On Thu, 2 Sep 2004, Niko Speybroeck wrote: > Thanks a lot for you answer Thomas. Do you have a reference which supports > this solution? Can you give an example of a weight that depends on > variables that shouldn't be in the model? >
Robert Baskin has answered some of this. Additional points 1) I don't have a reference, but the argument would be that the design variables affect the distribution of the outcome only through the weights. The situation where it might be preferable just to adjust for weights would be if the weights depended on a lot of variables (eg indicator variables for fifty states). 2) An example: suppose you were interested racial differences in heart disease. As race has a substantial effect on income and income may well have a substantial effect on health, you might want to fit models with and without income. If the survey weights depend on median income for a region you would be unable to fit models that did not include income. This illustrates the main situation when a variable can be strongly predictive but you don't want it in your model: when it is in the hypothesised causal pathway between an exposure you are interested in and the outcome. A less interesting situation is when you don't want to use a variable that is available in your data set because it won't be available in future data sets. -thomas ______________________________________________ [EMAIL PROTECTED] mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html