lme questions (was [R] difference between splus and R)

[Prof Brian Ripley]

> Date: Fri, 7 Apr 2000 09:55:08 -0400 (EDT)
> From: Faheem Mitha <faheem at email.unc.edu>

[I have given a more meaningful subject line.]

> On Fri, 7 Apr 2000, Prof Brian D Ripley wrote:
> 
> > On Thu, 6 Apr 2000, Faheem Mitha wrote:
> > 
> > > I'm running splus 5 on a solaris platform remotely, and running R on linux
> > > on my home machine.
...
> > 
> > Might I suggest you install nlme 3.x on Splus5 too? (nlme.stat.wisc.edu)
> > Then you won't have to use two different syntaxes.
> 
> Not an option, unfortunately. It is a univ mainframe, and I don't have
> those kinds of powers. I can ask them (the powers that be)  nicely to do
> it, I suppose. How should I convince them that nlme3.x is better than nlme
> 2.x which they presumably have installed?

Any user can have a private library on S-PLUS, and in the same way on R.
See V&R3 p.470 for S-PLUS (and set R_LIBS on R).  You need no special
privileges to install library sections or packages, so this should be an
option.


> > > The models are supposed to be identical, and my understanding of the L
> > > ratio and the p value is that they are the values corresponding to the
> > > null hypothesis that the smaller model is true ie. that the random
> > > effect due to age is zero. So a large p value in both cases
> > > corresponds to strong evidence in favour of the null hypothesis.
>  
> > > My understanding is that both R and Splus are doing exactly this. So
> > > why are they returning different value. Are the models somehow
> > > different? Another possibility is that one is using ordinary likelihood 
> > > and the other is using REML. I see from the R documentation that REML
> > > is indeed used here and I thought the same was true of Splus.
>  
> > (The fits for ran2 give the same statistics, so look both to be REML.) 
> > You should not be using anova on lme models fitted with REML. Although in
> > this case they are using the same fixed-effects model and so are on
> > comparable data, the supporting theory is for ML fits only, AFAIK.
> 
> I am only using the prepackaged function anova.lme from the package nlme.
> If you look at the documentation you will see that not doing anything
> unconventional with it. While I am not sure what likelihood ratio
> statistic is being used (the documentation does not say, but it appears
> that it is probably REML-based) if it is not a legitimate test, then why
> is it included in the package?

I *have* looked at the documentation.  It does not give a reference for
the validity of REML-based LRTs, so can you please supply one?
There is a warning note:

     Likelihood comparisons are not meaningful for objects fit using
     restricted maximum likelihood and with different fixed effects.

which does not say that the converse *is* meaningful. nlme2 even gives the
comparisons in the excluded case. Bill Venables' warning (V&R3 p.203) is
rather stronger.

S does not usually stop you doing non-meaningful statistics, so do not
assume that because it gives a result it is `legitimate'.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595

-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-help mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !)  To: r-help-request at stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._