never contain zero or one, which is,
I think, unreasonable too.
What about others' opinions?
Jun Xu
Ph.D. Candidate
Department of Sociology
Indiana University at Bloomington
http://mypage.iu.edu/~junxu/home
>From: "McKenna, Timothy" <[email protected]>
>Reply-To: [email protected]
>To: <[email protected]>
>Subject: st: RE: Help with confidence intervals generated with nlcom from
>survey data
>Date: Fri, 8 Jul 2005 11:18:07 -0400
>
>Confidence intervals that don't wander outside of prior known bounds can be
>tricky. The simplest way to approach this problem is to use a bootstrap.
>This will require more information about the survey design than you
>provide, but broadly bootstrapping works as long as you resample within
>sets of observations that are independent identically distributed.. The
>Stata bootstrap command has options that allow you to specify strata and
>clusters within which your observations are iid. I imagine there are
>survey designs that are complicated enough to defeat the above strategy,
>but for simple surveys it works well. I used this technique when computing
>binomial confidence intervals with survey data, as the CI provided by the
>survey commands in Stata (at least 8.2, not sure about 9) are ultimately
>based on normal approximations and as such the CI can wander outside of the
>0,1 interval.
>
>-Tim
>
>________________________________
>
>From: [email protected] on behalf of Bird, Tommy M "Mac"
>Sent: Thu 7/7/2005 12:50 PM
>To: [email protected]
>Subject: st: Help with confidence intervals generated with nlcom from
>survey data
>
>
>
>Hello,
>I am calculating rates of a particular disease in two different groups
>using survey data. I am using SVYRATIO for this and it is working
>nicely. Now I would like to create a rate ratio by dividing the rate
>from one group by the rate of the other group. I can do this with NLCOM
>and it works nicely. My problem is that the confidence intervals for my
>rate ratio are not logit transformed, thus some of my lower confidence
>intervals are negative. I really need a confidence interval that cannot
>go below zero. How can I do this? Is their a transformation that I can
>run on the ends of the confidence intervals? Is their an option that I
>am not aware of on NLCOM? Is there a better way to do this all
>together? Please keep in mind that I am neither a statistician, or a
>Stata expert. Thanks so much,
>
>Mac Bird, MS
>Research Associate
>Child Health Services Research Group
>Center for Applied Research and Evaluation
>Department of Pediatrics
>University of Arkansas for Medical Sciences
>
>
>*
>* 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/
>
>
><< winmail.dat >>
_________________________________________________________________
Don't just search. Find. Check out the new MSN Search!
http://search.msn.click-url.com/go/onm00200636ave/direct/01/
*
* 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/
<<winmail.dat>>
