PROC SURVEYMEANS vs. PROC GENMOD vs. Weighted mean & SD

[SteveDenham]

Aha.  If your variable cannot be negative, it most likely needs to use a different distribution.  What is the dependent variable--a count or perhaps a binomial response?  This will greatly affect the code.  Your checking on the poolability implies a distribution other than the default normal, so it should be accommodated.

If the data come from a distribution where there is a functional relationship between the mean and variance (poisson, binomial, etc.), then Stroup in Generaliized Linear Mixed Models addresses this extensively.  A GEE model is marginal, and tends to be biased, whereas a G-side model is conditional on the random effects and is less biased.  The trade-off is just what you see, broader confidence intervals.

Let us know what the dependent variable is, and I think we can get to a more robust answer.

Steve Denham