[R-sig-ME] model specification: fixed or random factor ?

[robert.espesser at lpl-aix.fr]

Dear all,

In a linguistic experiment, 2 participants involved in a dialog
are incitated to use 16 specific items. An item is a pair {name+ adjective}.
There are 2 short words (SW1,SW2) , 2 long words (LW1,lW2) and 4 
adjectives(A1,A4), leading to a total of 16 combinations or items:
{SW1,SW2}:{A1,..A4} and {LW1,LW2}:{A1,..A4} .
The # of occurrences of each item in a dialog are uncontrolled.
We can suppose there are 8 pairs of participants. Various measurements 
are done on these items.
For simplicity, I suppose there is one predictor of interest (predic).

The length word could  be  added to the model, but I am not interested 
in with the individual effect of word  and adjective. I think about the 
3 following models:

M1)  word(4 levels factor) and adjective(4 levels factor)  are 
considered as fixed factors (since they have not enough levels to be 
taken as random ) :
	
~ predic  + word + adjective
or
M2)
~ predic  + word * adjective


 From a linguistic point of view, each item (word+adjective) can be 
considered as a whole.
So the corresponding 16 levels factor  could be added to the model as 
random factor, and length word as a fixed factor, giving something like:

M3)
~ predictor + length + (1|word:adjective)

Is this model "correct" ,despite of the way the items are build ?
On the contrary, are M1 or M2 models  more valid models ?

Many thanks for any comments or suggestions.

Regards
R. Espesser
-- 
Robert Espesser
CNRS UMR 6057 - Université de Provence
5 Avenue Pasteur
13100 AIX-EN-PROVENCE

Tel: +33 (0)413 55 36 26