: Bug in Error() and the use of Error() for repeated measures with more than 2 fa ctors

Prof. Ripley,

I am sending this to you again, as the first one was undelivered.


Peter

ripley at stats.ox.ac.uk wrote:

There is no intended limit, but there was a bug fix in 1.5.0.  If this
is using 1.5.x, could you send me an example that malfunctions?
 

I have attached the .Rdata file with the dataframe 
simulation.6batches.aovdata. The scripts are the same ones used in the 
previous email

This analysis was carried out using:

R : Copyright 2002, The R Development Core Team
Version 1.5.1  (2002-06-17)



Peter

Brian

On Wed, 14 Aug 2002, Peter Ho wrote:

I have been trying to analyse an unreplicated repeated measures 
2-level
factorial design with 11 factors using aov() with Error(), similar to
that described in  "Notes on the use of R for psychology 
experiments and
questionnaires" by Jonathan Baron and  Yuelin Li (see the example of

   Hay's)

I have found that there seems to be a limit, in the number of 
factors in
Error() . For example,  the following worked :

aov(CV ~(Twb+Plb+Psb+tinsb+tseedb+Tmb+Mcb+Mom0b+Mom1b+Mom2b+Mom3b)3 +
Error(Batch /
(Twb+Plb+Psb+tinsb+tseedb+Tmb+Mcb)),simulation.6batches.aovdata)

but not:

aov(CV ~(Twb+Plb+Psb+tinsb+tseedb+Tmb+Mcb+Mom0b+Mom1b+Mom2b+Mom3b)3 +
Error(Batch /
(Twb+Plb+Psb+tinsb+tseedb+Tmb+Mcb+Mom0b+Mom1b)),simulation.6batches.aovdat 

a)

which I suppose should be the right way to code the strata.formula 
for
11 fixed factors. Returned the following reply after running this 
script:

Error in eval(expr, envir, enclos) : couldn't find function "Error"

The only difference between the two scripts is the second one 
contains
all 11 factors. Is this a bug? Or should I be using lme() or some 
other
function when the number of factors is greater than the "limit"?

I would also like to hear some views as to whether Error() should be
used for repeated measures, pro & cons,etc? I have found differing 
views
in books. For example, Neter et al  "applied linear statistical 
models "
uses simply the combined treatment by subject interaction mean square
when testing main effects and their interactions. They do state 
that the
relaxation of the assumptions of a repeated measures model can be 
made
by the inclusion of first-order interaction between each factor and
subjects in the model. However, this still falls short of the ideas
suggested with using Error() for testing.On the other hand,  
Mead's "the
design of experiments puts repeated measures in the same chapter with
split unit design, which are multistratum design. Jonathan Baron's 
and
Yuelin Li's document also references a number of books on this 
matter.

I know that this point is not really suitable for r-help, but it is
related indirectly with regard to the use of strata.formula in 
Error().


Any comments would be grateful


Peter Ho
ISR-Porto



-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-