[taking the liberty of cc'ing this to r-sig-phylo, which is a better place for such questions because (1) other people can answer (2) the answers are publicly viewable and get archived]
On 13-03-05 10:44 AM, Javier Lenzi wrote: > Dear Dr. Bolker, > > My name is Javier Lenzi and I am trying to perform a comparative analysis > with seabirds (gulls). I am trying to implement a model selection approach > using compar.gee in ape package and I have some problems. I was wondering > if I could ask for your advice to deal with them. In case you could help me > there some questions below that I would like to ask. I am sorry for > bothering you, my skills in statistics are basic and I have to be fully > autodidact in matters related to the comparative method. > > Thank you very much in advance, > Best regards, > Javier Lenzi > ____________________________ > > I would like to ask if it is possible to perform a model selection approach > using compar.gee, because the articles I have read that use compar.gee do > not use this approach when they have multiple explanatory variables. > > Then, I have a set of candidate models for each life history trait and when > I run compar.gee I get the QIC and the parameter estimates, but I can not > find an index of goodness of fit for the model and the P-value. Is there > any available? How should I proceed? > > I would also like to ask if there is any function similar to step() in > order to perform model selection using GEEs? > > Other question I would like to ask you is how can I get the amount of > variance that is explained by the phylogeny? I think it would be nice to > analyze and present it as a result. Unfortunately none of these are simple questions. Generalized estimating equations are not really designed for model selection and comparison approaches, as they don't operate in terms of likelihoods (which are the standard currency of such selection and comparison). Thus they don't have model-level p-values, just t-statistics and p-values for the individual parameters. If you are comfortable with QIC you can use them in a standard information-theoretic multi-model inference approach (i.e. decide to take the model with the best QIC, or compute QIC weights and get parameter weights or averaged predictions). I strongly recommend against stepwise model selection: google "stepwise regression problems". You can do it by hand if you need to ... Let's see if anyone else on the list has suggestions. good luck, Ben Bolker _______________________________________________ R-sig-phylo mailing list - [email protected] https://stat.ethz.ch/mailman/listinfo/r-sig-phylo Searchable archive at http://www.mail-archive.com/[email protected]/
