[R] lme v. aov?

[Prof Brian Ripley]

On Thu, 27 Nov 2003, Spencer Graves wrote:

>       Do you want to make inference about the specific subjects in your 
> study?  If yes, the subjects are a fixed effect.  If instead you want to 
> make inference about the societal processes that will generate the 
> subjects you will get in the future, that is a random effect.  The 
> function "lme" handles both fixed and random effects, as does 
> "varcomp".  The functions "aov" and "lm" are restricted to fixed effects 
> only.  You can use dummy coding for "lm" and "aov" as well. 

Have you considered Modern Applied Statistics with S (2002) by 
Venables and Ripley?  That has a very helpful chapter 10 which shows 
examples of doing random-effects and mixed-effects studies using aov()
and should help set straight some of your misconceptions.  

A lot of classical analyses of variance are for random effects and this 
used to be taught in all the basic statistic courses.  It is still in some 
of the better introductory books.  And all multistratum aov fits (think 
`Error() term') are mixed (since there is an overall mean which is a fixed 
effect).

-- 
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 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595