I try to compare some models obtained from glmmPQL. model1 <- glmmPQL(y~red*yellow+I(red^2)+I(yellow^2)+densite8+I(densite8^ 2)+freq8_4 +I(freq8_4^2), random=~1|num, binomial); model2 <- glmmPQL(y~red*yellow+I(red^2)+I(yellow^2)+densite8+I(densite8^ 2)+freq8_4 , random=~1|num, binomial); anova(model1, model2)
You try to compare models that differ in their fixed parts. This is not possible with the default method 'REML'. This would only be possible if you fitted your models using method 'ML' (see Pinheiro & Bates, 2000). In addition, if I understood a remark by Jos?? Pinheiro during one of his courses correctly, the anova comparisons are not save with a distribution other than normal. Thus, one should rely on the function intervals () to see whether confidence intervals of the parameters overlap zero or not. Perhaps someone else can comment on this issue? Regards, Lorenz - Lorenz Gygax Centre for proper housing of ruminants and pigs Swiss Federal Veterinary Office agroscope FAT T??nikon, CH-8356 Ettenhausen / Switzerland