On 18 April 2013 18:38, Thomas Foxley <thomasfox...@aol.com> wrote: > Rune, > > Thank you very much for your response. > > I don't actually have the models that failed to converge from the first > (glmulti) part as they were not saved with the confidence set. glmulti > generates thousands of models so it seems reasonable that a few of these may > not converge. > > The clmm() model I provided was just an example - not all models have 17 > parameters. There were only one or two that produced errors (the example I > gave being one of them), perhaps overparameterisation is the root of the > problem. > > Regarding incomplete data - there are only 103 (of 314) records where I have > data for every predictor. The number of observations included will obviously > vary for different models, models with fewer predictors will include more > observations. glmulti acts as a wrapper for another function, meaning (in > this case) na's are treated as they would be in clm(). Is there a way around > this (apart from filling in the missing data)? I believe its possible to > limit model complexity in the glmulti call - which may or may not increase > the number of observations - how would this affect interpretation of the > results?
Since the likelihood (and hence also AIC-like criteria) depends on the number of observations, I would make sure that only models with the same number of observations are compared using model selection criteria. This means that I would make a data.frame with complete observations either by just deleting all rows with one or more missing predictors or by imputing some data points. If one or a couple of variables are responsible for most of the missing observations, you could disregard these variables before deleting rows with NAs. As I said, I am no expert in model averaging or glmulti usage, so there might be better approaches or other opinions on this. Cheers, Rune ______________________________________________ R-help@r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.
