[R-sig-ME] Significant anova results for identical Models in glmer

[Ben Bolker]

burg4401 at uni-trier.de wrote:
> Dear List,
> 
> when fitting two identical lmer Models and comparing them with 
> anova() one Model has a significant p-value 
> although obviously all coefficients and variances and likelihoods remain the 
> same.
> 
> I think this comes from line 1000 in lmer.R (Rev 266)
> pchisq(0,0,lower=F) == 0
> 
> I'm not concerned about someone putting in twice the same model, but 
> apparently if two models with same degree of freedom are compared the pvalues 
> are calculated on a dchisq(df=0) basis...
> e.g. factor1 = SEX , factor2 = (higher than 1.78m)
> 
> Maybe I was looking in the wrong books, but I couldn't find an argument 
> against this comparison. Could it be that the Problem lies in non nested 
> parameter spaces?

  Yes.  If two models have the same df, they are either (1) identical
or (2) non-nested [and hence inappropriate for likelihood ratio tests,
which is what anova() implements].

  Whether one should try to catch these kinds of (arguably) user errors
depends on one's philosophy of software design ...

  cheers
    Ben Bolker