[R] lme and lmer df's and F-statistics again

[Peter Dalgaard]

Julia S. wrote:
> Hi there,
>
> thanks for your help. I did read Bates statement several times, and I am
> very glad and thankful that many statisticians spend much time on this. The
> problem is, as Dieter pointed it out, that many "end users" often have to
> use statistics without being able to fully understand the math behind it.
> Because if they would spend as much time on that as statisticians do, they
> wouldn't be able to do what they do where they use statistics for. 
> And, no, I don't expect that a "simple" answer exists, but it might be that
> somebody had a similar problem like me before and may have a convincing line
> for a referee at hands. I have problems reformulating what I read here in my
> own words.  
>
> Dieter: when you write:
> "but to use lme instead when possible" do you mean that when using lme the
> F-stats are correct? Because I assumed that the problem would be the same
> with lme. 
>
> Julia
>   
They aren't... And they can be badly wrong in some cases.

At this stage, I think the best one can do is to get a feeling for
whether the DF would be "large" and if so,  convince the referee to
accept an asymptotic chi-square test (Wald or LRT type).

I think that the rationale for requiring authors to state the DF is not
so much that journals believe in mighty SAS, but that they want to be
able to catch completely wrong analyses, like when people compare two
groups of each 5 rats and get a denominator DF of around 100 because
there were 10 (correlated) measurements on each rat and no between-rats
variation in the model.

As for figuring out whether or not you have large DF; if you have a
nearly balanced design. it might be worth looking into what aov() says
would be the DF for the same model with balanced data.

(And in any case, all DF-type corrections are in a sense wrong because
they depend on 3rd and 4th moments of the Gaussian distribution, and
your data probably aren't perfectly Gaussian.)

-- 
   O__  ---- Peter Dalgaard             Øster Farimagsgade 5, Entr.B
  c/ /'_ --- Dept. of Biostatistics     PO Box 2099, 1014 Cph. K
 (*) \(*) -- University of Copenhagen   Denmark      Ph:  (+45) 35327918
~~~~~~~~~~ - (p.dalgaard at biostat.ku.dk)              FAX: (+45) 35327907