[R-sig-ME] Problems in using lmer to fit a multilevel model

[Gregor Gorjanc]

...
> > The value of q is the total number of random effects and the value of
> > n is the number of observations.  I included that check because it did
> > not make sense to me to try to fit more random effects than you have
> > observations.
> 
> > I guess I could be persuaded that it would make sense
> > in some circumstances because the random effects are determined by a
> > penalized least squares optimization.
> >
> > What is the nature of the model that would require it to have more
> > random effects than observations?
> 
> Commonly, genetic models fit 2 or more random effects per individual, with 
> different prespecified covariance matrices (A, D, A*A, A*D...)

Yes, and it can even happen that there is more individuals in the pedigree
than there are phenotypic records! However, this is a bit special application. 
Perhaps, the warning might be more appropriate than the error.

Gregor