Sorry, I just posted the email below but realised I did not give a name or details, would it be possible to adjust before posting and send what is below, sorry again, first time user...
From: Joanne Lenehan [mailto:jlene...@une.edu.au] Sent: Tuesday, 1 December 2009 3:51 PM To: 'r-help@r-project.org' Subject: GLM Repeated measures test of assumptions: e.g. test for sphericity e.g. Bartletts and Levenes homogenous variances Hello and thanks in advance I am running a glm in R the code is as follows with residual diagnostic code below model4<-glm(Biomass~(Treatment+Time+Site)^2, data=bobB, family=quasi(link="log", variance="mu")) par(mfrow=c(2,2)) plot(model2) to test the effect of grazing exclusion of feral horses for a Phd with following factors: Treatment - 3 levels which are grazed controls (C), ungrazed horse exclosures (H)with 3 wire strand fence to allow other herbivores but not horses and ungrazed full netting exclosures to exclude all herbivores (A) Site - 6 levels for 6 Sites which are considered fixed Time - 6 levels also for 6 different monitoring or time periods which is where the repeated measures comes in In SPSS there is the option for Bartlett's test of sphericity and Levene's test for whether the variability is homogenous or not. Is it possible to do the same in R with glm? If not, what is the equivalent for testing these assumptions? Also, is it valid to use Tukey's post hocs to test main effects if necessary? If not, is there another method other than contrasts, something to do with multicomp? In a previous model where Site was random I used the following lme model model2<-lme(Anncov~Treat*Time, random=~1|Site, data=bob,control=list(opt="optim")) And checked residuals using: print(shapiro.test(model1$residuals)) print(qqnorm(model1)) #resids vs fitted plot(model1,resid(.,type="p")~fitted(.)|Site, abline=0) #boxplot of resids plot(model1,Anncov~fitted(.)|Site, abline=c(0,1)) And because you can't use Mauchly sphericity test used instead: library(ts) bob$resids<-model1$residuals[,2] site.level<-levels(bob$Site) par(mfrow=c(4,4), mar=c(2,4,0,0), oma=c(2,2,4,2)) for (i in seq(along=site.level)){ datain<-bob$Site==site.level[i] acf(bob[datain,"resids"],xlab="", ylab="", main=paste("Site", site.level[i]),las=1) } #end loop BUT WHEN I TRIED THE ABOVE FOR GLM IT WOULD NOT WORK And because can't use Tukeys used contrasts but heard recently about multicomp? I also checked and compared several variance covariance symmetry's using the following: compoundspec<-lme(Anncov~Treat*Time, random=~1|Site, data=bob,cor=corCompSymm(,form=~1|Site), control=list(opt="optim")) autoregressivesite<-lme(Anncov~Treat*Time, random=~1|Site, data=bob,cor=corAR1(,form=~1|Site), control=list(opt="optim")) autoheterosite<-lme(Anncov~Treat*Time, random=~1|Site, data=bob,cor=corAR1(,form=~1|Site),weight=varIdent(form=~1|Time), control=list(opt="optim")) unstructsite<-lme(Anncov~Treat*Time, random=~1|Site, data=bob,cor=corSymm(,form=~1|Site),weight=varIdent(form=~1|Time), control=list(opt="optim")) And then anova to compare the models SHOULD I DO THE SAME BUT IS THERE A BETTER WAY? SO IN SUMMARY - BACK TO ORIGINAL QUESTION HOW TO TEST ASSUMPTIONS OF GLM IN R? Many many thanks in advance for anyone who reads or responds. Kind regards Joanne Lenehan PhD Candidate University of New England Armidale NSW 02 6773 3723 [[alternative HTML version deleted]] ______________________________________________ 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.
