manon girard <mansof@VIDEOTRON.CA> wrote [in part]: > I encountered some problems with weighted data. I usually work with > populational data which contains weights and "design effects" (or variance > inflation due to sampling design - like stratification). In fact, each > sampled individual in the dataset is given a weight such that this guy > represents X persons in the total targeted population.

So this is not just 'weighted' data, but actually a probabilistic sample from a defined target population. [By the way, I don't like the term 'variance inflation', since we don't really inflate the variance, but use the sample design information to get a *correct* variance. And the usual simple random sample variance estimator is not always a good estimate of the true variance. So you shouldn't be using PROC GLM or PROC ANOVA anyway.]

> Usually I take sampling weights to readjust the sample size. By this I mean > to get n the same number of my sample size but individual may represents > less than 1 person (my weights would contain a fraction of individual).

Like I said, you should not be doing your analysis this way. There are risks of embarrassing errors from this kind of approach to sample survey data.

> Recently, I tried to make multiple comparisons with SAS-GLM (Anova) and all > the homogeneity of variance tests (underlying assumption of multiple > comparisons in ANOVA) (e.g. : HOVTEST=BF) could not be used with weights. > What can I do ? Bootstrapping ?

No, I recommend the survey analysis procs in SAS version 8. Try using PROC SURVEYREG and PROC SURVEYMEANS instead. They are both designed to handle survey data of the complexity you describe. They won't handle everything and they aren't designed to handle all possible survey designs, but they are a reasonable starting point.

David -- David Cassell, CSC Cassell.David@epa.gov Senior computing specialist mathematical statistician