Date:         Thu, 19 Dec 2002 10:55:54 -0500
Reply-To:     Steve Albert <salbert@AOL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Steve Albert <salbert@AOL.COM>
Subject:      Re: help on power analysis

David Cassell made a good point (as always!):  you have a problem in
sampling theory here.  You need to understand the sampling design, and the
issues involved.  (You also haven't described how your health care workers
are to be recruited, which may introduce some biases of its own, or how
they're told to select the n individuals from their communities, which may
introduce other biases, affect the correlation between observations in each
cluster, etc.)

I'd agree with David that you should consult a statistician who's used to
designing sampling schemes in order to both properly collect and properly
analyze your data.

To get background to help you talk to that statistician, some background
reading in sampling would be helpful.  If you don't have any background in
sampling, you might look for "A Sampler on Sampling" or check to see if
there is a booklet in one of the Sage series.  If you do have some, then
you might try reading about cluster sampling in Paul Levy and Stanley
Lemeshow's "Sampling of Populations: Methods and Applications", third
edition, starting with chapter 8; or for a shorter discussion, Zakkula
Govindarajulu's "Elements of Sampling Theory", chapter 9.  Govindarajulu
discusses cluster sampling for proportions (including sample size) on page
185.  A more mathematical treatment is in William Cochran's "Sampling
Techniques", 3rd edition (a classic on sampling), with specific discussion
of cluster sampling for proportions at the ends of chapters 3 and 9 (which
includes a brief mention of sample size).

To reiterate:  all of the above is for *background*; you still need to
consult with a statistician to decide an appropriate sampling scheme,
sample size, and analysis method.  Textbook treatments make assumptions
about the way the sample is drawn that may not be valid in your
circumstances, and you need a qualified statistician to design something
which is appropriate for *your* study, not for a textbook problem set.

Steve Albert
Director of Biostatistics
Spectrum Pharamceutical Research Corp.
San Antonio, TX
SAlbert at SpectrumCRO dot com