On Thu, 18 Dec 2008, Corey Sparks wrote:
> Dear list,
> I am using svyglm in the survey library to fit a binomial logistic
regression
> accounting for sample design. The documentation says the models are not
fit by
> maximum likelihood, so my question is what is the fitting method? Pseudo
> likelihood? Generalized least squares? Thanks for any insight.
They are fitted by maximizing the Horvitz--Thompson estimator of the population
loglikelihood. This may be what you mean by pseudo likelihood.
Operationally, the point estimates come from glm.fit() with the sampling weights
as weights (the standard errors are different, of course).
-thomas
Thomas Lumley Assoc. Professor, Biostatistics
tlumley at u.washington.edu University of Washington, Seattle