[R] AIC and anova, lme

[ian white]

Patrick,

The likelihoods of two models fitted using REML cannot be compared
unless the fixed effects are the same in the two models.  


On Tue, 2008-02-26 at 14:38 +0100, Patrick Giraudoux wrote:
> Dear listers,
> 
> Here we have a strange result we can hardly cope with. We want to 
> compare a null mixed model with a mixed model with one independent 
> variable.
> 
>  > lmmedt1<-lme(mediane~1, random=~1|site, na.action=na.omit, data=bdd2)
>  > lmmedt9<-lme(mediane~log(0.0001+transat), random=~1|site, 
> na.action=na.omit, data=bdd2)
> 
> Using the Akaike Criterion and selMod of the package pgirmess gives the 
> following output:
> 
>  > selMod(list(lmmedt1,lmmedt9))
>                  model       LL K  N2K       AIC  deltAIC  w_i      AICc 
> deltAICc w_ic
> 2 log(1e-04 + transat) 44.63758 4  7.5 -81.27516 0.000000 0.65 -79.67516 
> 0.000000 0.57
> 1                    1 43.02205 3 10.0 -80.04410 1.231069 0.35 -79.12102 
> 0.554146 0.43
> 
> The usual conclusion would be that the two models are equivalent and to 
> keep the null model for parsimony (!).
> 
> However, an anova shows that the variable 'log(1e-04 + transat)' is 
> significantly different from 0 in model 2 (lmmedt9)
> 
>  > anova(lmmedt9)
>                      numDF denDF   F-value p-value
> (Intercept)              1    20 289.43109  <.0001
> log(1e-04 + transat)     1    20  31.18446  <.0001
> 
> Has anyone an opinion about what looks like a paradox here ?
> 
> Patrick
> 
> 
> 
> ______________________________________________
> R-help at 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.