1
$\begingroup$

I am working with some CDC weighted survey data (specifically YRBSS). The CDC has a website that displays the proportion of respondents who respond positively to the survey questions. The results are broken down by sex, school grade, and race. In each cell there is a proportion and a confidence interval. CDC says that the confidence intervals were produced by SUrvey DAta ANalyzer (SUDAAN).

The task at hand is to reproduce these confidence intervals using the R survey package. I have used the svymean function, and it gives the correct proportions, but different confidence intervals. It is important that we get the same CIs up to 4 or 5 decimals places, and we are not close. I have noticed that the SUDAAN CIs are not symmetric about the mean. This reminds me of logistic regression, but I don't know where to take it from there.

Can anybody give me a hand?

$\endgroup$
  • $\begingroup$ Define your acronyms please. Otherwise, many of us, me especially, do not understand. $\endgroup$ – Carl Oct 30 '16 at 22:20
  • $\begingroup$ CDC is the Center for Disease Control. YRBSS is a particular survey, specifically, Youth Risk Behavior Surveillance System or similar. SUDAAN is a software package for survey data from RTI International. RTI stands for nothing, but used to stand for Research Triangle Institute. SUDAAN? Probabiy SUrvey DAta ANalyzer or the like. $\endgroup$ – Rex D Oct 30 '16 at 22:27
  • $\begingroup$ OK, help the reader by putting in links and using full text before citing acronyms as you will see shortly in my edit of your text: complete for CDC etc. There is acronym overload on this site. Yesterday, someone used "MLP", which unless you know a lot about machine learning looks like "my little pony." $\endgroup$ – Carl Oct 30 '16 at 22:43
  • $\begingroup$ Carl, thanks for your interest. The author of the survey package has suggested that I look at the svyciprop function. What is MLP? $\endgroup$ – Rex D Oct 31 '16 at 1:21
  • $\begingroup$ Yes, thus the acronym problem. "Multilayer perceptron" $\endgroup$ – Carl Oct 31 '16 at 2:25
2
$\begingroup$

In version 3.31-3 or later, I have added an option method="xlogit" to svyciprop() that appears to reproduce the SUDAAN results.

As far as I can tell, the SUDAAN formula is not documented, but published examples show it is the same as SPSS uses, and that is documented. The algorithm is

  1. Estimate the mean and standard error as usual
  2. Apply a logit transform to the mean and use the delta method to transform the standard error estimate
  3. Compute a confidence interval for the logit mean and back-transform to the probability scale.

The intervals should be almost identical to those from method="logit", which is available in earlier versions of the package.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.