Question: assuming that the distribution family and link function are
correctly specified, does the population-averaged generalized estimating
equation (GEE) approach have "full efficiency" when the working correlation
is unstructured, or when there are only two repeated measurements (panel
length / cluster size = 2) and the working correlation is exchangeable?
According to M. E. Stokes, C. S. Davis and G. G. Koch, _Categorical Data
Analysis Using the SAS System_, 2nd Edition, (Carey, N. Carolina: SAS
Institute, 2000), "For [situations where the requirements are met], weighted
least squares offers full efficiency, provides well-defined goodness-of-fit
statistics, accounts for all degrees of freedom, and is asymptotically
equivalent to maximum likelihood methods. You do lose these properties with
the GEE approach . . . (Page 430)."
Both weighted least-squares and PA-GEE are marginal approaches; I was under
the impression that GEE looses efficiency if the working correlation
structure is misspecified. It seems as if the working correlation couldn't
be misspecified when it's specified as unstructured (and convergence is
obtained), or when there are only pairs of observations and an exchangeable
correlation is used. Would -xtgee- still be less efficient than, say, PROC
CATMOD for, for example, Ricardo Ovaldia's glucose-meter reading problem
posted last week? ( www.hsph.harvard.edu/cgi-bin/lwgate/
STATALIST/archives/statalist.0310/date/article-595.html )
Joseph Coveney
*
* 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/
