[R-sig-ME] summary vs anova in lmerTest

[Alexandra Kuznetsova]

Dear Charlotte,

 1) As illustrated in the first part of the script, the new version does not perform the same elimination process as the old one: different elimination and rank and statistics
 I have looked at your example. Thank you for providing a reproducible example. Although I have tested on the lmerTest version 1.2-1 and the results agree with the new version 2.0-3. Which old version did you use?
  
 2) Then, in the new version, the anova function does not give the same statistic and denominator degrees of freedom as the step function for the first eliminated effect.
 WE have been here already - the reason lies in the REML comparing to ML

 3) Eventually, on a smaller dataset, when eliminating one of the two random effects (variance=0), the step function considers the model as fixed-effects only and does not do any effects selection. In the old version, with the same data, I got a selection.
 By default the program eliminates the random effects with 0 component. We will probably add a keep option in order to remain random effect anyways (I am not sure why it did not eliminate in the old version - probably the variance was not 0, as the new lme4 can give slightly different variance estimates comparing to the old one because of some computational changes)

 Again the results seem strange that you gave of the "old" version - as far as I have tested I do not get them. Again which version did you use? Or probably it is somehow related to the R version that you used as well... Also some remark: and probably here lies a reason that you got such results: you need to convert to a factor d$ringf <- as.factor(d$ringf) before you call model and step functions.

Hope that will help,

With regards,
Alexandra