Sacha Viquerat <tweedie-d <at> web.de> writes: > > Am 15.04.2011 20:14, schrieb Christian Hennig: > > Normality of the predictors doesn't belong to the assumptions of the > > GLM, so you don't have to check this. > > > > On Fri, 15 Apr 2011, Simone Santoro wrote: > > > >> I want to estimate the possible effects of some predictors on my > >> response variable that is n? of males and n? of females > >> (cbind(males,females)), so, it would be: > >> > >> fullmodel<-glm(cbind(males,females)~pred1+pred2+pred3, binomial) > >> > if you count no of males and females, shouldn't you choose the poisson > family? maybe whoever you told you to check for normality referred to > that, since count data are not normally distributed (neither are their > errors)! maybe thats all he/she wants!
I think the original model using the binomial distribution for the response seems entirely appropriate. I agree with the comment about tiny data sets: the usual rule of thumb is that (# parameters) should be <(effective N)/10 -- so in practice estimating anything more than a single binary or continuous predictor (both of which require a single parameter to estimate) would be pushing it. (Sad but true.) Ben Bolker ______________________________________________ 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.
