R help - Sep 2003 - probit analysis for correlated binary data

Dear all,

I have a question on the dose-response estimation with clustered/ correlated
binary data.
I would like to estimate the hit rate for a certain test at various
concentration levels. The test is used
on 5 subjects, and each subject is tested 20 times. If we assume that the 100
samples are
independent, the hit rate estimate is unbiased, but the variance is
under-estimated. The other
estimate of interest is the concentration that will give a 95% sensitivity of
the test. I suppose the
same problem would occur if a probit model is fitted to the data and assume all
the samples are
independent. Does anyone have any suggestions on how to analyze this kind of
data?

Thanks a lot in advance.

- Jacqueline

I can't tell you how to solve your problem but I can suggest where to look.
Alan Agresti released a second edition to his book on categorical data
analysis early this year. One of the big updates to that book is a chapter
on Generalized Linear Mixed Models (GLMM). The good news is that random
effects for Probit models are explicitly referenced in the index. The bad
news is that the section referenced is a mathematical exercise for the
reader. 

 

Charles E. White, Biostatistician

Walter Reed Army Institute of Research

503 Robert Grant Ave., Room 1w102

Silver Spring, MD 20910-1557

301 319-9781

WRAIR Home Page: http://wrair-www.army.mil/

 


	[[alternative HTML version deleted]]

On Wed, 24 Sep 2003, Law, Jacqueline {Regu~Pleasanton} wrote:
> Dear all,
>
> I have a question on the dose-response estimation with clustered/
> correlated binary data. I would like to estimate the hit rate for a
> certain test at various concentration levels. The test is used on 5
> subjects, and each subject is tested 20 times. If we assume that the 100
> samples are independent, the hit rate estimate is unbiased, but the
> variance is under-estimated. The other estimate of interest is the
> concentration that will give a 95% sensitivity of the test. I suppose
> the same problem would occur if a probit model is fitted to the data and
> assume all the samples are independent. Does anyone have any suggestions
> on how to analyze this kind of data?
>
I think you want a GEE, see the `gee' and `geepack' packages.

	-thomas